HERMES-GIELIE/api/routes.py

"""
Hermes Web UI -- Route handlers for GET and POST endpoints.
Extracted from server.py (Sprint 11) so server.py is a thin shell.
"""

import html as _html
import copy
import json
import logging
import os
import queue
import re
import platform
import shutil
import subprocess
import sys
import threading
import time
import uuid
import re
from pathlib import Path
from urllib.parse import parse_qs
from api.agent_sessions import MESSAGING_SOURCES

logger = logging.getLogger(__name__)

# Treat stalled/closed HTTP clients as normal disconnects.  Long-lived SSE
# connections often end this way when a browser tab sleeps, a phone switches
# networks, or Tailscale leaves the socket half-closed.  If these bubble to the
# request handler, the server logs 500s and can leave CLOSE-WAIT sockets around
# until the OS-level timeout fires.
_CLIENT_DISCONNECT_ERRORS = (
    BrokenPipeError,
    ConnectionResetError,
    ConnectionAbortedError,
    TimeoutError,
    OSError,
)

# ── Cron run tracking ────────────────────────────────────────────────────────
# Track job IDs currently being executed so the frontend can poll status.
_RUNNING_CRON_JOBS: dict[str, float] = {}  # job_id → start_timestamp
_RUNNING_CRON_LOCK = threading.Lock()
_CRON_OUTPUT_CONTENT_LIMIT = 8000
_CRON_OUTPUT_HEADER_CONTEXT = 200
_MESSAGING_RAW_SOURCES = {str(s).strip().lower() for s in MESSAGING_SOURCES}
_MESSAGING_SESSION_METADATA_CACHE: dict[str, object] = {
    "path": None,
    "mtime": None,
    "identity": {},
}
_MESSAGING_SESSION_METADATA_LOCK = threading.Lock()
_STALE_MESSAGING_END_REASONS = {"session_reset", "session_switch"}


def _normalize_messaging_source(raw_source) -> str:
    return str(raw_source or "").strip().lower()


def _is_known_messaging_source(raw_source) -> bool:
    return _normalize_messaging_source(raw_source) in _MESSAGING_RAW_SOURCES


def _safe_first(*values):
    for value in values:
        if value is None:
            continue
        text = str(value).strip()
        if text:
            return text
    return ""


def _gateway_session_metadata_path():
    try:
        from api.profiles import get_active_hermes_home
        hermes_home = Path(get_active_hermes_home()).expanduser().resolve()
    except Exception:
        hermes_home = Path(os.getenv("HERMES_HOME", str(Path.home() / ".hermes"))).expanduser().resolve()
    return hermes_home / "sessions" / "sessions.json"


def _load_gateway_session_identity_map() -> dict[str, dict]:
    path = _gateway_session_metadata_path()
    if not path.exists():
        return {}

    try:
        st = path.stat()
        cache = _MESSAGING_SESSION_METADATA_CACHE
        with _MESSAGING_SESSION_METADATA_LOCK:
            if cache["path"] == str(path) and cache["mtime"] == st.st_mtime:
                return cache["identity"].copy()
    except Exception:
        return {}

    try:
        raw_sessions = json.loads(path.read_text(encoding="utf-8"))
    except Exception as _json_err:
        logger.debug("Failed to parse gateway sessions metadata from %s: %s", path, _json_err)
        return {}

    mapping: dict[str, dict] = {}
    if isinstance(raw_sessions, dict):
        for _entry in raw_sessions.values():
            if not isinstance(_entry, dict):
                continue
            session_id = _safe_first(_entry.get("session_id"))
            if not session_id:
                continue
            origin = _entry.get("origin") if isinstance(_entry.get("origin"), dict) else {}
            platform = _safe_first(origin.get("platform"), _entry.get("platform"))
            mapping[session_id] = {
                "session_key": _safe_first(_entry.get("session_key"), _entry.get("key")),
                "chat_id": _safe_first(origin.get("chat_id"), _entry.get("chat_id")),
                "thread_id": _safe_first(origin.get("thread_id"), _entry.get("thread_id")),
                "chat_type": _safe_first(origin.get("chat_type"), _entry.get("chat_type")),
                "user_id": _safe_first(origin.get("user_id"), _entry.get("user_id")),
                "platform": platform,
                "raw_source": platform,
            }

    with _MESSAGING_SESSION_METADATA_LOCK:
        _MESSAGING_SESSION_METADATA_CACHE["path"] = str(path)
        _MESSAGING_SESSION_METADATA_CACHE["mtime"] = st.st_mtime
        _MESSAGING_SESSION_METADATA_CACHE["identity"] = mapping
    return mapping.copy()


def _mark_cron_running(job_id: str):
    with _RUNNING_CRON_LOCK:
        _RUNNING_CRON_JOBS[job_id] = time.time()


def _mark_cron_done(job_id: str):
    with _RUNNING_CRON_LOCK:
        _RUNNING_CRON_JOBS.pop(job_id, None)


def _is_cron_running(job_id: str) -> tuple[bool, float]:
    """Return (is_running, elapsed_seconds)."""
    with _RUNNING_CRON_LOCK:
        t = _RUNNING_CRON_JOBS.get(job_id)
        if t is None:
            return False, 0.0
        return True, time.time() - t


def _cron_response_marker_index(text: str) -> int:
    """Return the start index of a markdown Response heading, if present."""
    candidates = []
    for heading in ("## Response", "# Response"):
        if text.startswith(heading):
            candidates.append(0)
        idx = text.find(f"\n{heading}")
        if idx >= 0:
            candidates.append(idx + 1)
    return min(candidates) if candidates else -1


def _cron_output_content_window(text: str, limit: int = _CRON_OUTPUT_CONTENT_LIMIT) -> str:
    """Return a bounded cron output window that preserves useful response text.

    Cron output files can contain large skill dumps in the Prompt section. The
    UI already extracts ``## Response`` when present, so keep that section in
    the API payload instead of blindly returning the first ``limit`` chars.
    """
    if limit <= 0:
        return ""
    if len(text) <= limit:
        return text

    response_idx = _cron_response_marker_index(text)
    if response_idx >= 0:
        header = text[:min(_CRON_OUTPUT_HEADER_CONTEXT, response_idx)].rstrip()
        response = text[response_idx:].lstrip("\n")
        content = f"{header}\n...\n{response}" if header else response
        return content[:limit]

    return text[-limit:]


def _run_cron_tracked(job):
    """Wrapper that tracks running state around cron.scheduler.run_job."""
    from cron.scheduler import run_job  # import here — runs inside a worker thread
    from cron.jobs import mark_job_run, save_job_output

    job_id = job.get("id", "")
    try:
        success, output, final_response, error = run_job(job)
        save_job_output(job_id, output)

        # Match the scheduled cron path: an apparently successful run with no
        # final response should not leave the job looking healthy.
        if success and not final_response:
            success = False
            error = "Agent completed but produced empty response (model error, timeout, or misconfiguration)"

        mark_job_run(job_id, success, error)
    except Exception as e:
        logger.exception("Manual cron run failed for job %s", job_id)
        try:
            mark_job_run(job_id, False, str(e))
        except Exception:
            logger.debug("Failed to mark manual cron run failure for %s", job_id)
    finally:
        _mark_cron_done(job_id)

_PROVIDER_ALIASES = {
    "claude": "anthropic",
    "gpt": "openai",
    "gemini": "google",
    "openai-codex": "openai",
}

# OpenAI-compatible /v1/models endpoints for live model discovery.
# Used as fallback when hermes_cli.provider_model_ids() is unavailable or
# returns [] for a provider (#871).  Kept at module level so the dict is
# built once, not reconstructed per request.
_OPENAI_COMPAT_ENDPOINTS = {
    "zai": "https://api.z.ai/v1",
    "minimax": "https://api.minimax.chat/v1",
    "mistralai": "https://api.mistral.ai/v1",
    "xai": "https://api.x.ai/v1",
    "deepseek": "https://api.deepseek.com",
    "gemini": "https://generativelanguage.googleapis.com/v1beta/openai",
    "nvidia": "https://integrate.api.nvidia.com/v1",
}
# NOTE: "openai-codex" is excluded because it maps to the same endpoint as
# the base "openai" provider (api.openai.com/v1).  When both are configured
# the openai provider is already wired through provider_model_ids(); codex-
# specific model filtering happens downstream in hermes_cli.
#
_LIVE_MODELS_CACHE_TTL = 60.0
_LIVE_MODELS_CACHE: dict[tuple[str, str], tuple[float, dict]] = {}
_LIVE_MODELS_CACHE_LOCK = threading.RLock()


def _active_profile_for_live_models_cache() -> str:
    try:
        from api.profiles import get_active_profile_name

        return get_active_profile_name() or "default"
    except Exception as _e:
        # A transient profile-resolution error mis-scopes the cache for up to
        # 60s ("default" gets the wrong payload). Log so we can detect it; the
        # blast radius stays small because the TTL caps the bad-cache window.
        logger.debug("_active_profile_for_live_models_cache fell back to 'default': %s", _e)
        return "default"


def _live_models_cache_key(provider: str) -> tuple[str, str]:
    return (_active_profile_for_live_models_cache(), provider)


def _get_cached_live_models(key: tuple[str, str]) -> dict | None:
    now = time.monotonic()
    with _LIVE_MODELS_CACHE_LOCK:
        cached = _LIVE_MODELS_CACHE.get(key)
        if not cached:
            return None
        ts, payload = cached
        if now - ts >= _LIVE_MODELS_CACHE_TTL:
            _LIVE_MODELS_CACHE.pop(key, None)
            return None
        return copy.deepcopy(payload)


def _set_cached_live_models(key: tuple[str, str], payload: dict) -> None:
    with _LIVE_MODELS_CACHE_LOCK:
        _LIVE_MODELS_CACHE[key] = (time.monotonic(), copy.deepcopy(payload))


def _clear_live_models_cache() -> None:
    with _LIVE_MODELS_CACHE_LOCK:
        _LIVE_MODELS_CACHE.clear()

from api.config import (
    STATE_DIR,
    SESSION_DIR,
    DEFAULT_WORKSPACE,
    DEFAULT_MODEL,
    SESSIONS,
    SESSIONS_MAX,
    LOCK,
    STREAMS,
    STREAMS_LOCK,
    CANCEL_FLAGS,
    SERVER_START_TIME,
    _resolve_cli_toolsets,
    _INDEX_HTML_PATH,
    get_available_models,
    IMAGE_EXTS,
    MD_EXTS,
    MIME_MAP,
    MAX_FILE_BYTES,
    MAX_UPLOAD_BYTES,
    CHAT_LOCK,
    _get_session_agent_lock,
    SESSION_AGENT_LOCKS,
    SESSION_AGENT_LOCKS_LOCK,
    load_settings,
    save_settings,
    set_hermes_default_model,
    model_with_provider_context,
    get_reasoning_status,
    set_reasoning_display,
    set_reasoning_effort,
)
from api.helpers import (
    require,
    bad,
    safe_resolve,
    j,
    t,
    read_body,
    _security_headers,
    _sanitize_error,
    redact_session_data,
    _redact_text,
)


def _clear_stale_stream_state(session) -> bool:
    """Clear persisted streaming flags when the in-memory stream no longer exists.

    A server restart or worker crash can leave active_stream_id/pending_* in the
    session JSON while STREAMS is empty. The frontend then keeps reconnecting to
    a dead stream and shows a permanent running/thinking state.

    SAFETY (#1558): If ``session`` was loaded with ``metadata_only=True``, its
    ``messages`` array is empty by design and calling ``save()`` would
    atomically overwrite the on-disk JSON, wiping the conversation. In that
    case we re-load the full session before mutating, so the persisted
    write carries the real messages forward.
    """
    stream_id = getattr(session, "active_stream_id", None)
    if not stream_id:
        return False
    with STREAMS_LOCK:
        stream_alive = stream_id in STREAMS
    if stream_alive:
        return False

    # ── #1558 P0 safety: if we were handed a metadata-only stub, reload the
    # full session before touching persisted state. The original
    # metadata-only object is left untouched so the caller's read path is
    # unaffected.
    original_stub = session  # SHOULD-FIX #1 (Opus): keep reference so we can
                             # patch the caller's in-memory copy after a
                             # successful clear, avoiding one ghost SSE
                             # reconnect on the very next /api/session GET.
    if getattr(session, "_loaded_metadata_only", False):
        try:
            from api.models import get_session as _get_session
            session = _get_session(session.session_id, metadata_only=False)
        except Exception:
            # If we cannot upgrade to a full load (file gone, decode error,
            # etc.) bail without clearing — better to leave a stale
            # active_stream_id than to wipe the conversation.
            logger.warning(
                "_clear_stale_stream_state: refused to clear stale stream %s "
                "for session %s — full reload failed and we will not save a "
                "metadata-only stub. See #1558.",
                stream_id, getattr(session, "session_id", "?"),
            )
            return False
        if session is None:
            return False
        # The full-load path may have already repaired stale pending fields
        # via _repair_stale_pending(); only re-assert if still set.
        if not getattr(session, "active_stream_id", None):
            # Patch the caller's stub so its read path also sees the cleared
            # field (matches the Opus SHOULD-FIX #1 — without this, /api/session
            # would briefly return the stale active_stream_id and the frontend
            # would attempt one ghost SSE reconnect before recovering).
            try:
                original_stub.active_stream_id = None
                if hasattr(original_stub, "pending_user_message"):
                    original_stub.pending_user_message = None
                if hasattr(original_stub, "pending_attachments"):
                    original_stub.pending_attachments = []
                if hasattr(original_stub, "pending_started_at"):
                    original_stub.pending_started_at = None
            except Exception:
                pass
            return False

    # ── #1533 race fix: acquire the per-session lock and re-read
    # active_stream_id under it. A concurrent chat_start may have already
    # registered a new stream after our STREAMS_LOCK check above; in that
    # case we must NOT clobber its session.active_stream_id.
    with _get_session_agent_lock(session.session_id):
        if getattr(session, "active_stream_id", None) != stream_id:
            return False
        session.active_stream_id = None
        if hasattr(session, "pending_user_message"):
            session.pending_user_message = None
        if hasattr(session, "pending_attachments"):
            session.pending_attachments = []
        if hasattr(session, "pending_started_at"):
            session.pending_started_at = None
        try:
            session.save()
        except Exception:
            logger.exception(
                "_clear_stale_stream_state: save() failed for session %s",
                getattr(session, "session_id", "?"),
            )
    # Patch the caller's stub (if different from the full-load object) so
    # its in-memory active_stream_id matches what just got persisted.
    if original_stub is not session:
        try:
            original_stub.active_stream_id = None
            if hasattr(original_stub, "pending_user_message"):
                original_stub.pending_user_message = None
            if hasattr(original_stub, "pending_attachments"):
                original_stub.pending_attachments = []
            if hasattr(original_stub, "pending_started_at"):
                original_stub.pending_started_at = None
        except Exception:
            pass
    return True

# ── CSRF: validate Origin/Referer on POST ────────────────────────────────────
import re as _re


def _normalize_host_port(value: str) -> tuple[str, str | None]:
    """Split a host or host:port string into (hostname, port|None).
    Handles IPv6 bracket notation, e.g. [::1]:8080."""
    value = value.strip().lower()
    if not value:
        return '', None
    if value.startswith('['):
        end = value.find(']')
        if end != -1:
            host = value[1:end]
            rest = value[end + 1 :]
            if rest.startswith(':') and rest[1:].isdigit():
                return host, rest[1:]
            return host, None
    if value.count(':') == 1:
        host, port = value.rsplit(':', 1)
        if port.isdigit():
            return host, port
    return value, None


def _ports_match(origin_scheme: str, origin_port: str | None, allowed_port: str | None) -> bool:
    """Return True when two ports should be considered equivalent, scheme-aware.

    Treats an absent port as the scheme default: port 80 for http, port 443 for https.
    Port 80 is NOT treated as equivalent to 443 (different protocols = different origins).
    """
    if origin_port == allowed_port:
        return True
    # Determine the default port for the origin's scheme
    default = '443' if origin_scheme == 'https' else '80'
    if not origin_port and allowed_port == default:
        return True
    if not allowed_port and origin_port == default:
        return True
    return False


def _allowed_public_origins() -> set[str]:
    """Parse HERMES_WEBUI_ALLOWED_ORIGINS env var (comma-separated) into a set.

    Each entry must include the scheme, e.g. https://myapp.example.com:8000.
    Entries without a scheme are silently skipped and a warning is printed.
    """
    raw = os.getenv('HERMES_WEBUI_ALLOWED_ORIGINS', '')
    result = set()
    for value in raw.split(','):
        value = value.strip().rstrip('/').lower()
        if not value:
            continue
        if not (value.startswith('http://') or value.startswith('https://')):
            import sys
            print(
                f"[webui] WARNING: HERMES_WEBUI_ALLOWED_ORIGINS entry {value!r} is missing "
                f"the scheme (expected https://hostname or http://hostname). Entry ignored.",
                flush=True, file=sys.stderr,
            )
            continue
        result.add(value)
    return result


def _check_csrf(handler) -> bool:
    """Reject cross-origin POST requests. Returns True if OK."""
    origin = handler.headers.get("Origin", "")
    referer = handler.headers.get("Referer", "")
    host = handler.headers.get("Host", "")
    if not origin and not referer:
        return True  # non-browser clients (curl, agent) have no Origin
    target = origin or referer
    # Extract host:port from origin/referer
    m = _re.match(r"^https?://([^/]+)", target)
    if not m:
        return False
    origin_host = m.group(1)
    origin_scheme = m.group(0).split('://')[0].lower()  # 'http' or 'https'
    origin_name, origin_port = _normalize_host_port(origin_host)
    # Check against explicitly allowed public origins (env var)
    origin_value = m.group(0).rstrip('/').lower()
    if origin_value in _allowed_public_origins():
        return True
    # Allow same-origin: check Host, X-Forwarded-Host (reverse proxy), and
    # X-Real-Host against the origin. Reverse proxies (Caddy, nginx) set
    # X-Forwarded-Host to the client's original Host header.
    allowed_hosts = [
        h.strip()
        for h in [
            host,
            handler.headers.get("X-Forwarded-Host", ""),
            handler.headers.get("X-Real-Host", ""),
        ]
        if h.strip()
    ]
    for allowed in allowed_hosts:
        allowed_name, allowed_port = _normalize_host_port(allowed)
        if origin_name == allowed_name and _ports_match(origin_scheme, origin_port, allowed_port):
            return True
    return False


def _normalize_provider_id(value: str | None) -> str:
    raw = str(value or "").strip().lower()
    if not raw:
        return ""
    if raw in _PROVIDER_ALIASES:
        return _PROVIDER_ALIASES[raw]
    for prefix, normalized in (
        ("openai-codex", "openai"),
        ("openai", "openai"),
        ("anthropic", "anthropic"),
        ("claude", "anthropic"),
        ("google", "google"),
        ("gemini", "google"),
        ("openrouter", "openrouter"),
        ("custom", "custom"),
    ):
        if raw.startswith(prefix):
            return normalized
    # Unknown prefix — return empty so callers treat it as "no match" and pass
    # the model through unchanged rather than incorrectly stripping it.
    return "" 


def _catalog_provider_id_sets(catalog: dict) -> tuple[set[str], set[str]]:
    raw_provider_ids: set[str] = set()
    normalized_provider_ids: set[str] = set()
    for group in catalog.get("groups") or []:
        raw = str(group.get("provider_id") or "").strip().lower()
        if not raw:
            continue
        raw_provider_ids.add(raw)
        normalized = _normalize_provider_id(raw)
        if normalized:
            normalized_provider_ids.add(normalized)
    return raw_provider_ids, normalized_provider_ids


def _catalog_has_provider(
    provider_raw: str,
    provider_normalized: str,
    raw_provider_ids: set[str],
    normalized_provider_ids: set[str],
) -> bool:
    return (
        provider_raw in raw_provider_ids
        or (provider_normalized and provider_normalized in raw_provider_ids)
        or (provider_normalized and provider_normalized in normalized_provider_ids)
    )


def _model_matches_active_provider_family(
    model: str,
    active_provider: str,
) -> bool:
    model_lower = model.lower()
    for bare_prefix in ("gpt", "claude", "gemini"):
        if model_lower.startswith(bare_prefix):
            return _normalize_provider_id(bare_prefix) == active_provider
    return False


def _catalog_model_id_matches(candidate: str, model: str) -> bool:
    candidate = str(candidate or "").strip()
    if candidate.startswith("@") and ":" in candidate:
        candidate = candidate.rsplit(":", 1)[1]
    if "/" in candidate:
        candidate = candidate.split("/", 1)[1]
    return candidate.replace("-", ".").lower() == model.replace("-", ".").lower()


def _clean_session_model_provider(value: str | None) -> str | None:
    provider = str(value or "").strip().lower()
    if not provider or provider == "default":
        return None
    if provider.startswith("@"):
        provider = provider[1:]
    return provider or None


def _split_provider_qualified_model(model: str) -> tuple[str, str | None]:
    model = str(model or "").strip()
    if model.startswith("@") and ":" in model:
        provider_hint, bare_model = model[1:].rsplit(":", 1)
        provider = _clean_session_model_provider(provider_hint)
        bare = bare_model.strip()
        if provider and bare:
            return bare, provider
    return model, None


def _should_attach_codex_provider_context(model: str, raw_active_provider: str, catalog: dict) -> bool:
    """Return True when a bare Codex model needs separate provider context.

    OpenAI, OpenAI Codex, Copilot, and OpenRouter can all expose GPT-looking
    bare names. If a session stores only ``gpt-...`` while Codex is active, a
    later provider-list/default-model round trip can lose the user's Codex
    choice. Store the provider separately instead of converting the persisted
    model to ``@openai-codex:model``.
    """
    if raw_active_provider != "openai-codex":
        return False
    if not model.lower().startswith("gpt"):
        return False
    for group in catalog.get("groups") or []:
        if str(group.get("provider_id") or "").strip().lower() != "openai-codex":
            continue
        return any(
            _catalog_model_id_matches(entry.get("id"), model)
            for entry in group.get("models", [])
            if isinstance(entry, dict)
        )
    return False


def _resolve_compatible_session_model_state(
    model_id: str | None,
    model_provider: str | None = None,
) -> tuple[str, str | None, bool]:
    """Return (effective_model, effective_provider, model_was_normalized).

    Sessions can outlive provider changes. When an older session still points at
    a different provider namespace (for example `gemini/...` after switching the
    agent to OpenAI Codex), reusing that stale model causes chat startup to hit
    the wrong backend and fail. Normalize only obvious cross-provider mismatches.
    When a model has an explicit provider context, keep the model string itself
    in its picker/API shape and carry the provider as separate state.
    """
    catalog = get_available_models()
    default_model = str(catalog.get("default_model") or DEFAULT_MODEL or "").strip()
    model = str(model_id or "").strip()
    requested_provider = _clean_session_model_provider(model_provider)
    if not model:
        return default_model, requested_provider, bool(default_model)

    active_provider = _normalize_provider_id(catalog.get("active_provider"))
    # Also keep the raw active_provider slug for cross-provider detection with
    # non-listed providers (ollama-cloud, deepseek, xai, etc.) that _normalize_provider_id
    # returns "" for. If the raw provider is set but normalization returned "", we still
    # want to detect that a session model from a known provider (e.g. openai/gpt-5.4-mini)
    # is stale relative to this unknown active provider. (#1023)
    raw_active_provider = str(catalog.get("active_provider") or "").strip().lower()
    if not active_provider and not raw_active_provider:
        bare_model, explicit_provider = _split_provider_qualified_model(model)
        return model, explicit_provider or requested_provider, False

    bare_for_context, explicit_provider = _split_provider_qualified_model(model)
    if requested_provider and not explicit_provider:
        return model, requested_provider, False

    if model.startswith("@") and ":" in model:
        provider_raw = explicit_provider or ""
        provider_normalized = _normalize_provider_id(provider_raw)
        bare_model = bare_for_context.strip()
        if not provider_raw or not bare_model:
            return model, requested_provider, False

        raw_provider_ids, normalized_provider_ids = _catalog_provider_id_sets(catalog)
        hint_matches_active = (
            provider_raw == raw_active_provider
            or provider_raw == active_provider
            or (provider_normalized and provider_normalized == active_provider)
        )
        if hint_matches_active:
            # The @provider:model hint explicitly names the active provider, so this
            # selection is intentional — not a stale cross-provider artifact. Return
            # the full @provider:model string unchanged so downstream (resolve_model_provider
            # in config.py) can route through the correct provider. Stripping the prefix
            # here would collapse duplicate model IDs from different providers back to the
            # bare ID, causing the first matching provider to win on the next UI render
            # and the wrong provider to be used for the agent run. (#1253)
            return model, provider_raw, False

        if _catalog_has_provider(
            provider_raw,
            provider_normalized,
            raw_provider_ids,
            normalized_provider_ids,
        ):
            return model, provider_raw, False

        if _model_matches_active_provider_family(bare_model, active_provider):
            provider_context = (
                raw_active_provider
                if _should_attach_codex_provider_context(bare_model, raw_active_provider, catalog)
                else None
            )
            return bare_model, provider_context, True
        if default_model:
            provider_context = (
                raw_active_provider
                if _should_attach_codex_provider_context(default_model, raw_active_provider, catalog)
                else None
            )
            return default_model, provider_context, True
        return model, provider_raw, False

    slash = model.find("/")
    if slash < 0:
        model_lower = model.lower()
        for bare_prefix in ("gpt", "claude", "gemini"):
            if model_lower.startswith(bare_prefix):
                model_provider = _normalize_provider_id(bare_prefix)
                if model_provider and model_provider != active_provider and default_model:
                    provider_context = (
                        raw_active_provider
                        if _should_attach_codex_provider_context(default_model, raw_active_provider, catalog)
                        else None
                    )
                    return default_model, provider_context, True
                provider_context = (
                    raw_active_provider
                    if _should_attach_codex_provider_context(model, raw_active_provider, catalog)
                    else requested_provider
                )
                return model, provider_context, False
        return model, requested_provider, False

    model_provider = _normalize_provider_id(model[:slash])

    # For custom/openrouter active providers: only skip normalization when the
    # model's namespace prefix is actually routable by a group in the catalog.
    # A user who only has custom_providers configured (active_provider="custom")
    # with a stale session model like "openai/gpt-5.4-mini" would otherwise
    # never get cleaned up, causing "(unavailable)" to appear in the picker.
    if active_provider in {"custom", "openrouter"}:
        # These namespaces are always routable as-is — preserve them.
        if model_provider in {"", "custom", "openrouter"}:
            return model, requested_provider, False
        # Check if any catalog group can actually route this model's prefix.
        groups = catalog.get("groups") or []
        routable_provider_ids = {
            _normalize_provider_id(g.get("provider_id") or "") for g in groups
        }
        # openrouter group can route any provider/model namespace
        has_openrouter_group = any(
            (g.get("provider_id") or "") == "openrouter" for g in groups
        )
        if model_provider in routable_provider_ids or has_openrouter_group:
            return model, requested_provider, False
        # Model prefix is not routable — stale cross-provider reference, clear it.
        if default_model:
            return default_model, requested_provider, True
        return model, requested_provider, False

    # Skip normalization for models on custom/openrouter namespaces — these are
    # user-controlled and should never be silently replaced.
    # Also normalize when the model is from a known provider but the active provider
    # is an unlisted one (e.g. ollama-cloud) — active_provider is "" in that case
    # but raw_active_provider is set. If model_provider doesn't start with the raw
    # active provider name, the session model is stale. (#1023)
    _active_for_compare = active_provider or raw_active_provider
    if model_provider and model_provider not in {"", "custom", "openrouter"} and model_provider != _active_for_compare and default_model:
        return default_model, requested_provider, True
    return model, requested_provider, False


def _resolve_compatible_session_model(model_id: str | None) -> tuple[str, bool]:
    """Return (effective_model, model_was_normalized) for legacy callers."""
    effective_model, _provider, changed = _resolve_compatible_session_model_state(model_id)
    return effective_model, changed


def _normalize_session_model_in_place(session) -> str:
    original_model = getattr(session, "model", None) or ""
    original_provider = _clean_session_model_provider(
        getattr(session, "model_provider", None)
    )
    effective_model, effective_provider, changed = _resolve_compatible_session_model_state(
        original_model or None,
        original_provider,
    )
    provider_changed = effective_provider != original_provider
    # Only persist the correction if the session had an explicit model that needed changing.
    # Sessions with no model stored (empty/None) get the effective default returned without
    # a disk write — no need to rebuild the index for a fill-in-blank operation.
    if original_model and effective_model and (
        (changed and original_model != effective_model) or provider_changed
    ):
        if changed and original_model != effective_model:
            session.model = effective_model
        session.model_provider = effective_provider
        session.save(touch_updated_at=False)
    return effective_model


def _resolve_effective_session_model_for_display(session) -> str:
    """Resolve the model a session should display without mutating persisted state.

