Replace the hardcoded Skyly cancellation wording with the configured bot_name from settings, falling back to Hermes when unset.
Keep the client-side fallback in sync by using window._botName if the session refresh after cancellation fails.
Co-authored-by: Obryn 🐉 <obryn-ai@dotbeeps.dev>
Opus stage-360 review caught that the docstring at api/streaming.py:40-43
said 'around the entire agent run' which is no longer accurate after the
narrow-lock refactor. The lock is now held only briefly for the env-mutation
critical section; the agent runs outside the lock and the finally block
re-acquires to atomically restore env vars.
Docstring now points to both narrow-lock implementations as references:
- _run_agent_streaming at line ~2719 (the original pattern)
- profile_env_for_background_worker at api/profiles.py:715 (added stage-360)
#2299 introduced profile_env_for_background_worker() in api/profiles.py and
changed _ENV_LOCK from threading.Lock() to threading.RLock(). Both changes
were incorrect:
1. RLock masked rather than fixed the underlying deadlock. The QA
test_env_lock_is_non_reentrant test exists precisely to enforce
non-reentrance — RLock would let a single thread hold _ENV_LOCK across
nested critical sections, which hides bugs while still allowing
different-thread races.
2. The original context manager held _ENV_LOCK for the ENTIRE 'yield'
duration, meaning the lock was held for the full background worker's
runtime (title generation, compression, update summary — possibly
many seconds). That blocked ALL other sessions on _ENV_LOCK, which
the QA test_third_message_completes runtime test caught as a timeout
on the third sequential message.
Fix: mirror the narrow-lock pattern from _run_agent_streaming:
- Acquire _ENV_LOCK only for env mutation (set runtime_env + patch
skill modules)
- Release immediately, yield to worker (no lock held)
- Reacquire in finally to restore env + skill modules
Restored _ENV_LOCK back to threading.Lock(). All 20 QA tests now pass,
including test_third_message_completes (was timing out, now 35s).
The FakeAgent in test_issue1857_usage_overwrite returned only 2 messages
(user + assistant) without the conversation history. The real agent always
returns the full history plus new messages. This mismatch caused the new
_has_new_assistant_reply helper (which checks only messages beyond the
pre-turn offset) to see len(result)==len(prev) and incorrectly flag the
turn as a silent failure.
Fix: prepend conversation_history to the FakeAgent's response so the
message list mirrors production behavior.
When a provider error (401/429/rate-limit) causes the agent to return
without producing a new assistant reply, the WebUI should emit an
apperror event so the user sees an inline error. However, the detection
logic scanned ALL messages in result['messages'] — which includes the
full conversation history. If any prior turn had an assistant response,
_assistant_added would be True and the apperror would be silently
skipped, leaving the user staring at a blank response.
Extract a helper _has_new_assistant_reply(all_messages, prev_count)
that only inspects messages beyond the pre-turn history offset. Apply
it to both the main detection path and the self-heal/retry path.
Tests: 15 new cases covering history masking, empty content, whitespace,
edge-case shrinks, and multi-assistant scenarios.
Opus identified that PR #2227's preservation block had two related bugs in
the parent_session_id handling:
1. During preservation save: code did
_old_parent = s.parent_session_id
s.parent_session_id = None
s.save(touch_updated_at=False, skip_index=True)
s.parent_session_id = _old_parent
The save persisted parent=None to disk. The in-memory restoration didn't
reach the disk copy. Result: a /branch fork session that subsequently
compressed lost its 'Forked from X' badge on the preserved old snapshot.
2. Stamping the continuation: code did
if not s.parent_session_id:
s.parent_session_id = old_sid
The 'if not' guard skipped the stamp when the session already had a
parent_session_id from a prior fork. Result: fork-of-fork compression
broke lineage — the continuation jumped back to the original fork parent
instead of the just-preserved immediate predecessor snapshot.
Fix (matches Opus's recommendation):
- Remove the parent clearing during preservation save (preserve as-is)
- Drop the 'if not' guard; always stamp continuation to old_sid
This makes the lineage chain consistent: new → old → old.parent → ... root.
Traversal from the continuation always walks through the just-preserved
snapshot to get to its parent's parent, never jumping over the snapshot.
Two new regression tests pin both invariants:
- test_parent_session_id_stamped_unconditionally (no 'if not' guard)
- test_old_session_parent_preserved_during_archive_save (no parent=None)
Both pass against the fix. All 8 tests in the file pass.
The previous implementation renamed old_sid.json → new_sid.json during
context compression, destroying the only persistent copy of the full
conversation history. If the summarisation LLM call also failed, the
user was left with zero recoverable messages.
Fix:
- Remove the destructive old_path.rename(new_path) call
- Preserve old_sid.json as an immutable pre-compression archive
- Create new_sid.json as a fresh file via s.save()
- Set parent_session_id on the continuation session for lineage
- Save in-memory messages to old_sid.json if they're newer than disk
Test: test_issue2223_compression_no_rename.py (6 tests, all passing)
Refs #2215 Fix B: remove the mid-response stripping hazard without losing leading multi-line wrapper cleanup.
The pattern now strips only a leading 'the user is asking' wrapper line and preserves the visible answer that follows. Add regression coverage for both the leading-wrapper and mid-response prose cases.
Opus flagged that PR #2151's cancel-handler partial-dedup loop used a
substring check that was too broad: any short prior assistant reply
('OK', 'Here is the answer:') would dedup a longer new partial containing
it, silently dropping the partial and resurrecting the #893 data-loss bug.
Tightened to only dedup against actual prior _partial=True markers with
exact (whitespace-stripped) content match. Three new regression tests
added (short-non-partial-prefix-does-not-dedup, exact-partial-match-still-
dedups, same-content-non-partial-does-not-dedup).
10/10 partial-cancel tests pass after the fix. Also updated CHANGELOG with
the conflict-resolution notes for #2151 vs #2136 and the #2178 test-fix.
fix: clarify cancelled chat turn status (Jordan-SkyLF)
Conflict resolution on api/streaming.py:4549-4567 (the cancel-handler
ownership guard). Both this PR and the already-shipped PR #2136 add a
guard at the same site against stale stream writebacks, from different
angles:
- PR #2136 (HEAD): _stream_writeback_is_current(_cs, stream_id) — strictly
dominates by checking the active_stream_id token equality.
- PR #2151: 'worker won the race' check via (active_stream_id != stream_id
and not pending_user_message), with _emit_cancel_event = False to suppress
the terminal cancel event.
Resolution merges both: keep #2136's strictly-stronger condition for skip
detection, and adopt #2151's _emit_cancel_event = False semantic so the
cancel event isn't emitted in addition to skipping the writeback (when
client may have already received the successful done payload).
55/55 tests pass across cancelled-turn-status + stale-stream-writeback +
the four cancel/data-loss sibling test files.
Providers like Xiaomi MiMo, DeepSeek, and Kimi require reasoning_content
to be echoed back on every assistant message in multi-turn conversations
with tool calls. Omitting it causes HTTP 400: 'The reasoning_content in
the thinking mode must be passed back to the API.'
The WebUI's _sanitize_messages_for_api() strips all fields not in
_API_SAFE_MSG_KEYS before sending conversation history to the LLM API.
reasoning_content was not in this whitelist, so it was silently dropped.
The CLI path (run_agent.py) is unaffected because it has its own
_copy_reasoning_content_for_api() logic that operates on raw message
dicts without going through this filter. This is why the same session
works from CLI but fails from WebUI with HTTP 400.
The fix adds 'reasoning_content' to _API_SAFE_MSG_KEYS so the field
passes through sanitization intact.
fix: guard stale stream writebacks (LumenYoung)
Prevents stale WebUI stream workers from writing old results into a session
after that session has already moved on to another stream. Adds new helper
_stream_writeback_is_current() (a token equality check against the session's
active_stream_id) and short-circuits the two finalize/cancel paths when the
worker no longer owns the session writeback.
Opus advisor pass on stage-341 found three surgical items:
1. static/i18n.js:it — PR #2064 branched before stage-340 landed the 'it'
locale (#2067), missing 9 session_*worktree* keys. Mechanical mirror of
en/ja position. Italian falls back to English silently without this fix.
2. api/streaming.py — PR #2107's new break short-circuit was silent in both
the aux and agent title-generation paths. Added logger.debug calls before
each break so production logs surface the exit shape.
