Release v0.51.224 — Release GR (stage-p6 — profile tool/skill config authoritative on streaming worker #3294) (#3456)

* fix: respect profile toolset/skill config on WebUI streaming worker (#3294)

The streaming agent runs on a detached worker thread that does not inherit
the per-request thread-local profile context (set from the hermes_profile
cookie on the HTTP handler thread). On that worker, the ambient get_config()
resolves via get_active_profile_name() which falls back to the process-global
_active_profile (usually 'default'). A session under a non-default profile
with an empty platform_toolsets.cli therefore loaded the DEFAULT profile's
full toolset list, inflating a tools-disabled profile's prompt from ~400 to
~15K input tokens.

Add api.config.get_config_for_profile_home() — a race-free direct disk read
of an explicit profile home's config.yaml (no shared-cache mutation), which
defers to get_config() when the requested home matches the ambient path so
in-memory test overrides are preserved. The streaming worker now resolves
_cfg from the session's own profile home, fixing toolsets, prefill context,
and fallback chains in one place.

Closes #3294

Co-authored-by: gottipx <gottipx@users.noreply.github.com>

* docs(changelog): v0.51.224 — #3294 profile toolset config (dropped #3405 to hold per Codex stale-model-repair finding)

---------

Co-authored-by: nesquena-hermes <[email protected]>
Co-authored-by: gottipx <gottipx@users.noreply.github.com>
This commit is contained in:
nesquena-hermes 2026-06-02 18:33:22 -07:00 committed by GitHub
parent b850a88e8b
commit bb5aa02523
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 191 additions and 4 deletions

View file

@ -3,7 +3,12 @@
## [Unreleased]
## [v0.51.223] — 2026-06-02 — Release GQ (stage-p5 — openai-api picker provider + MiniMax-M3)
## [v0.51.224] — 2026-06-03 — Release GR (stage-p6 — profile tool/skill config authoritative on the streaming worker)
### Fixed
- Profile tool/skill restrictions are now respected for WebUI chats even when the per-session "Tool Restrictions" field is left blank. The streaming agent runs on a detached worker thread that does not inherit the per-request thread-local profile context, so the ambient `get_config()` resolved the process-global `default` profile and loaded its `platform_toolsets.cli` (all tools) instead of the session profile's configured list — inflating a tools-disabled profile's prompt from ~400 to ~15K input tokens. The worker now reads the session's own profile config explicitly via a new `get_config_for_profile_home()` helper (a race-free direct disk read with no shared-cache mutation), so toolsets, prefill context, and fallback chains all match the profile the session actually runs under (#3294, @nesquena-hermes).
## [v0.51.223] — 2026-06-02 — Release GQ (stage-p5 — openai-api first-class picker provider + MiniMax-M3)
### Fixed
- GPT models now appear in the model picker when hermes-agent exposes its built-in OpenAI provider under the `openai-api` slug (the one activated by `OPENAI_API_KEY` / `OPENAI_BASE_URL`, distinct from `openai-codex`). `openai-api` is now a first-class picker provider in `_PROVIDER_DISPLAY` / `_PROVIDER_MODELS` rather than an alias of `openai` — an alias would have fixed the display but broken the send path, since the agent registry has `openai-api` and not `openai`. Env detection for `OPENAI_API_KEY` was also corrected to surface `openai-api` instead of a bare `openai` the agent registry can't resolve (#3443, @rodboev).

View file

@ -382,6 +382,43 @@ def _load_yaml_config_file(config_path: Path) -> dict:
return {}
def get_config_for_profile_home(profile_home: "Path | str | None") -> dict:
"""Return the config dict for an explicit profile home directory.
The streaming agent runs on a detached worker thread that does NOT inherit
the per-request thread-local profile context (set from the ``hermes_profile``
cookie on the HTTP handler thread). On that worker, the ambient
``get_config()`` resolves through ``get_active_profile_name()`` which falls
back to the process-global ``_active_profile`` (usually ``default``) — so a
session running under a non-default profile would silently read the
**default** profile's ``config.yaml`` for toolsets, prefill context, and
fallback chains (issue #3294).
This helper reads the config for a *known* profile home directly off disk,
bypassing the thread-local resolver entirely. When ``profile_home`` matches
the path the ambient resolver would pick (the common single-profile case),
we return the cached ``get_config()`` to preserve in-memory overrides used
by tests and runtime callers. Only when the session's profile home diverges
from the ambient path do we read the session profile's file directly — a
pure read with no global cache mutation, so it is race-free across
concurrent sessions on different profiles.
"""
if not profile_home:
return get_config()
try:
target = Path(profile_home).expanduser()
except Exception:
return get_config()
# If the ambient resolver already points at this profile home, defer to
# get_config() so in-memory overrides (monkeypatched cfg) are honored.
try:
if _get_config_path().parent == target:
return get_config()
except Exception:
pass
return _load_yaml_config_file(target / "config.yaml")
def _save_yaml_config_file(config_path: Path, config_data: dict) -> None:
try:
import yaml as _yaml

View file

@ -4892,9 +4892,20 @@ def _run_agent_streaming(
resolved_provider, resolved_api_key, resolved_base_url
)
# Read per-profile config at call time (not module-level snapshot)
from api.config import get_config as _get_config
_cfg = _get_config()
# Read per-profile config at call time (not module-level snapshot).
# The streaming worker is a detached thread that does NOT inherit the
# per-request thread-local profile context, so the ambient
# get_config() would resolve the process-global (default) profile and
# leak the wrong profile's toolsets / prefill / fallback config into
# this run (issue #3294). Read the SESSION's own profile home
# explicitly so toolsets and context match the profile the session
# actually runs under.
from api.config import get_config_for_profile_home as _get_config_for_home
try:
_cfg = _get_config_for_home(_profile_home)
except Exception:
from api.config import get_config as _get_config
_cfg = _get_config()
_prefill_context = _load_webui_prefill_context(_cfg)
_prefill_messages = _prefill_messages_with_webui_context(_prefill_context, _cfg)
put('context_status', {