    `GET /api/session` should stay side-effect free. If a stale persisted model
    needs normalization for the current provider configuration, return the
    effective model for the response payload only and leave disk state alone.
    """
    original_model = getattr(session, "model", None) or ""
    effective_model, _provider, _changed = _resolve_compatible_session_model_state(
        original_model or None,
        getattr(session, "model_provider", None),
    )
    return effective_model or original_model

def _resolve_effective_session_model_provider_for_display(session) -> str | None:
    original_model = getattr(session, "model", None) or ""
    _model, provider, _changed = _resolve_compatible_session_model_state(
        original_model or None,
        getattr(session, "model_provider", None),
    )
    return provider


def _session_model_state_from_request(
    model: str | None,
    requested_provider: str | None,
    current_provider: str | None = None,
) -> tuple[str | None, str | None]:
    model_value = str(model).strip() if model is not None else None
    provider = (
        _clean_session_model_provider(requested_provider)
        if requested_provider is not None
        else None
    )
    if model_value:
        _bare, explicit_provider = _split_provider_qualified_model(model_value)
        if explicit_provider:
            provider = explicit_provider
        elif requested_provider is None:
            provider = _clean_session_model_provider(current_provider)
        model_value, provider, _changed = _resolve_compatible_session_model_state(
            model_value,
            provider,
        )
    return model_value, provider


def _lookup_gateway_session_identity(session_id: str) -> dict:
    if not session_id:
        return {}
    metadata = _load_gateway_session_identity_map().get(str(session_id))
    return metadata if isinstance(metadata, dict) else {}


def _lookup_cli_session_metadata(session_id: str) -> dict:
    if not session_id:
        return {}
    try:
        for row in get_cli_sessions():
            if row.get("session_id") == session_id:
                return row
    except Exception:
        return {}
    return {}


def _messaging_session_identity(session: dict, raw_source: str) -> str:
    metadata = _lookup_gateway_session_identity(session.get("session_id"))
    session_key = _safe_first(
        metadata.get("session_key"),
        session.get("session_key"),
        session.get("gateway_session_key"),
    )
    if session_key:
        return f"{raw_source}|session_key:{session_key}"

    chat_id = _safe_first(
        metadata.get("chat_id"),
        session.get("chat_id"),
        session.get("origin_chat_id"),
    )
    thread_id = _safe_first(metadata.get("thread_id"), session.get("thread_id"))
    chat_type = _safe_first(metadata.get("chat_type"), session.get("chat_type"))
    user_id = _safe_first(
        metadata.get("user_id"),
        session.get("user_id"),
        session.get("origin_user_id"),
    )

    identity_parts = []
    if chat_type:
        identity_parts.append(f"chat_type:{chat_type}")
    if chat_id:
        identity_parts.append(f"chat_id:{chat_id}")
    if thread_id:
        identity_parts.append(f"thread_id:{thread_id}")
    if user_id:
        identity_parts.append(f"user_id:{user_id}")

    if identity_parts:
        return f"{raw_source}|" + "|".join(identity_parts)
    return raw_source


def _session_messaging_raw_source(session: dict) -> str:
    raw = _safe_first(
        session.get("raw_source"),
        session.get("source_tag"),
        session.get("source"),
        session.get("platform"),
    )
    if not raw:
        raw = session.get("source_label") or "messaging"
    return _normalize_messaging_source(raw)


def _has_durable_messaging_identity(session: dict) -> bool:
    metadata = _lookup_gateway_session_identity(session.get("session_id"))
    return bool(_safe_first(
        metadata.get("session_key"),
        session.get("session_key"),
        session.get("gateway_session_key"),
        metadata.get("chat_id"),
        session.get("chat_id"),
        session.get("origin_chat_id"),
        metadata.get("thread_id"),
        session.get("thread_id"),
    ))


def _numeric_count(value) -> int:
    try:
        return int(float(_safe_first(value, 0) or 0))
    except (TypeError, ValueError):
        return 0


def _should_hide_stale_messaging_session(
    session: dict,
    active_gateway_session_ids: set[str],
    active_gateway_sources: set[str],
) -> bool:
    """Hide stale Gateway-owned internal rows after an external chat moved on.

    Hermes Gateway keeps the external conversation identity in sessions.json.
    Compression/session-reset can leave old Agent state.db rows behind; those
    rows are implementation segments, not distinct conversations users chose.
    Only apply this aggressive hiding when Gateway is currently advertising an
    active session for the same messaging source. Without that source-of-truth
    file we keep the old fallback behavior.
    """
    raw_source = _session_messaging_raw_source(session)
    if not _is_known_messaging_source(raw_source):
        return False
    if not active_gateway_session_ids or raw_source not in active_gateway_sources:
        return False

    sid = _safe_first(session.get("session_id"))
    if sid and sid in active_gateway_session_ids:
        return False

    if _safe_first(session.get("end_reason")) in _STALE_MESSAGING_END_REASONS:
        return True

    if not _has_durable_messaging_identity(session):
        return True

    if session.get("parent_session_id"):
        return True

    message_count = _numeric_count(session.get("message_count"))
    actual_count = _numeric_count(session.get("actual_message_count"))
    if message_count <= 0 and actual_count <= 0:
        return True

    return False


def _is_messaging_session_record(session) -> bool:
    """Return true for sessions backed by external messaging channels."""
    if not session:
        return False
    if (
        (getattr(session, "session_source", None) if not isinstance(session, dict) else session.get("session_source")) == "messaging"
    ):
        return True
    raw = _safe_first(
        getattr(session, "raw_source", None) if not isinstance(session, dict) else session.get("raw_source"),
        getattr(session, "source_tag", None) if not isinstance(session, dict) else session.get("source_tag"),
        getattr(session, "source", None) if not isinstance(session, dict) else session.get("source"),
        session.get("source_label") if isinstance(session, dict) else None,
    )
    return _is_known_messaging_source(raw)


def _is_messaging_session_id(sid: str) -> bool:
    """Detect messaging-backed sessions from WebUI metadata or Agent rows."""
    try:
        session = Session.load(sid)
        if _is_messaging_session_record(session):
            return True
    except Exception:
        pass
    return _is_messaging_session_record(_lookup_cli_session_metadata(sid))


def _session_sort_timestamp(session: dict) -> float:
    return float(
        _safe_first(
            session.get("last_message_at"),
            session.get("updated_at"),
            session.get("created_at"),
            session.get("started_at"),
            0,
        ) or 0
    ) or 0.0


def _merge_cli_sidebar_metadata(ui_session: dict, cli_meta: dict) -> dict:
    """Merge source-of-truth CLI metadata into a sidebar session row.

    Preserve UI-owned state (archived/pinned) while replacing metadata that can
    legitimately drift in WebUI snapshots.
    """
    if not ui_session:
        return ui_session
    if not cli_meta:
        return dict(ui_session)
    merged = dict(ui_session)
    merged["is_cli_session"] = True
    for key in (
        "source_tag",
        "raw_source",
        "session_source",
        "source_label",
        "user_id",
        "chat_id",
        "chat_type",
        "thread_id",
        "session_key",
        "platform",
        "parent_session_id",
        "end_reason",
        "actual_message_count",
        "_lineage_root_id",
        "_lineage_tip_id",
        "_compression_segment_count",
    ):
        value = _safe_first(cli_meta.get(key))
        if value:
            merged[key] = value

    if cli_meta.get("created_at") is not None:
        merged["created_at"] = cli_meta["created_at"]
    if cli_meta.get("updated_at") is not None:
        merged["updated_at"] = cli_meta["updated_at"]
    if cli_meta.get("last_message_at") is not None:
        merged["last_message_at"] = cli_meta["last_message_at"]
    if cli_meta.get("message_count") is not None:
        merged["message_count"] = cli_meta["message_count"]
    elif cli_meta.get("actual_message_count") is not None:
        merged["message_count"] = cli_meta["actual_message_count"]

    if cli_meta.get("title"):
        current_title = merged.get("title")
        if not current_title or current_title == "Untitled":
            merged["title"] = cli_meta["title"]

    if cli_meta.get("model"):
        if not merged.get("model") or merged.get("model") == "unknown":
            merged["model"] = cli_meta["model"]
    return merged


def _messaging_source_key(session: dict) -> str | None:
    raw = _session_messaging_raw_source(session)
    if not _is_known_messaging_source(raw):
        return None
    return _messaging_session_identity(session, raw)


def _keep_latest_messaging_session_per_source(sessions: list[dict]) -> list[dict]:
    """Keep only the newest sidebar row per messaging session identity."""
    gateway_metadata = _load_gateway_session_identity_map()
    active_gateway_session_ids = {str(sid) for sid in gateway_metadata.keys() if sid}
    active_gateway_sources = {
        _normalize_messaging_source(_safe_first(meta.get("raw_source"), meta.get("platform")))
        for meta in gateway_metadata.values()
        if isinstance(meta, dict)
    }
    active_gateway_sources = {source for source in active_gateway_sources if _is_known_messaging_source(source)}

    kept_sources: set[str] = set()
    best_by_source: dict[str, dict] = {}
    kept: list[dict] = []
    for session in sessions:
        key = _messaging_source_key(session)
        if not key:
            kept.append(session)
            continue
        if _should_hide_stale_messaging_session(session, active_gateway_session_ids, active_gateway_sources):
            continue
        if key in kept_sources:
            kept_sources.add(key)
            current = best_by_source.get(key)
            if current is None or _session_sort_timestamp(session) > _session_sort_timestamp(current):
                best_by_source[key] = session
            continue
        kept_sources.add(key)
        best_by_source[key] = session

    kept.extend(best_by_source.values())
    kept.sort(key=_session_sort_timestamp, reverse=True)
    return kept


from api.models import (
    Session,
    get_session,
    new_session,
    all_sessions,
    title_from,
    _write_session_index,
    SESSION_INDEX_FILE,
    load_projects,
    save_projects,
    import_cli_session,
    get_cli_sessions,
    get_cli_session_messages,
    ensure_cron_project,
    is_cron_session,
)
from api.workspace import (
    load_workspaces,
    save_workspaces,
    get_last_workspace,
    set_last_workspace,
    list_dir,
    list_workspace_suggestions,
    read_file_content,
    safe_resolve_ws,
    resolve_trusted_workspace,
    validate_workspace_to_add,
    _is_blocked_system_path,
    _workspace_blocked_roots,
)
from api.upload import handle_upload, handle_upload_extract, handle_transcribe
from api.streaming import _sse, _run_agent_streaming, cancel_stream
from api.providers import get_providers, set_provider_key, remove_provider_key
from api.onboarding import (
    apply_onboarding_setup,
    get_onboarding_status,
    complete_onboarding,
    probe_provider_endpoint,
)

# Approval system (optional -- graceful fallback if agent not available)
try:
    from tools.approval import (
        submit_pending as _submit_pending_raw,
        approve_session,
        approve_permanent,
        save_permanent_allowlist,
        is_approved,
        _pending,
        _lock,
        _permanent_approved,
        resolve_gateway_approval,
        enable_session_yolo,
        disable_session_yolo,
        is_session_yolo_enabled,
    )
except ImportError:
    _submit_pending_raw = lambda *a, **k: None
    approve_session = lambda *a, **k: None
    approve_permanent = lambda *a, **k: None
    save_permanent_allowlist = lambda *a, **k: None
    is_approved = lambda *a, **k: True
    resolve_gateway_approval = lambda *a, **k: 0
    enable_session_yolo = lambda *a, **k: None
    disable_session_yolo = lambda *a, **k: None
    is_session_yolo_enabled = lambda *a, **k: False
    _pending = {}
    _lock = threading.Lock()
    _permanent_approved = set()


# ── Approval SSE subscribers (long-connection push) ──────────────────────────
_approval_sse_subscribers: dict[str, list[queue.Queue]] = {}


def _approval_sse_subscribe(session_id: str) -> queue.Queue:
    """Register an SSE subscriber for approval events on a given session."""
    q = queue.Queue(maxsize=16)
    with _lock:
        _approval_sse_subscribers.setdefault(session_id, []).append(q)
    return q


def _approval_sse_unsubscribe(session_id: str, q: queue.Queue) -> None:
    """Remove an SSE subscriber."""
    with _lock:
        subs = _approval_sse_subscribers.get(session_id)
        if subs and q in subs:
            subs.remove(q)
            if not subs:
                _approval_sse_subscribers.pop(session_id, None)


def _approval_sse_notify_locked(session_id: str, head: dict | None, total: int) -> None:
    """Push an approval event to all SSE subscribers for a session.

    CALLER MUST HOLD `_lock`. Snapshots the subscriber list under the held
    lock and then calls `q.put_nowait()` on each (which is itself thread-safe).

    `head` is the approval entry currently at the head of the queue (the one
    the UI should display) — NOT the just-appended entry. With multiple
    parallel approvals (#527), the just-appended entry is at the TAIL, but
    `/api/approval/pending` always returns the HEAD, so SSE must match.

    `total` is the total number of pending approvals.

    Pass `head=None` and `total=0` when the queue has just been emptied (e.g.
    `_handle_approval_respond` popped the last entry) so the client knows to
    hide its approval card.
    """
    payload = {"pending": dict(head) if head else None, "pending_count": total}
    subs = _approval_sse_subscribers.get(session_id, ())
    for q in subs:
        try:
            q.put_nowait(payload)
        except queue.Full:
            pass  # drop if subscriber is slow (bounded queue prevents memory leak)


def _approval_sse_notify(session_id: str, head: dict | None, total: int) -> None:
    """Convenience wrapper that takes `_lock` itself.

    Use only from contexts that don't already hold `_lock`. Production call
    sites (submit_pending, _handle_approval_respond) MUST hold the lock and
    call `_approval_sse_notify_locked` directly to avoid a notify-ordering
    race where a later append's notify can fire before an earlier append's
    notify (resulting in stale `pending_count`).
    """
    with _lock:
        _approval_sse_notify_locked(session_id, head, total)


def submit_pending(session_key: str, approval: dict) -> None:
    """Append a pending approval to the per-session queue.

    Wraps the agent's submit_pending to:
    - Add a stable approval_id (uuid4 hex) so the respond endpoint can target
      a specific entry even when multiple approvals are queued simultaneously.
    - Change the storage from a single overwriting dict value to a list, so
      parallel tool calls each get their own approval slot (fixes #527).
    - Notify any connected SSE subscribers immediately.
    """
    entry = dict(approval)
    entry.setdefault("approval_id", uuid.uuid4().hex)
    with _lock:
        queue_list = _pending.setdefault(session_key, [])
        # Replace a legacy non-list value if the agent version uses the old pattern.
        if not isinstance(queue_list, list):
            _pending[session_key] = [queue_list]
            queue_list = _pending[session_key]
        queue_list.append(entry)
        total = len(queue_list)
        head = queue_list[0]  # /api/approval/pending always returns head
        # Push to SSE subscribers from inside _lock so two parallel
        # submit_pending calls can't deliver out-of-order (T2's later
        # notify arriving before T1's earlier notify with a stale count).
        _approval_sse_notify_locked(session_key, head, total)
    # NOTE: We do NOT call _submit_pending_raw here — that function overwrites
    # _pending[session_key] with a single dict, which would undo the list we just
    # built. The gateway blocking path uses _gateway_queues (a separate mechanism
    # managed by check_all_command_guards / register_gateway_notify), which is
    # unaffected by _pending. The _pending dict is only used for UI polling.

# Clarify prompts (optional -- graceful fallback if agent not available)
try:
    from api.clarify import (
        submit_pending as submit_clarify_pending,
        get_pending as get_clarify_pending,
        resolve_clarify,
        sse_subscribe as clarify_sse_subscribe,
        sse_unsubscribe as clarify_sse_unsubscribe,
    )
except ImportError:
    submit_clarify_pending = lambda *a, **k: None
    get_clarify_pending = lambda *a, **k: None
    clarify_sse_subscribe = None
    resolve_clarify = lambda *a, **k: 0


# ── Login page locale strings ─────────────────────────────────────────────────
# Add entries here to support more languages on the login page.
# The key must match the 'language' setting value (from static/i18n.js LOCALES).
_LOGIN_LOCALE = {
    "en": {
        "lang": "en",
        "title": "Sign in",
        "subtitle": "Enter your password to continue",
        "placeholder": "Password",
        "btn": "Sign in",
        "invalid_pw": "Invalid password",
        "conn_failed": "Connection failed",
    },
    "es": {
        "lang": "es-ES",
        "title": "Iniciar sesi\u00f3n",
        "subtitle": "Introduce tu contrase\u00f1a para continuar",
        "placeholder": "Contrase\u00f1a",
        "btn": "Entrar",
        "invalid_pw": "Contrase\u00f1a inv\u00e1lida",
        "conn_failed": "Error de conexi\u00f3n",
    },
    "de": {
        "lang": "de-DE",
        "title": "Anmelden",
        "subtitle": "Geben Sie Ihr Passwort ein, um fortzufahren",
        "placeholder": "Passwort",
        "btn": "Anmelden",
        "invalid_pw": "Ung\u00fcltiges Passwort",
        "conn_failed": "Verbindung fehlgeschlagen",
    },
    "ru": {
        "lang": "ru-RU",
        "title": "\u0412\u043e\u0439\u0442\u0438",
        "subtitle": "\u0412\u0432\u0435\u0434\u0438\u0442\u0435 \u043f\u0430\u0440\u043e\u043b\u044c, \u0447\u0442\u043e\u0431\u044b \u043f\u0440\u043e\u0434\u043e\u043b\u0436\u0438\u0442\u044c",
        "placeholder": "\u041f\u0430\u0440\u043e\u043b\u044c",
        "btn": "\u0412\u043e\u0439\u0442\u0438",
        "invalid_pw": "\u041d\u0435\u0432\u0435\u0440\u043d\u044b\u0439 \u043f\u0430\u0440\u043e\u043b\u044c",
        "conn_failed": "\u041d\u0435 \u0443\u0434\u0430\u043b\u043e\u0441\u044c \u043f\u043e\u0434\u043a\u043b\u044e\u0447\u0438\u0442\u044c\u0441\u044f",
    },
    "zh": {
        "lang": "zh-CN",
        "title": "\u767b\u5f55",
        "subtitle": "\u8f93\u5165\u5bc6\u7801\u7ee7\u7eed\u4f7f\u7528",
        "placeholder": "\u5bc6\u7801",
        "btn": "\u767b\u5f55",
        "invalid_pw": "\u5bc6\u7801\u9519\u8bef",
        "conn_failed": "\u8fde\u63a5\u5931\u8d25",
    },
    "zh-Hant": {
        "lang": "zh-TW",
        "title": "\u767b\u5f55",
        "subtitle": "\u8f38\u5165\u5bc6\u78bc\u7e7c\u7e8c\u4f7f\u7528",
        "placeholder": "\u5bc6\u78bc",
        "btn": "\u767b\u5f55",
        "invalid_pw": "\u5bc6\u78bc\u932f\u8aa4",
        "conn_failed": "\u9023\u63a5\u5931\u6557",
    },
    # Strings mirror static/i18n.js login_* keys for the corresponding locale.
    # See issue #1442. When adding a new locale to LOCALES in i18n.js, also add
    # the matching entry here — tests/test_login_locale_parity.py enforces this.
    "ja": {
        "lang": "ja-JP",
        "title": "\u30b5\u30a4\u30f3\u30a4\u30f3",
        "subtitle": "\u30d1\u30b9\u30ef\u30fc\u30c9\u3092\u5165\u529b\u3057\u3066\u7d9a\u884c",
        "placeholder": "\u30d1\u30b9\u30ef\u30fc\u30c9",
        "btn": "\u30b5\u30a4\u30f3\u30a4\u30f3",
        "invalid_pw": "\u30d1\u30b9\u30ef\u30fc\u30c9\u304c\u7121\u52b9\u3067\u3059",
        "conn_failed": "\u63a5\u7d9a\u5931\u6557",
    },
    "pt": {
        "lang": "pt-BR",
        "title": "Entrar",
        "subtitle": "Digite sua senha para continuar",
        "placeholder": "Senha",
        "btn": "Entrar",
        "invalid_pw": "Senha inv\u00e1lida",
        "conn_failed": "Falha na conex\u00e3o",
    },
    "ko": {
        "lang": "ko-KR",
        "title": "\ub85c\uadf8\uc778",
        "subtitle": "\uacc4\uc18d\ud558\ub824\uba74 \ube44\ubc00\ubc88\ud638\ub97c \uc785\ub825\ud558\uc138\uc694",
        "placeholder": "\ube44\ubc00\ubc88\ud638",
        "btn": "\ub85c\uadf8\uc778",
        "invalid_pw": "\ube44\ubc00\ubc88\ud638\uac00 \uc62c\ubc14\ub974\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4",
        "conn_failed": "\uc5f0\uacb0 \uc2e4\ud328",
    },
}


def _resolve_login_locale_key(raw_lang: str | None) -> str:
    """Resolve settings.language to a known _LOGIN_LOCALE key."""
    if not raw_lang:
        return "en"
    lang = str(raw_lang).strip()
    if not lang:
        return "en"
    if lang in _LOGIN_LOCALE:
        return lang

    normalized = lang.replace("_", "-")
    lower = normalized.lower()

    # Case-insensitive direct key match first.
    for key in _LOGIN_LOCALE:
        if key.lower() == lower:
            return key

    # Common Chinese aliases.
    if lower == "zh" or lower.startswith("zh-cn") or lower.startswith("zh-sg") or lower.startswith("zh-hans"):
        return "zh"
    if lower.startswith("zh-tw") or lower.startswith("zh-hk") or lower.startswith("zh-mo") or lower.startswith("zh-hant"):
        return "zh-Hant" if "zh-Hant" in _LOGIN_LOCALE else "zh"

    # Fallback to base language subtag (e.g. en-US -> en).
    base = lower.split("-", 1)[0]
    for key in _LOGIN_LOCALE:
        if key.lower() == base:
            return key
    return "en"

# ── Login page (self-contained, no external deps) ────────────────────────────
_LOGIN_PAGE_HTML = """<!doctype html>
<html lang="{{LANG}}"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
<title>{{BOT_NAME}} — {{LOGIN_TITLE}}</title>
<style>
*{box-sizing:border-box;margin:0;padding:0}
body{background:#1a1a2e;color:#e8e8f0;font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",system-ui,sans-serif;
  height:100vh;display:flex;align-items:center;justify-content:center}
.card{background:#16213e;border:1px solid rgba(255,255,255,.08);border-radius:16px;padding:36px 32px;
  width:320px;text-align:center;box-shadow:0 8px 32px rgba(0,0,0,.3)}
.logo{width:48px;height:48px;border-radius:12px;background:linear-gradient(145deg,#e8a030,#e94560);
  display:flex;align-items:center;justify-content:center;font-weight:800;font-size:20px;color:#fff;
  margin:0 auto 12px;box-shadow:0 2px 12px rgba(233,69,96,.3)}
h1{font-size:18px;font-weight:600;margin-bottom:4px}
.sub{font-size:12px;color:#8888aa;margin-bottom:24px}
input{width:100%;padding:10px 14px;border-radius:10px;border:1px solid rgba(255,255,255,.1);
  background:rgba(255,255,255,.04);color:#e8e8f0;font-size:14px;outline:none;margin-bottom:14px;
  transition:border-color .15s}
input:focus{border-color:rgba(124,185,255,.5);box-shadow:0 0 0 3px rgba(124,185,255,.1)}
button{width:100%;padding:10px;border-radius:10px;border:none;background:rgba(124,185,255,.15);
  border:1px solid rgba(124,185,255,.3);color:#7cb9ff;font-size:14px;font-weight:600;cursor:pointer;
  transition:all .15s}
button:hover{background:rgba(124,185,255,.25)}
.err{color:#e94560;font-size:12px;margin-top:10px;display:none}
</style></head><body>
<div class="card">
  <div class="logo">{{BOT_NAME_INITIAL}}</div>
  <h1>{{BOT_NAME}}</h1>
  <p class="sub">{{LOGIN_SUBTITLE}}</p>
  <form id="login-form" data-invalid-pw="{{LOGIN_INVALID_PW}}" data-conn-failed="{{LOGIN_CONN_FAILED}}">
    <input type="password" id="pw" placeholder="{{LOGIN_PLACEHOLDER}}" autofocus>
    <button type="submit">{{LOGIN_BTN}}</button>
  </form>
  <div class="err" id="err"></div>
</div>
<script src="/static/login.js"></script>
</body></html>"""

# ── Insights endpoint ──────────────────────────────────────────────────────────

def _handle_insights(handler, parsed) -> bool:
    """Return usage analytics from local WebUI session data."""
    import collections
    import time as _time

    query = parse_qs(parsed.query)
    try:
        days = min(max(int(query.get("days", ["30"])[0]), 1), 365)
    except (ValueError, TypeError):
        days = 30

    now = _time.time()
    cutoff = now - (days * 86400)

    # Walk session index (fast, no full JSON parse)
    sessions_data = []
    idx_path = SESSION_DIR / "_index.json"
    if idx_path.exists():
        try:
            idx = json.loads(idx_path.read_text(encoding="utf-8"))
        except Exception:
            idx = []
    else:
        idx = []

    for entry in idx:
        created = entry.get("created_at", 0) or 0
        updated = entry.get("updated_at", 0) or 0
        # Session is relevant if it was created or updated within the window
        if max(created, updated) < cutoff:
            continue
        sessions_data.append(entry)

    # Aggregate
    total_sessions = len(sessions_data)
    total_messages = 0
    total_input_tokens = 0
    total_output_tokens = 0
    total_cost = 0.0
    model_counts = collections.Counter()
    # Activity by day of week (0=Mon .. 6=Sun)
    dow_activity = collections.Counter()
    # Activity by hour of day (0-23)
    hod_activity = collections.Counter()

    for s in sessions_data:
        total_messages += max(s.get("message_count", 0) or 0, 0)
        total_input_tokens += max(s.get("input_tokens", 0) or 0, 0)
        total_output_tokens += max(s.get("output_tokens", 0) or 0, 0)
        cost = s.get("estimated_cost")
        if cost is not None:
            try:
                total_cost += float(cost)
            except (ValueError, TypeError):
                pass
        model = s.get("model") or "unknown"
        if model:
            model_counts[model] += 1
        # Activity patterns
        ts = s.get("updated_at", s.get("created_at", 0)) or 0
        if ts:
            try:
                dt = _time.localtime(ts)
                dow_activity[dt.tm_wday] += 1
                hod_activity[dt.tm_hour] += 1
            except Exception:
                pass

    # Build model breakdown
    models_breakdown = []
    for model, count in model_counts.most_common():
        models_breakdown.append({"model": model, "sessions": count})

    # Day-of-week labels
    dow_labels = ["Mon", "Tue", "Wed", "Thu", "Fri", "Sat", "Sun"]
    dow_data = [{"day": dow_labels[i], "sessions": dow_activity.get(i, 0)} for i in range(7)]

    # Hour-of-day data
    hod_data = [{"hour": h, "sessions": hod_activity.get(h, 0)} for h in range(24)]

    return j(handler, {
        "period_days": days,
        "total_sessions": total_sessions,
        "total_messages": total_messages,
        "total_input_tokens": total_input_tokens,
        "total_output_tokens": total_output_tokens,
        "total_tokens": total_input_tokens + total_output_tokens,
        "total_cost": round(total_cost, 6),
        "models": models_breakdown,
        "activity_by_day": dow_data,
        "activity_by_hour": hod_data,
    })


# ── GET routes ────────────────────────────────────────────────────────────────


def handle_get(handler, parsed) -> bool:
    """Handle all GET routes. Returns True if handled, False for 404."""

    if parsed.path.startswith("/session/static/"):
        # Strip the leading "/session" so _serve_static() sees a path that
        # starts with "/static/" (its required prefix). _serve_static enforces
        # its own path-traversal sandbox via Path.resolve()+relative_to().
        stripped = parsed._replace(path=parsed.path[len("/session"):])
        return _serve_static(handler, stripped)

    if parsed.path in ("/", "/index.html") or parsed.path.startswith("/session/"):
        from urllib.parse import quote
        from api.updates import WEBUI_VERSION
        version_token = quote(WEBUI_VERSION, safe="")
        from api.extensions import inject_extension_tags

        html = _INDEX_HTML_PATH.read_text(encoding="utf-8").replace("__WEBUI_VERSION__", version_token)
        return t(
            handler,
            inject_extension_tags(html),
            content_type="text/html; charset=utf-8",
        )

    if parsed.path == "/login":
        _settings = load_settings()
        _bn = _html.escape(_settings.get("bot_name") or "Hermes")
        _lang = _settings.get("language", "en")
        _login_strings = _LOGIN_LOCALE[
            _resolve_login_locale_key(_lang)
        ]
        _page = (
            _LOGIN_PAGE_HTML.replace("{{BOT_NAME}}", _bn)
            .replace("{{BOT_NAME_INITIAL}}", _bn[0].upper())
            .replace("{{LANG}}", _html.escape(_login_strings["lang"]))
            .replace("{{LOGIN_TITLE}}", _html.escape(_login_strings["title"]))
            .replace("{{LOGIN_SUBTITLE}}", _html.escape(_login_strings["subtitle"]))
            .replace(
                "{{LOGIN_PLACEHOLDER}}", _html.escape(_login_strings["placeholder"])
            )
            .replace("{{LOGIN_BTN}}", _html.escape(_login_strings["btn"]))
            .replace("{{LOGIN_INVALID_PW}}", _html.escape(_login_strings["invalid_pw"]))
            .replace(
                "{{LOGIN_CONN_FAILED}}", _html.escape(_login_strings["conn_failed"])
            )
        )
        return t(handler, _page, content_type="text/html; charset=utf-8")

    if parsed.path == "/api/auth/status":
        from api.auth import is_auth_enabled, parse_cookie, verify_session

        logged_in = False
        if is_auth_enabled():
            cv = parse_cookie(handler)
            logged_in = bool(cv and verify_session(cv))
        return j(handler, {"auth_enabled": is_auth_enabled(), "logged_in": logged_in})

    if parsed.path in ("/manifest.json", "/manifest.webmanifest"):
        static_root = Path(__file__).parent.parent / "static"
        manifest_path = (static_root / "manifest.json").resolve()
        if manifest_path.exists():
            data = manifest_path.read_bytes()
            handler.send_response(200)
            handler.send_header("Content-Type", "application/manifest+json; charset=utf-8")
            handler.send_header("Cache-Control", "no-store")
            handler.send_header("Content-Length", str(len(data)))
            handler.end_headers()
            handler.wfile.write(data)
            return True
        return j(handler, {"error": "not found"}, status=404)

    if parsed.path == "/sw.js":
        static_root = Path(__file__).parent.parent / "static"
        sw_path = (static_root / "sw.js").resolve()
        if sw_path.exists():
            # Inject the current git-derived version as the cache name so the
            # service worker cache busts automatically on every new deploy.
            from urllib.parse import quote
            from api.updates import WEBUI_VERSION
            version_token = quote(WEBUI_VERSION, safe="")
            text = sw_path.read_text(encoding="utf-8").replace(
                "__WEBUI_VERSION__", version_token
            )
            data = text.encode("utf-8")
            handler.send_response(200)
            handler.send_header("Content-Type", "application/javascript; charset=utf-8")
            handler.send_header("Cache-Control", "no-store")
            handler.send_header("Service-Worker-Allowed", "/")
            handler.send_header("Content-Length", str(len(data)))
            handler.end_headers()
            handler.wfile.write(data)
            return True
        return j(handler, {"error": "not found"}, status=404)

    if parsed.path == "/favicon.ico":
        static_root = Path(__file__).parent.parent / "static"
        ico_path = (static_root / "favicon.ico").resolve()
        if ico_path.exists() and ico_path.is_file():
            data = ico_path.read_bytes()
            handler.send_response(200)
            handler.send_header("Content-Type", "image/x-icon")
            handler.send_header("Content-Length", str(len(data)))
            handler.send_header("Cache-Control", "public, max-age=86400")
            handler.end_headers()
            handler.wfile.write(data)
        else:
            handler.send_response(204)
            handler.end_headers()
        return True