3. api/streaming.py — Expanded _title_should_skip_remaining_attempts docstring
to document the membership criterion explicitly (vs the implicit
reasoning-only-burn case it ships with today). Future additions
(llm_safety_blocked, llm_oauth_quota) have a clear inclusion test.
CHANGELOG updated under the Stage-341 maintainer fixes section to mirror
the stage-340 pattern. All targeted tests pass (57/57 in the affected
modules).
Reasoning models (Qwen3-thinking via LM Studio, DeepSeek-R1, Kimi-K2,
etc.) can burn their entire output budget on hidden reasoning tokens and
emit no visible content. The previous title-generation retry path
classified that as llm_length and doubled the budget — but the second
call produces the same shape, so the retry only doubled the GPU/credit
burn. Repeated across the two prompts in _title_prompts() this came to
~3000 reasoning tokens of GPU work per new chat. On local LM Studio
servers behind a custom: provider (where is_lmstudio=False means
reasoning_effort: none never reaches the model) it manifested as the GPU
never going idle after a prompt.
Fix:
- _extract_title_response: classify reasoning-bearing empty responses
as llm_empty_reasoning regardless of finish_reason. The presence of
reasoning_content is the diagnostic signal, not finish_reason.
- _title_retry_status: drop llm_empty_reasoning from the retry set.
Length-truncated responses WITHOUT reasoning still retry (those are
legitimately recoverable by a larger budget).
- Add _title_should_skip_remaining_attempts() and break out of the
prompt-iteration loop on empty-reasoning. A second prompt against
the same model would produce the same shape.
- Falls through to _fallback_title_from_exchange for a local-summary
title.
Tests updated to invert the previous reasoning-retry assertions:
- test_aux_short_circuits_on_empty_reasoning_without_retrying
- test_aux_still_retries_finish_length_without_reasoning
- test_agent_route_short_circuits_on_empty_reasoning_without_retrying
- test_agent_route_still_retries_finish_length_without_reasoning
Companion agent-side work (LM Studio classifier for custom: providers)
is tracked separately on the hermes-agent side; this WebUI fix is the
belt-and-braces guard so the loop stops regardless of agent classifier
state.
Reported by @darkopetrovic. Closes#2083.
Co-authored-by: darkopetrovic <darkopetrovic@users.noreply.github.com>
(cherry picked from commit efeae4a86e377069c0f09d140429ecb111a8dd1a)
When context compression fires, the agent rotates to a new session_id.
The compression migration block correctly migrates the session lock,
SESSION_AGENT_CACHE, SESSIONS dict, and the session file rename, but
does not ensure s.profile is set on the continuation session.
On the next request, _run_agent_streaming resolves the profile via:
get_hermes_home_for_profile(getattr(s, 'profile', None))
With s.profile == None this falls back to the default profile's
HERMES_HOME. Memory tool calls then read and write the wrong profile's
MEMORY.md — confirmed by investigation: session 0dfefb (continuation
after compression from a troubleshooting profile session) read memory
at 16% / 1,184 chars with 4 entries, while the troubleshooting profile's
actual state was 72-77% / 5,000+ chars. That reading could only come
from the default profile's bank. Subsequent replace operations failed
because the target entries existed only in the troubleshooting profile.
There are two failure paths:
1. In-memory: if s.profile was None from the start (legacy session or
one created before this fix), the continuation session object carries
null through the current request.
2. Persistence: s.save() persists "profile": null to the continuation
session's JSON file (profile is in METADATA_FIELDS, models.py ~408).
On the next request, Session.load(new_sid) reads it back as null and
get_hermes_home_for_profile(None) falls back to the default profile.
Fix: capture _resolved_profile_name at request entry (~line 2019),
immediately after profile home resolution. This is the only point where
profile context is reliable: s.profile if already set, otherwise
get_active_profile_name() — which at that point reads thread-local
storage (_tls.profile) correctly set by the HTTP handler thread via
set_request_profile(). Calling get_active_profile_name() at compression
time instead would be unsafe: the streaming thread is a separate
threading.Thread, does not inherit TLS, and the call would fall back to
the process-global _active_profile which may belong to a different
concurrent tab.
Stamp s.profile in the compression migration block immediately after
s.session_id = new_sid. Guarded by `if not s.profile` so sessions that
already have a profile set are unaffected. A logger.info line records
when the stamp fires, making future investigation straightforward.
Fixes: memory writes bleeding into default profile after compression
Reproduces: reliably on any long non-default profile session that hits
the compression threshold (default: 0.80 context fill)
Per-request profile switches (process_wide=False, introduced in #1700)
update os.environ['HERMES_HOME'] but skip _set_hermes_home(), which is
responsible for monkeypatching module-level caches.
Both tools/skills_tool.py and tools/skill_manager_tool.py set
HERMES_HOME and SKILLS_DIR once at import time. When a non-default
profile is active in the WebUI, os.environ['HERMES_HOME'] is correctly
updated per-turn in the _ENV_LOCK block, but the module-level
constants still point at the root profile. All agent-side skill
operations — skills_list(), skill_view(), skill_manage() — read and
write to the wrong directory.
Add the same monkeypatching that _set_hermes_home() already performs
(profiles.py line ~620) to the per-turn env setup block in
streaming.py, covering both skills_tool and skill_manager_tool.
The WebUI display half was already fixed in #1917 via
_active_skills_dir() in routes.py. This patch fixes the agent-side
half so the running agent resolves skills from the correct profile.
Issue #1968: switching to a non-default profile in the WebUI dropdown
had no effect on which MCP servers were available. Every chat session,
regardless of profile, only saw the default profile's mcp_servers from
~/.hermes/config.yaml. Non-default profile MCP servers (postgres, custom
stdio servers, anything in <profile>/config.yaml) never registered.
Root cause: api/streaming.py:1922 called discover_mcp_tools() at the
TOP of _run_agent_streaming(), about 100 lines BEFORE the per-session
'os.environ["HERMES_HOME"] = _profile_home' mutation at line 2053.
discover_mcp_tools() reads ~/.hermes/config.yaml via get_hermes_home(),
which uses os.environ['HERMES_HOME']. So at the call site, HERMES_HOME
was still whatever the WebUI server process had at startup — the default
profile, every time.
Fix: relocate the discover_mcp_tools() call past the _ENV_LOCK block so
get_hermes_home() resolves to the session's actual profile home. Same
try/except wrapping is preserved; same idempotency semantics on
already-connected servers; same lazy-import pattern.
Caveat (out of scope, agent-side): _servers in tools/mcp_tool.py is a
process-global Dict[str, MCPServerTask] keyed only by server name. So
once profile A registers a server named e.g. 'postgres', profile B's
discovery sees 'postgres' as already connected and skips it — even if
B's config points at a different binary or DB. Concurrent multi-profile
WebUI processes will still hit 'first profile wins per server name'.
Fully fixing that requires keying _servers by (profile_home, name)
upstream in hermes-agent. This PR ships layer 1 only — fixes the
single-non-default-profile case (the headline symptom).
Tests: tests/test_issue1968_mcp_profile_discovery.py — 4 static tests
pinning the lexical ordering invariants. Verified mutation-safety: a
proof-of-concept revert (re-adding a discover call before the
HERMES_HOME mutation) makes the 'only called once' test fail.
Test suite: 5047 passed, 4 skipped, 3 xpassed, 0 regressions.
Closes#1968
CRITICAL: #1951 PENDING_GOAL_CONTINUATION race
Removes `PENDING_GOAL_CONTINUATION.discard(session_id)` from the
streaming worker's `finally` cleanup block. The marker is set inside
the SAME function call (line ~3328 on `goal_continue`) and the discard
in the `finally` (line ~3553) almost always raced ahead of the
frontend's SSE-receive → POST /api/chat/start round-trip, erasing
the marker before the consumer in routes.py could read it. The
consumer (`_start_chat_stream_for_session` in routes.py:6522) already
discards atomically when consuming, so removing the streaming-side
discard preserves single-use semantics and unblocks the
goal-continuation chain.
Adds tests/test_stage326_pending_goal_continuation_race.py with 5
regression guards:
1. streaming.py's finally must NOT discard PENDING_GOAL_CONTINUATION
2. routes.py consumer must check + set + discard atomically
3. PENDING_GOAL_CONTINUATION must be a set (GIL-safe single-op)
4. STREAM_GOAL_RELATED.pop must be keyed by stream_id, not session_id
5. PENDING_GOAL_CONTINUATION.add must precede the goal_continue SSE
emission in source ordering
HARDENING: #1956 composer-draft input validation
Per Opus, the POST /api/session/draft handler accepted unbounded /
arbitrary-typed text and files inputs. With the 400ms debounced
auto-save firing on every keystroke, a misbehaving client could
persist multi-MB strings into the session JSON. Adds:
- text: coerced to str if not already; clamped to 50_000 chars
- files: coerced to list if not already; clamped to 50 entries
Validation runs BEFORE the session lock acquire / save.