View file

@ -0,0 +1,134 @@
"""Regression coverage for #3294 — profile tool configuration not respected in WebUI.
When a non-default profile leaves the per-session "Tool Restrictions"
(`enabled_toolsets`) blank, the streaming worker must resolve toolsets (and the
rest of the per-profile config: prefill context, fallback chains) from the
*session's own profile* config.yaml — not from the process-global ``default``
profile.
Root cause: the streaming agent runs on a detached ``threading.Thread`` that
does NOT inherit the per-request thread-local profile context (set from the
``hermes_profile`` cookie on the HTTP handler thread). On that worker the
ambient ``get_config()`` resolves through ``get_active_profile_name()`` which
falls back to the process-global ``_active_profile`` (usually ``default``). A
profile B with an empty ``platform_toolsets.cli`` would therefore load the
default profile's full toolset list, inflating the prompt from ~400 to ~15K
tokens (the reporter's symptom).
Fix: ``api.config.get_config_for_profile_home(profile_home)`` reads the config
for an explicit profile home directly off disk, bypassing the thread-local
resolver. The streaming worker passes the session's own resolved profile home.
"""
from __future__ import annotations
import sys
from pathlib import Path
import pytest
REPO = Path(__file__).resolve().parent.parent
if str(REPO) not in sys.path:
sys.path.insert(0, str(REPO))
def _write_cfg(home: Path, cli_toolsets) -> None:
home.mkdir(parents=True, exist_ok=True)
import yaml
home.joinpath("config.yaml").write_text(
yaml.safe_dump({"platform_toolsets": {"cli": cli_toolsets}}, sort_keys=False),
encoding="utf-8",
)
@pytest.fixture
def _two_profiles(tmp_path, monkeypatch):
"""default profile = all tools; profile B = empty toolset list."""
from api import config as cfg
default_home = tmp_path / "default"
profile_b_home = tmp_path / "profiles" / "b"
_write_cfg(default_home, ["terminal", "file", "web", "search", "browser"])
_write_cfg(profile_b_home, [])
# Point the ambient resolver at the DEFAULT profile, simulating a worker
# thread that fell back to the process-global default profile.
monkeypatch.setenv("HERMES_CONFIG_PATH", str(default_home / "config.yaml"))
cfg.reload_config()
yield cfg, default_home, profile_b_home
def test_ambient_get_config_reads_default_profile(_two_profiles):
"""Baseline: the ambient resolver sees the DEFAULT profile (the bug source)."""
cfg, default_home, profile_b_home = _two_profiles
ambient = cfg.get_config()
assert ambient.get("platform_toolsets", {}).get("cli") == [
"terminal",
"file",
"web",
"search",
"browser",
]
def test_get_config_for_profile_home_reads_session_profile(_two_profiles):
"""The fix: explicitly resolving profile B's home reads B's empty toolsets,
even though the ambient resolver still points at the default profile."""
cfg, default_home, profile_b_home = _two_profiles
profile_b_cfg = cfg.get_config_for_profile_home(profile_b_home)
assert profile_b_cfg.get("platform_toolsets", {}).get("cli") == []
# And the toolset resolver run against profile B's config must yield a
# strictly smaller set than the default profile's — the token-inflation
# the reporter observed comes from resolving against the wrong (default)
# profile's full list.
default_toolsets = set(cfg._resolve_cli_toolsets(cfg.get_config()))
profile_b_toolsets = set(cfg._resolve_cli_toolsets(profile_b_cfg))
assert profile_b_toolsets < default_toolsets
assert "terminal" in default_toolsets
assert "terminal" not in profile_b_toolsets
assert "browser" not in profile_b_toolsets
def test_matching_home_defers_to_get_config(_two_profiles):
"""When the requested home matches the ambient path, defer to get_config()
so in-memory overrides (monkeypatched cfg / runtime overrides) are honored."""
cfg, default_home, profile_b_home = _two_profiles
same = cfg.get_config_for_profile_home(default_home)
# Same dict identity path as get_config() (cached) — verify it returns the
# default profile's config, not a fresh disk read that could miss overrides.
assert same.get("platform_toolsets", {}).get("cli") == [
"terminal",
"file",
"web",
"search",
"browser",
]
def test_empty_or_none_home_falls_back_to_get_config(_two_profiles):
"""A missing/empty profile home (ImportError fallback path in the worker)
must not crash — it falls back to the ambient get_config()."""
cfg, default_home, profile_b_home = _two_profiles
assert cfg.get_config_for_profile_home(None) == cfg.get_config()
assert cfg.get_config_for_profile_home("") == cfg.get_config()
def test_nonexistent_profile_home_returns_empty_not_default(_two_profiles):
"""A profile home that diverges from ambient but has no config.yaml yet
returns an empty dict (the profile's own absent config) rather than silently
inheriting the default profile's tools."""
cfg, default_home, profile_b_home = _two_profiles
ghost = profile_b_home.parent / "ghost"
result = cfg.get_config_for_profile_home(ghost)
assert result == {}
# Critically: it does NOT return the default profile's full toolset list.
assert result.get("platform_toolsets", {}).get("cli") != [
"terminal",
"file",
"web",
"search",
"browser",
]