    # ── Insights ──
    if parsed.path == "/api/insights":
        return _handle_insights(handler, parsed)

    if parsed.path == "/health":
        with STREAMS_LOCK:
            n_streams = len(STREAMS)
        return j(
            handler,
            {
                "status": "ok",
                "sessions": len(SESSIONS),
                "active_streams": n_streams,
                "uptime_seconds": round(time.time() - SERVER_START_TIME, 1),
            },
        )

    if parsed.path == "/api/models":
        return j(handler, get_available_models())

    if parsed.path == "/api/models/live":
        return _handle_live_models(handler, parsed)

    # ── Providers (GET) ──
    if parsed.path == "/api/providers":
        return j(handler, get_providers())

    if parsed.path == "/api/settings":
        settings = load_settings()
        # Never expose the stored password hash to clients
        settings.pop("password_hash", None)
        # Surface env-var precedence so the UI can disable the password field
        # instead of silently no-oping the save (#1560). The setting takes
        # precedence in api.auth.get_password_hash(), but until now the UI
        # had no way to know — see issue #1139 / #1560.
        settings["password_env_var"] = bool(
            os.getenv("HERMES_WEBUI_PASSWORD", "").strip()
        )
        # Inject the running version so the UI badge stays in sync with git tags
        # without any manual release step.
        try:
            from api.updates import WEBUI_VERSION
            settings["webui_version"] = WEBUI_VERSION
        except Exception:
            pass
        return j(handler, settings)

    if parsed.path == "/api/reasoning":
        # Current reasoning config (shared source of truth with the CLI —
        # reads display.show_reasoning and agent.reasoning_effort from
        # the active profile's config.yaml).
        return j(handler, get_reasoning_status())

    if parsed.path == "/api/onboarding/status":
        return j(handler, get_onboarding_status())

    if parsed.path.startswith("/extensions/"):
        from api.extensions import serve_extension_static

        return serve_extension_static(handler, parsed)

    if parsed.path.startswith("/static/"):
        return _serve_static(handler, parsed)

    if parsed.path == "/api/session":
        import time as _time
        _t0 = _time.monotonic()
        _debug_slow = os.environ.get("HERMES_DEBUG_SLOW", "")
        query = parse_qs(parsed.query)
        sid = query.get("session_id", [""])[0]
        if not sid:
            return j(handler, {"error": "session_id is required"}, status=400)
        # ?messages=0 skips the message payload for fast session switching.
        # The frontend uses this when switching conversations in the sidebar
        # (only needs metadata). The full message array is loaded lazily
        # via ?messages=1 when the message panel opens.
        load_messages = query.get("messages", ["1"])[0] != "0"
        resolve_model_default = "1" if load_messages else "0"
        resolve_model = query.get("resolve_model", [resolve_model_default])[0] != "0"
        # ?msg_limit=N returns only the last N messages (tail window).
        # Used by the frontend for fast session switching — avoids serialising
        # and sending hundreds of messages when the user only sees the most
        # recent exchange.  Older messages are loaded on-demand via scrolling.
        _msg_limit = query.get("msg_limit", [None])[0]
        try:
            msg_limit = max(1, int(_msg_limit)) if _msg_limit else None
        except (ValueError, TypeError):
            msg_limit = None
        # ?msg_before=N — 0-based index into the full message array.
        # Returns messages before this index (for scroll-to-top lazy loading).
        # Combined with msg_limit for paging.
        _msg_before = query.get("msg_before", [None])[0]
        try:
            msg_before = int(_msg_before) if _msg_before else None
        except (ValueError, TypeError):
            msg_before = None
        try:
            _t1 = _time.monotonic()
            s = get_session(sid, metadata_only=(not load_messages))
            _clear_stale_stream_state(s)
            cli_meta = _lookup_cli_session_metadata(sid)
            is_messaging_session = _is_messaging_session_record(s) or _is_messaging_session_record(cli_meta)
            cli_messages = []
            if is_messaging_session:
                cli_messages = get_cli_session_messages(sid)
            _t2 = _time.monotonic()
            effective_model = (
                _resolve_effective_session_model_for_display(s)
                if resolve_model
                else None
            )
            effective_provider = (
                _resolve_effective_session_model_provider_for_display(s)
                if resolve_model
                else None
            )
            _t3 = _time.monotonic()
            if load_messages:
                if is_messaging_session and cli_messages:
                    _all_msgs = cli_messages
                else:
                    _all_msgs = s.messages
            else:
                _all_msgs = []
            if load_messages:
                if msg_before is not None:
                    # Scroll-to-top paging: msg_before is a 0-based index into
                    # the full message list. Return the msg_limit messages that
                    # appear *before* this index (i.e. older messages).
                    # Using index instead of timestamp avoids issues with
                    # duplicate/missing timestamps.
                    _before_idx = max(0, min(int(msg_before), len(_all_msgs)))
                    _slice = _all_msgs[:_before_idx]
                    _truncated_msgs = _slice[-msg_limit:] if msg_limit else _slice
                elif msg_limit and len(_all_msgs) > msg_limit:
                    _truncated_msgs = _all_msgs[-msg_limit:]
                else:
                    _truncated_msgs = _all_msgs
            else:
                _truncated_msgs = _all_msgs
            # Resolve effective context_length with model-metadata fallback so
            # older sessions (pre-#1318) that have context_length=0 persisted
            # still render a meaningful indicator on load.  Mirrors the
            # SSE-path fallback in api/streaming.py:2333-2342.  Fixes #1436.
            _persisted_cl = getattr(s, "context_length", 0) or 0
            if not _persisted_cl:
                _model_for_lookup = (
                    getattr(s, "model", "") or effective_model or ""
                ).strip()
                if _model_for_lookup:
                    try:
                        from agent.model_metadata import get_model_context_length as _get_cl
                        _fb_cl = _get_cl(_model_for_lookup, "") or 0
                        if _fb_cl:
                            _persisted_cl = _fb_cl
                    except Exception:
                        pass
            raw = s.compact() | {
                "messages": _truncated_msgs,
                "tool_calls": getattr(s, "tool_calls", []) if load_messages else [],
                "active_stream_id": getattr(s, "active_stream_id", None),
                "pending_user_message": getattr(s, "pending_user_message", None),
                "pending_attachments": getattr(s, "pending_attachments", []) if load_messages else [],
                "pending_started_at": getattr(s, "pending_started_at", None),
                "context_length": _persisted_cl,
                "threshold_tokens": getattr(s, "threshold_tokens", 0) or 0,
                "last_prompt_tokens": getattr(s, "last_prompt_tokens", 0) or 0,
            }
            if cli_meta and _is_messaging_session_record(cli_meta):
                raw = _merge_cli_sidebar_metadata(raw, cli_meta)
            # Signal to the frontend that older messages were omitted.
            # For msg_before paging, compare against the filtered set,
            # not the full list — otherwise we signal truncation even when
            # all older messages were returned.
            if msg_before is not None:
                _truncated = load_messages and msg_limit is not None and len(_slice) > msg_limit
            else:
                _truncated = load_messages and msg_limit is not None and len(_all_msgs) > msg_limit
            raw["_messages_truncated"] = _truncated
            # Index of the first returned message in the full message array.
            # Frontend uses this as cursor for scroll-to-top paging.
            if msg_before is not None:
                raw["_messages_offset"] = max(0, _before_idx - len(_truncated_msgs))
            else:
                raw["_messages_offset"] = max(0, len(_all_msgs) - len(_truncated_msgs))
            _t4 = _time.monotonic()
            if effective_model:
                raw["model"] = effective_model
            if effective_provider:
                raw["model_provider"] = effective_provider
            redact = redact_session_data(raw)
            _t5 = _time.monotonic()
            resp = j(handler, {"session": redact})
            _t6 = _time.monotonic()
            if _debug_slow:
                logger.warning(
                    "[SLOW] session_id=%s get_session=%.1fms model_resolve=%.1fms "
                    "compact=%.1fms redact=%.1fms json_write=%.1fms total=%.1fms",
                    sid,
                    (_t2-_t1)*1000, (_t3-_t2)*1000, (_t4-_t3)*1000,
                    (_t5-_t4)*1000, (_t6-_t5)*1000, (_t6-_t0)*1000,
                )
            return resp
        except KeyError:
            # Not a WebUI session -- try CLI store
            cli_meta = _lookup_cli_session_metadata(sid)
            msgs = get_cli_session_messages(sid)
            if msgs:
                sess = {
                    "session_id": sid,
                    "title": (cli_meta or {}).get("title", "CLI Session"),
                    "workspace": (cli_meta or {}).get("workspace", ""),
                    "model": (cli_meta or {}).get("model", "unknown"),
                    "message_count": len(msgs),
                    "created_at": (cli_meta or {}).get("created_at", 0),
                    "updated_at": (cli_meta or {}).get("updated_at", 0),
                    "last_message_at": (cli_meta or {}).get("last_message_at")
                    or (cli_meta or {}).get("updated_at", 0)
                    or (msgs[-1] if msgs else {"timestamp": 0}).get("timestamp", 0),
                    "pinned": False,
                    "archived": False,
                    "project_id": None,
                    "profile": (cli_meta or {}).get("profile"),
                    "is_cli_session": True,
                    "source_tag": (cli_meta or {}).get("source_tag"),
                    "raw_source": (cli_meta or {}).get("raw_source"),
                    "session_source": (cli_meta or {}).get("session_source"),
                    "source_label": (cli_meta or {}).get("source_label"),
                    "messages": msgs,
                    "tool_calls": [],
                }
                sess = _merge_cli_sidebar_metadata(sess, cli_meta)
                return j(handler, {"session": redact_session_data(sess)})
            return bad(handler, "Session not found", 404)

    if parsed.path == "/api/session/status":
        sid = parse_qs(parsed.query).get("session_id", [""])[0]
        if not sid:
            return bad(handler, "Missing session_id")
        try:
            from api.session_ops import session_status
            _clear_stale_stream_state(get_session(sid, metadata_only=True))
            return j(handler, session_status(sid))
        except KeyError:
            return bad(handler, "Session not found", 404)

    if parsed.path == "/api/session/yolo":
        sid = parse_qs(parsed.query).get("session_id", [""])[0]
        if not sid:
            return bad(handler, "Missing session_id")
        return j(handler, {"yolo_enabled": is_session_yolo_enabled(sid)})

    if parsed.path == "/api/session/usage":
        sid = parse_qs(parsed.query).get("session_id", [""])[0]
        if not sid:
            return bad(handler, "Missing session_id")
        try:
            from api.session_ops import session_usage
            return j(handler, session_usage(sid))
        except KeyError:
            return bad(handler, "Session not found", 404)

    if parsed.path == "/api/background/status":
        sid = parse_qs(parsed.query).get("session_id", [""])[0]
        if not sid:
            return bad(handler, "Missing session_id")
        from api.background import get_results
        return j(handler, {"results": get_results(sid)})

    if parsed.path == "/api/sessions":
        webui_sessions = all_sessions()
        settings = load_settings()
        if settings.get("show_cli_sessions"):
            cli = get_cli_sessions()
            cli_by_id = {s["session_id"]: s for s in cli}
            for s in webui_sessions:
                meta = cli_by_id.get(s.get("session_id"))
                if not meta:
                    continue
                if _is_messaging_session_record(meta):
                    s.update(_merge_cli_sidebar_metadata(s, meta))
                    if s.get("session_id") != meta.get("session_id"):
                        s["session_id"] = meta.get("session_id")
                else:
                    for key in ("source_tag", "raw_source", "session_source", "source_label"):
                        if not s.get(key) and meta.get(key):
                            s[key] = meta[key]
            webui_ids = {s["session_id"] for s in webui_sessions}
            from api.models import _hide_from_default_sidebar as _cron_hide
            deduped_cli = [s for s in cli
                           if s["session_id"] not in webui_ids
                           and not _cron_hide(s)]
        else:
            deduped_cli = []
        merged = webui_sessions + deduped_cli
        merged.sort(
            key=lambda s: s.get("last_message_at") or s.get("updated_at", 0) or 0,
            reverse=True,
        )
        merged = _keep_latest_messaging_session_per_source(merged)
        safe_merged = []
        for s in merged:
            item = dict(s)
            if isinstance(item.get("title"), str):
                item["title"] = _redact_text(item["title"])
            safe_merged.append(item)
        return j(handler, {
            "sessions": safe_merged,
            "cli_count": len(deduped_cli),
            "server_time": time.time(),
            "server_tz": time.strftime("%z"),
        })

    if parsed.path == "/api/projects":
        return j(handler, {"projects": load_projects()})

    if parsed.path == "/api/session/export":
        return _handle_session_export(handler, parsed)

    if parsed.path == "/api/workspaces":
        return j(
            handler, {"workspaces": load_workspaces(), "last": get_last_workspace()}
        )

    if parsed.path == "/api/workspaces/suggest":
        qs = parse_qs(parsed.query)
        prefix = qs.get("prefix", [""])[0]
        return j(
            handler,
            {
                "suggestions": list_workspace_suggestions(prefix),
                "prefix": prefix,
            },
        )

    if parsed.path == "/api/sessions/search":
        return _handle_sessions_search(handler, parsed)

    if parsed.path == "/api/list":
        return _handle_list_dir(handler, parsed)

    if parsed.path == "/api/personalities":
        # Read personalities from config.yaml agent.personalities section
        # (matches hermes-agent CLI behavior, not filesystem SOUL.md approach)
        from api.config import reload_config as _reload_cfg

        _reload_cfg()  # pick up config.yaml changes without server restart
        from api.config import get_config as _get_cfg

        _cfg = _get_cfg()
        agent_cfg = _cfg.get("agent", {})
        raw_personalities = agent_cfg.get("personalities", {})
        personalities = []
        if isinstance(raw_personalities, dict):
            for name, value in raw_personalities.items():
                desc = ""
                if isinstance(value, dict):
                    desc = value.get("description", "")
                elif isinstance(value, str):
                    desc = value[:80] + ("..." if len(value) > 80 else "")
                personalities.append({"name": name, "description": desc})
        return j(handler, {"personalities": personalities})

    if parsed.path == "/api/git-info":
        qs = parse_qs(parsed.query)
        sid = qs.get("session_id", [""])[0]
        if not sid:
            return bad(handler, "session_id required")
        try:
            s = get_session(sid)
        except KeyError:
            return bad(handler, "Session not found", 404)
        from api.workspace import git_info_for_workspace

        info = git_info_for_workspace(Path(s.workspace))
        return j(handler, {"git": info})

    if parsed.path == "/api/commands":
        from api.commands import list_commands
        return j(handler, {"commands": list_commands()})

    if parsed.path == "/api/updates/check":
        settings = load_settings()
        if not settings.get("check_for_updates", True):
            return j(handler, {"disabled": True})
        qs = parse_qs(parsed.query)
        force = qs.get("force", ["0"])[0] == "1"
        # ?simulate=1 returns fake behind counts for UI testing (localhost only)
        if (
            qs.get("simulate", ["0"])[0] == "1"
            and handler.client_address[0] == "127.0.0.1"
        ):
            return j(
                handler,
                {
                    "webui": {
                        "name": "webui",
                        "behind": 3,
                        "current_sha": "abc1234",
                        "latest_sha": "def5678",
                        "branch": "master",
                    },
                    "agent": {
                        "name": "agent",
                        "behind": 1,
                        "current_sha": "aaa0001",
                        "latest_sha": "bbb0002",
                        "branch": "master",
                    },
                    "checked_at": 0,
                },
            )
        from api.updates import check_for_updates

        return j(handler, check_for_updates(force=force))

    if parsed.path == "/api/chat/stream/status":
        stream_id = parse_qs(parsed.query).get("stream_id", [""])[0]
        return j(handler, {"active": stream_id in STREAMS, "stream_id": stream_id})

    if parsed.path == "/api/chat/cancel":
        stream_id = parse_qs(parsed.query).get("stream_id", [""])[0]
        if not stream_id:
            return bad(handler, "stream_id required")
        cancelled = cancel_stream(stream_id)
        return j(handler, {"ok": True, "cancelled": cancelled, "stream_id": stream_id})

    if parsed.path == "/api/chat/stream":
        return _handle_sse_stream(handler, parsed)

    if parsed.path == "/api/terminal/output":
        return _handle_terminal_output(handler, parsed)

    if parsed.path == '/api/sessions/gateway/stream':
        return _handle_gateway_sse_stream(handler, parsed)

    if parsed.path == "/api/media":
        return _handle_media(handler, parsed)

    if parsed.path == "/api/file/raw":
        return _handle_file_raw(handler, parsed)

    if parsed.path == "/api/file":
        return _handle_file_read(handler, parsed)

    if parsed.path == "/api/approval/pending":
        return _handle_approval_pending(handler, parsed)

    if parsed.path == "/api/approval/stream":
        return _handle_approval_sse_stream(handler, parsed)

    if parsed.path == "/api/approval/inject_test":
        # Loopback-only: used by automated tests; blocked from any remote client
        if handler.client_address[0] != "127.0.0.1":
            return j(handler, {"error": "not found"}, status=404)
        return _handle_approval_inject(handler, parsed)

    if parsed.path == "/api/clarify/pending":
        return _handle_clarify_pending(handler, parsed)

    if parsed.path == "/api/clarify/stream":
        return _handle_clarify_sse_stream(handler, parsed)

    if parsed.path == "/api/clarify/inject_test":
        # Loopback-only: used by automated tests; blocked from any remote client
        if handler.client_address[0] != "127.0.0.1":
            return j(handler, {"error": "not found"}, status=404)
        return _handle_clarify_inject(handler, parsed)

    # ── OAuth (Codex device-code) ──
    if parsed.path == "/api/oauth/codex/start":
        """Start Codex device-code OAuth flow. Returns user_code + verification_uri."""
        try:
            from api.oauth import start_codex_device_code
            result = start_codex_device_code()
            return j(handler, result)
        except Exception as e:
            return j(handler, {"error": str(e)}, status=500)

    if parsed.path == "/api/oauth/codex/poll":
        """SSE endpoint for polling Codex OAuth token."""
        qs = parse_qs(parsed.query)
        device_code = qs.get("device_code", [""])[0]
        if not device_code:
            return j(handler, {"error": "device_code required"}, status=400)
        handler.send_response(200)
        handler.send_header("Content-Type", "text/event-stream")
        handler.send_header("Cache-Control", "no-cache")
        handler.send_header("Connection", "keep-alive")
        handler.end_headers()
        try:
            from api.oauth import poll_codex_token
            for event in poll_codex_token(device_code):
                handler.wfile.write(f"data: {json.dumps(event)}\n\n".encode())
                handler.wfile.flush()
                if event.get("status") in ("success", "error"):
                    break
        except Exception as e:
            handler.wfile.write(f"data: {json.dumps({'status': 'error', 'error': str(e)})}\n\n".encode())
            handler.wfile.flush()
        return  # SSE handled, no JSON response

    # ── Cron API (GET) ──
    if parsed.path == "/api/crons":
        from cron.jobs import list_jobs

        return j(handler, {"jobs": list_jobs(include_disabled=True)})

    if parsed.path == "/api/crons/output":
        return _handle_cron_output(handler, parsed)

    if parsed.path == "/api/crons/history":
        return _handle_cron_history(handler, parsed)

    if parsed.path == "/api/crons/run":
        return _handle_cron_run_detail(handler, parsed)

    if parsed.path == "/api/crons/recent":
        return _handle_cron_recent(handler, parsed)

    if parsed.path == "/api/crons/status":
        return _handle_cron_status(handler, parsed)

    # ── Skills API (GET) ──
    if parsed.path == "/api/skills":
        from tools.skills_tool import skills_list as _skills_list

        raw = _skills_list()
        data = json.loads(raw) if isinstance(raw, str) else raw
        return j(handler, {"skills": data.get("skills", [])})

    if parsed.path == "/api/skills/content":
        from tools.skills_tool import skill_view as _skill_view, SKILLS_DIR

        qs = parse_qs(parsed.query)
        name = qs.get("name", [""])[0]
        if not name:
            return j(handler, {"error": "name required"}, status=400)
        file_path = qs.get("file", [""])[0]
        if file_path:
            # Serve a linked file from the skill directory
            import re as _re

            if _re.search(r"[*?\[\]]", name):
                return bad(handler, "Invalid skill name", 400)
            skill_dir = None
            for p in SKILLS_DIR.rglob(name):
                if p.is_dir():
                    skill_dir = p
                    break
            if not skill_dir:
                return bad(handler, "Skill not found", 404)
            target = (skill_dir / file_path).resolve()
            try:
                target.relative_to(skill_dir.resolve())
            except ValueError:
                return bad(handler, "Invalid file path", 400)
            if not target.exists() or not target.is_file():
                return bad(handler, "File not found", 404)
            return j(
                handler,
                {"content": target.read_text(encoding="utf-8"), "path": file_path},
            )
        raw = _skill_view(name)
        data = json.loads(raw) if isinstance(raw, str) else raw
        if not isinstance(data.get("linked_files"), dict):
            data["linked_files"] = {}
        return j(handler, data)

    # ── Memory API (GET) ──
    if parsed.path == "/api/memory":
        return _handle_memory_read(handler)

    # ── Profile API (GET) ──
    if parsed.path == "/api/profiles":
        from api.profiles import list_profiles_api, get_active_profile_name

        return j(
            handler,
            {"profiles": list_profiles_api(), "active": get_active_profile_name()},
        )

    if parsed.path == "/api/profile/active":
        from api.profiles import get_active_profile_name, get_active_hermes_home

        return j(
            handler,
            {"name": get_active_profile_name(), "path": str(get_active_hermes_home())},
        )

    # ── Gateway Status (GET) ──
    if parsed.path == "/api/gateway/status":
        import datetime
        identity_map = _load_gateway_session_identity_map()
        sessions_path = _gateway_session_metadata_path()
        running = bool(identity_map)
        platforms_set: set[str] = set()
        for meta in identity_map.values():
            raw = meta.get("raw_source") or meta.get("platform") or ""
            norm = _normalize_messaging_source(raw)
            if norm:
                platforms_set.add(norm)
        _PLATFORM_LABELS = {
            "telegram": "Telegram",
            "discord": "Discord",
            "slack": "Slack",
            "web": "Web",
            "api": "API",
        }
        platforms = sorted(
            [{"name": p, "label": _PLATFORM_LABELS.get(p, p.title())} for p in platforms_set],
            key=lambda x: x["label"],
        )
        last_active = ""
        if running and sessions_path.exists():
            try:
                mtime = sessions_path.stat().st_mtime
                last_active = datetime.datetime.fromtimestamp(mtime).isoformat()
            except Exception:
                pass
        return j(handler, {
            "running": running,
            "platforms": platforms,
            "last_active": last_active,
            "session_count": len(identity_map),
        })

    # ── MCP Servers (GET) ──
    if parsed.path == "/api/mcp/servers":
        return _handle_mcp_servers_list(handler)

    # ── Checkpoints / Rollback (GET) ──
    if parsed.path == "/api/rollback/list":
        qs = parse_qs(parsed.query)
        workspace = qs.get("workspace", [""])[0]
        if not workspace:
            return bad(handler, "workspace query parameter is required")
        try:
            from api.rollback import list_checkpoints
            return j(handler, list_checkpoints(workspace))
        except ValueError as e:
            return bad(handler, str(e))
        except Exception as e:
            logger.exception("rollback/list failed")
            return bad(handler, str(e), status=500)

    if parsed.path == "/api/rollback/diff":
        qs = parse_qs(parsed.query)
        workspace = qs.get("workspace", [""])[0]
        checkpoint = qs.get("checkpoint", [""])[0]
        if not workspace or not checkpoint:
            return bad(handler, "workspace and checkpoint query parameters are required")
        try:
            from api.rollback import get_checkpoint_diff
            return j(handler, get_checkpoint_diff(workspace, checkpoint))
        except ValueError as e:
            return bad(handler, str(e))
        except Exception as e:
            logger.exception("rollback/diff failed")
            return bad(handler, str(e), status=500)

    return False  # 404


# ── GET route helpers


def handle_post(handler, parsed) -> bool:
    """Handle all POST routes. Returns True if handled, False for 404."""
    # CSRF: reject cross-origin browser requests
    if not _check_csrf(handler):
        return j(handler, {"error": "Cross-origin request rejected"}, status=403)

    if parsed.path == "/api/upload":
        return handle_upload(handler)
    if parsed.path == "/api/upload/extract":
        return handle_upload_extract(handler)

    if parsed.path == "/api/transcribe":
        return handle_transcribe(handler)

    body = read_body(handler)

    if parsed.path == "/api/session/new":
        try:
            workspace = str(resolve_trusted_workspace(body.get("workspace"))) if body.get("workspace") else None
        except ValueError as e:
            return bad(handler, str(e))
        model, model_provider = _session_model_state_from_request(
            body.get("model"),
            body.get("model_provider"),
        )
        # Use the profile sent by the client tab (if any) so that two tabs on
        # different profiles never clobber each other via the process-level global.
        s = new_session(
            workspace=workspace,
            model=model,
            model_provider=model_provider,
            profile=body.get("profile") or None,
            project_id=body.get("project_id") or None,
        )
        return j(handler, {"session": s.compact() | {"messages": s.messages}})

    if parsed.path == "/api/session/duplicate":
        try:
            sid = body.get("session_id")
            if not sid:
                return bad(handler, "session_id is required")

            session = Session.load(sid)
            if not session:
                # 404, not 400 — missing resource, not a malformed request.
                return bad(handler, "Session not found", status=404)

            # Deep-copy mutable lists so the duplicate is *actually* independent.
            # `Session.__init__` does `self.messages = messages or []` — plain
            # assignment, no copy. Without deepcopy, both sessions share the same
            # list object in memory; appending to one mutates the other.
            # Items inside `messages` are dicts with mutable values (tool_calls,
            # content arrays), so a shallow `list(...)` is not enough.
            copied_session = Session(
                session_id=uuid.uuid4().hex[:12],
                # Defensive: legacy sessions may have title=None on disk; fall back to 'Untitled'
                # so `+ " (copy)"` doesn't TypeError.
                title=(session.title or "Untitled") + " (copy)",
                workspace=session.workspace,
                model=session.model,
                model_provider=session.model_provider,
                messages=copy.deepcopy(session.messages),
                tool_calls=copy.deepcopy(session.tool_calls),
                # Reset ephemeral / per-session-instance flags. Duplicating an
                # archived conversation should produce a visible (un-archived)
                # copy; pinned status doesn't transfer either.
                pinned=False,
                archived=False,
                project_id=session.project_id,
                profile=session.profile,
                input_tokens=session.input_tokens,
                output_tokens=session.output_tokens,
                estimated_cost=session.estimated_cost,
                # Per-session settings the user may have customized — carry them over
                # so the duplicate behaves identically until further edits. Compression
                # anchor + last_prompt_tokens are intentionally NOT carried — those
                # re-derive on the next turn.
                personality=session.personality,
                enabled_toolsets=getattr(session, "enabled_toolsets", None),
                context_length=getattr(session, "context_length", None),
                threshold_tokens=getattr(session, "threshold_tokens", None),
                created_at=time.time(),
                updated_at=time.time(),
            )

            with LOCK:
                SESSIONS[copied_session.session_id] = copied_session
                SESSIONS.move_to_end(copied_session.session_id)
                while len(SESSIONS) > SESSIONS_MAX:
                    SESSIONS.popitem(last=False)
            # Persist immediately. The pre-PR flow (/api/session/new + /api/session/rename)
            # accidentally avoided this because `/api/session/rename` calls `s.save()`.
            # Without this explicit save, the duplicate is in-memory only — if the user
            # refreshes before sending a turn, the duplicate vanishes.
            copied_session.save()

            return j(handler, {"session": copied_session.compact() | {"messages": copied_session.messages}})
        except Exception as e:
            return bad(handler, str(e))

    if parsed.path == "/api/default-model":
        try:
            return j(handler, set_hermes_default_model(body.get("model")))
        except ValueError as e:
            return bad(handler, str(e))
        except RuntimeError as e:
            return bad(handler, str(e), 500)