Adds tests/test_stage326_composer_draft_validation.py with 5 guards.
Verdict from Opus advisor on stage-326: SHIP-WITH-FIXES.
This commit applies the required + recommended fixes; #1957 hardening
fixed in a prior stage commit.
_build_native_multimodal_message() unconditionally embedded images as
native image_url parts, bypassing the agent's image_input_mode config.
Add _resolve_image_input_mode(cfg) helper mirroring the agent's
decide_image_input_mode logic, and wire it into
_build_native_multimodal_message with a new cfg parameter.
When mode resolves to 'text' (explicit aux vision config, or
image_input_mode: text), returns plain string so the agent's
existing text-mode pipeline (vision_analyze) handles images.
Closes#1959
The goal evaluation hook was firing on every completed assistant turn
when a goal was active, even for unrelated messages like "what time is
it". This burned the goal budget, triggered continuation prompts that
interrupted unrelated conversations, and made /goal status numbers
misleading.
Add STREAM_GOAL_RELATED and PENDING_GOAL_CONTINUATION flags to gate
the evaluate_goal_after_turn() call in the streaming loop. Only streams
started from goal kickoff (/goal <text>) or goal continuation are
marked as goal-related. Normal user messages skip the hook entirely.
The two get_model_context_length() fallback callsites in api/streaming.py
(session save + SSE usage payload) were calling the resolver with only
model + base_url. When the agent's compressor reports 0 (fresh/cached/
transitioning agent), resolution fell through to the 256K DEFAULT_FALLBACK
even when users had set model.context_length: 1048576 in config.yaml.
For LCM users on 1M-context models, the wrong window cascaded into a
session-killing failure: auto-compression triggered at ~25% of the wrong
value, floods of compress requests, 429s, credential pool exhaustion,
fallback 429s, then 'API call failed after 3 retries'.
Reported by @AvidFuturist on Discord with deepseek-v4-flash. Reproduced 5x.
Both callsites now pass config_context_length, provider, and
custom_providers. The resolver consults these BEFORE probing, so the
config override wins. Both are wrapped in except TypeError blocks that
retry with the legacy 2-arg form for older hermes-agent builds whose
get_model_context_length signature pre-dates these kwargs.
Tests: 7 source-string regressions guarding both call shapes, the safe
config parse, the legacy fallback, and the per-profile config source.
Also bumped the line-distance assertion in test_pr1341 (the test
explicitly invites bumping when a new pre-save mutation block is added).
Closes#1896
Co-authored-by: Hermes Agent <agent@hermes.local>
Same-session profile switches reused cached AIAgent from previous profile,
silently leaking the old persona's SOUL.md / system prompt into the new
profile's turns. session_id stays stable across profile switches, and the
signature didn't include the active profile home, so every signature input
matched and the stale agent was returned from SESSION_AGENT_CACHE.
Append _profile_home to the signature blob so profile switches force a
cache miss and a fresh agent build under the new HERMES_HOME (which
triggers a fresh load_soul_md() call).
Tests: 3 source-string regressions guarding the signature contract,
ordering, and empty-home fallback.
Closes#1897
Co-authored-by: Hermes Agent <agent@hermes.local>
Read agent.max_turns when constructing streaming WebUI AIAgent instances, pass it as max_iterations when supported, and include it in the per-session agent cache signature so budget changes take effect.
Add regression coverage for the config read, constructor kwarg, and cache key.
Closes#1695.
@Patrick-81 reported the bare "AIAgent not available -- check that
hermes-agent is on sys.path" error on a symlinked install (~/Programmes/hermes-agent
linked to ~/hermes-agent). The maintainer's response — three diagnostic
commands plus `pip install -e .` in the agent dir — fixed it for them.
This PR captures both halves of that learning so the next user with the
same shape doesn't have to file a new issue:
1. **Error message diagnostic block.** New helper
`_aiagent_import_error_detail()` in api/streaming.py builds a multi-line
diagnostic when the import fails, including:
- the running Python interpreter
- HERMES_WEBUI_AGENT_DIR (set value, or "(not set)")
- sys.path entries that mention hermes/agent (or "no entries mention..."
— itself a strong diagnostic signal)
- the most-common fix (`pip install -e .` in the agent dir)
- a pointer to docs/troubleshooting.md
The original error message string is preserved as the FIRST line so
existing log scrapers and docs-search keep matching.
Helper is kept as a separate function so it stays out of the hot path
until we actually need to raise — building it on every successful import
would be wasted work.
2. **New docs/troubleshooting.md.** Symptom → Why → Diagnostic commands →
Fix → When-to-file-a-bug template, with one entry to start: the
"AIAgent not available" flow Patrick-81 walked through. Future
recurring failure modes follow the same template. Required a one-line
addition to .gitignore — docs/* is gitignored with an allowlist, and
the new file needed `!docs/troubleshooting.md` to be tracked.
3. **README link.** docs/troubleshooting.md added to the `## Docs` section
so users know where to look first.
13 regression tests in tests/test_1695_aiagent_import_error_detail.py:
9 for the helper output shape (preserves original message line, includes
running python, shows HERMES_WEBUI_AGENT_DIR set/unset both ways, includes
pip-install-e hint, points at troubleshooting doc, lists relevant sys.path
entries when present, says "no entries..." when absent, output is multi-line)
plus 4 for the docs-presence regression (file exists, has the AIAgent
section, includes pip install -e ., describes the diagnostic chain with
readlink + agent/__init__.py verification).
190 streaming/aiagent tests pass after the change. ast.parse on
api/streaming.py clean.
CI failure on prior push was due to the docs/* gitignore swallowing the
new troubleshooting.md file silently — this commit adds the allowlist
entry so the file is tracked.
Switch the per-turn duration fallback from `is not None` to a truthy check so
None, missing-attr, and an explicit 0 all uniformly fall back to time.time().
Without this, a 0 timestamp (e.g. via a buggy migration or manual file edit)
would yield `time.time() - 0` ≈ wall-clock-since-epoch, displaying nonsense
like 'Done in 56 years 4 months ...'. In practice pending_started_at is always
set via int(time.time()) so this is a hardening fix, not a live-bug fix.
Also drop the brittle source-string assertion in the regression test that
pinned the literal expression. The behavioural test
test_done_handler_persists_duration_on_last_assistant_message already proves
the duration field is set; pinning the source line broke twice during the
v0.50.290 release pipeline alone (Opus tightening + maintainer revert).
Fixes#1595
Signed-off-by: Sanjay Santhanam <51058514+Sanjays2402@users.noreply.github.com>
Spliced from #1531 by @Asunfly: take Change-1 only (the actual bug fix +
cache signature inclusion) and skip Change-2 (auxiliary title-route
extra_body change) which is a separate scope concern.
## What
Two surgical fixes in api/streaming.py:
1. Line 1820 — `_cfg.cfg.get(...)` → `_cfg.get(...)`. `get_config()` returns
a plain dict (not a wrapper exposing `.cfg`). The buggy line raised
AttributeError that the surrounding try/except swallowed, so
`_reasoning_config` was always None regardless of what `/reasoning
<level>` had been set to. Verified locally — `api/streaming.py:1959`
already correctly used `_cfg.get(...)` in the same function, so the
same `_cfg` was being read two different ways in one file.
2. Line 1888 — added `_reasoning_config or {}` to `_sig_blob`. Without
this, switching effort mid-session would fail to take effect because
the per-session agent cache key would still match the old entry.
Mirrors how `resolved_provider` / `resolved_base_url` already
participate in the signature.
## Why splice instead of merge #1531 directly
@Asunfly force-pushed a Change-2 onto #1531 after the original review
that removes `extra_body={"reasoning": {"enabled": False}}` from
`generate_title_raw_via_aux` (the auxiliary title-generation route).
That intent is reasonable (let operator-configured `extra_body.reasoning`
flow through to the title route) but it touches a different surface and
deserves its own PR.
The narrow concern is operators who selected a reasoning-capable
auxiliary title model without explicitly setting
`reasoning.enabled=False` in the task config — pre-Change-2 the WebUI
defended against accidental reasoning on the title hot path; post-Change-2
those configs would reason on every new conversation`s title, with cost
and latency implications.
## What is NOT in this PR
- The `generate_title_raw_via_aux` extra_body refactor (Change-2 from #1531).
- The `test_does_not_override_configured_reasoning_extra_body` test (guards
Change-2). Asunfly can re-open that as its own focused PR.
## Tests
Two new R17b/R17c regression assertions in tests/test_regressions.py:
- `test_streaming_reads_reasoning_effort_from_config_dict` — static-source
guard: `_cfg.cfg` must not return to streaming.py
- `test_streaming_agent_cache_signature_includes_reasoning_config` —
catches removal of `_reasoning_config` from `_sig_blob`
## Closes
- Closes#1531 (the Change-1 portion ships here; Asunfly can re-open
Change-2 as a separate PR if desired)
Co-authored-by: Asunfly <[email protected]>
Read configured max_tokens from config.yaml, pass it into WebUI-created AIAgent instances when supported, and include it in the agent cache signature. Also classify OpenRouter quota phrasing such as more credits, can only afford, and fewer max_tokens.
Adds regression coverage for max_tokens propagation, cache signature isolation, and quota error classification.
PR #1421 (SessionDB WAL handle leak fix on cached-agent reuse path) had a
sibling leak at the LRU eviction site that I caught during pre-review:
api/streaming.py SESSION_AGENT_CACHE.popitem(last=False) was discarding
the evicted entry with `evicted_sid, _ = ...`. The agent's _session_db
was dropped on the floor and only released when GC eventually finalized
the agent — which on a long-running server may be never (cyclic refs,
extension types holding C handles, etc.).
Same fix shape as #1421: capture the evicted entry, call
_evicted_agent._session_db.close() explicitly. SessionDB.close() is
idempotent + thread-safe (with self._lock: if self._conn:), so the
double-close-is-benign property still holds.
5 regression tests in test_v050259_sessiondb_fd_leak.py:
- Source-level: cached-agent reuse path closes before replace
- Source-level: LRU eviction path captures + closes evicted agent
- Behavioral: SessionDB.close() is idempotent (3 calls safe)
- Behavioral: cached-agent reuse with mock — close called exactly once
- Behavioral: LRU eviction with mock — only evicted agent's DB closes
Full suite: 3615 passed, 0 failed.
Nathan explicitly authorized 'just go ahead and merge it as a small release'
since the PR is 9 LOC, focused, has Opus pre-release follow-up + tests, and
matches the empirically-confirmed leak shape (73-handle leak at EMFILE).
SessionDB WAL handles leak when streaming.py creates a new SessionDB
instance per request and replaces the cached agent's _session_db without
closing the old one. Each orphaned connection holds 2 FDs (.db +
.db-wal), causing FD exhaustion and EMFILE crashes after ~73 messages.
Fix: close the previous _session_db before replacing it on cached
agents, mirroring the close-before-replace pattern used elsewhere in the
codebase.
Opus pre-release advisor caught a 5th issue not covered by my initial
follow-up sweep, this one CRITICAL: PR #1402#493 per-session toolset
override silently no-op'd every time.
Bug: api/streaming.py:1755 called _session_meta.get('enabled_toolsets') on
the result of Session.load_metadata_only(). It returns a Session INSTANCE,
not a dict. .get() raised AttributeError, which the surrounding bare
except swallowed silently. The toolset chip in the UI saved correctly to
disk, but the streaming agent always ran with global toolsets.
Fix: use getattr(_session_meta, 'enabled_toolsets', None).
Two new regression tests:
- Source-level: forbid the .get() / [] dict-access shape.
- Runtime: Session.load_metadata_only must return a Session instance.
Full suite: 3604 passed, 0 failed.
Persist session model_provider separately from model IDs so active/default provider selections like gpt-5.5 remain bare while routing through OpenAI Codex. Keep @provider:model for picker disambiguation and runtime bridging, and preserve explicit OpenRouter plus custom/proxy base_url routing.
Opus pass-2 review of v0.50.251 caught a critical regression in PR
#1375:
The cancel-partial message stored captured tool calls under the
'tool_calls' key. That key is whitelisted by _API_SAFE_MSG_KEYS so
_sanitize_messages_for_api forwarded the entries to the next-turn
LLM call. But the captured entries use the WebUI internal shape
({name, args, done, duration, is_error}) — they don't have the
OpenAI/Anthropic id + function: {name, arguments} envelope. Strict
providers (OpenAI, Anthropic, Z.AI/GLM) would 400 on the malformed
entries. Net effect: the very cancel-then-continue scenario PR
#1375 aimed to improve becomes a hard fail.
Fix:
- Rename the persisted key to '_partial_tool_calls' (underscore-
prefixed private key NOT in _API_SAFE_MSG_KEYS, so sanitize
correctly strips it).
- Update static/messages.js hasMessageToolMetadata check to also
recognize _partial_tool_calls for UI rendering.
- Update test_issue1361_cancel_data_loss.py assertion to check
_partial_tool_calls (and tool_calls as legacy fallback).
Plus 2 NIT fixes from the same Opus review:
NIT 1 (api/profiles.py:153): re.match → re.fullmatch for consistency
with other _PROFILE_ID_RE callers in the codebase. The trailing-
newline footgun ($ matches before final \n in re.match) is now
closed. Without #1373's is_dir() guard, a name like 'valid\n' would
have created a directory named 'valid\n' on Linux. Doesn't escape
<HERMES_HOME>/profiles/ via Path joining, but unintended.
NIT 2 (test_issue798.py): R19j coverage gaps — added trailing-
newline tests, length-boundary tests (64-char valid, 65-char
rejected), single-char minimum, and non-ASCII / Unicode-trick tests.
New regression test (tests/test_pr1375_partial_tool_calls_sanitize.py):
- test_partial_tool_calls_field_not_forwarded_to_llm: pins that
sanitize-for-API strips _partial_tool_calls + reasoning + does
NOT have tool_calls on a partial message
- test_legitimate_tool_calls_are_preserved_for_completed_turns:
pins that real OpenAI-shape tool_calls on completed turns survive
sanitize unchanged
Tests: 3486 passing (3484 → 3486, +2 sanitize tests).
Three distinct data-loss paths fixed:
§A — Reasoning text was accumulated in a thread-local _reasoning_text
inside _run_agent_streaming. cancel_stream() never saw it because it
went out of scope when the thread was interrupted. Now mirrored to a
new shared dict STREAM_REASONING_TEXT keyed by stream_id, populated
in on_reasoning() and the reasoning branch of on_tool(), read in
cancel_stream().
§B — Live tool calls in thread-local _live_tool_calls were similarly
invisible to cancel_stream(). Now mirrored to STREAM_LIVE_TOOL_CALLS
on tool.started + tool.completed.
§C — Reasoning-only streams produced no partial message because the
thinking-block regex strip returned empty string and the `if _stripped:`
guard skipped the append. Now appends the partial message when EITHER
content text, reasoning trace, OR tool calls exist.
Mirrors the existing STREAM_PARTIAL_TEXT pattern from #893 exactly:
same dict creation in _run_agent_streaming, same _live_config fallback
in cancel_stream, same cleanup in _periodic_checkpoint.
8 regression tests in tests/test_issue1361_cancel_data_loss.py
covering all three sections plus tools+text combinations.
Co-authored-by: bergeouss <bergeouss@users.noreply.github.com>
- api/streaming.py SSE payload now falls back to agent.model_metadata.get_model_context_length when compressor doesn't supply context_length (mirrors the session-save fallback shipped in v0.50.247).
- api/streaming.py also falls back to s.last_prompt_tokens to avoid using the cumulative input_tokens counter.
- static/ui.js tracks rawPct separately from pct and shows '(context exceeded)' tooltip when rawPct > 100 instead of misleading '100% used (0% left)'.
- static/messages.js clears 'Uploading...' composer status after upload completes.
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
* fix(streaming): fallback to model_metadata for context_length when compressor missing (#1318 follow-up)
PR #1318 (shipped in v0.50.246 via PR #1341 + commit a5c10d5) persisted
context_length on the session so the context-ring indicator survives
page reloads. But the writer only fired when agent.context_compressor
was present and reported a non-zero value. Fresh agents, interrupted
streams, or compressors without the attribute would still leave
s.context_length=0 — and the indicator would still show 0% on reload.