    # ── Providers (POST) ──
    if parsed.path == "/api/providers":
        provider_id = (body.get("provider") or "").strip().lower()
        api_key = body.get("api_key")
        if not provider_id:
            return bad(handler, "provider is required")
        if api_key is not None:
            api_key = str(api_key).strip() or None
        result = set_provider_key(provider_id, api_key)
        if not result.get("ok"):
            return bad(handler, result.get("error", "Unknown error"))
        return j(handler, result)

    if parsed.path == "/api/providers/delete":
        provider_id = (body.get("provider") or "").strip().lower()
        if not provider_id:
            return bad(handler, "provider is required")
        result = remove_provider_key(provider_id)
        if not result.get("ok"):
            return bad(handler, result.get("error", "Unknown error"))
        return j(handler, result)

    if parsed.path == "/api/reasoning":
        # CLI-parity /reasoning handler — writes to the same config.yaml keys
        # the CLI uses (display.show_reasoning, agent.reasoning_effort) so a
        # preference set via WebUI is honoured in the terminal REPL and vice
        # versa.  Body is one of:
        #   {"display": "show"|"hide"|"on"|"off"}   → display.show_reasoning
        #   {"effort":  "none"|"minimal"|"low"|"medium"|"high"|"xhigh"}
        #                                            → agent.reasoning_effort
        try:
            display = body.get("display")
            effort = body.get("effort")
            if display is not None:
                flag = str(display).strip().lower()
                if flag in ("show", "on", "true", "1"):
                    return j(handler, set_reasoning_display(True))
                if flag in ("hide", "off", "false", "0"):
                    return j(handler, set_reasoning_display(False))
                return bad(handler, f"display must be show|hide|on|off (got '{display}')")
            if effort is not None:
                return j(handler, set_reasoning_effort(effort))
            return bad(handler, "reasoning: must supply 'display' or 'effort'")
        except ValueError as e:
            return bad(handler, str(e))
        except RuntimeError as e:
            return bad(handler, str(e), 500)

    if parsed.path == "/api/admin/reload":
        # Hot-reload api.models module to pick up code changes without restart.
        import importlib
        from api import models as _models
        importlib.reload(_models)
        # Also re-expose get_session from the reloaded module so routes.py
        # continues to work (routes.py imported it at module level).
        import api.routes as _routes
        _routes.get_session = _models.get_session
        _routes.Session = _models.Session
        _routes.compact = _models.compact
        return j(handler, {"status": "ok", "reloaded": "api.models"})

    if parsed.path == "/api/sessions/cleanup":
        return _handle_sessions_cleanup(handler, body, zero_only=False)

    if parsed.path == "/api/sessions/cleanup_zero_message":
        return _handle_sessions_cleanup(handler, body, zero_only=True)

    if parsed.path == "/api/session/rename":
        try:
            require(body, "session_id", "title")
        except ValueError as e:
            return bad(handler, str(e))
        try:
            s = get_session(body["session_id"])
        except KeyError:
            return bad(handler, "Session not found", 404)
        with _get_session_agent_lock(body["session_id"]):
            s.title = str(body["title"]).strip()[:80] or "Untitled"
            s.save()
        return j(handler, {"session": s.compact()})

    if parsed.path == "/api/personality/set":
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        if "name" not in body:
            return bad(handler, "Missing required field: name")
        sid = body["session_id"]
        name = body["name"].strip()
        try:
            s = get_session(sid)
        except KeyError:
            return bad(handler, "Session not found", 404)
        # Resolve personality from config.yaml agent.personalities section
        # (matches hermes-agent CLI behavior)
        prompt = ""
        if name:
            from api.config import reload_config as _reload_cfg2

            _reload_cfg2()  # pick up config changes without restart
            from api.config import get_config as _get_cfg2

            _cfg2 = _get_cfg2()
            agent_cfg = _cfg2.get("agent", {})
            raw_personalities = agent_cfg.get("personalities", {})
            if not isinstance(raw_personalities, dict) or name not in raw_personalities:
                return bad(
                    handler, f'Personality "{name}" not found in config.yaml', 404
                )
            value = raw_personalities[name]
            # Resolve prompt using the same logic as hermes-agent cli.py
            if isinstance(value, dict):
                parts = [value.get("system_prompt", "") or value.get("prompt", "")]
                if value.get("tone"):
                    parts.append(f"Tone: {value['tone']}")
                if value.get("style"):
                    parts.append(f"Style: {value['style']}")
                prompt = "\n".join(p for p in parts if p)
            else:
                prompt = str(value)
        with _get_session_agent_lock(sid):
            s.personality = name if name else None
            s.save()
        return j(handler, {"ok": True, "personality": s.personality, "prompt": prompt})

    if parsed.path == "/api/session/toolsets":
        """Set or clear per-session toolset override (#493).

        POST body: { session_id, toolsets: [...] | null }
        - toolsets: list of toolset names to restrict the session to, or null to clear.
        """
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        sid = body["session_id"]
        toolsets = body.get("toolsets")
        # Validate: if not None, must be a non-empty list of strings
        if toolsets is not None:
            if not isinstance(toolsets, list) or not toolsets:
                return bad(handler, "toolsets must be a non-empty list or null")
            if not all(isinstance(t, str) and t for t in toolsets):
                return bad(handler, "each toolset must be a non-empty string")
        try:
            s = get_session(sid)
        except KeyError:
            return bad(handler, "Session not found", 404)
        with _get_session_agent_lock(sid):
            s.enabled_toolsets = toolsets
            s.save()
        return j(handler, {"ok": True, "enabled_toolsets": s.enabled_toolsets})

    if parsed.path == "/api/session/update":
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        try:
            s = get_session(body["session_id"])
        except KeyError:
            return bad(handler, "Session not found", 404)
        old_ws = getattr(s, "workspace", "")
        try:
            new_ws = str(resolve_trusted_workspace(body.get("workspace", s.workspace)))
        except ValueError as e:
            return bad(handler, str(e))
        with _get_session_agent_lock(body["session_id"]):
            s.workspace = new_ws
            if "model" in body or "model_provider" in body:
                model, provider = _session_model_state_from_request(
                    body.get("model", s.model),
                    body.get("model_provider") if "model_provider" in body else None,
                    getattr(s, "model_provider", None),
                )
                if model is not None:
                    s.model = model
                s.model_provider = provider
            s.save()
        if str(old_ws or "") != str(new_ws or ""):
            try:
                from api.terminal import close_terminal
                close_terminal(body["session_id"])
            except Exception:
                logger.debug("Failed to close workspace terminal after workspace update")
        set_last_workspace(new_ws)
        return j(handler, {"session": s.compact() | {"messages": s.messages}})

    if parsed.path == "/api/session/delete":
        sid = body.get("session_id", "")
        if not sid:
            return bad(handler, "session_id is required")
        if not all(c in '0123456789abcdefghijklmnopqrstuvwxyz_' for c in sid):
            return bad(handler, "Invalid session_id", 400)
        is_messaging_session = _is_messaging_session_id(sid)
        # Delete from WebUI session store
        with LOCK:
            SESSIONS.pop(sid, None)
        try:
            SESSION_INDEX_FILE.unlink(missing_ok=True)
        except Exception:
            logger.debug("Failed to unlink session index")
        # Evict cached agent so turn count doesn't leak into a recycled session
        from api.config import _evict_session_agent
        _evict_session_agent(sid)
        try:
            p = (SESSION_DIR / f"{sid}.json").resolve()
            p.relative_to(SESSION_DIR.resolve())
        except Exception:
            return bad(handler, "Invalid session_id", 400)
        try:
            p.unlink(missing_ok=True)
        except Exception:
            logger.debug("Failed to unlink session file %s", p)
        # Prune the per-session agent lock so deleted sessions don't leak
        # Lock entries in SESSION_AGENT_LOCKS forever.
        with SESSION_AGENT_LOCKS_LOCK:
            SESSION_AGENT_LOCKS.pop(sid, None)
        try:
            from api.terminal import close_terminal
            close_terminal(sid)
        except Exception:
            logger.debug("Failed to close workspace terminal for deleted session %s", sid)
        # Also delete from CLI state.db for CLI sessions shown in sidebar,
        # but never erase external messaging channel memory via WebUI delete.
        if not is_messaging_session:
            try:
                from api.models import delete_cli_session

                delete_cli_session(sid)
            except Exception:
                logger.debug("Failed to delete CLI session %s", sid)
        return j(handler, {"ok": True})

    if parsed.path == "/api/session/clear":
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        try:
            s = get_session(body["session_id"])
        except KeyError:
            return bad(handler, "Session not found", 404)
        with _get_session_agent_lock(body["session_id"]):
            s.messages = []
            s.tool_calls = []
            s.title = "Untitled"
            s.save()
            # Evict cached agent — cleared session is a fresh conversation
            from api.config import _evict_session_agent
            _evict_session_agent(body["session_id"])
        return j(handler, {"ok": True, "session": s.compact()})

    if parsed.path == "/api/session/truncate":
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        if body.get("keep_count") is None:
            return bad(handler, "Missing required field(s): keep_count")
        try:
            s = get_session(body["session_id"])
        except KeyError:
            return bad(handler, "Session not found", 404)
        keep = int(body["keep_count"])
        with _get_session_agent_lock(body["session_id"]):
            s.messages = s.messages[:keep]
            s.save()
        return j(
            handler, {"ok": True, "session": s.compact() | {"messages": s.messages}}
        )

    if parsed.path == "/api/session/branch":
        # Fork a conversation from any message point (#465).
        # Accepts: {session_id, keep_count?, title?}
        #   keep_count: number of messages to copy (0=empty, undefined=full history)
        #   title: custom title (defaults to "<original title> (fork)")
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        # Reject non-string session_id explicitly so the failure surfaces as a
        # 400 instead of a generic 500 from get_session() raising TypeError.
        # (Opus pre-release follow-up.)
        if not isinstance(body["session_id"], str):
            return bad(handler, "session_id must be a string")
        try:
            source = get_session(body["session_id"])
        except KeyError:
            return bad(handler, "Session not found", 404)

        keep_count = body.get("keep_count")
        if keep_count is not None:
            try:
                keep_count = int(keep_count)
            except (ValueError, TypeError):
                return bad(handler, "keep_count must be an integer")
            # Negative slice (`messages[:-N]`) returns "all but last N", which
            # is a confusing fork semantic. Reject explicitly so the user
            # doesn't accidentally fork a session with the tail truncated when
            # they meant to copy the prefix. (Opus pre-release follow-up.)
            if keep_count < 0:
                return bad(handler, "keep_count must be non-negative")

        custom_title = body.get("title")
        if custom_title:
            custom_title = str(custom_title).strip()[:80] or None

        # Build messages slice
        source_messages = source.messages or []
        if keep_count is not None:
            forked_messages = source_messages[:keep_count]
        else:
            forked_messages = list(source_messages)

        # Derive title
        if custom_title:
            branch_title = custom_title
        else:
            source_title = source.title or "Untitled"
            branch_title = f"{source_title} (fork)"

        # Create new session inheriting workspace/model/profile
        branch = Session(
            workspace=source.workspace,
            model=source.model,
            profile=getattr(source, "profile", None),
            title=branch_title,
            messages=forked_messages,
            parent_session_id=source.session_id,
        )
        with LOCK:
            SESSIONS[branch.session_id] = branch
            SESSIONS.move_to_end(branch.session_id)
            while len(SESSIONS) > SESSIONS_MAX:
                SESSIONS.popitem(last=False)

        # Persist only if there are messages (matches new_session pattern)
        if forked_messages:
            branch.save()

        return j(handler, {
            "session_id": branch.session_id,
            "title": branch_title,
            "parent_session_id": source.session_id,
        })

    if parsed.path == "/api/session/compress":
        return _handle_session_compress(handler, body)

    if parsed.path == "/api/session/conversation-rounds":
        return _handle_conversation_rounds(handler, body)

    if parsed.path == "/api/session/handoff-summary":
        return _handle_handoff_summary(handler, body)

    if parsed.path == "/api/session/retry":
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        try:
            from api.session_ops import retry_last
            result = retry_last(body["session_id"])
            return j(handler, {"ok": True, **result})
        except KeyError:
            return bad(handler, "Session not found", 404)
        except ValueError as e:
            return j(handler, {"error": str(e)})

    if parsed.path == "/api/session/undo":
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        try:
            from api.session_ops import undo_last
            result = undo_last(body["session_id"])
            return j(handler, {"ok": True, **result})
        except KeyError:
            return bad(handler, "Session not found", 404)
        except ValueError as e:
            return j(handler, {"error": str(e)})

    # ── YOLO mode toggle (POST) ──
    # Session-scoped only — stored in-memory on the server side.
    # Important lifecycle notes:
    #   • Page reload: state PERSISTS (frontend re-fetches via GET endpoint)
    #   • Cross-tab: state is SHARED (same server-side flag per session)
    #   • Server restart: state is LOST (in-memory only)
    #   • Cross-session: isolated (each session has its own flag)
    # Fixes #467
    if parsed.path == "/api/session/yolo":
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        sid = body["session_id"]
        enabled = bool(body.get("enabled", True))
        if enabled:
            enable_session_yolo(sid)
            # Also resolve any pending approvals for this session so the
            # agent doesn't stay stuck waiting on an already-dismissed card.
            try:
                from tools.approval import _pending as _p, _lock as _l
                with _l:
                    _p.pop(sid, None)
            except Exception:
                pass
            resolve_gateway_approval(sid, "once", resolve_all=True)
        else:
            disable_session_yolo(sid)
        return j(handler, {"ok": True, "yolo_enabled": enabled})

    if parsed.path == "/api/btw":
        return _handle_btw(handler, body)

    if parsed.path == "/api/background":
        return _handle_background(handler, body)

    if parsed.path == "/api/chat/start":
        return _handle_chat_start(handler, body)

    if parsed.path == "/api/chat":
        return _handle_chat_sync(handler, body)

    if parsed.path == "/api/chat/steer":
        from api.streaming import _handle_chat_steer
        return _handle_chat_steer(handler, body)

    if parsed.path == "/api/terminal/start":
        return _handle_terminal_start(handler, body)

    if parsed.path == "/api/terminal/input":
        return _handle_terminal_input(handler, body)

    if parsed.path == "/api/terminal/resize":
        return _handle_terminal_resize(handler, body)

    if parsed.path == "/api/terminal/close":
        return _handle_terminal_close(handler, body)

    # ── Cron API (POST) ──
    if parsed.path == "/api/crons/create":
        return _handle_cron_create(handler, body)

    if parsed.path == "/api/crons/update":
        return _handle_cron_update(handler, body)

    if parsed.path == "/api/crons/delete":
        return _handle_cron_delete(handler, body)

    if parsed.path == "/api/crons/run":
        return _handle_cron_run(handler, body)

    if parsed.path == "/api/crons/pause":
        return _handle_cron_pause(handler, body)

    if parsed.path == "/api/crons/resume":
        return _handle_cron_resume(handler, body)

    # ── File ops (POST) ──
    if parsed.path == "/api/file/delete":
        return _handle_file_delete(handler, body)

    if parsed.path == "/api/file/save":
        return _handle_file_save(handler, body)

    if parsed.path == "/api/file/create":
        return _handle_file_create(handler, body)

    if parsed.path == "/api/file/rename":
        return _handle_file_rename(handler, body)

    if parsed.path == "/api/file/create-dir":
        return _handle_create_dir(handler, body)

    if parsed.path == "/api/file/reveal":
        return _handle_file_reveal(handler, body)

    # ── Workspace management (POST) ──
    if parsed.path == "/api/workspaces/add":
        return _handle_workspace_add(handler, body)

    if parsed.path == "/api/workspaces/remove":
        return _handle_workspace_remove(handler, body)

    if parsed.path == "/api/workspaces/rename":
        return _handle_workspace_rename(handler, body)

    if parsed.path == "/api/workspaces/reorder":
        return _handle_workspace_reorder(handler, body)

    # ── Approval (POST) ──
    if parsed.path == "/api/approval/respond":
        return _handle_approval_respond(handler, body)

    # ── Clarify (POST) ──
    if parsed.path == "/api/clarify/respond":
        return _handle_clarify_respond(handler, body)

    # ── Skills (POST) ──
    if parsed.path == "/api/skills/save":
        return _handle_skill_save(handler, body)

    if parsed.path == "/api/skills/delete":
        return _handle_skill_delete(handler, body)

    # ── Memory (POST) ──
    if parsed.path == "/api/memory/write":
        return _handle_memory_write(handler, body)

    # ── Profile API (POST) ──
    if parsed.path == "/api/profile/switch":
        name = body.get("name", "").strip()
        if not name:
            return bad(handler, "name is required")
        try:
            from api.profiles import switch_profile, _validate_profile_name
            from api.helpers import build_profile_cookie
            if name != 'default':
                _validate_profile_name(name)
            # process_wide=False: don't mutate the process-global _active_profile.
            # Per-client profile is managed via cookie + thread-local (#798).
            result = switch_profile(name, process_wide=False)
            # Invalidate the models cache so the very next /api/models request
            # rebuilds from the new profile's config.yaml rather than returning
            # the old profile's cached model list (#1200 — profile-switch model bug).
            from api.config import invalidate_models_cache
            invalidate_models_cache()
            return j(handler, result, extra_headers={
                'Set-Cookie': build_profile_cookie(name),
            })
        except (ValueError, FileNotFoundError) as e:
            return bad(handler, _sanitize_error(e), 404)
        except RuntimeError as e:
            return bad(handler, str(e), 409)

    if parsed.path == "/api/profile/create":
        name = body.get("name", "").strip()
        if not name:
            return bad(handler, "name is required")
        import re as _re

        if not _re.match(r"^[a-z0-9][a-z0-9_-]{0,63}$", name):
            return bad(
                handler,
                "Invalid profile name: lowercase letters, numbers, hyphens, underscores only",
            )
        clone_from = body.get("clone_from")
        if clone_from is not None:
            clone_from = str(clone_from).strip()
            if not _re.match(r"^[a-z0-9][a-z0-9_-]{0,63}$", clone_from):
                return bad(handler, "Invalid clone_from name")
        base_url = body.get("base_url", "").strip() if body.get("base_url") else None
        api_key = body.get("api_key", "").strip() if body.get("api_key") else None
        if base_url and not base_url.startswith(("http://", "https://")):
            return bad(handler, "base_url must start with http:// or https://")
        try:
            from api.profiles import create_profile_api

            result = create_profile_api(
                name,
                clone_from=clone_from,
                clone_config=bool(body.get("clone_config", False)),
                base_url=base_url,
                api_key=api_key,
            )
            return j(handler, {"ok": True, "profile": result})
        except (ValueError, FileExistsError, RuntimeError) as e:
            return bad(handler, str(e))

    if parsed.path == "/api/profile/delete":
        name = body.get("name", "").strip()
        if not name:
            return bad(handler, "name is required")
        try:
            from api.profiles import delete_profile_api, _validate_profile_name

            _validate_profile_name(name)
            result = delete_profile_api(name)
            return j(handler, result)
        except (ValueError, FileNotFoundError) as e:
            return bad(handler, _sanitize_error(e))
        except RuntimeError as e:
            return bad(handler, str(e), 409)

    # ── Settings (POST) ──
    if parsed.path == "/api/settings":
        from api.auth import (
            create_session,
            is_auth_enabled,
            parse_cookie,
            set_auth_cookie,
            verify_session,
        )

        if "bot_name" in body:
            body["bot_name"] = (str(body["bot_name"]) or "").strip() or "Hermes"

        auth_enabled_before = is_auth_enabled()
        current_cookie = parse_cookie(handler)
        logged_in_before = bool(current_cookie and verify_session(current_cookie))
        requested_password = bool(
            isinstance(body.get("_set_password"), str)
            and body.get("_set_password", "").strip()
        )
        requested_clear_password = bool(body.get("_clear_password"))

        # #1560: HERMES_WEBUI_PASSWORD env var takes precedence in
        # api.auth.get_password_hash(), so writing password_hash to settings.json
        # has no effect on auth. Refuse loudly with 409 instead of silently
        # succeeding — the previous behaviour returned 200 + a green save toast
        # while every subsequent login still required the env-var password.
        if requested_password or requested_clear_password:
            if os.getenv("HERMES_WEBUI_PASSWORD", "").strip():
                return bad(
                    handler,
                    "HERMES_WEBUI_PASSWORD env var is set — it overrides the settings password. "
                    "Unset the env var and restart the server before changing the password here.",
                    409,
                )

        saved = save_settings(body)
        saved.pop("password_hash", None)  # never expose hash to client

        auth_enabled_after = is_auth_enabled()
        auth_just_enabled = bool(
            requested_password and auth_enabled_after and not auth_enabled_before
        )
        logged_in_after = logged_in_before
        new_cookie = None

        if auth_just_enabled and not logged_in_before:
            new_cookie = create_session()
            logged_in_after = True

        saved["auth_enabled"] = auth_enabled_after
        saved["logged_in"] = logged_in_after
        saved["auth_just_enabled"] = auth_just_enabled

        if not new_cookie:
            return j(handler, saved)

        response_body = json.dumps(saved, ensure_ascii=False, indent=2).encode("utf-8")
        handler.send_response(200)
        handler.send_header("Content-Type", "application/json; charset=utf-8")
        handler.send_header("Content-Length", str(len(response_body)))
        handler.send_header("Cache-Control", "no-store")
        set_auth_cookie(handler, new_cookie)
        _security_headers(handler)
        handler.end_headers()
        handler.wfile.write(response_body)
        return True

    if parsed.path == "/api/onboarding/setup":
        # Writing API keys to disk - restrict to local/private networks unless auth is active.
        # In Docker, requests arrive from the bridge network (172.x.x.x), not 127.0.0.1,
        # even when the user accesses via localhost:8787 on the host.
        # Behind a reverse proxy (nginx/Caddy/Traefik) or SSH tunnel, X-Forwarded-For
        # carries the real origin IP — read it first before falling back to the raw socket addr.
        # HERMES_WEBUI_ONBOARDING_OPEN=1 lets operators on remote servers explicitly bypass
        # the check when they control network access themselves (e.g. firewall + VPN).
        from api.auth import is_auth_enabled
        import os as _os
        if not is_auth_enabled() and not _os.getenv("HERMES_WEBUI_ONBOARDING_OPEN"):
            import ipaddress
            try:
                # Prefer forwarded headers set by reverse proxies
                _xff = handler.headers.get("X-Forwarded-For", "").split(",")[0].strip()
                _xri = handler.headers.get("X-Real-IP", "").strip()
                _raw = handler.client_address[0]
                _ip_str = _xff or _xri or _raw
                addr = ipaddress.ip_address(_ip_str)
                is_local = addr.is_loopback or addr.is_private
            except ValueError:
                is_local = False
            if not is_local:
                return bad(handler, "Onboarding setup is only available from local networks when auth is not enabled. To bypass this on a remote server, set HERMES_WEBUI_ONBOARDING_OPEN=1.", 403)
        try:
            return j(handler, apply_onboarding_setup(body))
        except ValueError as e:
            return bad(handler, str(e))
        except RuntimeError as e:
            return bad(handler, str(e), 500)

    if parsed.path == "/api/onboarding/complete":
        return j(handler, complete_onboarding())

    if parsed.path == "/api/onboarding/probe":
        # Probe a self-hosted provider endpoint (#1499).  Validates the
        # configured base URL is reachable + parses /models, returns the
        # model catalog so the wizard can populate its dropdown.
        # Read-only: no config.yaml or .env writes happen here.  Same local-
        # network gate as /api/onboarding/setup (also writing-adjacent in
        # spirit because it carries an api_key the user typed).
        from api.auth import is_auth_enabled
        import os as _os
        if not is_auth_enabled() and not _os.getenv("HERMES_WEBUI_ONBOARDING_OPEN"):
            import ipaddress
            try:
                _xff = handler.headers.get("X-Forwarded-For", "").split(",")[0].strip()
                _xri = handler.headers.get("X-Real-IP", "").strip()
                _raw = handler.client_address[0]
                _ip_str = _xff or _xri or _raw
                addr = ipaddress.ip_address(_ip_str)
                is_local = addr.is_loopback or addr.is_private
            except ValueError:
                is_local = False
            if not is_local:
                return bad(handler, "Onboarding probe is only available from local networks when auth is not enabled. To bypass this on a remote server, set HERMES_WEBUI_ONBOARDING_OPEN=1.", 403)
        provider = str((body or {}).get("provider") or "").strip().lower()
        base_url = str((body or {}).get("base_url") or "")
        api_key = str((body or {}).get("api_key") or "").strip() or None
        try:
            return j(handler, probe_provider_endpoint(provider, base_url, api_key))
        except Exception as e:
            return bad(handler, f"probe failed: {e}", 500)

    # ── Session pin (POST) ──
    if parsed.path == "/api/session/pin":
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        try:
            s = get_session(body["session_id"])
        except KeyError:
            return bad(handler, "Session not found", 404)
        with _get_session_agent_lock(body["session_id"]):
            s.pinned = bool(body.get("pinned", True))
            s.save()
        return j(handler, {"ok": True, "session": s.compact()})

    # ── Session archive (POST) ──
    if parsed.path == "/api/session/archive":
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        sid = body["session_id"]
        try:
            s = get_session(sid)
        except KeyError:
            cli_meta = _lookup_cli_session_metadata(sid)
            if not cli_meta:
                return bad(handler, "Session not found", 404)
            if _is_messaging_session_record(cli_meta):
                s = Session(
                    session_id=sid,
                    title=cli_meta.get("title") or title_from(get_cli_session_messages(sid), "CLI Session"),
                    workspace=get_last_workspace(),
                    messages=[],
                    model=cli_meta.get("model") or "unknown",
                    created_at=cli_meta.get("created_at"),
                    updated_at=cli_meta.get("updated_at"),
                )
                s.is_cli_session = True
                s.source_tag = cli_meta.get("source_tag")
                s.raw_source = cli_meta.get("raw_source") or cli_meta.get("source_tag")
                s.session_source = cli_meta.get("session_source")
                s.source_label = cli_meta.get("source_label")
                s.user_id = cli_meta.get("user_id")
                s.chat_id = cli_meta.get("chat_id")
                s.chat_type = cli_meta.get("chat_type")
                s.thread_id = cli_meta.get("thread_id")
                s.session_key = cli_meta.get("session_key")
                s.platform = cli_meta.get("platform")
                s.save(touch_updated_at=False)
            else:
                msgs = get_cli_session_messages(sid)
                if not msgs:
                    return bad(handler, "Session not found", 404)
                s = import_cli_session(
                    sid,
                    cli_meta.get("title") or title_from(msgs, "CLI Session"),
                    msgs,
                    cli_meta.get("model") or "unknown",
                    profile=cli_meta.get("profile"),
                    created_at=cli_meta.get("created_at"),
                    updated_at=cli_meta.get("updated_at"),
                )
                s.is_cli_session = True
                s.source_tag = cli_meta.get("source_tag")
                s.raw_source = cli_meta.get("raw_source") or cli_meta.get("source_tag")
                s.session_source = cli_meta.get("session_source")
                s.source_label = cli_meta.get("source_label")
                s.user_id = cli_meta.get("user_id")
                s.chat_id = cli_meta.get("chat_id")
                s.chat_type = cli_meta.get("chat_type")
                s.thread_id = cli_meta.get("thread_id")
                s.session_key = cli_meta.get("session_key")
                s.platform = cli_meta.get("platform")
        with _get_session_agent_lock(sid):
            s.archived = bool(body.get("archived", True))
            s.save(touch_updated_at=False)
        return j(handler, {"ok": True, "session": s.compact()})

    # ── Session move to project (POST) ──
    if parsed.path == "/api/session/move":
        try:
            require(body, "session_id")
        except ValueError as e:
            return bad(handler, str(e))
        try:
            s = get_session(body["session_id"])
        except KeyError:
            return bad(handler, "Session not found", 404)
        with _get_session_agent_lock(body["session_id"]):
            s.project_id = body.get("project_id") or None
            s.save()
        return j(handler, {"ok": True, "session": s.compact()})

    # ── Project CRUD (POST) ──
    if parsed.path == "/api/projects/create":
        try:
            require(body, "name")
        except ValueError as e:
            return bad(handler, str(e))
        import re as _re

        name = body["name"].strip()[:128]
        if not name:
            return bad(handler, "name required")
        color = body.get("color")
        if color and not _re.match(r"^#[0-9a-fA-F]{3,8}$", color):
            return bad(handler, "Invalid color format")
        projects = load_projects()
        proj = {
            "project_id": uuid.uuid4().hex[:12],
            "name": name,
            "color": color,
            "created_at": time.time(),
        }
        projects.append(proj)
        save_projects(projects)
        return j(handler, {"ok": True, "project": proj})

    if parsed.path == "/api/projects/rename":
        try:
            require(body, "project_id", "name")
        except ValueError as e:
            return bad(handler, str(e))
        import re as _re

        projects = load_projects()
        proj = next(
            (p for p in projects if p["project_id"] == body["project_id"]), None
        )
        if not proj:
            return bad(handler, "Project not found", 404)
        proj["name"] = body["name"].strip()[:128]
        if "color" in body:
            color = body["color"]
            if color and not _re.match(r"^#[0-9a-fA-F]{3,8}$", color):
                return bad(handler, "Invalid color format")
            proj["color"] = color
        save_projects(projects)
        return j(handler, {"ok": True, "project": proj})

    if parsed.path == "/api/projects/delete":
        try:
            require(body, "project_id")
        except ValueError as e:
            return bad(handler, str(e))
        projects = load_projects()
        proj = next(
            (p for p in projects if p["project_id"] == body["project_id"]), None
        )
        if not proj:
            return bad(handler, "Project not found", 404)
        projects = [p for p in projects if p["project_id"] != body["project_id"]]
        save_projects(projects)
        # Unassign all sessions that belonged to this project
        if SESSION_INDEX_FILE.exists():
            try:
                index = json.loads(SESSION_INDEX_FILE.read_text(encoding="utf-8"))
                for entry in index:
                    if entry.get("project_id") == body["project_id"]:
                        try:
                            s = get_session(entry["session_id"])
                            s.project_id = None
                            s.save()
                        except Exception:
                            logger.debug("Failed to update session %s", entry.get("session_id"))
            except Exception:
                logger.debug("Failed to load session index for project unlink")
        return j(handler, {"ok": True})