This follow-up adds a fallback that calls
agent.model_metadata.get_model_context_length(model, base_url) when the
compressor didn't populate the value. The function returns a sensible
static context window for any known model (with a 256K default for
unknown models). Wrapped in a broad try/except because older
hermes-agent builds may not expose the helper.
Sourced from PR #1344 (@jasonjcwu) — extracted into this focused
follow-up after #1344 was closed as superseded by #1341.
Adds 6 structural tests covering: import + call presence, falsy-gate,
agent.model/base_url passing, exception swallowing, save() ordering,
result assignment.
Closes the data-flow gap in #1318 for the compressor-missing case.
* test: relax pr1341 block-size assertion to accommodate the new fallback
---------
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Pre-release Opus + nesquena review on v0.50.246 caught that PR #1341
added the data-structure scaffolding (Session.__init__ accepts the 3
fields, save() persists them, compact() exposes them, GET /api/session
returns them) but did NOT add the writer that actually populates them.
Without a writer, the user-visible bug (context-ring shows 0% after
page reload) was NOT fixed by #1341 alone — the fields stayed None
forever because nothing wrote to s.context_length anywhere.
Adds the writer at api/streaming.py:2188 (post-merge per-turn save block,
before s.save()) so the values from agent.context_compressor land on
disk and survive page reloads.
Also moves the SSE usage payload comment to clarify that the live SSE
payload and the session-level persistence are now distinct paths
(payload below, persistence above).
Adds tests/test_pr1341_context_window_persistence.py — 6 structural +
round-trip tests covering Session __init__/save/compact, the routes
response, and the streaming.py writer placement.
Closes#1318 (the actual user-visible bug, not just the scaffolding).
Pre-release Opus review on v0.50.246 caught a SHOULD-FIX in PR #1338's
cancel_stream synthesis: the symmetric substring guard
(_pending_user in _last_content OR _last_content in _pending_user) was too
loose. Common confirmation replies ("ok", "yes", "go") in the prior turn
would match longer follow-up prompts ("ok please continue"), the synthesis
would be skipped, and the user's typed text would be lost — exactly the
data-loss bug #1298 was supposed to fix.
The fix: gate the substring check on a timestamp comparison. Only treat
the latest user turn as 'already merged by the streaming thread' if its
timestamp is at or after pending_started_at. Earlier turns whose content
happens to be a substring of the pending must not short-circuit synthesis.
Also drops the symmetric (_last_content in _pending_user) branch — that
direction was the false-positive vector. Keeps the equality and prefix
match (workspace-prefix tolerance from the streaming thread).
Adds tests/test_issue1298_cancel_and_activity.py::
test_cancel_synthesizes_when_prior_turn_content_is_substring_of_pending —
regression for the exact 'ok' → 'ok please continue' scenario.
From PR #1338. Already independently APPROVED by nesquena before being absorbed into v0.50.246.
CHANGELOG entries from this PR were dropped during squash (the v0.50.245 section is already
shipped); they will be re-added under [v0.50.246] in the release commit.
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
The agent cache stores full AIAgent instances (each holding complete
conversation history) without size limit. Long-running servers with
many sessions can accumulate unbounded memory usage.
Changes:
- Replace dict with OrderedDict for LRU tracking
- Add SESSION_AGENT_CACHE_MAX = 50 limit
- Evict least-recently-used entries when cache exceeds limit
- Call move_to_end() on cache hits to maintain LRU order
This prevents memory exhaustion on servers with many active sessions.
* Shorten session sidebar relative time labels
* feat: adaptive session title refresh based on conversation evolution
Addresses #869 — the 'Optional' part: adapt session names to current
conversation context instead of only generating once from the first exchange.
Backend (api/streaming.py):
- Add _latest_exchange_snippets() to extract last user+assistant pair
- Add _count_exchanges() to count user messages
- Add _get_title_refresh_interval() to read the setting
- Add _run_background_title_refresh() — refreshes title from latest exchange
with LLM, skips if title is unchanged or user manually renamed
- Add _maybe_schedule_title_refresh() — checks exchange count and schedules
refresh after stream_end (non-blocking)
Config (api/config.py):
- Add auto_title_refresh_every setting (default '0' = off)
- Enum validation: {'0', '5', '10', '20'}
Frontend:
- Settings UI dropdown (static/index.html)
- Wire up load/save in panels.js
- i18n keys for all 6 locales (en/ru/es/de/zh/zh-Hant)
Default: off. Opt-in via Settings > Conversation > Adaptive title refresh.
* test: add 37 tests for adaptive title refresh helpers
Covers all five new functions introduced in this PR:
_count_exchanges, _latest_exchange_snippets, _get_title_refresh_interval,
_run_background_title_refresh, _maybe_schedule_title_refresh
Co-authored-by: bergeouss <bergeouss@users.noreply.github.com>
* fix(settings): show selected state on theme/skin/font-size picker cards
The CSS rule `#mainSettings .theme-pick-btn { border-color: var(--border) !important }` was
overriding the inline `style.borderColor = "var(--accent)"` set by `_syncThemePicker()` and
siblings — `!important` beats inline styles. Active cards showed no visual highlight.
Fix: move to `.active` CSS class with `border-color:var(--accent)!important` so the active
rule wins over the base rule, and clear the stale inline borderColor/boxShadow from the
sync functions. 5 regression tests added.
Closes#1057
* fix: rename test file to match PR number, fix stale issue reference
* docs: v0.50.211 release notes and version bump
Compact sidebar timestamps, adaptive title refresh (opt-in), settings picker fix.
* docs(changelog): correct settings tab for adaptive title refresh
The v0.50.211 entry for #1058 said "Settings → Appearance" but the
toggle is actually rendered inside settingsPanePreferences (the
Preferences tab) per static/index.html:604+. The commit message also
had the wrong tab ("Conversation"). Updated CHANGELOG to match the
actual UI surface so users can find the toggle.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix: create state dir before writing settings file
save_settings() called SETTINGS_FILE.write_text() without ensuring the
parent directory exists. In fresh environments (CI, first run without
HERMES_WEBUI_STATE_DIR set) this raised FileNotFoundError.
Add mkdir(parents=True, exist_ok=True) before the write.
---------
Co-authored-by: Pavol Biely <biely@webtec.sk>
Co-authored-by: bergeouss <bergeouss@users.noreply.github.com>
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(models): add gpt-5.5 to openai, openai-codex, copilot catalogs
Adds GPT-5.5 and GPT-5.5 Mini entries to the static _PROVIDER_MODELS
catalog so they appear in the model picker for the openai, openai-codex,
and copilot providers.
Signed-off-by: Pix (PiClaw, claude-opus-4-7) via Hermes Agent
* fix(models): add gpt-5.5-mini to copilot provider catalog
* fix(renderer): suppress Mermaid Google Fonts CSP violation via fontFamily inherit (#1044)
Mermaid's built-in 'dark' and 'default' themes inject an @import for
fonts.googleapis.com/Manrope into every generated SVG. The CSP style-src
only allows cdn.jsdelivr.net, so this request is blocked on every diagram
render, filling the console with CSP errors.
Fix: pass fontFamily:'inherit' (and fontSize:'14px') in the themeVariables
block of mermaid.initialize() in renderMermaidBlocks(). This suppresses
Mermaid's external font import and uses the page's existing font stack.
Avoids adding fonts.googleapis.com to the CSP — no new external dependency,
no font FOUT, consistent with the rest of the UI typography.
3 regression tests added in tests/test_1044_mermaid_csp_font.py.
2215/2215 tests passing.
* fix(onboarding): non-standard provider/path cluster (#1029)
* fix(bfcache): restore full layout on tab/session restore — rail, topbar, panels (#1045)
The pageshow handler added for #822 only cleared the session search filter
and re-rendered the session list. This left the rest of the layout chrome
(topbar, rail icons, workspace panel, resize handles, gateway SSE) in the
stale bfcache DOM state, causing a broken layout (oversized search icon,
uninitialized rail) that required a hard refresh to fix.
Fix: extend the pageshow handler to re-run the full set of layout sync calls
that the boot IIFE runs on a fresh page load:
syncTopbar() — restores model chip, title, topbar state
syncWorkspacePanelState() — restores workspace panel open/closed
_initResizePanels() — reattaches panel resize drag listeners
startGatewaySSE() — reconnects the gateway SSE watcher
(bfcache-persisted connections are dead)
All four calls are typeof-guarded for safe degradation if a helper is not
yet defined. The existing #822 fixes (sessionSearch clear +
renderSessionListFromCache) are preserved unchanged.
loadSession() is intentionally NOT re-called — it would cause message
flicker; the sync calls above are sufficient to restore visual state.
7 regression tests added in tests/test_1045_bfcache_layout_restore.py.
2219/2219 tests passing.
* fix(bfcache): also close open dropdowns on bfcache restore (#1045)
Additional symptom noted in issue #1045: bfcache freezes the DOM including
any open dropdown/popover state. The thinking-level selector (and other
composer dropdowns) left open when navigating away would appear open without
user interaction on tab restore.
Extend the pageshow handler to call all four named close functions before
the layout sync:
closeModelDropdown() — composer model selector
closeReasoningDropdown() — thinking/reasoning effort selector
closeWsDropdown() — workspace chip dropdown
closeProfileDropdown() — profile switcher dropdown
All calls are typeof-guarded, matching the style of the layout sync calls
already in the handler.
2 new tests (9 total in test_1045_bfcache_layout_restore.py):
- pageshow closes all four named dropdowns
- dropdown closes appear before layout sync calls (clean state first)
2221/2221 tests passing.
* fix(bfcache): remove _initResizePanels() — bfcache preserves listeners
* fix(bfcache): remove _initResizePanels from pageshow — bfcache preserves listeners; update test
* fix(sessions): use cron job name as session title when available (#1032)
* fix(test): add id column to messages table in cron title test fixture
* fix(merge): inject cron title lookup into read_importable loop, remove stale sqlite3 block
* fix(pwa): redirect to /login client-side on 401 — fixes iOS PWA auth expiry trap (#1038)
When an auth session expires, the server returns a 302→/login for page
requests. In a normal browser this works fine, but in an iOS PWA running
in standalone mode the redirect navigates out of the PWA shell into Safari,
leaving the app permanently stuck on 'Authentication required' with no
recovery path.
Fix: intercept 401 responses client-side before surfacing any error.
- workspace.js api(): check res.status===401 first; call
window.location.href='/login' and return immediately (no throw)
- ui.js: add _redirectIfUnauth() helper; wire into all direct fetch()
calls that bypass api() — api/models, api/models/live, api/upload
All fetch paths that could receive a 401 now redirect cleanly within
the PWA frame rather than opening Safari.
6 regression tests added in tests/test_1038_pwa_auth_redirect.py.
2175/2175 tests passing.
* fix(pwa): preserve current URL in ?next= param on 401 redirect
* fix(test): update 401-redirect assertion to accept ?next= URL format
* feat(pwa): add _safeNextPath() to login.js so ?next= param is honored after re-login
Addresses reviewer suggestion: the ?next= URL set on 401 redirect was ignored by
the login success handler (always redirected to ./). _safeNextPath() validates and
returns the ?next= param with open-redirect guards: rejects non-path-absolute inputs,
// protocol-relative URLs, backslash variants, and control characters.
4 new regression tests added.
* Implement session agent cache for AIAgent reuse
Added session agent cache to reuse AIAgent across messages.
* Implement agent caching for session management
* Implement session agent eviction on session deletion
Added session agent eviction to prevent turn count leakage in recycled sessions.
* docs: v0.50.210 release notes — 7 PRs, 2239 tests (+27)
* docs(changelog): drop stale [Unreleased] entries duplicated by v0.50.210
Three entries in the [Unreleased] section are duplicates of items now
listed under v0.50.210:
- Mermaid CSP font fix (#1044) → v0.50.210 / Mermaid Google Fonts CSP
- bfcache layout restore (#1045) → v0.50.210 / bfcache layout and dropdown restore
- iOS PWA auth redirect (#1038) → v0.50.210 / Login redirects back to original URL
The original drafts landed in [Unreleased] when individual PRs (#1047,
#1048, #1043) were approved; the v0.50.210 release-notes commit then
added the same items under the version section without removing the
[Unreleased] copies. Drop the duplicates so users reading the CHANGELOG
don't see the same fix listed twice.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Signed-off-by: Pix (PiClaw, claude-opus-4-7) via Hermes Agent
Co-authored-by: Pix (Hermes) <aliceisjustplaying@users.noreply.github.com>
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Co-authored-by: qxxaa <mrhanoi@outlook.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix: remove orphaned i18n keys from top-level LOCALES object
Three Traditional Chinese translation keys (cmd_status, memory_saved,
profile_delete_title) were placed outside any locale block between the
en and ru blocks in static/i18n.js. They became top-level properties
of the LOCALES object, causing them to appear as invalid language
options in the Settings > Preferences dropdown.
The correct translations already exist in the zh-Hant locale block.
Fixes#1008
* fix: block stale SSE events from polluting new session's DOM
- appendThinking(): guard with !S.session||!S.activeStreamId to drop
events from a previous session's SSE stream during a session switch
- appendLiveToolCard(): same guard for consistency
- finalizeThinkingCard(): scroll thinking-card-body to top when
scroll is pinned, so completed response is immediately visible
- appendThinking(): auto-scroll thinking card body to bottom while
streaming if user is watching (scroll pinned)
* Fix empty agent sessions in sidebar
* fix: resolve cron UI UX issues — icon ambiguity, toast overlap, running status
Fixes#995 — three sub-issues in the Cron Jobs UI:
1. Dual play icons ambiguous: Resume button now shows a distinct
play+bar icon (play triangle + vertical line) instead of the
identical triangle used by Run now.
2. Toast notification overlapping header buttons: Added
position:relative; z-index:10 to .main-view-header so it
stacks above the fixed toast (z-index:100 within its layer).
3. No running status after trigger: After triggering a job, the
status badge immediately shows 'running…' with a CSS spinner
animation, and polls the cron list every 3s (up to 30s) to
refresh when the job completes.
- Added cron_status_running i18n key in all 5 locales (en, es, de, ru, zh, zh-Hant)
- Added .detail-badge.running CSS class with spinner animation
- New functions: _setCronDetailStatus(), _startCronRunningPoll()
* fix(#1011): address review feedback — poll cleanup, badge persistence, 30s fallback
- _clearCronDetail() now clears _cronRunningPoll interval on navigation
- Poll re-applies 'running' badge after loadCrons() re-render (prevents flicker)
- When poll ends (30s max), detail re-renders with actual status as fallback
* feat: create folder and add space directly from UI (#782)
- After creating a folder via the file tree New folder button, offer to add it as a space via confirm dialog
- Add Create folder if it doesnt exist checkbox in the New Space form
- Backend: support create flag in /api/workspaces/add to mkdir before validation
- i18n: 4 new keys (folder_add_as_space_title/msg/btn, workspace_auto_create_folder) in all 6 locales
* fix: validate workspace path before mkdir to prevent orphan directories
Review feedback (critical): the previous code called mkdir() before
validate_workspace_to_add(), which meant a rejected path (e.g. system dir)
would leave an orphan directory on disk.
New flow:
1. Resolve path and check against blocked system roots BEFORE any mutation
2. mkdir() only if path passes the blocklist check
3. Full validation (exists, is_dir) after mkdir
Also imports _workspace_blocked_roots for the pre-mutation blocklist check.
* fix(#1014): classify model-not-found errors with helpful message
- Add model_not_found error type to streaming.py exception classifier
- Detect 404, 'not found', 'does not exist', 'invalid model' patterns
- Strip HTML tags from provider error messages (nginx 404 pages, etc.)
- Add model_not_found branch to apperror handler in messages.js
- Add i18n key model_not_found_label in all 6 locales
- 15 tests covering detection, sanitization, frontend, and i18n
* feat(ui): add live TPS stat to header
Adds a TPS (Tokens Per Second) chip to the right of the header title bar
that updates live while AI output is streaming.