    # ── Session import from JSON (POST) ──
    if parsed.path == "/api/session/import":
        return _handle_session_import(handler, body)

    # ── Self-update (POST) ──
    if parsed.path == "/api/updates/apply":
        target = body.get("target", "")
        if target not in ("webui", "agent"):
            return bad(handler, 'target must be "webui" or "agent"')
        from api.updates import apply_update

        return j(handler, apply_update(target))

    if parsed.path == "/api/updates/force":
        target = body.get("target", "")
        if target not in ("webui", "agent"):
            return bad(handler, 'target must be "webui" or "agent"')
        from api.updates import apply_force_update

        return j(handler, apply_force_update(target))

    # ── CLI session import (POST) ──
    if parsed.path == "/api/session/import_cli":
        return _handle_session_import_cli(handler, body)

    # ── Auth endpoints (POST) ──
    if parsed.path == "/api/auth/login":
        from api.auth import (
            verify_password,
            create_session,
            set_auth_cookie,
            is_auth_enabled,
        )
        from api.auth import _check_login_rate, _record_login_attempt

        if not is_auth_enabled():
            return j(handler, {"ok": True, "message": "Auth not enabled"})
        client_ip = handler.client_address[0]
        if not _check_login_rate(client_ip):
            return j(
                handler,
                {"error": "Too many attempts. Try again in a minute."},
                status=429,
            )
        password = body.get("password", "")
        if not verify_password(password):
            _record_login_attempt(client_ip)
            return bad(handler, "Invalid password", 401)
        cookie_val = create_session()
        handler.send_response(200)
        handler.send_header("Content-Type", "application/json")
        handler.send_header("Cache-Control", "no-store")
        _security_headers(handler)
        set_auth_cookie(handler, cookie_val)
        handler.end_headers()
        handler.wfile.write(json.dumps({"ok": True}).encode())
        return True

    if parsed.path == "/api/auth/logout":
        from api.auth import clear_auth_cookie, invalidate_session, parse_cookie

        cookie_val = parse_cookie(handler)
        if cookie_val:
            invalidate_session(cookie_val)
        handler.send_response(200)
        handler.send_header("Content-Type", "application/json")
        handler.send_header("Cache-Control", "no-store")
        _security_headers(handler)
        clear_auth_cookie(handler)
        handler.end_headers()
        handler.wfile.write(json.dumps({"ok": True}).encode())
        return True

    # ── Checkpoints / Rollback (POST) ──
    if parsed.path == "/api/rollback/restore":
        if not body:
            return bad(handler, "request body is required")
        workspace = body.get("workspace", "")
        checkpoint = body.get("checkpoint", "")
        if not workspace or not checkpoint:
            return bad(handler, "workspace and checkpoint are required")
        try:
            from api.rollback import restore_checkpoint
            return j(handler, restore_checkpoint(workspace, checkpoint))
        except ValueError as e:
            return bad(handler, str(e))
        except Exception as e:
            logger.exception("rollback/restore failed")
            return bad(handler, str(e), status=500)

    return False  # 404

# ── GET route helpers ─────────────────────────────────────────────────────────

# MIME types for static file serving. Hoisted to module scope to avoid
# rebuilding the dict on every request.
_STATIC_MIME = {
    "css": "text/css",
    "js": "application/javascript",
    "html": "text/html",
    "svg": "image/svg+xml",
    "png": "image/png",
    "jpg": "image/jpeg",
    "jpeg": "image/jpeg",
    "ico": "image/x-icon",
    "gif": "image/gif",
    "webp": "image/webp",
    "woff": "font/woff",
    "woff2": "font/woff2",
}
# MIME types that are text-based and should carry charset=utf-8
_TEXT_MIME_TYPES = {"text/css", "application/javascript", "text/html", "image/svg+xml", "text/plain"}


def _serve_static(handler, parsed):
    static_root = (Path(__file__).parent.parent / "static").resolve()
    # Strip the leading '/static/' prefix, then resolve and sandbox
    rel = parsed.path[len("/static/") :]
    static_file = (static_root / rel).resolve()
    try:
        static_file.relative_to(static_root)
    except ValueError:
        return j(handler, {"error": "not found"}, status=404)
    if not static_file.exists() or not static_file.is_file():
        return j(handler, {"error": "not found"}, status=404)
    ext = static_file.suffix.lower()
    ct = _STATIC_MIME.get(ext.lstrip("."), "text/plain")
    ct_header = f"{ct}; charset=utf-8" if ct in _TEXT_MIME_TYPES else ct
    handler.send_response(200)
    handler.send_header("Content-Type", ct_header)
    handler.send_header("Cache-Control", "no-store")
    raw = static_file.read_bytes()
    handler.send_header("Content-Length", str(len(raw)))
    handler.end_headers()
    handler.wfile.write(raw)
    return True


def _handle_session_export(handler, parsed):
    sid = parse_qs(parsed.query).get("session_id", [""])[0]
    if not sid:
        return bad(handler, "session_id is required")
    try:
        s = get_session(sid)
    except KeyError:
        return bad(handler, "Session not found", 404)
    safe = redact_session_data(s.__dict__)
    payload = json.dumps(safe, ensure_ascii=False, indent=2)
    handler.send_response(200)
    handler.send_header("Content-Type", "application/json; charset=utf-8")
    handler.send_header(
        "Content-Disposition", f'attachment; filename="hermes-{sid}.json"'
    )
    handler.send_header("Content-Length", str(len(payload.encode("utf-8"))))
    handler.send_header("Cache-Control", "no-store")
    handler.end_headers()
    handler.wfile.write(payload.encode("utf-8"))
    return True


def _handle_sessions_search(handler, parsed):
    qs = parse_qs(parsed.query)
    q = qs.get("q", [""])[0].lower().strip()
    content_search = qs.get("content", ["1"])[0] == "1"
    depth = int(qs.get("depth", ["5"])[0])
    if not q:
        safe_sessions = []
        for s in all_sessions():
            item = dict(s)
            if isinstance(item.get("title"), str):
                item["title"] = _redact_text(item["title"])
            safe_sessions.append(item)
        return j(handler, {"sessions": safe_sessions})
    results = []
    for s in all_sessions():
        title_match = q in (s.get("title") or "").lower()
        if title_match:
            item = dict(s, match_type="title")
            if isinstance(item.get("title"), str):
                item["title"] = _redact_text(item["title"])
            results.append(item)
            continue
        if content_search:
            try:
                sess = get_session(s["session_id"])
                msgs = sess.messages[:depth] if depth else sess.messages
                for m in msgs:
                    c = m.get("content") or ""
                    if isinstance(c, list):
                        c = " ".join(
                            p.get("text", "")
                            for p in c
                            if isinstance(p, dict) and p.get("type") == "text"
                        )
                    if q in str(c).lower():
                        item = dict(s, match_type="content")
                        if isinstance(item.get("title"), str):
                            item["title"] = _redact_text(item["title"])
                        results.append(item)
                        break
            except (KeyError, Exception):
                pass
    return j(handler, {"sessions": results, "query": q, "count": len(results)})


def _handle_list_dir(handler, parsed):
    qs = parse_qs(parsed.query)
    sid = qs.get("session_id", [""])[0]
    if not sid:
        return bad(handler, "session_id is required")
    try:
        s = get_session(sid)
        workspace = s.workspace
    except KeyError:
        # Fallback for CLI sessions not loaded in WebUI memory
        try:
            cli_meta = None
            for cs in get_cli_sessions():
                if cs["session_id"] == sid:
                    cli_meta = cs
                    break
            if not cli_meta:
                return bad(handler, "Session not found", 404)
            workspace = cli_meta.get("workspace", "")
        except Exception:
            return bad(handler, "Session not found", 404)
    try:
        return j(
            handler,
            {
                "entries": list_dir(Path(workspace), qs.get("path", ["."])[0]),
                "path": qs.get("path", ["."])[0],
            },
        )
    except (FileNotFoundError, ValueError) as e:
        return bad(handler, _sanitize_error(e), 404)


def _handle_sse_stream(handler, parsed):
    stream_id = parse_qs(parsed.query).get("stream_id", [""])[0]
    q = STREAMS.get(stream_id)
    if q is None:
        return j(handler, {"error": "stream not found"}, status=404)
    handler.send_response(200)
    handler.send_header("Content-Type", "text/event-stream; charset=utf-8")
    handler.send_header("Cache-Control", "no-cache")
    handler.send_header("X-Accel-Buffering", "no")
    handler.send_header("Connection", "keep-alive")
    handler.end_headers()
    try:
        while True:
            try:
                event, data = q.get(timeout=30)
            except queue.Empty:
                handler.wfile.write(b": heartbeat\n\n")
                handler.wfile.flush()
                continue
            _sse(handler, event, data)
            if event in ("stream_end", "error", "cancel"):
                break
    except _CLIENT_DISCONNECT_ERRORS:
        pass
    return True


def _terminal_session_and_workspace(body_or_query):
    sid = str(body_or_query.get("session_id", "")).strip()
    if not sid:
        raise ValueError("session_id required")
    try:
        s = get_session(sid)
    except KeyError:
        raise KeyError("Session not found")
    workspace = resolve_trusted_workspace(getattr(s, "workspace", "") or "")
    return sid, workspace


def _handle_terminal_start(handler, body):
    try:
        sid, workspace = _terminal_session_and_workspace(body)
        from api.terminal import start_terminal
        term = start_terminal(
            sid,
            workspace,
            rows=int(body.get("rows") or 24),
            cols=int(body.get("cols") or 80),
            restart=bool(body.get("restart")),
        )
        return j(
            handler,
            {
                "ok": True,
                "session_id": sid,
                "workspace": term.workspace,
                "running": term.is_alive(),
            },
        )
    except KeyError as e:
        return bad(handler, str(e), 404)
    except ValueError as e:
        return bad(handler, str(e), 400)
    except Exception as e:
        return bad(handler, _sanitize_error(e), 500)


def _handle_terminal_input(handler, body):
    try:
        require(body, "session_id")
        data = str(body.get("data", ""))
        if len(data) > 8192:
            return bad(handler, "input too large", 413)
        from api.terminal import write_terminal
        write_terminal(body["session_id"], data)
        return j(handler, {"ok": True})
    except KeyError as e:
        return bad(handler, str(e), 404)
    except ValueError as e:
        return bad(handler, str(e), 400)
    except Exception as e:
        return bad(handler, _sanitize_error(e), 500)


def _handle_terminal_resize(handler, body):
    try:
        require(body, "session_id")
        from api.terminal import resize_terminal
        resize_terminal(
            body["session_id"],
            rows=int(body.get("rows") or 24),
            cols=int(body.get("cols") or 80),
        )
        return j(handler, {"ok": True})
    except KeyError as e:
        return bad(handler, str(e), 404)
    except ValueError as e:
        return bad(handler, str(e), 400)
    except Exception as e:
        return bad(handler, _sanitize_error(e), 500)


def _handle_terminal_close(handler, body):
    try:
        require(body, "session_id")
        from api.terminal import close_terminal
        closed = close_terminal(body["session_id"])
        return j(handler, {"ok": True, "closed": closed})
    except ValueError as e:
        return bad(handler, str(e), 400)


def _handle_terminal_output(handler, parsed):
    qs = parse_qs(parsed.query)
    sid = qs.get("session_id", [""])[0]
    if not sid:
        return bad(handler, "session_id required")
    from api.terminal import get_terminal
    term = get_terminal(sid)
    if term is None:
        return j(handler, {"error": "terminal not running"}, status=404)

    handler.send_response(200)
    handler.send_header("Content-Type", "text/event-stream; charset=utf-8")
    handler.send_header("Cache-Control", "no-cache")
    handler.send_header("X-Accel-Buffering", "no")
    handler.send_header("Connection", "keep-alive")
    handler.end_headers()
    try:
        while True:
            try:
                event, data = term.output.get(timeout=25)
            except queue.Empty:
                handler.wfile.write(b": terminal heartbeat\n\n")
                handler.wfile.flush()
                if term.closed.is_set() and term.output.empty():
                    _sse(handler, "terminal_closed", {"exit_code": term.proc.poll()})
                    break
                continue
            _sse(handler, event, data)
            if event in ("terminal_closed", "terminal_error"):
                break
    except (BrokenPipeError, ConnectionResetError, ConnectionAbortedError):
        pass
    return True


def _gateway_sse_probe_payload(settings, watcher):
    enabled = bool(settings.get('show_cli_sessions'))
    # Use the public is_alive() accessor where available (current GatewayWatcher);
    # fall back to the private _thread check for any older in-memory instance
    # that might still be hanging around mid-upgrade, and for test doubles that
    # don't implement the full public API.
    if watcher is None:
        watcher_alive = False
    elif hasattr(watcher, 'is_alive') and callable(getattr(watcher, 'is_alive')):
        watcher_alive = bool(watcher.is_alive())
    else:
        _t = getattr(watcher, '_thread', None)
        watcher_alive = _t is not None and _t.is_alive()
    payload = {
        'enabled': enabled,
        'fallback_poll_ms': 30000,
        'ok': enabled and watcher_alive,
        'watcher_running': watcher_alive,
    }
    if not enabled:
        payload['error'] = 'agent sessions not enabled'
        return payload, 404
    if not watcher_alive:
        payload['error'] = 'watcher not started'
        return payload, 503
    return payload, 200


def _handle_gateway_sse_stream(handler, parsed):
    """SSE endpoint for real-time gateway session updates.
    Streams change events from the gateway watcher background thread.
    Only active when show_cli_sessions (show_agent_sessions) setting is enabled.
    """
    settings = load_settings()

    from api.gateway_watcher import get_watcher
    watcher = get_watcher()

    probe = parse_qs(parsed.query).get('probe', [''])[0].lower() in {'1', 'true', 'yes'}
    if probe:
        payload, status = _gateway_sse_probe_payload(settings, watcher)
        return j(handler, payload, status=status)

    # Check if the feature is enabled
    if not settings.get('show_cli_sessions'):
        return j(handler, {'error': 'agent sessions not enabled'}, status=404)

    # Same watcher_alive semantics as the probe path — centralised via
    # the helper so both branches stay in sync.
    _probe_body, _probe_status = _gateway_sse_probe_payload(settings, watcher)
    if not _probe_body['watcher_running']:
        return j(handler, {'error': 'watcher not started'}, status=503)

    handler.send_response(200)
    handler.send_header('Content-Type', 'text/event-stream; charset=utf-8')
    handler.send_header('Cache-Control', 'no-cache')
    handler.send_header('X-Accel-Buffering', 'no')
    handler.send_header('Connection', 'keep-alive')
    handler.end_headers()

    q = watcher.subscribe()
    try:
        # Send initial snapshot immediately
        from api.models import get_cli_sessions
        initial = get_cli_sessions()
        _sse(handler, 'sessions_changed', {'sessions': initial})

        while True:
            try:
                event_data = q.get(timeout=30)
            except queue.Empty:
                handler.wfile.write(b': keepalive\n\n')
                handler.wfile.flush()
                continue
            if event_data is None:
                break  # watcher is stopping
            _sse(handler, event_data.get('type', 'sessions_changed'), event_data)
    except _CLIENT_DISCONNECT_ERRORS:
        pass
    finally:
        watcher.unsubscribe(q)
    return True


def _content_disposition_value(disposition: str, filename: str) -> str:
    """Build a latin-1-safe Content-Disposition value with RFC 5987 filename*."""
    import urllib.parse as _up

    safe_name = Path(filename).name.replace("\r", "").replace("\n", "")
    ascii_fallback = "".join(
        ch if 32 <= ord(ch) < 127 and ch not in {'"', '\\'} else "_"
        for ch in safe_name
    ).strip(" .")
    if not ascii_fallback:
        suffix = Path(safe_name).suffix
        ascii_suffix = "".join(
            ch if 32 <= ord(ch) < 127 and ch not in {'"', '\\'} else "_"
            for ch in suffix
        )
        ascii_fallback = f"download{ascii_suffix}" if ascii_suffix else "download"
    quoted_name = _up.quote(safe_name, safe="")
    return (
        f'{disposition}; filename="{ascii_fallback}"; '
        f"filename*=UTF-8''{quoted_name}"
    )


def _parse_range_header(range_header: str, file_size: int) -> tuple[int, int] | None:
    """Parse a single HTTP bytes range into inclusive start/end offsets."""
    if not range_header or not range_header.startswith("bytes=") or file_size < 1:
        return None
    spec = range_header.split("=", 1)[1].strip()
    if "," in spec or "-" not in spec:
        return None
    start_s, end_s = spec.split("-", 1)
    try:
        if start_s == "":
            # suffix range: bytes=-500
            suffix_len = int(end_s)
            if suffix_len <= 0:
                return None
            start = max(0, file_size - suffix_len)
            end = file_size - 1
        else:
            start = int(start_s)
            end = int(end_s) if end_s else file_size - 1
            if start < 0:
                return None
            end = min(end, file_size - 1)
        if start > end or start >= file_size:
            return None
        return start, end
    except ValueError:
        return None


def _serve_file_bytes(handler, target: Path, mime: str, disposition: str, cache_control: str, *, csp: str | None = None):
    """Serve a file with correct MIME/disposition and optional byte-range support."""
    try:
        file_size = target.stat().st_size
    except PermissionError:
        return bad(handler, "Permission denied", 403)
    except Exception:
        return bad(handler, "Could not stat file", 500)

    byte_range = _parse_range_header(handler.headers.get("Range", ""), file_size)
    if handler.headers.get("Range") and byte_range is None:
        handler.send_response(416)
        handler.send_header("Content-Range", f"bytes */{file_size}")
        handler.send_header("Accept-Ranges", "bytes")
        _security_headers(handler)
        handler.end_headers()
        return True

    start, end = byte_range if byte_range else (0, max(0, file_size - 1))
    content_length = end - start + 1 if file_size else 0
    handler.send_response(206 if byte_range else 200)
    handler.send_header("Content-Type", mime)
    handler.send_header("Content-Length", str(content_length))
    handler.send_header("Accept-Ranges", "bytes")
    if byte_range:
        handler.send_header("Content-Range", f"bytes {start}-{end}/{file_size}")
    handler.send_header("Cache-Control", cache_control)
    handler.send_header("Content-Disposition", _content_disposition_value(disposition, target.name))
    if csp:
        handler.send_header("Content-Security-Policy", csp)
    _security_headers(handler)
    handler.end_headers()

    if content_length:
        try:
            with target.open("rb") as f:
                f.seek(start)
                remaining = content_length
                while remaining:
                    chunk = f.read(min(1024 * 1024, remaining))
                    if not chunk:
                        break
                    handler.wfile.write(chunk)
                    remaining -= len(chunk)
        except PermissionError:
            return True
    return True


def _handle_media(handler, parsed):
    """Serve a local file by absolute path for inline display in the chat.

    Security:
    - Path must resolve to an allowed root (hermes home, /tmp, common dirs)
    - Auth-gated when auth is enabled
    - Only image MIME types are served inline; all others force download
    - SVG always served as attachment (XSS risk)
    - No path traversal: resolved path must stay within an allowed root
    """
    import os as _os
    from api.auth import is_auth_enabled, parse_cookie, verify_session
    _HOME = Path(_os.path.expanduser("~"))
    _HERMES_HOME = Path(_os.getenv("HERMES_HOME", str(_HOME / ".hermes"))).expanduser()

    # Auth check
    if is_auth_enabled():
        cv = parse_cookie(handler)
        if not (cv and verify_session(cv)):
            handler.send_response(401)
            handler.send_header("Content-Type", "application/json")
            handler.end_headers()
            handler.wfile.write(b'{"error":"Authentication required"}')
            return

    qs = parse_qs(parsed.query)
    raw_path = qs.get("path", [""])[0].strip()
    if not raw_path:
        return bad(handler, "path parameter required", 400)

    # Resolve the path and check it is within an allowed root
    try:
        target = Path(raw_path).resolve()
    except Exception:
        return bad(handler, "Invalid path", 400)

    # Allowed roots: hermes home, /tmp, and active workspace.
    # Intentionally NOT the entire home dir — that would expose ~/.ssh,
    # ~/.aws, browser profiles, etc. to any authenticated user.
    allowed_roots = [
        _HERMES_HOME.resolve(),
        Path("/tmp").resolve(),
        (_HOME / ".hermes").resolve(),
    ]
    # Also allow the active workspace directory (where screenshots land)
    try:
        from api.workspace import get_last_workspace
        ws = Path(get_last_workspace()).resolve()
        if ws.is_dir():
            allowed_roots.append(ws)
    except Exception:
        pass
    within_allowed = any(
        _os.path.commonpath([str(target), str(root)]) == str(root)
        for root in allowed_roots
        if root.exists()
    )
    if not within_allowed:
        return bad(handler, "Path not in allowed location", 403)

    if not target.exists() or not target.is_file():
        return j(handler, {"error": "not found"}, status=404)

    # Determine MIME type
    ext = target.suffix.lower()
    mime = MIME_MAP.get(ext, "application/octet-stream")

    # Only serve safe media/PDF types inline when explicitly requested. Everything
    # else remains a download. SVG is always a download (XSS risk).
    _INLINE_IMAGE_TYPES = {
        "image/png", "image/jpeg", "image/gif", "image/webp",
        "image/x-icon", "image/bmp",
    }
    _INLINE_PREVIEW_TYPES = _INLINE_IMAGE_TYPES | {
        "audio/mpeg", "audio/wav", "audio/x-wav", "audio/mp4", "audio/aac",
        "audio/ogg", "audio/opus", "audio/flac",
        "video/mp4", "video/quicktime", "video/webm", "video/ogg",
        "application/pdf",
    }
    _DOWNLOAD_TYPES = {"image/svg+xml"}  # SVG: XSS risk, force download
    inline_preview = qs.get("inline", [""])[0] == "1"
    disposition = "inline" if (
        mime not in _DOWNLOAD_TYPES and (
            mime in _INLINE_IMAGE_TYPES or (inline_preview and mime in _INLINE_PREVIEW_TYPES)
        )
    ) else "attachment"
    return _serve_file_bytes(handler, target, mime, disposition, "private, max-age=3600")


def _handle_file_raw(handler, parsed):
    qs = parse_qs(parsed.query)
    sid = qs.get("session_id", [""])[0]
    if not sid:
        return bad(handler, "session_id is required")
    try:
        s = get_session(sid)
    except KeyError:
        return bad(handler, "Session not found", 404)
    rel = qs.get("path", [""])[0]
    force_download = qs.get("download", [""])[0] == "1"
    target = safe_resolve(Path(s.workspace), rel)
    if not target.exists() or not target.is_file():
        return j(handler, {"error": "not found"}, status=404)
    ext = target.suffix.lower()
    mime = MIME_MAP.get(ext, "application/octet-stream")
    # Security: force download for dangerous MIME types to prevent XSS.
    # Exception: ?inline=1 permits text/html to be served inline for the
    # sandboxed workspace HTML preview iframe (sandbox="allow-scripts" with no
    # allow-same-origin, so the iframe cannot access parent cookies/storage).
    inline_preview = qs.get("inline", [""])[0] == "1"
    dangerous_types = {"text/html", "application/xhtml+xml", "image/svg+xml"}
    html_inline_ok = inline_preview and mime == "text/html"
    disposition = "attachment" if force_download or (mime in dangerous_types and not html_inline_ok) else "inline"
    # Defense-in-depth for ?inline=1 HTML: even though the workspace.js iframe
    # sets sandbox="allow-scripts", a user could be tricked into opening the
    # ?inline=1 URL directly in a top-level tab (e.g. via a chat link), which
    # would render the HTML in the WebUI's origin without iframe sandbox. The
    # CSP sandbox directive applies the same isolation server-side: without
    # allow-same-origin, the document is treated as a unique opaque origin and
    # cannot read WebUI cookies, localStorage, or postMessage to the parent.
    csp = "sandbox allow-scripts" if html_inline_ok else None
    # _serve_file_bytes sends Content-Security-Policy when csp is set.
    return _serve_file_bytes(handler, target, mime, disposition, "no-store", csp=csp)


def _handle_file_read(handler, parsed):
    qs = parse_qs(parsed.query)
    sid = qs.get("session_id", [""])[0]
    if not sid:
        return bad(handler, "session_id is required")
    try:
        s = get_session(sid)
    except KeyError:
        return bad(handler, "Session not found", 404)
    rel = qs.get("path", [""])[0]
    if not rel:
        return bad(handler, "path is required")
    try:
        return j(handler, read_file_content(Path(s.workspace), rel))
    except (FileNotFoundError, ValueError) as e:
        return bad(handler, _sanitize_error(e), 404)


def _handle_approval_pending(handler, parsed):
    sid = parse_qs(parsed.query).get("session_id", [""])[0]
    with _lock:
        queue = _pending.get(sid)
        # Support both the new list format and a legacy single-dict value.
        if isinstance(queue, list):
            p = queue[0] if queue else None
            total = len(queue)
        elif queue:
            p = queue
            total = 1
        else:
            p = None
            total = 0
    if p:
        return j(handler, {"pending": dict(p), "pending_count": total})
    return j(handler, {"pending": None, "pending_count": 0})


def _handle_approval_sse_stream(handler, parsed):
    """SSE endpoint for real-time approval notifications.

    Long-lived connection that pushes approval events the moment they arrive,
    replacing the 1.5s polling loop.  The frontend uses EventSource and falls
    back to HTTP polling if the connection fails.
    """
    sid = parse_qs(parsed.query).get("session_id", [""])[0]
    if not sid:
        return bad(handler, "session_id is required")

    # Subscribe AND snapshot atomically under a single _lock acquisition so a
    # submit_pending() that fires between the two cannot be lost. If we
    # snapshot first then subscribe (the naive ordering), an approval that
    # arrives in the gap is appended to _pending (after our snapshot) AND
    # notified to subscribers (before we joined) — leaving the client unaware
    # until the next event arrives.
    q = queue.Queue(maxsize=16)
    initial_pending = None
    initial_count = 0
    with _lock:
        _approval_sse_subscribers.setdefault(sid, []).append(q)
        q_list = _pending.get(sid)
        if isinstance(q_list, list):
            initial_pending = dict(q_list[0]) if q_list else None
            initial_count = len(q_list)
        elif q_list:
            initial_pending = dict(q_list)
            initial_count = 1

    handler.send_response(200)
    handler.send_header('Content-Type', 'text/event-stream; charset=utf-8')
    handler.send_header('Cache-Control', 'no-cache')
    handler.send_header('X-Accel-Buffering', 'no')
    handler.send_header('Connection', 'keep-alive')
    handler.end_headers()

    from api.streaming import _sse

    # Push initial state immediately so the client doesn't miss anything.
    _sse(handler, 'initial', {"pending": initial_pending, "pending_count": initial_count})

    try:
        while True:
            try:
                payload = q.get(timeout=30)
            except queue.Empty:
                # Keepalive — SSE comment line prevents proxy/CDN timeout.
                handler.wfile.write(b': keepalive\n\n')
                handler.wfile.flush()
                continue
            if payload is None:
                break  # signal to close
            _sse(handler, 'approval', payload)
    except _CLIENT_DISCONNECT_ERRORS:
        pass  # client went away — normal for long-lived connections
    finally:
        _approval_sse_unsubscribe(sid, q)


def _handle_approval_inject(handler, parsed):
    """Inject a fake pending approval -- loopback-only, used by automated tests."""
    qs = parse_qs(parsed.query)
    sid = qs.get("session_id", [""])[0]
    key = qs.get("pattern_key", ["test_pattern"])[0]
    cmd = qs.get("command", ["rm -rf /tmp/test"])[0]
    if sid:
        submit_pending(
            sid,
            {
                "command": cmd,
                "pattern_key": key,
                "pattern_keys": [key],
                "description": "test pattern",
            },
        )
        return j(handler, {"ok": True, "session_id": sid})
    return j(handler, {"error": "session_id required"}, status=400)


def _handle_clarify_pending(handler, parsed):
    sid = parse_qs(parsed.query).get("session_id", [""])[0]
    pending = get_clarify_pending(sid)
    if pending:
        return j(handler, {"pending": pending})
    return j(handler, {"pending": None})


def _handle_clarify_sse_stream(handler, parsed):
    """SSE endpoint for real-time clarify notifications.