Metering (api/metering.py)
- Tracks per-session output + reasoning tokens via GlobalMeter singleton
- Per-session TPS = total_tokens / elapsed_time
- Global TPS = average of active sessions' TPS values
- HIGH/LOW are max/min of global_tps snapshots over a 60-minute rolling
window (only recorded when > 0, so idle periods are excluded)
- Thread-safe with a single lock
Metering events emitted from streaming.py
- Throttled at 100ms from token/reasoning/tool callbacks so the display
updates rapidly during fast token streams
- 1Hz ticker as fallback for slow streams (exits when no active sessions)
- Final stats emitted on stream end
Routes (api/routes.py)
- Removed POST /api/metering/interval endpoint (dynamic interval via
focus/blur was replaced with simple always-1s-when-active approach)
UI (static/messages.js, index.html, style.css)
- TPS chip in titlebar: shows 'N.N t/s . N.N high . N.N low'
- Default: '0.0 t/s . 0.0 high' when idle
- Display updates on every metering SSE event (throttled to 100ms)
* feat: session restore speed + title gen reasoning hardening (#1025, #1026)
PR #1025 (@franksong2702): Speed up large session restore paths
- GET /api/session?messages=0 now parses only metadata before the messages array
- Metadata-only loads no longer populate the full-session LRU cache
- Frontend lazy fetch uses resolve_model=0 to avoid cold model-catalog lookup
- Hard reload no longer waits for populateModelDropdown() before restoring session
PR #1026 (@franksong2702): Harden auto title generation for reasoning models
- Raises title-gen completion budget to 512 tokens (reasoning-safe)
- Retries once with 1024 tokens on empty content / finish_reason:length
- Applies retry to both auxiliary and active-agent fallback routes
- Preserves underlying failure reason in title_status on local fallback
Co-authored-by: Frank Song <franksong2702@gmail.com>
* feat: session attention indicators in right slot + last_message_at timestamps (#1024)
PR #1024 (@franksong2702): Polish session attention indicators
- Streaming spinners and unread dots now reuse the right-side actions slot
- Running/unread rows hide timestamps; idle/read rows keep right-aligned timestamps
- Date group carets point down when expanded, right when collapsed
- Pinned group no longer repeats pinned-star icon per row
- Running indicators appear immediately after send (local busy state while /api/sessions catches up)
- Sidebar sorting/grouping/timestamps now prefer last_message_at (derived from last real message)
so metadata-only saves don't make old sessions appear under Today
Co-authored-by: Frank Song <franksong2702@gmail.com>
* docs: v0.50.207 release notes — 10 PRs, 2169 tests (+36)
---------
Co-authored-by: bergeouss <bergeouss@users.noreply.github.com>
Co-authored-by: Josh <josh@fyul.link>
Co-authored-by: Frank Song <franksong2702@gmail.com>
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
- Backend: save session JSON with metadata fields before messages array
so load_metadata_only() reads only ~1KB without parsing the full session
- Backend: add GET /api/session?messages=0 for metadata-only responses
(~1KB vs ~400KB), enabling instant sidebar switching
- Backend: add POST /api/admin/reload to hot-reload models without restart
- Backend: gzip compress JSON API responses (>1KB) for 70-80% bandwidth reduction
- Frontend: show Loading indicator immediately on session switch, replacing
old DOM before API call to prevent stale content flash
- Frontend: clear S.messages before API call so _ensureMessagesLoaded
always fetches fresh data for the target session
- Frontend: wrap both Phase 1 (messages=0) and Phase 2 (_ensureMessagesLoaded)
in try/catch to prevent permanently stuck loading state on network/server errors
Handle DeepSeek DSML variants including truncated and spaced tag forms, and sanitize thinking-card text so leaked XML fragments never render. Add regression tests for DSML edge cases and thinking-card sanitization.
Made-with: Cursor
Co-authored-by: bsgdigital <bsg@bsgdigital.com>
Rebased onto master after #931 (aux title routing) to resolve streaming.py conflict.
All changes from both PRs are cleanly integrated.
2088 tests passing (2065 master + 23 from #931).
Co-authored-by: bergeouss <bergeouss@gmail.com>
- _aux_title_configured(): returns True when provider/model/base_url is set
- _aux_title_timeout(): reads configured timeout, falls back to 15.0s default
- _generate_llm_session_title_via_aux: use_agent_model kwarg preserves old behavior
- Missing llm_invalid_aux fallback now triggers agent-model retry
- 23 new tests in tests/test_title_aux_routing.py — all pass
Co-authored-by: starship-s <starship-s@users.noreply.github.com>
* fix(cancel): preserve partial streamed response on Stop Generation (#893)
* docs(cancel): fix misleading comment — partial message is NOT _error=True
The outer comment block claimed `_error=True so _sanitize_messages_for_api()
strips it from future conversation history`, but the actual append call
sets only `_partial=True` (correctly matching the inner comment six lines
below and the PR description). Updated the outer comment to match reality
so a future reader doesn't try to "fix" the code to match the wrong comment.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
fix: correct message ordering after task cancellation — v0.50.163 (#883)
Fixes the message-ordering glitch from #882: clicking Cancel while the
agent is responding could cause a subsequent response to render above
the "*Task cancelled.*" marker.
Root cause: the cancel handler pushed the marker only to local S.messages
without persisting to the server. When the done event fired shortly after
and replaced S.messages from server state, the marker disappeared from
client state while the next response anchored to the server-authoritative
position.
Fix has three parts:
- Server (cancel_stream): append *Task cancelled.* to session.messages
with _error:True + timestamp, then save. _error ensures
_sanitize_messages_for_api() strips it from conversation_history on
the next agent turn, so the LLM never sees it as a prior assistant
turn. Precedent: same flag used for the apperror marker at line 1343.
- Client (SSE cancel handler): fetch /api/session instead of pushing
locally (same pattern as the done handler). Falls back to local push
if the fetch fails.
- Tests: fix test window width for cancel handler (1200→dynamic); add
two regression tests pinning _error flag and _sanitize invariant.
1941 tests passing.
Co-authored-by: piliang <piliang1@jd.com>
Pass gateway_session_key=session_id to AIAgent from streaming.py so Honcho per-session strategy pins to stable WebUI session ID rather than creating a new Honcho session each turn.
Closes#461
Adds full /reasoning CLI parity to the WebUI slash command system:
- /reasoning show|on → window._showThinking = true; writes display.show_reasoning to config.yaml (same key as CLI); mirrors to settings.json for boot.js
- /reasoning hide|off → same in reverse; re-renders immediately
- /reasoning none|minimal|low|medium|high|xhigh → POST /api/reasoning → writes agent.reasoning_effort to config.yaml; takes effect next turn (matching CLI semantics)
- /reasoning (no args) → GET /api/reasoning → live status toast from config.yaml
- Autocomplete shows all 8 options: show|hide|none|minimal|low|medium|high|xhigh
- Profile-isolated: _get_config_path() is thread-local so per-profile settings never bleed across
- Boot hydration: window._showThinking initialised from settings.json show_thinking on page load
- Inspect.signature guard in streaming.py so older hermes-agent builds don't TypeError
28 new tests, 1708/1708 total passing. Full browser QA on port 8789 with isolated state. CLI/config.yaml sync verified with hermes_constants.parse_reasoning_effort().
Fixes the multi-client profile isolation bug (#798).
- get_hermes_home_for_profile(): pure path resolver, validates name against
_PROFILE_ID_RE (rejects path traversal), never mutates os.environ or globals
- new_session() accepts explicit profile= param from POST body (S.activeProfile),
short-circuits the process-level _active_profile global
- streaming handler resolves HERMES_HOME from s.profile instead of the global
- sessions.js sends profile: S.activeProfile in every new-session POST
10 tests in tests/test_issue798.py including concurrency and traversal coverage.
Co-authored-by: nesquena <nesquena@users.noreply.github.com>
cancel_stream() now pops STREAMS/CANCEL_FLAGS/AGENT_INSTANCES and clears session.active_stream_id immediately after signalling cancel. Fixes sessions permanently stuck at 409 when the agent thread is blocked in a bad tool call. Session cleanup runs outside STREAMS_LOCK to preserve lock ordering.
Fixes#653
Co-authored-by: bergeouss <bergeouss@users.noreply.github.com>
- quota_exhausted error type: distinguishes credit exhaustion from rate limits
- Streaming errors persisted to session file so they survive page reload
- _error flag excludes persisted errors from subsequent LLM API calls
- stream_end and title SSE events use original session_id (not s.session_id which rotates during context compaction)
Fixes#739, #652, #653
Squash merge of PR #717 — rebased on behalf of @franksong2702.
## What it does
Fixes#680. Footer chrome (timestamps, copy, edit, regenerate) is now hover-only for both user and assistant message rows, consistent throughout the conversation. The last assistant turn keeps cumulative usage visible at rest; timestamp and actions are revealed inline on hover in the same row.