    Long-lived connection that pushes clarify events the moment they arrive,
    replacing the 1.5s polling loop.  The frontend uses EventSource and falls
    back to HTTP polling if the connection fails.
    """
    if clarify_sse_subscribe is None:
        return bad(handler, "clarify SSE not available")

    sid = parse_qs(parsed.query).get("session_id", [""])[0]
    if not sid:
        return bad(handler, "session_id is required")

    # Subscribe AND snapshot atomically.  We import clarify's _lock so that
    # subscribe and the snapshot read happen under the same mutex — same
    # pattern as the approval SSE handler.
    #
    # NOTE: We must NOT call clarify.get_pending() here — it acquires _lock
    # internally, which would deadlock since clarify._lock is a non-reentrant
    # threading.Lock.  Instead, read _gateway_queues / _pending inline under
    # the lock we already hold.
    from api.clarify import (
        _lock as _clarify_lock,
        _clarify_sse_subscribers as _clarify_subs,
        _gateway_queues as _clarify_gateway_queues,
        _pending as _clarify_pending,
    )
    q = queue.Queue(maxsize=16)
    initial_pending = None
    initial_count = 0
    with _clarify_lock:
        _clarify_subs.setdefault(sid, []).append(q)
        gw_q = _clarify_gateway_queues.get(sid) or []
        if gw_q:
            initial_pending = dict(gw_q[0].data)
            initial_count = len(gw_q)
        else:
            _legacy = _clarify_pending.get(sid)
            if _legacy:
                initial_pending = dict(_legacy)
                initial_count = 1

    handler.send_response(200)
    handler.send_header('Content-Type', 'text/event-stream; charset=utf-8')
    handler.send_header('Cache-Control', 'no-cache')
    handler.send_header('X-Accel-Buffering', 'no')
    handler.send_header('Connection', 'keep-alive')
    handler.end_headers()

    from api.streaming import _sse

    # Push initial state immediately so the client doesn't miss anything.
    _sse(handler, 'initial', {"pending": initial_pending, "pending_count": initial_count})

    try:
        while True:
            try:
                payload = q.get(timeout=30)
            except queue.Empty:
                handler.wfile.write(b': keepalive\n\n')
                handler.wfile.flush()
                continue
            if payload is None:
                break
            _sse(handler, 'clarify', payload)
    except _CLIENT_DISCONNECT_ERRORS:
        pass
    finally:
        clarify_sse_unsubscribe(sid, q)


def _handle_clarify_inject(handler, parsed):
    """Inject a fake pending clarify prompt -- loopback-only, used by automated tests."""
    qs = parse_qs(parsed.query)
    sid = qs.get("session_id", [""])[0]
    question = qs.get("question", ["Which option?"])[0]
    choices = qs.get("choices", [])
    if sid:
        submit_clarify_pending(
            sid,
            {
                "question": question,
                "choices_offered": choices,
                "session_id": sid,
                "kind": "clarify",
            },
        )
        return j(handler, {"ok": True, "session_id": sid})
    return j(handler, {"error": "session_id required"}, status=400)


def _handle_live_models(handler, parsed):
    """Return the live model list for a provider.

    Delegates to the agent's provider_model_ids() which handles:
    - OpenRouter: live fetch from /api/v1/models
    - Anthropic: live fetch from /v1/models (API key or OAuth token)
    - Copilot: live fetch from api.githubcopilot.com/models with correct headers
    - openai-codex: Codex OAuth endpoint + local ~/.codex/ cache fallback
    - Nous: live fetch from inference-api.nousresearch.com/v1/models
    - DeepSeek, kimi-coding, opencode-zen/go, custom: generic OpenAI-compat /v1/models
    - ZAI, MiniMax, Google/Gemini: fall back to static list (non-standard endpoints)
    - All others: static _PROVIDER_MODELS fallback

    The agent already maintains all provider-specific auth and endpoint logic
    in one place; the WebUI inherits it rather than duplicating it.

    Query params:
        provider  (optional) — provider ID; defaults to active profile provider
    """
    qs = parse_qs(parsed.query)
    provider = (qs.get("provider", [""])[0] or "").lower().strip()

    try:
        from api.config import get_config as _gc
        cfg = _gc()
        if not provider:
            provider = cfg.get("model", {}).get("provider") or ""
        if not provider:
            return j(handler, {"error": "no_provider", "models": []})

        # Normalize provider alias so 'z.ai' -> 'zai', 'x.ai' -> 'xai', etc.
        # The browser sends whatever active_provider the static endpoint returned;
        # without normalization, provider_model_ids() misses the alias and returns [].
        # Uses the WebUI-owned table (api/config._resolve_provider_alias) which
        # works even when hermes_cli is not on sys.path.
        from api.config import _resolve_provider_alias
        provider = _resolve_provider_alias(provider)

        cache_key = _live_models_cache_key(provider)
        cached = _get_cached_live_models(cache_key)
        if cached is not None:
            return j(handler, cached)

        def _finish(payload: dict):
            _set_cached_live_models(cache_key, payload)
            return j(handler, payload)

        # Delegate to the agent's live-fetch + fallback resolver.
        # provider_model_ids() tries live endpoints first and falls back to
        # the static _PROVIDER_MODELS list — it never raises.
        try:
            import sys as _sys
            import os as _os
            _agent_dir = _os.path.join(_os.path.dirname(_os.path.dirname(_os.path.abspath(__file__))),
                                       "..", "..", ".hermes", "hermes-agent")
            _agent_dir = _os.path.normpath(_agent_dir)
            if _agent_dir not in _sys.path:
                _sys.path.insert(0, _agent_dir)
            from hermes_cli.models import provider_model_ids as _pmi
            ids = _pmi(provider)
        except Exception as _import_err:
            logger.debug("provider_model_ids import failed for %s: %s", provider, _import_err)
            ids = []

        if not ids:
            # For 'custom' provider, provider_model_ids() returns [] because
            # 'custom' isn't a real endpoint.  Fall back to the custom_providers
            # entries from config.yaml so the live-model enrichment step can
            # add any models that weren't already in the static list.
            if provider == "custom":
                try:
                    _cp_entries = cfg.get("custom_providers", [])
                    if isinstance(_cp_entries, list):
                        ids = [
                            _cp.get("model", "")
                            for _cp in _cp_entries
                            if isinstance(_cp, dict) and _cp.get("model", "")
                        ]
                except Exception:
                    pass
            
            # If still no ids, try fetching from model.base_url directly (OpenAI-compat endpoint)
            if not ids and provider == "custom":
                _base_url = cfg.get("model", {}).get("base_url")
                _api_key = cfg.get("model", {}).get("api_key")
                if _base_url and _api_key:
                    try:
                        import urllib.request
                        import json
                        
                        # Build the models endpoint URL
                        # AxonHub and similar OpenAI-compat endpoints serve /v1/models
                        _ep = _base_url.rstrip("/")
                        # If base_url already ends with /v1, use /models; otherwise add /v1/models
                        if _ep.endswith("/v1"):
                            _models_url = f"{_ep}/models"
                        else:
                            _models_url = f"{_ep}/v1/models"
                        
                        _req = urllib.request.Request(
                            _models_url,
                            headers={"Authorization": f"Bearer {_api_key}"},
                        )
                        
                        with urllib.request.urlopen(_req, timeout=8) as _resp:
                            _body = json.loads(_resp.read())
                        
                        # Parse response: {"data": [{"id": "model1", ...}, ...]}
                        if isinstance(_body, dict):
                            _data = _body.get("data", [])
                            if isinstance(_data, list):
                                ids = [m.get("id", "") for m in _data if m.get("id")]
                        elif isinstance(_body, list):
                            ids = [m.get("id", m) if isinstance(m, dict) else m for m in _body]
                        
                        if ids:
                            logger.debug("Live-fetched %d models from custom provider %s", len(ids), _base_url)
                        else:
                            logger.debug("Custom provider returned no models from %s", _base_url)
                    
                    except Exception as _fetch_err:
                        logger.debug("Live fetch from custom provider failed: %s", _fetch_err)

        # ── OpenAI-compat live fetch fallback ──────────────────────────────────
        # When provider_model_ids() is unavailable or returns [] for a provider
        # that exposes a standard /v1/models endpoint, fetch directly.  This
        # eliminates the need to keep _PROVIDER_MODELS in sync for providers
        # that have a discoverable API (#871).
        #
        # WARNING: This uses synchronous urllib.request which blocks the worker
        # thread for up to 8 seconds on timeout. This is acceptable because:
        #  (a) the server uses threading (not async), so other requests continue;
        #  (b) the frontend shows the static list immediately and enriches in
        #      the background via _fetchLiveModels(), so the user never waits.
        if not ids:
            _ep = _OPENAI_COMPAT_ENDPOINTS.get(provider)
            if _ep:
                try:
                    import urllib.request
                    _providers_cfg = cfg.get("providers", {})
                    _prov = _providers_cfg.get(provider, {}) if isinstance(_providers_cfg, dict) else {}
                    # Only use provider-scoped key — never fall back to a top-level
                    # api_key which may belong to a different provider.
                    _key = _prov.get("api_key") if isinstance(_prov, dict) else None
                    if not _key:
                        _key = cfg.get("model", {}).get("api_key")
                    if _key:
                        _req = urllib.request.Request(
                            f"{_ep}/models",
                            headers={"Authorization": f"Bearer {_key}"},
                        )
                        with urllib.request.urlopen(_req, timeout=8) as _resp:
                            _body = json.loads(_resp.read())
                        ids = [m.get("id", "") for m in _body.get("data", []) if m.get("id")]
                        logger.debug("Live-fetched %d models from %s /v1/models", len(ids), provider)
                except Exception as _fetch_err:
                    logger.debug("Live fetch from %s failed: %s", provider, _fetch_err)
                    # Fall through to static list below

        # Static fallback — only reached when live fetch also failed.
        if not ids:
            from api.config import _PROVIDER_MODELS as _pm
            ids = [m["id"] for m in _pm.get(provider, [])]
        if not ids:
            return _finish({"provider": provider, "models": [], "count": 0})

        # For Nous Portal, apply the same featured-set cap that
        # /api/models uses so background enrichment via _fetchLiveModels()
        # doesn't undo the dropdown trim — otherwise a 397-model catalog
        # would still flood the picker after the initial render finished
        # the cap. The full list is returned via the main /api/models
        # endpoint's extra_models field for /model autocomplete; the live
        # endpoint is purely a dropdown-enrichment surface, so it should
        # match the dropdown's visibility budget. (#1567)
        if provider == "nous":
            try:
                from api.config import _build_nous_featured_set
                _default_model = (cfg.get("model", {}) or {}).get("model") if isinstance(cfg.get("model"), dict) else None
                _featured, _ = _build_nous_featured_set(ids, selected_model_id=_default_model)
                ids = _featured
            except Exception:
                logger.debug("Failed to apply Nous featured-set cap for /api/models/live")

        # Normalise to {id, label} — provider_model_ids() returns plain string IDs.
        # For ollama-cloud use the shared Ollama formatter (handles `:variant` suffix).
        # For all other providers use a simpler hyphen-split capitaliser.
        from api.config import _format_ollama_label as _fmt_ollama

        def _make_label(mid):
            """Best-effort human label from a model ID string."""
            if provider in ("ollama", "ollama-cloud"):
                return _fmt_ollama(mid)
            # Preserve slashes for router IDs like "anthropic/claude-sonnet-4.6"
            display = mid.split("/")[-1] if "/" in mid else mid
            parts = display.split("-")
            result = []
            for p in parts:
                pl = p.lower()
                if pl == "gpt":
                    result.append("GPT")
                elif pl in ("claude", "gemini", "gemma", "llama", "mistral",
                            "qwen", "deepseek", "grok", "kimi", "glm"):
                    result.append(p.capitalize())
                elif p[:1].isdigit():
                    result.append(p)  # version numbers: 5.4, 3.5, 4.6 — unchanged
                else:
                    result.append(p.capitalize())
            label = " ".join(result)
            # Restore well-known uppercase tokens that title-casing breaks
            for orig in ("GPT", "GLM", "API", "AI", "XL", "MoE"):
                label = label.replace(orig.title(), orig)
            return label

        models_out = [{"id": mid, "label": _make_label(mid)} for mid in ids if mid]
        return _finish({"provider": provider, "models": models_out,
                        "count": len(models_out)})

    except Exception as _e:
        logger.debug("_handle_live_models failed for %s: %s", provider, _e)
        return j(handler, {"error": str(_e), "models": []})


def _handle_cron_history(handler, parsed):
    """List cron run output files with metadata (no content).

    Returns lightweight file listing so the frontend can render a run history
    without fetching full output for every run.
    """
    from cron.jobs import OUTPUT_DIR as CRON_OUT
    import re as _re

    qs = parse_qs(parsed.query)
    job_id = qs.get("job_id", [""])[0]
    if not job_id:
        return j(handler, {"error": "job_id required"}, status=400)
    # Defense-in-depth: cron job_ids are 12-char hex from the agent's scheduler.
    # Without validation, a job_id of "../<other>" would let an authenticated
    # caller enumerate .md filenames in adjacent directories under CRON_OUT's
    # parent. Mirror the rollback checkpoint id regex shape.
    # (Opus pre-release advisor finding.)
    if not _re.fullmatch(r"[A-Za-z0-9_-][A-Za-z0-9_.-]{0,63}", job_id) or job_id in (".", ".."):
        return j(handler, {"error": "invalid job_id"}, status=400)
    # Reject malformed offset/limit instead of letting int() raise ValueError
    # and surface as a confusing 500. Clamp to safe ranges.
    try:
        offset = max(0, int(qs.get("offset", ["0"])[0]))
        limit = max(1, min(500, int(qs.get("limit", ["50"])[0])))
    except (ValueError, TypeError):
        return j(handler, {"error": "offset and limit must be integers"}, status=400)
    out_dir = CRON_OUT / job_id
    runs = []
    total = 0
    if out_dir.exists():
        all_files = sorted(out_dir.glob("*.md"), key=lambda f: f.stat().st_mtime, reverse=True)
        total = len(all_files)
        page = all_files[offset:offset + limit]
        for f in page:
            try:
                st = f.stat()
                runs.append({
                    "filename": f.name,
                    "size": st.st_size,
                    "modified": st.st_mtime,
                })
            except OSError:
                logger.debug("Failed to stat cron output file %s", f)
    return j(handler, {"job_id": job_id, "runs": runs, "total": total, "offset": offset})


def _handle_cron_run_detail(handler, parsed):
    """Return full content of a single cron run output file."""
    from cron.jobs import OUTPUT_DIR as CRON_OUT
    import re as _re

    qs = parse_qs(parsed.query)
    job_id = qs.get("job_id", [""])[0]
    filename = qs.get("filename", [""])[0]
    if not job_id or not filename:
        return j(handler, {"error": "job_id and filename required"}, status=400)
    # Validate job_id shape (defense-in-depth even though the resolve+is_relative_to
    # check below catches traversal — fail-closed at the parameter boundary so
    # malformed job_ids return a 400 from the validator rather than a 400 from
    # the path resolver).
    if not _re.fullmatch(r"[A-Za-z0-9_-][A-Za-z0-9_.-]{0,63}", job_id) or job_id in (".", ".."):
        return j(handler, {"error": "invalid job_id"}, status=400)
    # Prevent path traversal — resolve and verify it stays within the job's output dir
    fpath = (CRON_OUT / job_id / filename).resolve()
    if not fpath.is_relative_to(CRON_OUT.resolve()):
        return j(handler, {"error": "invalid filename"}, status=400)
    if not fpath.exists():
        return j(handler, {"error": "run not found"}, status=404)
    try:
        content = fpath.read_text(encoding="utf-8", errors="replace")
        snippet = _cron_output_snippet(content)
        return j(handler, {"job_id": job_id, "filename": filename,
                           "content": content, "snippet": snippet})
    except Exception as e:
        return j(handler, {"error": str(e)}, status=500)


def _cron_output_snippet(text: str, limit: int = 600) -> str:
    """Extract the response body from a cron output .md file for preview.

    Contract: cron output files use markdown front-matter followed by a
    ``## Response`` (or ``# Response``) heading that marks the start of the
    agent's reply.  This function locates that heading and returns everything
    after it (up to *limit* chars).  If no heading is found the entire text
    is returned — callers should be aware that front-matter fields (model,
    timestamp, …) may appear in the snippet.
    """
    lines = text.split("\n")
    response_idx = -1
    for i, line in enumerate(lines):
        if line.startswith("## Response") or line.startswith("# Response"):
            response_idx = i
            break
    body = ("\n".join(lines[response_idx + 1:]) if response_idx >= 0 else "\n".join(lines)).strip()
    return body[:limit] or "(empty)"


def _handle_cron_output(handler, parsed):
    from cron.jobs import OUTPUT_DIR as CRON_OUT

    qs = parse_qs(parsed.query)
    job_id = qs.get("job_id", [""])[0]
    limit = int(qs.get("limit", ["5"])[0])
    if not job_id:
        return j(handler, {"error": "job_id required"}, status=400)
    out_dir = CRON_OUT / job_id
    outputs = []
    if out_dir.exists():
        files = sorted(out_dir.glob("*.md"), key=lambda f: f.stat().st_mtime, reverse=True)[:limit]
        for f in files:
            try:
                txt = f.read_text(encoding="utf-8", errors="replace")
                outputs.append({"filename": f.name, "content": _cron_output_content_window(txt)})
            except Exception:
                logger.debug("Failed to read cron output file %s", f)
    return j(handler, {"job_id": job_id, "outputs": outputs})


def _handle_cron_status(handler, parsed):
    """Return running status for one or all cron jobs."""
    qs = parse_qs(parsed.query)
    job_id = qs.get("job_id", [""])[0]
    if job_id:
        running, elapsed = _is_cron_running(job_id)
        return j(handler, {"job_id": job_id, "running": running, "elapsed": round(elapsed, 1)})
    # Return status for all running jobs
    with _RUNNING_CRON_LOCK:
        all_running = {jid: round(time.time() - t, 1) for jid, t in _RUNNING_CRON_JOBS.items()}
    return j(handler, {"running": all_running})


def _handle_cron_recent(handler, parsed):
    """Return cron jobs that have completed since a given timestamp."""
    import datetime

    qs = parse_qs(parsed.query)
    since = float(qs.get("since", ["0"])[0])
    try:
        from cron.jobs import list_jobs

        jobs = list_jobs(include_disabled=True)
        completions = []
        for job in jobs:
            last_run = job.get("last_run_at")
            if not last_run:
                continue
            if isinstance(last_run, str):
                try:
                    ts = datetime.datetime.fromisoformat(
                        last_run.replace("Z", "+00:00")
                    ).timestamp()
                except (ValueError, TypeError):
                    continue
            else:
                ts = float(last_run)
            if ts > since:
                completions.append(
                    {
                        "job_id": job.get("id", ""),
                        "name": job.get("name", "Unknown"),
                        "status": job.get("last_status", "unknown"),
                        "completed_at": ts,
                    }
                )
        return j(handler, {"completions": completions, "since": since})
    except ImportError:
        return j(handler, {"completions": [], "since": since})


def _handle_memory_read(handler):
    try:
        from api.profiles import get_active_hermes_home

        mem_dir = get_active_hermes_home() / "memories"
    except ImportError:
        mem_dir = Path.home() / ".hermes" / "memories"
    mem_file = mem_dir / "MEMORY.md"
    user_file = mem_dir / "USER.md"
    memory = (
        mem_file.read_text(encoding="utf-8", errors="replace")
        if mem_file.exists()
        else ""
    )
    user = (
        user_file.read_text(encoding="utf-8", errors="replace")
        if user_file.exists()
        else ""
    )
    return j(
        handler,
        {
            "memory": _redact_text(memory),
            "user": _redact_text(user),
            "memory_path": str(mem_file),
            "user_path": str(user_file),
            "memory_mtime": mem_file.stat().st_mtime if mem_file.exists() else None,
            "user_mtime": user_file.stat().st_mtime if user_file.exists() else None,
        },
    )


# ── POST route helpers ────────────────────────────────────────────────────────


def _handle_sessions_cleanup(handler, body, zero_only=False):
    cleaned = 0
    for p in SESSION_DIR.glob("*.json"):
        if p.name.startswith("_"):
            continue
        try:
            s = Session.load(p.stem)
            if zero_only:
                should_delete = s and len(s.messages) == 0
            else:
                should_delete = s and s.title == "Untitled" and len(s.messages) == 0
            if should_delete:
                with LOCK:
                    SESSIONS.pop(p.stem, None)
                p.unlink(missing_ok=True)
                cleaned += 1
        except Exception:
            logger.debug("Failed to clean up session file %s", p)
    if SESSION_INDEX_FILE.exists():
        SESSION_INDEX_FILE.unlink(missing_ok=True)
    return j(handler, {"ok": True, "cleaned": cleaned})


def _handle_btw(handler, body):
    """POST /api/btw — ephemeral side question using session context.

    Creates a temporary hidden session, streams the answer via SSE, then
    discards the session. The parent session is not modified.
    """
    try:
        require(body, "session_id")
        require(body, "question")
    except ValueError as e:
        return bad(handler, str(e))
    try:
        s = get_session(body["session_id"])
    except KeyError:
        return bad(handler, "Session not found", 404)
    question = str(body["question"]).strip()
    if not question:
        return bad(handler, "question is required")
    # Duplicate-stream guard (same pattern as chat/start)
    current_stream_id = getattr(s, "active_stream_id", None)
    if current_stream_id:
        with STREAMS_LOCK:
            if current_stream_id in STREAMS:
                return j(handler, {"error": "session already has an active stream"}, status=409)
        s.active_stream_id = None
    # Create ephemeral hidden session inheriting context
    from api.models import new_session as _new_session
    model_provider = getattr(s, 'model_provider', None)
    ephemeral = _new_session(
        workspace=s.workspace,
        model=s.model,
        model_provider=model_provider,
        profile=getattr(s, 'profile', None),
    )
    # Copy conversation history for context (agent reads from messages)
    ephemeral.messages = list(s.messages or [])
    ephemeral.title = f"btw: {question[:60]}"
    ephemeral.save()
    stream_id = uuid.uuid4().hex
    ephemeral.active_stream_id = stream_id
    ephemeral.save()
    q = queue.Queue()
    with STREAMS_LOCK:
        STREAMS[stream_id] = q
    from api.background import track_btw
    track_btw(body["session_id"], ephemeral.session_id, stream_id, question)
    thr = threading.Thread(
        target=_run_agent_streaming,
        args=(ephemeral.session_id, question, s.model, s.workspace, stream_id, None),
        kwargs={"ephemeral": True, "model_provider": model_provider},
        daemon=True,
    )
    thr.start()
    return j(handler, {"stream_id": stream_id, "session_id": ephemeral.session_id, "parent_session_id": body["session_id"]})


def _handle_background(handler, body):
    """POST /api/background — run prompt in parallel background agent.

    Creates a hidden session, starts streaming in a daemon thread.
    Frontend polls /api/background/status for completed results.
    """
    try:
        require(body, "session_id")
        require(body, "prompt")
    except ValueError as e:
        return bad(handler, str(e))
    try:
        s = get_session(body["session_id"])
    except KeyError:
        return bad(handler, "Session not found", 404)
    prompt = str(body["prompt"]).strip()
    if not prompt:
        return bad(handler, "prompt is required")
    from api.models import new_session as _new_session
    model_provider = getattr(s, 'model_provider', None)
    bg = _new_session(
        workspace=s.workspace,
        model=s.model,
        model_provider=model_provider,
        profile=getattr(s, 'profile', None),
    )
    bg.title = f"bg: {prompt[:60]}"
    bg.save()
    stream_id = uuid.uuid4().hex
    bg.active_stream_id = stream_id
    bg.save()
    q = queue.Queue()
    with STREAMS_LOCK:
        STREAMS[stream_id] = q
    task_id = uuid.uuid4().hex[:8]
    from api.background import track_background, complete_background
    parent_sid = body["session_id"]
    bg_sid = bg.session_id
    track_background(parent_sid, bg_sid, stream_id, task_id, prompt)

    def _run_bg_and_notify():
        """Run the background agent, then mark the tracked task `done` with the
        last assistant reply so `/api/background/status` can surface it.  Without
        this, `complete_background()` is never called and the result is lost —
        `get_results()` would see a forever-`running` task and return nothing.
        """
        try:
            _run_agent_streaming(
                bg_sid,
                prompt,
                s.model,
                s.workspace,
                stream_id,
                None,
                model_provider=model_provider,
            )
            # Reload the bg session from disk and extract the final assistant reply.
            try:
                from api.models import Session as _Session
                reloaded = _Session.load(bg_sid)
                _answer = ""
                for _m in reversed((reloaded.messages if reloaded else None) or []):
                    if not isinstance(_m, dict) or _m.get("role") != "assistant":
                        continue
                    if _m.get("_error"):
                        continue
                    _content = str(_m.get("content") or "").strip()
                    if _content:
                        _answer = _content
                        break
                complete_background(parent_sid, task_id, _answer or "(no answer produced)")
            except Exception:
                complete_background(parent_sid, task_id, "(background task failed)")
            # Best-effort cleanup of the hidden bg session file so it doesn't
            # clutter the sidebar or SESSION_DIR. The index is pruned on the
            # next rebuild via _index_entry_exists().
            try:
                (SESSION_DIR / f"{bg_sid}.json").unlink(missing_ok=True)
            except Exception:
                pass
        except Exception:
            try:
                complete_background(parent_sid, task_id, "(background task failed)")
            except Exception:
                pass

    thr = threading.Thread(target=_run_bg_and_notify, daemon=True)
    thr.start()
    return j(handler, {"task_id": task_id, "stream_id": stream_id, "session_id": bg.session_id})


def _handle_chat_start(handler, body):
    try:
        require(body, "session_id")
    except ValueError as e:
        return bad(handler, str(e))
    try:
        s = get_session(body["session_id"])
    except KeyError:
        return bad(handler, "Session not found", 404)
    msg = str(body.get("message", "")).strip()
    if not msg:
        return bad(handler, "message is required")
    attachments = _normalize_chat_attachments(body.get("attachments") or [])[:20]
    try:
        workspace = str(resolve_trusted_workspace(body.get("workspace") or s.workspace))
    except ValueError as e:
        return bad(handler, str(e))
    requested_model = body.get("model") or s.model
    requested_provider = (
        body.get("model_provider")
        if "model_provider" in body
        else getattr(s, "model_provider", None)
    )
    model, model_provider, normalized_model = _resolve_compatible_session_model_state(
        requested_model,
        requested_provider,
    )
    # Prevent duplicate runs in the same session while a stream is still active.
    # This commonly happens after page refresh/reconnect races and can produce
    # duplicated clarify cards for what appears to be a single user request.
    current_stream_id = getattr(s, "active_stream_id", None)
    if current_stream_id:
        with STREAMS_LOCK:
            current_active = current_stream_id in STREAMS
        if current_active:
            return j(
                handler,
                {
                    "error": "session already has an active stream",
                    "active_stream_id": current_stream_id,
                },
                status=409,
            )
        # Stale stream id from a previous run; clear and continue.
        _clear_stale_stream_state(s)
    stream_id = uuid.uuid4().hex
    with _get_session_agent_lock(s.session_id):
        s.workspace = workspace
        s.model = model
        s.model_provider = model_provider
        s.active_stream_id = stream_id
        s.pending_user_message = msg
        s.pending_attachments = attachments
        s.pending_started_at = time.time()
        s.save()
    set_last_workspace(workspace)
    q = queue.Queue()
    with STREAMS_LOCK:
        STREAMS[stream_id] = q
    thr = threading.Thread(
        target=_run_agent_streaming,
        args=(s.session_id, msg, model, workspace, stream_id, attachments),
        kwargs={"model_provider": model_provider},
        daemon=True,
    )
    thr.start()
    response = {"stream_id": stream_id, "session_id": s.session_id}
    if normalized_model:
        response["effective_model"] = model
    if model_provider:
        response["effective_model_provider"] = model_provider
    return j(handler, response)


def _normalize_chat_attachments(raw_attachments):
    """Normalize attachment payloads from the browser.

    Older clients send a list of filenames. Newer clients send upload result
    objects containing name/path/mime/size so image attachments can be supplied
    to Hermes as native multimodal inputs for the current turn.
    """
    normalized = []
    if not isinstance(raw_attachments, list):
        return normalized
    for item in raw_attachments:
        if isinstance(item, dict):
            name = str(item.get("name") or item.get("filename") or "").strip()
            path = str(item.get("path") or "").strip()
            mime = str(item.get("mime") or "").strip()
            att = {"name": name or path, "path": path, "mime": mime}
            size = item.get("size")
            if isinstance(size, int):
                att["size"] = size
            is_image = item.get("is_image")
            if isinstance(is_image, bool):
                att["is_image"] = is_image
            normalized.append(att)
        else:
            value = str(item).strip()
            if value:
                normalized.append({"name": value, "path": "", "mime": ""})
    return normalized


def _handle_chat_sync(handler, body):
    """Fallback synchronous chat endpoint (POST /api/chat). Not used by frontend."""
    s = get_session(body["session_id"])
    msg = str(body.get("message", "")).strip()
    if not msg:
        return j(handler, {"error": "empty message"}, status=400)
    try:
        workspace = str(resolve_trusted_workspace(body.get("workspace") or s.workspace))
    except ValueError as e:
        return bad(handler, str(e))
    with _get_session_agent_lock(s.session_id):
        s.workspace = workspace
        model, model_provider = _resolve_compatible_session_model_state(
            body.get("model") or s.model,
            body.get("model_provider") if "model_provider" in body else getattr(s, "model_provider", None),
        )[:2]
        s.model = model
        s.model_provider = model_provider
    from api.streaming import _ENV_LOCK

    with _ENV_LOCK:
        old_cwd = os.environ.get("TERMINAL_CWD")
        os.environ["TERMINAL_CWD"] = str(workspace)
        old_exec_ask = os.environ.get("HERMES_EXEC_ASK")
        old_session_key = os.environ.get("HERMES_SESSION_KEY")
        os.environ["HERMES_EXEC_ASK"] = "1"
        os.environ["HERMES_SESSION_KEY"] = s.session_id
    try:
        from run_agent import AIAgent

        with CHAT_LOCK:
            from api.config import resolve_model_provider

            _model, _provider, _base_url = resolve_model_provider(
                model_with_provider_context(s.model, getattr(s, "model_provider", None))
            )
            # Resolve API key via Hermes runtime provider (matches gateway behaviour)
            _api_key = None
            try:
                from hermes_cli.runtime_provider import resolve_runtime_provider

                _rt = resolve_runtime_provider(requested=_provider)
                _api_key = _rt.get("api_key")
                # Also use runtime provider/base_url if the webui config didn't resolve them
                if not _provider:
                    _provider = _rt.get("provider")
                if not _base_url:
                    _base_url = _rt.get("base_url")
            except Exception as _e:
                print(
                    f"[webui] WARNING: resolve_runtime_provider failed: {_e}",
                    flush=True,
                )
            agent = AIAgent(
                model=_model,
                provider=_provider,
                base_url=_base_url,
                api_key=_api_key,
                # Identify browser-originated sessions as WebUI so Hermes Agent
                # does not inject CLI-specific terminal/output guidance.
                platform="webui",
                quiet_mode=True,
                enabled_toolsets=_resolve_cli_toolsets(),
                session_id=s.session_id,
            )
            workspace_ctx = f"[Workspace: {s.workspace}]\n"
            workspace_system_msg = (
                f"Active workspace at session start: {s.workspace}\n"
                "Every user message is prefixed with [Workspace: /absolute/path] indicating the "
                "workspace the user has selected in the web UI at the time they sent that message. "
                "This tag is the single authoritative source of the active workspace and updates "
                "with every message. It overrides any prior workspace mentioned in this system "
                "prompt, memory, or conversation history. Always use the value from the most recent "
                "[Workspace: ...] tag as your default working directory for ALL file operations: "
                "write_file, read_file, search_files, terminal workdir, and patch. "
                "Never fall back to a hardcoded path when this tag is present."
            )
            from api.streaming import (
                _merge_display_messages_after_agent_result,
                _restore_reasoning_metadata,
                _sanitize_messages_for_api,
                _session_context_messages,
            )

            _previous_messages = list(s.messages or [])
            _previous_context_messages = list(_session_context_messages(s))

            result = agent.run_conversation(
                user_message=workspace_ctx + msg,
                system_message=workspace_system_msg,
                conversation_history=_sanitize_messages_for_api(_previous_context_messages),
                task_id=s.session_id,
                persist_user_message=msg,
            )
    finally:
        with _ENV_LOCK:
            if old_cwd is None:
                os.environ.pop("TERMINAL_CWD", None)
            else:
                os.environ["TERMINAL_CWD"] = old_cwd
            if old_exec_ask is None:
                os.environ.pop("HERMES_EXEC_ASK", None)
            else:
                os.environ["HERMES_EXEC_ASK"] = old_exec_ask
            if old_session_key is None:
                os.environ.pop("HERMES_SESSION_KEY", None)
            else:
                os.environ["HERMES_SESSION_KEY"] = old_session_key
    with _get_session_agent_lock(s.session_id):
        _result_messages = result.get("messages") or _previous_context_messages
        _next_context_messages = _restore_reasoning_metadata(
            _previous_context_messages,
            _result_messages,
        )
        s.context_messages = _next_context_messages
        s.messages = _merge_display_messages_after_agent_result(
            _previous_messages,
            _previous_context_messages,
            _restore_reasoning_metadata(_previous_messages, _result_messages),
            msg,
        )
        # Only auto-generate title when still default; preserves user renames
        if s.title == "Untitled":
            s.title = title_from(s.messages, s.title)
        s.save()
    # Sync to state.db for /insights (opt-in setting)
    try:
        if load_settings().get("sync_to_insights"):
            from api.state_sync import sync_session_usage

            sync_session_usage(
                session_id=s.session_id,
                input_tokens=s.input_tokens or 0,
                output_tokens=s.output_tokens or 0,
                estimated_cost=s.estimated_cost,
                model=s.model,
                title=s.title,
                message_count=len(s.messages),
            )
    except Exception:
        logger.debug("Failed to update session cost tracking")
    return j(
        handler,
        {
            "answer": result.get("final_response") or "",
            "status": "done" if result.get("completed", True) else "partial",
            "session": s.compact() | {"messages": s.messages},
            "result": {k: v for k, v in result.items() if k != "messages"},
        },
    )


def _handle_cron_create(handler, body):
    try:
        require(body, "prompt", "schedule")
    except ValueError as e:
        return bad(handler, str(e))
    try:
        from cron.jobs import create_job

        job = create_job(
            prompt=body["prompt"],
            schedule=body["schedule"],
            name=body.get("name") or None,
            deliver=body.get("deliver") or "local",
            skills=body.get("skills") or [],
            model=body.get("model") or None,
        )
        return j(handler, {"ok": True, "job": job})
    except Exception as e:
        return j(handler, {"error": str(e)}, status=400)


def _handle_cron_update(handler, body):
    try:
        require(body, "job_id")
    except ValueError as e:
        return bad(handler, str(e))
    from cron.jobs import update_job

    updates = {k: v for k, v in body.items() if k != "job_id" and v is not None}
    job = update_job(body["job_id"], updates)
    if not job:
        return bad(handler, "Job not found", 404)
    return j(handler, {"ok": True, "job": job})


def _handle_cron_delete(handler, body):
    try:
        require(body, "job_id")
    except ValueError as e:
        return bad(handler, str(e))
    from cron.jobs import remove_job

    ok = remove_job(body["job_id"])
    if not ok:
        return bad(handler, "Job not found", 404)
    return j(handler, {"ok": True, "job_id": body["job_id"]})


def _handle_cron_run(handler, body):
    job_id = body.get("job_id", "")
    if not job_id:
        return bad(handler, "job_id required")
    from cron.jobs import get_job

    job = get_job(job_id)
    if not job:
        return bad(handler, "Job not found", 404)
    # Prevent double-run: reject if the job is already tracked as running
    already_running, elapsed = _is_cron_running(job_id)
    if already_running:
        return j(handler, {"ok": False, "job_id": job_id, "status": "already_running",
                            "elapsed": round(elapsed, 1)})
    _mark_cron_running(job_id)
    threading.Thread(target=_run_cron_tracked, args=(job,), daemon=True).start()
    return j(handler, {"ok": True, "job_id": job_id, "status": "running"})


def _handle_cron_pause(handler, body):
    job_id = body.get("job_id", "")
    if not job_id:
        return bad(handler, "job_id required")
    from cron.jobs import pause_job

    result = pause_job(job_id, reason=body.get("reason"))
    if result:
        return j(handler, {"ok": True, "job": result})
    return bad(handler, "Job not found", 404)


def _handle_cron_resume(handler, body):
    job_id = body.get("job_id", "")
    if not job_id:
        return bad(handler, "job_id required")
    from cron.jobs import resume_job

    result = resume_job(job_id)
    if result:
        return j(handler, {"ok": True, "job": result})
    return bad(handler, "Job not found", 404)


def _handle_file_delete(handler, body):
    try:
        require(body, "session_id", "path")
    except ValueError as e:
        return bad(handler, str(e))
    try:
        s = get_session(body["session_id"])
    except KeyError:
        return bad(handler, "Session not found", 404)
    try:
        target = safe_resolve(Path(s.workspace), body["path"])
        if not target.exists():
            return bad(handler, "File not found", 404)
        if target.is_dir():
            if not body.get("recursive"):
                return bad(handler, "Set recursive=true to delete directories")
            shutil.rmtree(target)
        else:
            target.unlink()
        return j(handler, {"ok": True, "path": body["path"]})
    except (ValueError, PermissionError) as e:
        return bad(handler, _sanitize_error(e))


def _handle_file_save(handler, body):
    try:
        require(body, "session_id", "path")
    except ValueError as e:
        return bad(handler, str(e))
    try:
        s = get_session(body["session_id"])
    except KeyError:
        return bad(handler, "Session not found", 404)
    try:
        target = safe_resolve(Path(s.workspace), body["path"])
        if not target.exists():
            return bad(handler, "File not found", 404)
        if target.is_dir():
            return bad(handler, "Cannot save: path is a directory")
        target.write_text(body.get("content", ""), encoding="utf-8")
        return j(
            handler, {"ok": True, "path": body["path"], "size": target.stat().st_size}
        )
    except (ValueError, PermissionError) as e:
        return bad(handler, _sanitize_error(e))


def _handle_file_create(handler, body):
    try:
        require(body, "session_id", "path")
    except ValueError as e:
        return bad(handler, str(e))
    try:
        s = get_session(body["session_id"])
    except KeyError:
        return bad(handler, "Session not found", 404)
    try:
        target = safe_resolve(Path(s.workspace), body["path"])
        if target.exists():
            return bad(handler, "File already exists")
        target.parent.mkdir(parents=True, exist_ok=True)
        target.write_text(body.get("content", ""), encoding="utf-8")
        return j(
            handler, {"ok": True, "path": str(target.relative_to(Path(s.workspace)))}
        )
    except (ValueError, PermissionError) as e:
        return bad(handler, _sanitize_error(e))


def _handle_file_rename(handler, body):
    try:
        require(body, "session_id", "path", "new_name")
    except ValueError as e:
        return bad(handler, str(e))
    try:
        s = get_session(body["session_id"])
    except KeyError:
        return bad(handler, "Session not found", 404)
    try:
        source = safe_resolve(Path(s.workspace), body["path"])
        if not source.exists():
            return bad(handler, "File not found", 404)
        new_name = body["new_name"].strip()
        if not new_name or "/" in new_name or ".." in new_name:
            return bad(handler, "Invalid file name")
        dest = source.parent / new_name
        if dest.exists():
            return bad(handler, f'A file named "{new_name}" already exists')
        source.rename(dest)
        new_rel = str(dest.relative_to(Path(s.workspace)))
        return j(handler, {"ok": True, "old_path": body["path"], "new_path": new_rel})
    except (ValueError, PermissionError, OSError) as e:
        return bad(handler, _sanitize_error(e))


def _handle_create_dir(handler, body):
    try:
        require(body, "session_id", "path")
    except ValueError as e:
        return bad(handler, str(e))
    try:
        s = get_session(body["session_id"])
    except KeyError:
        return bad(handler, "Session not found", 404)
    try:
        target = safe_resolve(Path(s.workspace), body["path"])
        if target.exists():
            return bad(handler, "Path already exists")
        target.mkdir(parents=True)
        return j(
            handler, {"ok": True, "path": str(target.relative_to(Path(s.workspace)))}
        )
    except (ValueError, PermissionError, OSError) as e:
        return bad(handler, _sanitize_error(e))


def _handle_file_reveal(handler, body):
    try:
        require(body, "session_id", "path")
    except ValueError as e:
        return bad(handler, str(e))
    try:
        s = get_session(body["session_id"])
    except KeyError:
        return bad(handler, "Session not found", 404)
    try:
        target = safe_resolve(Path(s.workspace), body["path"])
        if not target.exists():
            return bad(handler, "File not found", 404)

        system = platform.system()
        if system == "Darwin":
            subprocess.Popen(["open", "-R", str(target)])
        elif system == "Windows":
            subprocess.Popen(["explorer.exe", "/select," + str(target)])
        else:
            # Linux / other — open parent directory
            subprocess.Popen(["xdg-open", str(target.parent)])

        return j(handler, {"ok": True, "path": body["path"]})
    except (ValueError, PermissionError, OSError) as e:
        return bad(handler, _sanitize_error(e))


def _handle_workspace_add(handler, body):
    path_str = body.get("path", "").strip()
    name = body.get("name", "").strip()
    auto_create = body.get("create", False)
    if not path_str:
        return bad(handler, "path is required")
    # Validate the path is NOT a blocked system root BEFORE any filesystem mutation.
    # This prevents creating orphan directories on rejected paths (#782 review).
    # _is_blocked_system_path honours user-tmp carve-outs (e.g. /var/folders on
    # macOS) so pytest's tmp_path_factory paths and other legit user-tmp dirs
    # still register cleanly.
    candidate = Path(path_str).expanduser().resolve()
    if _is_blocked_system_path(candidate):
        return bad(handler, f"Path points to a system directory: {candidate}")
    # Now safe to create the directory if requested
    if auto_create:
        try:
            candidate.mkdir(parents=True, exist_ok=True)
        except (OSError, PermissionError) as e:
            return bad(handler, f"Could not create directory: {_sanitize_error(e)}")
    # Full validation (exists, is_dir) — should pass now that dir exists
    try:
        p = validate_workspace_to_add(path_str)
    except ValueError as e:
        return bad(handler, str(e))
    wss = load_workspaces()
    if any(w["path"] == str(p) for w in wss):
        return bad(handler, "Workspace already in list")
    wss.append({"path": str(p), "name": name or p.name})
    save_workspaces(wss)
    return j(handler, {"ok": True, "workspaces": wss})


def _handle_workspace_remove(handler, body):
    path_str = body.get("path", "").strip()
    if not path_str:
        return bad(handler, "path is required")
    wss = load_workspaces()
    wss = [w for w in wss if w["path"] != path_str]
    save_workspaces(wss)
    return j(handler, {"ok": True, "workspaces": wss})


def _handle_workspace_rename(handler, body):
    path_str = body.get("path", "").strip()
    name = body.get("name", "").strip()
    if not path_str or not name:
        return bad(handler, "path and name are required")
    wss = load_workspaces()
    for w in wss:
        if w["path"] == path_str:
            w["name"] = name
            break
    else:
        return bad(handler, "Workspace not found", 404)
    save_workspaces(wss)
    return j(handler, {"ok": True, "workspaces": wss})


def _handle_workspace_reorder(handler, body):
    """Reorder workspaces by providing an ordered list of paths.

    Accepts {"paths": ["path1", "path2", ...]}. The workspaces list is
    rewritten so that entries appear in the given order. Any workspace
    not included in the request is appended at the end (preserves data).
    """
    paths = body.get("paths", [])
    if not paths or not isinstance(paths, list):
        return bad(handler, "paths is required and must be a list")
    wss = load_workspaces()
    by_path = {w["path"]: w for w in wss}
    # Build reordered list: given order first, then any omitted entries
    reordered = []
    seen = set()
    for p in paths:
        p = p.strip()
        if p in by_path and p not in seen:
            reordered.append(by_path[p])
            seen.add(p)
    # Append any workspaces not mentioned (safety net)
    for w in wss:
        if w["path"] not in seen:
            reordered.append(w)
    save_workspaces(reordered)
    return j(handler, {"ok": True, "workspaces": reordered})


def _handle_approval_respond(handler, body):
    sid = body.get("session_id", "")
    if not sid:
        return bad(handler, "session_id is required")
    choice = body.get("choice", "deny")
    if choice not in ("once", "session", "always", "deny"):
        return bad(handler, f"Invalid choice: {choice}")
    approval_id = body.get("approval_id", "")

    # Pop the targeted entry from the pending queue by approval_id.
    # Falls back to popping the first entry for backward-compat with old clients.
    pending = None
    with _lock:
        queue = _pending.get(sid)
        if isinstance(queue, list):
            if approval_id:
                # Find and remove the specific entry by approval_id.
                for i, entry in enumerate(queue):
                    if entry.get("approval_id") == approval_id:
                        pending = queue.pop(i)
                        break
                else:
                    # approval_id not found -- fall back to oldest entry.
                    pending = queue.pop(0) if queue else None
            else:
                pending = queue.pop(0) if queue else None
            if not queue:
                _pending.pop(sid, None)
        elif queue:
            # Legacy single-dict value.
            pending = _pending.pop(sid, None)
        # Notify SSE subscribers of the new head (or empty state) so the UI
        # surfaces any trailing approvals that were queued behind this one
        # without waiting for the next submit_pending. Without this, a parallel
        # tool-call scenario (#527) would leave the second approval invisible
        # in the SSE path until the next event ever fired (the agent thread
        # would be parked indefinitely from the user's perspective).
        if isinstance(_pending.get(sid), list) and _pending[sid]:
            _approval_sse_notify_locked(sid, _pending[sid][0], len(_pending[sid]))
        else:
            _approval_sse_notify_locked(sid, None, 0)

    if pending:
        keys = pending.get("pattern_keys") or [pending.get("pattern_key", "")]
        if choice in ("once", "session"):
            for k in keys:
                approve_session(sid, k)
        elif choice == "always":
            for k in keys:
                approve_session(sid, k)
                approve_permanent(k)
            save_permanent_allowlist(_permanent_approved)
    # Unblock the agent thread waiting in the gateway approval queue.
    # This is the primary signal when streaming is active — the agent
    # thread is parked in entry.event.wait() and needs to be woken up.
    resolve_gateway_approval(sid, choice, resolve_all=False)
    return j(handler, {"ok": True, "choice": choice})


def _handle_clarify_respond(handler, body):
    sid = body.get("session_id", "")
    if not sid:
        return bad(handler, "session_id is required")
    response = body.get("response")
    if response is None:
        response = body.get("answer")
    if response is None:
        response = body.get("choice")
    response = str(response or "").strip()
    if not response:
        return bad(handler, "response is required")
    resolve_clarify(sid, response, resolve_all=False)
    return j(handler, {"ok": True, "response": response})


def _handle_session_compress(handler, body):
    def _visible_messages_for_anchor(messages):
        out = []
        for m in messages or []:
            if not isinstance(m, dict):
                continue
            role = m.get("role")
            if not role or role == "tool":
                continue
            content = m.get("content", "")
            has_attachments = bool(m.get("attachments"))
            if role == "assistant":
                tool_calls = m.get("tool_calls")
                has_tool_calls = isinstance(tool_calls, list) and len(tool_calls) > 0
                has_tool_use = False
                has_reasoning = bool(m.get("reasoning"))
                if isinstance(content, list):
                    for p in content:
                        if not isinstance(p, dict):
                            continue
                        if p.get("type") == "tool_use":
                            has_tool_use = True
                        if p.get("type") in {"thinking", "reasoning"}:
                            has_reasoning = True
                    text = "\n".join(
                        str(p.get("text") or p.get("content") or "")
                        for p in content
                        if isinstance(p, dict) and p.get("type") == "text"
                    ).strip()
                else:
                    text = str(content or "").strip()
                if text or has_attachments or has_tool_calls or has_tool_use or has_reasoning:
                    out.append(m)
                continue
            if isinstance(content, list):
                text = "\n".join(
                    str(p.get("text") or p.get("content") or "")
                    for p in content
                    if isinstance(p, dict) and p.get("type") == "text"
                ).strip()
            else:
                text = str(content or "").strip()
            if text or has_attachments:
                out.append(m)
        return out

    def _anchor_message_key(m):
        if not isinstance(m, dict):
            return None
        role = str(m.get("role") or "")
        if not role or role == "tool":
            return None
        content = m.get("content", "")
        if isinstance(content, list):
            text = "\n".join(
                str(p.get("text") or p.get("content") or "")
                for p in content
                if isinstance(p, dict) and p.get("type") == "text"
            )
        else:
            text = str(content or "")
        norm = " ".join(text.split()).strip()[:160]
        ts = m.get("_ts") or m.get("timestamp")
        attachments = m.get("attachments")
        attach_count = len(attachments) if isinstance(attachments, list) else 0
        if not norm and not attach_count and not ts:
            return None
        return {"role": role, "ts": ts, "text": norm, "attachments": attach_count}

    try:
        require(body, "session_id")
    except ValueError as e:
        return bad(handler, str(e))

    sid = str(body.get("session_id") or "").strip()
    if not sid:
        return bad(handler, "session_id is required")

    # Cap focus_topic to 500 chars — matches the defensive input-size pattern
    # used elsewhere (session title :80, first-exchange snippets :500) and
    # prevents a user from forwarding an unbounded string into the compressor
    # prompt path. No privilege boundary here (user prompting themself), just
    # cheap bound-checking.
    focus_topic = str(body.get("focus_topic") or body.get("topic") or "").strip()[:500] or None

    try:
        s = get_session(sid)
    except KeyError:
        return bad(handler, "Session not found", 404)

    if getattr(s, "active_stream_id", None):
        return bad(handler, "Session is still streaming; wait for the current turn to finish.", 409)

    try:
        from api.streaming import _sanitize_messages_for_api

        messages = _sanitize_messages_for_api(s.messages)
        if len(messages) < 4:
            return bad(handler, "Not enough conversation to compress (need at least 4 messages).")

        def _fallback_estimate_messages_tokens_rough(msgs):
            """Fallback heuristic token estimate when runtime metadata helpers are absent.

            Uses whitespace token-like word counting only. This intentionally
            over/under-estimates BPE token counts (roughly around x3/x4 scale),
            and is only for resilient fallback behavior.
            """
            total = 0
            for m in msgs or []:
                if not isinstance(m, dict):
                    continue
                content = m.get("content", "")
                if isinstance(content, list):
                    content_text = "\n".join(
                        str(p.get("text") or p.get("content") or "")
                        for p in content
                        if isinstance(p, dict)
                    )
                else:
                    content_text = str(content or "")
                total += len(content_text.split())
            return max(1, total)

        def _fallback_summarize_manual_compression(original_messages, compressed_messages, before_tokens, after_tokens, focus_topic=None):
            """Lightweight fallback summary to keep /session/compress usable in tests/runtime."""
            after_tokens = after_tokens if after_tokens is not None else _fallback_estimate_messages_tokens_rough(compressed_messages)
            headline = f"Compressed: {len(original_messages)} \u2192 {len(compressed_messages)} messages"
            summary = {
                "headline": headline,
                "token_line": f"Rough transcript estimate: ~{before_tokens} \u2192 ~{after_tokens} tokens",
                "note": f"Focus: {focus_topic}" if focus_topic else None,
            }
            summary["reference_message"] = (
                f"[CONTEXT COMPACTION \u2014 REFERENCE ONLY] {headline}\n"
                f"{summary['token_line']}\n"
                + (summary["note"] + "\n" if summary.get("note") else "")
                + "Compression completed."
            )
            return summary

        def _estimate_messages_tokens_rough(msgs):
            try:
                from agent.model_metadata import estimate_messages_tokens_rough

                return estimate_messages_tokens_rough(msgs)
            except Exception:
                return _fallback_estimate_messages_tokens_rough(msgs)

        def _summarize_manual_compression(
            original_messages,
            compressed_messages,
            before_tokens,
            after_tokens,
            focus_topic=None,
        ):
            try:
                from agent.manual_compression_feedback import summarize_manual_compression

                return summarize_manual_compression(
                    original_messages,
                    compressed_messages,
                    before_tokens,
                    after_tokens,
                )
            except Exception:
                return _fallback_summarize_manual_compression(
                    original_messages,
                    compressed_messages,
                    before_tokens,
                    after_tokens,
                    focus_topic,
                )

        import api.config as _cfg
        import hermes_cli.runtime_provider as _runtime_provider
        import run_agent as _run_agent

        resolved_model, resolved_provider, resolved_base_url = _cfg.resolve_model_provider(
            _cfg.model_with_provider_context(s.model, getattr(s, "model_provider", None))
        )

        resolved_api_key = None
        try:
            _rt = _runtime_provider.resolve_runtime_provider(requested=resolved_provider)
            resolved_api_key = _rt.get("api_key")
            if not resolved_provider:
                resolved_provider = _rt.get("provider")
            if not resolved_base_url:
                resolved_base_url = _rt.get("base_url")
        except Exception as _e:
            logger.warning("resolve_runtime_provider failed for compression: %s", _e)

        if not resolved_api_key:
            return bad(handler, "No provider configured -- cannot compress.")

        # Compute compression *outside* the lock — the LLM round-trip can take
        # many seconds and we must not block cancel_stream or other writers.
        # Lock contract: hold for the in-memory mutation only, never across
        # network I/O.
        original_messages = list(messages)
        approx_tokens = _estimate_messages_tokens_rough(original_messages)

        agent = _run_agent.AIAgent(
            model=resolved_model,
            provider=resolved_provider,
            base_url=resolved_base_url,
            api_key=resolved_api_key,
            # Identify browser-originated sessions as WebUI so Hermes Agent
            # does not inject CLI-specific terminal/output guidance.
            platform="webui",
            quiet_mode=True,
            enabled_toolsets=_resolve_cli_toolsets(),
            session_id=sid,
        )
        compressed = agent.context_compressor.compress(
            original_messages,
            current_tokens=approx_tokens,
            focus_topic=focus_topic,
        )
        new_tokens = _estimate_messages_tokens_rough(compressed)
        summary = _summarize_manual_compression(
            original_messages,
            compressed,
            approx_tokens,
            new_tokens,
            focus_topic=focus_topic,
        )

        with _cfg._get_session_agent_lock(sid):
            # Re-read messages to detect concurrent edits during the LLM call.
            # If the history changed, the compression result is stale — abort.
            if _sanitize_messages_for_api(s.messages) != original_messages:
                return bad(handler, "Session was modified during compression; please retry.", 409)

            s.messages = compressed
            s.context_messages = compressed
            s.tool_calls = []
            s.active_stream_id = None
            s.pending_user_message = None
            s.pending_attachments = []
            s.pending_started_at = None
            visible_after = _visible_messages_for_anchor(compressed)
            s.compression_anchor_visible_idx = max(0, len(visible_after) - 1) if visible_after else None
            s.compression_anchor_message_key = _anchor_message_key(visible_after[-1]) if visible_after else None
            s.save()

        session_payload = redact_session_data(
            s.compact() | {
                "messages": s.messages,
                "tool_calls": s.tool_calls,
                "active_stream_id": s.active_stream_id,
                "pending_user_message": s.pending_user_message,
                "pending_attachments": s.pending_attachments,
                "pending_started_at": s.pending_started_at,
                "compression_anchor_visible_idx": getattr(s, "compression_anchor_visible_idx", None),
                "compression_anchor_message_key": getattr(s, "compression_anchor_message_key", None),
            }
        )
        return j(
            handler,
            {
                "ok": True,
                "session": session_payload,
                "summary": summary,
                "focus_topic": focus_topic,
            },
        )
    except Exception as e:
        logger.warning("Manual session compression failed: %s", e)
        return bad(handler, f"Compression failed: {_sanitize_error(e)}")


def _handle_conversation_rounds(handler, body):
    """Return conversation-round count for a gateway session.

    Request body::

        { "session_id": "...", "since": <unix_ts_or_iso> }

    Response::

        { "ok": true, "rounds": 12, "threshold": 10, "should_show": true }
    """
    try:
        require(body, "session_id")
    except ValueError as e:
        return bad(handler, str(e))

    sid = str(body.get("session_id") or "").strip()
    if not sid:
        return bad(handler, "session_id is required")

    since = body.get("since")
    if since is not None:
        try:
            since = float(since)
        except (TypeError, ValueError):
            return bad(handler, "since must be a unix timestamp (number)")

    from api.models import count_conversation_rounds, CONVERSATION_ROUND_THRESHOLD

    rounds = count_conversation_rounds(sid, since=since)
    return j(handler, {
        "ok": True,
        "rounds": rounds,
        "threshold": CONVERSATION_ROUND_THRESHOLD,
        "should_show": rounds >= CONVERSATION_ROUND_THRESHOLD,
    })


def _build_handoff_summary_tool_message(
    sid: str,
    summary: str,
    channel: str | None,
    rounds: int | None = None,
    fallback: bool = False,
) -> dict:
    """Build a compact tool-role transcript marker for persistence."""
    now = time.time()
    return {
        "role": "tool",
        # Keep this intentionally empty so API-history sanitization drops it from
        # model context (it is display-only data).
        "tool_call_id": "",
        "name": "handoff_summary",
        "timestamp": now,
        "_ts": now,
        "content": json.dumps({
            "_handoff_summary_card": True,
            "session_id": sid,
            "summary": str(summary or "").strip(),
            "channel": (str(channel or "").strip() or None),
            "rounds": rounds,
            "fallback": bool(fallback),
            "generated_at": now,
        }, ensure_ascii=False),
    }


def _extract_handoff_summary_payload(message: dict) -> dict | None:
    """Return a normalized handoff-summary payload if *message* is a tool marker."""
    if not isinstance(message, dict):
        return None
    if message.get("role") != "tool" or message.get("name") != "handoff_summary":
        return None

    content = message.get("content")
    if isinstance(content, dict):
        payload = content
    else:
        try:
            payload = json.loads(content or "")
        except Exception:
            return None

    if not isinstance(payload, dict) or not payload.get("_handoff_summary_card"):
        return None
    if payload.get("session_id") is None:
        return None
    return {
        "session_id": str(payload.get("session_id")),
        "summary": str(payload.get("summary", "")),
        "channel": payload.get("channel"),
        "rounds": payload.get("rounds"),
        "fallback": bool(payload.get("fallback")),
        "_handoff_summary_card": True,
    }


def _is_matching_handoff_summary_message(existing: dict, target: dict) -> bool:
    """Return True when two message payloads represent the same handoff summary."""
    existing_payload = _extract_handoff_summary_payload(existing)
    target_payload = _extract_handoff_summary_payload(target)
    if not existing_payload or not target_payload:
        return False
    return (
        existing_payload.get("session_id") == target_payload.get("session_id") and
        existing_payload.get("summary") == target_payload.get("summary") and
        existing_payload.get("channel") == target_payload.get("channel") and
        existing_payload.get("rounds") == target_payload.get("rounds") and
        existing_payload.get("fallback") == target_payload.get("fallback") and
        existing_payload.get("_handoff_summary_card") == target_payload.get("_handoff_summary_card")
    )


def _is_matching_handoff_summary_content(content: object, target_payload: dict | None) -> bool:
    """Return True if DB content JSON matches an expected handoff summary payload."""
    if target_payload is None:
        return False
    try:
        payload = json.loads(content or "")
    except Exception:
        return False
    if not isinstance(payload, dict):
        return False
    if payload.get("session_id") is None:
        return False
    return (
        payload.get("_handoff_summary_card") is True and
        str(payload.get("session_id")) == str(target_payload.get("session_id")) and
        str(payload.get("summary", "")) == str(target_payload.get("summary", "")) and
        payload.get("channel") == target_payload.get("channel") and
        payload.get("rounds") == target_payload.get("rounds") and
        bool(payload.get("fallback")) == bool(target_payload.get("fallback"))
    )


def _persist_handoff_summary_locally(sid: str, message: dict) -> bool:
    """Persist a handoff summary marker into a local WebUI session file."""
    try:
        from api.models import get_session

        s = get_session(sid)
    except KeyError:
        return False

    try:
        if s.messages and _is_matching_handoff_summary_message(s.messages[-1], message):
            return True
        s.messages.append(message)
        s.save()
        return True
    except Exception as e:
        logger.warning("Failed to persist handoff summary marker in local session %s: %s", sid, e)
        return False


def _persist_handoff_summary_to_state_db(sid: str, message: dict) -> bool:
    """Persist a handoff summary marker into CLI sessions state.db.

    This keeps summary cards available after hard-refresh for imported gateway
    sessions that are not in local session JSON yet.
    """
    import os

    try:
        import sqlite3
    except ImportError:
        return False

    try:
        from api.profiles import get_active_hermes_home

        hermes_home = Path(get_active_hermes_home()).expanduser().resolve()
    except Exception:
        hermes_home = Path(os.getenv("HERMES_HOME", str(Path.home() / ".hermes"))).expanduser().resolve()

    db_path = hermes_home / "state.db"
    if not db_path.exists():
        return False

    ts = message.get("timestamp", time.time())
    content = message.get("content", "")
    if not isinstance(content, str):
        content = json.dumps(content, ensure_ascii=False)

    marker_payload = _extract_handoff_summary_payload(message)
    try:
        with sqlite3.connect(str(db_path)) as conn:
            try:
                if marker_payload is not None:
                    cur = conn.execute(
                        "SELECT content FROM messages WHERE session_id = ? AND role = 'tool' "
                        "ORDER BY rowid DESC LIMIT 1",
                        (sid,),
                    )
                    row = cur.fetchone()
                    if row is not None and _is_matching_handoff_summary_content(row[0], marker_payload):
                        return True
            except Exception:
                # If tail-read fails, continue with a best-effort write.
                logger.debug("Unable to read tail handoff marker from state.db for %s", sid)

            conn.execute(
                "INSERT INTO messages (session_id, role, content, timestamp) "
                "VALUES (?, 'tool', ?, ?)",
                (sid, content, ts),
            )
            # Keep session row message_count/last-activity aligned with displayed
            # transcript length. session rows are optional in some test DBs, so
            # this update is best-effort.
            conn.execute(
                "UPDATE sessions SET message_count = COALESCE(message_count, 0) + 1 "
                "WHERE id = ?",
                (sid,),
            )
            conn.commit()
        return True
    except Exception as e:
        logger.warning("Failed to persist handoff summary marker in state.db for %s: %s", sid, e)
        return False


def _persist_handoff_summary(sid: str, summary: str, channel: str | None, rounds: int | None, fallback: bool = False) -> dict:
    """Persist a handoff summary marker across local/session backends."""
    marker = _build_handoff_summary_tool_message(sid, summary, channel, rounds, fallback)
    is_messaging_session = _is_messaging_session_id(sid)
    if is_messaging_session:
        _persist_handoff_summary_to_state_db(sid, marker)
        _persist_handoff_summary_locally(sid, marker)
        return marker
    persisted_local = _persist_handoff_summary_locally(sid, marker)
    if persisted_local:
        return marker
    return marker if _persist_handoff_summary_to_state_db(sid, marker) else marker


def _handle_handoff_summary(handler, body):
    """Generate an on-demand handoff summary for a gateway session.

    Request body::

        { "session_id": "...", "since": <unix_ts_or_iso> }

    Uses the session's configured model to produce a concise summary of
    recent conversation activity.  Returns the summary text so the caller
    can display it in a tool-card.
    """
    try:
        require(body, "session_id")
    except ValueError as e:
        return bad(handler, str(e))

    sid = str(body.get("session_id") or "").strip()
    if not sid:
        return bad(handler, "session_id is required")

    since = body.get("since")
    if since is not None:
        try:
            since = float(since)
        except (TypeError, ValueError):
            return bad(handler, "since must be a unix timestamp (number)")

    from api.models import get_cli_session_messages, count_conversation_rounds, CONVERSATION_ROUND_THRESHOLD

    rounds = count_conversation_rounds(sid, since=since)
    if rounds < CONVERSATION_ROUND_THRESHOLD:
        return bad(handler, "Not enough conversation rounds to generate a summary.", 400)

    # Filter messages by ``since``.
    all_msgs = get_cli_session_messages(sid)
    if since is not None:
        import datetime as _dt
        filtered = []
        for m in all_msgs:
            ts_raw = m.get("timestamp")
            if ts_raw is None:
                continue
            try:
                if isinstance(ts_raw, (int, float)):
                    ts_val = float(ts_raw)
                else:
                    ts_val = _dt.datetime.fromisoformat(
                        str(ts_raw).replace("Z", "+00:00")
                    ).timestamp()
                if ts_val > since:
                    filtered.append(m)
            except Exception:
                pass
        msgs = filtered
    else:
        msgs = all_msgs

    # Cap to last 50 messages.
    msgs = msgs[-50:]

    if len(msgs) < 2:
        return bad(handler, "Not enough messages to summarize.", 400)

    def _extract_handoff_text(raw_content):
        if isinstance(raw_content, list):
            return " ".join(
                str(p.get("text") or p.get("content") or "")
                for p in raw_content
                if isinstance(p, dict)
            ).strip()
        return str(raw_content or "").strip()

    def _contains_chinese(text):
        return any("\u4e00" <= ch <= "\u9fff" for ch in str(text))

    transcript_is_chinese = any(
        _contains_chinese(_extract_handoff_text(m.get("content")))
        for m in msgs
    )
    # Build a lightweight conversation transcript for the LLM.
    lines = []
    for m in msgs:
        role = m.get("role", "")
        content = _extract_handoff_text(m.get("content"))
        content = str(content or "").strip()[:1000]
        if role in ("user", "assistant") and content:
            lines.append(content)
    transcript = "\n".join(lines)

    def _fallback_handoff_summary(items):
        """Return a deterministic summary when LLM summary generation is unavailable."""
        user_points = []
        assistant_points = []

        def _summarize_snippet(raw_text, max_len=78):
            text = " ".join(str(raw_text or "").split()).strip()
            if not text:
                return ""
            if len(text) <= max_len:
                return text
            return text[: max_len - 1].rstrip() + "…"

        for m in items:
            role = m.get("role", "")
            content = _summarize_snippet(_extract_handoff_text(m.get("content")), 82)
            if role in ("user", "assistant") and content:
                if role == "user":
                    user_points.append(content)
                else:
                    assistant_points.append(content)
        if not user_points and not assistant_points:
            return (
                "近期可读文本不足，无法生成更完整的交接摘要，请补充一条消息后重试。"
                if transcript_is_chinese
                else "Not enough readable text to create a useful handoff summary; please send one more message and retry."
            )

        if transcript_is_chinese:
            bullets = []
            if user_points:
                bullets.append(f"- 你刚讨论了：{user_points[-1]}。")
            if assistant_points:
                bullets.append(f"- 助手已回复：{assistant_points[-1]}。")
            if len(user_points) + len(assistant_points) >= 2:
                bullets.append("- 当前对话存在尚未确认的后续动作。")
            else:
                bullets.append("- 当前信息偏少，建议补充关键点后再切换。")
            return "\n".join(bullets)

        bullets = []
        if user_points:
            bullets.append(f"- You asked: {user_points[-1]}.")
        if assistant_points:
            bullets.append(f"- The assistant responded: {assistant_points[-1]}.")
        if len(user_points) + len(assistant_points) >= 2:
            bullets.append("- There is pending context to continue next.")
        else:
            bullets.append("- The conversation is still short; add one more turn before summarizing.")
        return "\n".join(bullets)

    def _summary_output_incomplete(text):
        """Best-effort guard for truncated summaries when LLM signals are unavailable."""
        if not isinstance(text, str):
            text = str(text or "")
        text = text.strip()
        if not text:
            return True
        if text.endswith("...") or text.endswith("…"):
            return True
        lines = [line.strip() for line in text.splitlines() if line.strip()]
        if not lines:
            return True
        last_line = lines[-1]
        if re.search(r"[。！？；!?.；]$", last_line):
            return False
        if len(last_line) >= 56 and not re.search(r"\b(and|or|so|then|because|if|when|but|so|as)\b$", last_line, re.IGNORECASE):
            return True
        return bool(re.search(r"\b(and|or|but|so|because|if|when)$", last_line, re.IGNORECASE))

    def _agent_summary_incomplete(summary_result):
        if not isinstance(summary_result, dict):
            return True
        reason = (summary_result.get("finish_reason") or "").strip().lower()
        if reason == "length":
            return True
        stop_reason = (summary_result.get("stop_reason") or "").strip().lower()
        if stop_reason in {"max_tokens", "length"}:
            return True
        return _summary_output_incomplete(summary_result.get("text", ""))

    def _resolve_handoff_channel_label():
        channel_label = None
        try:
            from api.models import get_session as _get_session, get_cli_sessions

            session_meta = _get_session(sid)
            channel_label = (
                session_meta.source_label
                or session_meta.raw_source
                or session_meta.source_tag
                or session_meta.session_source
            )
            if not channel_label:
                for candidate in get_cli_sessions():
                    if candidate.get("session_id") == sid:
                        channel_label = (
                            candidate.get("source_label")
                            or candidate.get("raw_source")
                            or candidate.get("source_tag")
                            or candidate.get("source")
                        )
                        break
        except Exception:
            pass
        return channel_label

    def _agent_text_completion(agent, system_prompt, user_text, max_tokens=700):
        """Use the current Hermes Agent transport without mutating conversation history."""
        api_messages = [
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": user_text},
        ]
        result = {
            "text": "",
            "finish_reason": None,
            "stop_reason": None,
            "incomplete": True,
        }
        disabled_reasoning = {"enabled": False}
        previous_reasoning = getattr(agent, "reasoning_config", None)
        try:
            agent.reasoning_config = disabled_reasoning
            if getattr(agent, "api_mode", "") == "codex_responses":
                codex_kwargs = agent._build_api_kwargs(api_messages)
                codex_kwargs.pop("tools", None)
                codex_kwargs["max_output_tokens"] = max_tokens
                resp = agent._run_codex_stream(codex_kwargs)
                assistant_message, _ = agent._normalize_codex_response(resp)
                result["text"] = str((assistant_message.content or "") if assistant_message else "").strip()
                result["incomplete"] = _summary_output_incomplete(result["text"])
                return result

            if getattr(agent, "api_mode", "") == "anthropic_messages":
                from agent.anthropic_adapter import build_anthropic_kwargs, normalize_anthropic_response

                ant_kwargs = build_anthropic_kwargs(
                    model=agent.model,
                    messages=api_messages,
                    tools=None,
                    max_tokens=max_tokens,
                    reasoning_config=disabled_reasoning,
                    is_oauth=getattr(agent, "_is_anthropic_oauth", False),
                    preserve_dots=agent._anthropic_preserve_dots(),
                    base_url=getattr(agent, "_anthropic_base_url", None),
                )
                resp = agent._anthropic_messages_create(ant_kwargs)
                assistant_message, _ = normalize_anthropic_response(
                    resp,
                    strip_tool_prefix=getattr(agent, "_is_anthropic_oauth", False),
                )
                result["text"] = str((assistant_message.content or "") if assistant_message else "").strip()
                result["incomplete"] = _summary_output_incomplete(result["text"])
                return result

            api_kwargs = agent._build_api_kwargs(api_messages)
            api_kwargs.pop("tools", None)
            api_kwargs["temperature"] = 0.2
            api_kwargs["timeout"] = 30.0
            if "max_completion_tokens" in api_kwargs:
                api_kwargs["max_completion_tokens"] = max_tokens
            else:
                api_kwargs["max_tokens"] = max_tokens
            resp = agent._ensure_primary_openai_client(reason="handoff_summary").chat.completions.create(
                **api_kwargs,
            )
            choice = (getattr(resp, "choices", None) or [None])[0]
            msg = getattr(choice, "message", None) if choice is not None else None
            result["text"] = str(getattr(msg, "content", "") or "").strip()
            result["finish_reason"] = getattr(choice, "finish_reason", None)
            result["stop_reason"] = getattr(choice, "stop_reason", None)
            result["incomplete"] = _agent_summary_incomplete(result)
            return result
        finally:
            agent.reasoning_config = previous_reasoning

        # Call LLM for summary.
    try:
        import api.config as _cfg
        import hermes_cli.runtime_provider as _runtime_provider
        import run_agent as _run_agent

        # Try to resolve model from an existing session, fall back to default.
        resolved_model = None
        resolved_provider = None
        resolved_base_url = None
        try:
            from api.models import get_session
            s_obj = get_session(sid)
            resolved_model = getattr(s_obj, "model", None)
        except Exception:
            pass

        resolved_model, resolved_provider, resolved_base_url = _cfg.resolve_model_provider(resolved_model)

        resolved_api_key = None
        try:
            _rt = _runtime_provider.resolve_runtime_provider(requested=resolved_provider)
            resolved_api_key = _rt.get("api_key")
            if not resolved_provider:
                resolved_provider = _rt.get("provider")
            if not resolved_base_url:
                resolved_base_url = _rt.get("base_url")
        except Exception as _e:
            logger.warning("resolve_runtime_provider failed for handoff summary: %s", _e)

        if not resolved_api_key:
            summary_text = _fallback_handoff_summary(msgs)
            try:
                _persist_handoff_summary(
                    sid,
                    summary_text,
                    _resolve_handoff_channel_label(),
                    rounds,
                    fallback=True,
                )
            except Exception:
                pass
            return j(handler, {
                "ok": True,
                "summary": summary_text,
                "message_count": len(msgs),
                "rounds": rounds,
                "fallback": True,
            })

        agent = _run_agent.AIAgent(
            model=resolved_model,
            provider=resolved_provider,
            base_url=resolved_base_url,
            api_key=resolved_api_key,
            platform="webui",
            quiet_mode=True,
            enabled_toolsets=[],
            session_id=sid,
        )

        summary_system_prompt = (
            "You are summarizing an external-channel conversation so a Web UI reader "
            "can quickly catch up after switching contexts.\n\n"
            "Only use the latest messages, and never copy raw transcript lines.\n"
            "Do not output role labels (no “你:” / “assistant:” / “user:” / “assistant”).\n"
            "Use direct 2–5 bullet points in the conversation language.\n"
            "English: speak using “you”.\n"
            "中文: 使用“你”。\n\n"
            "Focus on:\n"
            "- Unfinished tasks or action items\n"
            "- Pending questions that need replies\n"
            "- Key decisions made\n"
            "- Open disagreements or TBD items\n\n"
            "If the conversation is purely casual with no actionable items, "
            "say so in one sentence."
        )
        summary_user_text = f"Conversation transcript:\n{transcript}"

        try:
            first_pass = _agent_text_completion(
                agent,
                summary_system_prompt,
                summary_user_text,
                max_tokens=700,
            )
            summary_text = first_pass.get("text") if isinstance(first_pass, dict) else ""
            if _agent_summary_incomplete(first_pass):
                second_pass = _agent_text_completion(
                    agent,
                    summary_system_prompt,
                    summary_user_text,
                    max_tokens=1400,
                )
                summary_text = second_pass.get("text") if isinstance(second_pass, dict) else ""
                if _agent_summary_incomplete(second_pass):
                    summary_text = _fallback_handoff_summary(msgs)
                    fallback = True
                else:
                    fallback = False
            else:
                fallback = False
        finally:
            try:
                agent.release_clients()
            except Exception:
                pass
        if not summary_text:
            summary_text = _fallback_handoff_summary(msgs)
            fallback = True
        elif _summary_output_incomplete(summary_text):
            if not fallback:
                fallback = True

        channel_label = _resolve_handoff_channel_label()
        _persist_handoff_summary(
            sid,
            summary_text,
            channel_label,
            rounds,
            fallback=fallback,
        )

        return j(handler, {
            "ok": True,
            "summary": summary_text,
            "message_count": len(msgs),
            "rounds": rounds,
            "fallback": fallback,
        })
    except Exception as e:
        logger.warning("Handoff summary generation failed: %s", e)
        summary_text = _fallback_handoff_summary(msgs)
        try:
            _persist_handoff_summary(
                sid,
                summary_text,
                _resolve_handoff_channel_label(),
                rounds,
                fallback=True,
            )
        except Exception:
            pass
        return j(handler, {
            "ok": True,
            "summary": summary_text,
            "message_count": len(msgs),
            "rounds": rounds,
            "fallback": True,
            "warning": f"Summary generation used local fallback: {_sanitize_error(e)}",
        })


def _handle_skill_save(handler, body):
    try:
        require(body, "name", "content")
    except ValueError as e:
        return bad(handler, str(e))
    skill_name = body["name"].strip().lower().replace(" ", "-")
    if not skill_name or "/" in skill_name or ".." in skill_name:
        return bad(handler, "Invalid skill name")
    category = body.get("category", "").strip()
    if category and ("/" in category or ".." in category):
        return bad(handler, "Invalid category")
    from tools.skills_tool import SKILLS_DIR

    if category:
        skill_dir = SKILLS_DIR / category / skill_name
    else:
        skill_dir = SKILLS_DIR / skill_name
    # Validate resolved path stays within SKILLS_DIR
    try:
        skill_dir.resolve().relative_to(SKILLS_DIR.resolve())
    except ValueError:
        return bad(handler, "Invalid skill path")
    skill_dir.mkdir(parents=True, exist_ok=True)
    skill_file = skill_dir / "SKILL.md"
    skill_file.write_text(body["content"], encoding="utf-8")
    return j(handler, {"ok": True, "name": skill_name, "path": str(skill_file)})


def _handle_skill_delete(handler, body):
    try:
        require(body, "name")
    except ValueError as e:
        return bad(handler, str(e))
    from tools.skills_tool import SKILLS_DIR
    import shutil

    matches = list(SKILLS_DIR.rglob(f"{body['name']}/SKILL.md"))
    if not matches:
        return bad(handler, "Skill not found", 404)
    skill_dir = matches[0].parent
    shutil.rmtree(str(skill_dir))
    return j(handler, {"ok": True, "name": body["name"]})


def _handle_memory_write(handler, body):
    try:
        require(body, "section", "content")
    except ValueError as e:
        return bad(handler, str(e))
    try:
        from api.profiles import get_active_hermes_home

        mem_dir = get_active_hermes_home() / "memories"
    except ImportError:
        mem_dir = Path.home() / ".hermes" / "memories"
    mem_dir.mkdir(parents=True, exist_ok=True)
    section = body["section"]
    if section == "memory":
        target = mem_dir / "MEMORY.md"
    elif section == "user":
        target = mem_dir / "USER.md"
    else:
        return bad(handler, 'section must be "memory" or "user"')
    target.write_text(body["content"], encoding="utf-8")
    return j(handler, {"ok": True, "section": section, "path": str(target)})


def _normalize_message_for_import_refresh(message: object) -> object:
    """Normalize message payloads for import refresh prefix checks.

    The strict dict comparison previously failed when existing messages held
    integer timestamps while refreshed messages held floating-point timestamps.
    Strip timing keys before comparison so we can safely treat semantic
    prefixes as equivalent.
    """
    if not isinstance(message, dict):
        return message
    normalized = dict(message)
    normalized.pop("timestamp", None)
    normalized.pop("_ts", None)
    return normalized


def _is_messages_refresh_prefix_match(existing_messages: list, fresh_messages: list) -> bool:
    """Return True when existing_messages is a prefix of fresh_messages by value.

    This is a semantic comparison intended for import refresh, not deep
    structural equality. It intentionally ignores timing fields that may differ
    in type/precision between storage layers.
    """
    if not isinstance(existing_messages, list) or not isinstance(fresh_messages, list):
        return False
    if len(existing_messages) > len(fresh_messages):
        return False
    for idx, existing_message in enumerate(existing_messages):
        fresh_message = fresh_messages[idx]
        if _normalize_message_for_import_refresh(existing_message) != _normalize_message_for_import_refresh(fresh_message):
            return False
    return True


def _handle_session_import_cli(handler, body):
    """Import a single CLI session into the WebUI store."""
    try:
        require(body, "session_id")
    except ValueError as e:
        return bad(handler, str(e))

    sid = str(body["session_id"])

    # Check if already imported — refresh messages from CLI store if new ones arrived
    existing = Session.load(sid)
    if existing:
        fresh_msgs = get_cli_session_messages(sid)
        changed = False
        cli_meta = None
        for cs in list(get_cli_sessions()):
            if cs["session_id"] == sid:
                cli_meta = cs
                break
        if fresh_msgs and len(fresh_msgs) > len(existing.messages):
            # Prefix-equality guard: only extend if existing messages are a prefix of
            # the fresh CLI messages. Prevents silently dropping WebUI-added messages
            # on hybrid sessions (user sent messages via WebUI while CLI continued).
            if _is_messages_refresh_prefix_match(existing.messages, fresh_msgs):
                existing.messages = fresh_msgs
                changed = True
        if cli_meta:
            updates = {
                "is_cli_session": True,
                "source_tag": existing.source_tag or cli_meta.get("source_tag"),
                "raw_source": existing.raw_source or cli_meta.get("raw_source") or cli_meta.get("source_tag"),
                "session_source": existing.session_source or cli_meta.get("session_source"),
                "source_label": existing.source_label or cli_meta.get("source_label"),
            }
            for attr, value in updates.items():
                if getattr(existing, attr, None) != value:
                    setattr(existing, attr, value)
                    changed = True
        if changed:
            existing.save(touch_updated_at=False)
        return j(
            handler,
            {
                "session": existing.compact()
                | {
                    "messages": existing.messages,
                    "is_cli_session": True,
                },
                "imported": False,
            },
        )

    # Fetch messages from CLI store
    msgs = get_cli_session_messages(sid)
    if not msgs:
        return bad(handler, "Session not found in CLI store", 404)

    # Get profile, model, timestamps, and title from CLI session metadata
    profile = None
    created_at = None
    updated_at = None
    cli_title = None
    cli_source_tag = None
    model = "unknown"
    cli_raw_source = None
    cli_session_source = None
    cli_source_label = None
    cli_user_id = None
    cli_chat_id = None
    cli_chat_type = None
    cli_thread_id = None
    cli_session_key = None
    cli_platform = None
    for cs in get_cli_sessions():
        if cs["session_id"] == sid:
            profile = cs.get("profile")
            model = cs.get("model", "unknown")
            created_at = cs.get("created_at")
            updated_at = cs.get("updated_at")
            cli_title = cs.get("title")
            cli_source_tag = cs.get("source_tag")
            cli_raw_source = cs.get("raw_source")
            cli_session_source = cs.get("session_source")
            cli_source_label = cs.get("source_label")
            cli_user_id = cs.get("user_id")
            cli_chat_id = cs.get("chat_id")
            cli_chat_type = cs.get("chat_type")
            cli_thread_id = cs.get("thread_id")
            cli_session_key = cs.get("session_key")
            cli_platform = cs.get("platform")
            break

    # Use the CLI session title if available (e.g., cron job name), otherwise derive from messages
    title = cli_title or title_from(msgs, "CLI Session")

    # Auto-assign cron sessions to the dedicated "Cron Jobs" project (#1079)
    cron_project_id = None
    if is_cron_session(sid, cli_source_tag):
        cron_project_id = ensure_cron_project()

    s = import_cli_session(
        sid,
        title,
        msgs,
        model,
        profile=profile,
        created_at=created_at,
        updated_at=updated_at,
    )
    if cron_project_id:
        s.project_id = cron_project_id
    s.is_cli_session = True
    s.source_tag = cli_source_tag
    s.raw_source = cli_raw_source or cli_source_tag
    s.session_source = cli_session_source
    s.source_label = cli_source_label
    s.user_id = cli_user_id
    s.chat_id = cli_chat_id
    s.chat_type = cli_chat_type
    s.thread_id = cli_thread_id
    s.session_key = cli_session_key
    s.platform = cli_platform
    s._cli_origin = sid
    s.save(touch_updated_at=False)
    return j(
        handler,
        {
            "session": s.compact()
            | {
                "messages": msgs,
                "is_cli_session": True,
            },
            "imported": True,
        },
    )


def _handle_session_import(handler, body):
    """Import a session from a JSON export. Creates a new session with a new ID."""
    if not body or not isinstance(body, dict):
        return bad(handler, "Request body must be a JSON object")
    messages = body.get("messages")
    if not isinstance(messages, list):
        return bad(handler, 'JSON must contain a "messages" array')
    title = body.get("title", "Imported session")
    workspace = body.get("workspace", str(DEFAULT_WORKSPACE))
    model = body.get("model", DEFAULT_MODEL)
    s = Session(
        title=title,
        workspace=workspace,
        model=model,
        messages=messages,
        tool_calls=body.get("tool_calls", []),
    )
    s.pinned = body.get("pinned", False)
    with LOCK:
        SESSIONS[s.session_id] = s
        SESSIONS.move_to_end(s.session_id)
        while len(SESSIONS) > SESSIONS_MAX:
            SESSIONS.popitem(last=False)
    s.save()
    return j(handler, {"ok": True, "session": s.compact() | {"messages": s.messages}})


# ── MCP Server helpers ──
from api.config import get_config, _save_yaml_config_file, _get_config_path, reload_config

def _mask_secrets(obj):
    """Mask sensitive values in env vars and headers."""
    if not isinstance(obj, dict):
        return obj
    sensitive = ("auth", "token", "key", "secret", "password", "credential")
    masked = {}
    for k, v in obj.items():
        if isinstance(v, str) and any(s in k.lower() for s in sensitive):
            masked[k] = "••••••"
        elif isinstance(v, dict):
            masked[k] = _mask_secrets(v)
        else:
            masked[k] = v
    return masked


def _server_summary(name, cfg):
    """Return a safe summary of an MCP server config."""
    out = {"name": name}
    if "url" in cfg:
        out["transport"] = "http"
        # Mask auth headers
        if "headers" in cfg:
            out["headers"] = _mask_secrets(cfg["headers"])
        out["url"] = cfg["url"]
    else:
        out["transport"] = "stdio"
        out["command"] = cfg.get("command", "")
        out["args"] = cfg.get("args", [])
        if "env" in cfg:
            out["env"] = _mask_secrets(cfg["env"])
    out["timeout"] = cfg.get("timeout", 120)
    return out


def _handle_mcp_servers_list(handler):
    """List all configured MCP servers."""
    cfg = get_config()
    servers = cfg.get("mcp_servers", {})
    if not isinstance(servers, dict):
        servers = {}
    result = [_server_summary(name, scfg) for name, scfg in servers.items()]
    return j(handler, {"servers": result})


def _handle_mcp_server_delete(handler, name):
    """Delete an MCP server by name."""
    from urllib.parse import unquote
    name = unquote(name)
    if not name:
        return bad(handler, "name is required")
    cfg = get_config()
    servers = cfg.get("mcp_servers", {})
    if not isinstance(servers, dict):
        servers = {}
    if name not in servers:
        return bad(handler, f"MCP server '{name}' not found", 404)
    del servers[name]
    cfg["mcp_servers"] = servers
    _save_yaml_config_file(_get_config_path(), cfg)
    reload_config()
    return j(handler, {"ok": True, "deleted": name})


_MASKED_PLACEHOLDER = "••••••"


def _strip_masked_values(submitted, existing):
    """Remove masked placeholder values from submitted dict, keeping originals."""
    if not isinstance(submitted, dict) or not isinstance(existing, dict):
        return submitted
    cleaned = {}
    for k, v in submitted.items():
        if isinstance(v, str) and v == _MASKED_PLACEHOLDER:
            if k in existing and isinstance(existing[k], str):
                cleaned[k] = existing[k]  # preserve original real value
                continue
        elif isinstance(v, dict) and k in existing and isinstance(existing[k], dict):
            cleaned[k] = _strip_masked_values(v, existing[k])
        else:
            cleaned[k] = v
    return cleaned


def _handle_mcp_server_update(handler, name, body):
    """Add or update an MCP server."""
    from urllib.parse import unquote
    name = unquote(name)
    if not name:
        return bad(handler, "name is required")
    # Validate: must have url (http) or command (stdio)
    server_cfg = {}
    cfg = get_config()
    servers = cfg.get("mcp_servers", {})
    if not isinstance(servers, dict):
        servers = {}
    existing_cfg = servers.get(name, {})
    if body.get("url"):
        server_cfg["url"] = body["url"].strip()
        if body.get("headers"):
            server_cfg["headers"] = _strip_masked_values(body["headers"], existing_cfg.get("headers", {}))
    elif body.get("command"):
        server_cfg["command"] = body["command"].strip()
        if body.get("args"):
            server_cfg["args"] = body["args"] if isinstance(body["args"], list) else [body["args"]]
        if body.get("env"):
            server_cfg["env"] = _strip_masked_values(body["env"], existing_cfg.get("env", {}))
    else:
        return bad(handler, "url or command is required")
    if body.get("timeout") is not None:
        try:
            server_cfg["timeout"] = int(body["timeout"])
        except (ValueError, TypeError):
            pass
    servers[name] = server_cfg
    cfg["mcp_servers"] = servers
    _save_yaml_config_file(_get_config_path(), cfg)
    reload_config()
    return j(handler, {"ok": True, "server": _server_summary(name, server_cfg)})