Key changes:
- `static/ui.js`: new `_formatMessageFooterTimestamp()` (local timezone, cross-day fuller format); `timeHtml` no longer gated to user-only; last assistant usage moved from separate `.msg-usage` div to inline `.msg-usage-inline` span in the footer
- `static/style.css`: `.msg-foot-with-usage` class + rules; assistant footer opacity changed from 0.45 to 0 (hover-only); `:focus-within` alongside `:hover` for keyboard users
- `api/streaming.py`: `_restore_reasoning_metadata()` now preserves `_ts`/`timestamp` for unchanged historical messages
- `tests/test_sprint49.py`: 8 new tests covering rendering contract, hover CSS, timestamp preservation
Tests: 1518 passed. QA: 20/20. Browser verified. Reviewed and approved by @nesquena and @aronprins.
Strips <function_calls> XML from assistant messages before rendering, adds workspace file panel empty-state messages, and changes notification description from 'tab' to 'app'. 16 new tests. Fixes#702, #703, #704.
Fixes <|turn|>thinking delimiter (was wrong as <|turn>thinking) in api/streaming.py, static/messages.js, and static/ui.js. Adds 13 regression tests. Independent review by @nesquena.
Squash-merges PR #611 (@franksong2702). Fixes two edge cases in auto-generated session titles.
1. Strip Markdown labels (`**Session Title:**`, `Title:`) from sanitizer output — these were being persisted verbatim when the LLM emitted them.
2. Skip empty assistant tool-call placeholder messages when extracting the first exchange for title generation — previously the empty row could be latched onto instead of the first real answer.
Also tightens the title prompt to explicitly forbid Markdown, bullets, and label prefixes.
1371 tests passing, QA harness green.
Co-authored-by: Frank Song <franksong2702@gmail.com>
Forwards `api_mode`, `acp_command`, `acp_args`, and `credential_pool` from the resolved runtime provider into `AIAgent.__init__()` in the WebUI streaming path. Fixes Codex account switching and credential pool support for WebUI sessions. Also adds 6 defensive variable initializations to prevent NameError in cleanup paths.
Tests: 1329 passed, 0 skipped. Full TestRuntimeRouteInjection suite passes.
PR by @suinia. Rebased and CHANGELOG added by maintainer.
Co-authored-by: suinia <suinia@users.noreply.github.com>
Squash-merges PR #578 (rebased from #574 by @renheqiang + #575 by @nesquena-hermes). MCP server toolsets now included in WebUI sessions; onboarding wizard no longer fires for non-standard providers. 1331 tests pass. Nathan override applied for self-built #575.
BUG-1 (CRITICAL): messages.js line 522 — mismatched quote in
setComposerStatus('Reconnecting…') caused JS syntax error on the
reconnect path.
BUG-2 (HIGH): messages.js line 491 — broken template literal
'\\n\\n*{d.hint}*' restored to '\n\n*${d.hint}*'. Error hint
text was non-functional (missing $ prefix and escaped newlines).
BUG-3 (HIGH): messages.js — showApprovalCard(pending, pendingCount),
_approvalCurrentId, and approval_id in respondApproval() were removed,
regressing the simultaneous approval queue fix from PR #546. Restored
all three, including the '1 of N pending' counter and poll passthrough.
BUG-4 (LOW): api/streaming.py — MiniMax thinking delimiter regex
missing closing pipe: <|channel> -> <|channel|> in both
_strip_thinking_markup() and _looks_invalid_generated_title().
ALSO: test_issue487b.py docstring changed to raw string to fix
DeprecationWarning for invalid escape sequence '\s'.
Extends _sanitize_messages_for_api() with a two-pass approach:
1. Collect all tool_call_ids declared in assistant messages (handles
both OpenAI 'id' and Anthropic 'call_id' field names).
2. Drop any tool-role messages whose tool_call_id was not declared
by a preceding assistant message.
Strictly-conformant providers (Mercury-2/Inception, newer OpenAI
models) reject histories with orphaned tool results with a 400 error:
'Message has tool role, but there was no previous assistant message
with a tool call.' This can happen when histories are edited, when
switching between providers, or when partial messages are stored.
Adds 13 regression tests covering: valid roundtrip preservation,
multiple tool calls, partial orphan filtering, Anthropic call_id,
edge cases (None tool_calls, missing tool_call_id, non-dict entries).
Removes /root from _BLOCKED_SYSTEM_ROOTS in api/workspace.py, allowing
Hermes running as root (e.g. Docker, VPS) to use /root as a workspace
without a 'system directory' rejection.
Fixes a fragile string split in api/streaming.py: base_text extraction
now guards against msg_text that contains no '[Attached files:' marker,
preventing the split from producing empty-string on those messages.
Fixes: #510, partial fix from #521 (workspace + split guard only).
Co-authored-by: ccqqlo <ccqqlo@users.noreply.github.com>
* fix: preserve live session output across chat switches
(cherry picked from commit 401e3b643d25e8dad8c06883b478b3c3073f07a5)
* fix: preserve todo state after session reload
(cherry picked from commit 7ee093ba19978af23b79148df2f2347e2f1e5bde)
* fix: preserve live assistant anchor across rerenders
* fix: stream live reasoning and tool progress
* fix: recover inflight session state after reload
* fix: add loadInflightState stub + CHANGELOG v0.50.21
- static/ui.js: add loadInflightState() function (currently returns null —
the typeof guard in sessions.js means reload recovery works via the
else-path attachLiveStream call; this stub satisfies the guard cleanly
and documents the extension point for future localStorage-backed state)
- CHANGELOG.md: v0.50.21 entry; 960 tests (up from 949)
---------
Co-authored-by: Jordan SkyLF <jordan@skylinkfiber.net>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: silent errors, stale models, live model fetching (#373, #374, #375)
- api/streaming.py: detect empty agent response (_assistant_added check),
emit apperror(type='no_response' or 'auth_mismatch') instead of silent done
- api/streaming.py: add _token_sent flag so guard works for streaming agents
- static/messages.js: done handler belt-and-suspenders guard for zero replies
- static/messages.js: apperror handler labels 'no_response' type distinctly
- api/config.py: remove gpt-4o and o3 from _FALLBACK_MODELS and
_PROVIDER_MODELS['openai'] (superseded by gpt-5.4-mini and o4-mini)
- api/routes.py: new /api/models/live?provider= endpoint, fetches /v1/models
from provider API with B310 scheme check + SSRF guard
- static/ui.js: _fetchLiveModels() background fetch after static list loads,
appends new models to dropdown, caches per session, skips unsupported providers
Other:
- tests/test_issues_373_374_375.py: 25 new structural tests
- tests/test_regressions.py: extend done-handler window 1500->2500 chars
- CHANGELOG.md: v0.50.19 entry; 947 tests (up from 922)
* fix: SSRF hostname bypass + auth detection operator precedence
1. routes.py: SSRF guard used substring matching (any(k in hostname))
which allows bypass via hostnames like evil-ollama.attacker.com.
Changed to exact hostname matching against a fixed set of known
local hostnames (localhost, 127.0.0.1, 0.0.0.0, ::1).
2. streaming.py: _is_auth detection had a Python operator precedence
bug on the ternary expression. The line:
'AuthenticationError' in type(...).__name__ if _last_err else False
parsed as the ternary absorbing the rest of the or-chain when
_last_err was falsy. Fixed to: (_last_err and 'AuthenticationError' in ...)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* docs: fix v0.50.20 CHANGELOG version number and test count (949 tests)
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: inject SessionDB into AIAgent for WebUI sessions
session_search tool requires a SessionDB instance passed via the
session_db parameter. The CLI and gateway paths already do this,
but the WebUI streaming path was missing it, causing every
session_search call to return 'Session database not available'.
Initialize SessionDB before creating the AIAgent and pass it through.
Failure is non-fatal — a warning is printed and session_search
gracefully degrades.
* fix: inject SessionDB into AIAgent for WebUI sessions (enables session_search) (#356)
- api/streaming.py: initialize SessionDB() before AIAgent construction and
pass session_db= kwarg so session_search works in WebUI sessions
- tests/test_sprint42.py: 7 new tests covering SessionDB injection, try/except
guard, WARNING log, ordering, and AST lock-safety check
- CHANGELOG.md: v0.50.13 entry; 822 tests total (up from 815)
---------
Co-authored-by: 王昌旭 <wangchangxu@xiaohongshu.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
PR #301 changes:
- api/streaming.py: guard title_from() with s.title == 'Untitled' check
- api/routes.py: same guard in sync/non-streaming path
PR #302 changes (cleaned — restores accidentally-removed features):
- static/boot.js: PANEL_MAX 500 -> 1200
- static/boot.js: clearPreview() calls renderBreadcrumb() to restore dir view
- static/style.css: responsive .messages-inner breakpoints (1400px/1800px)
- static/workspace.js: renderFileBreadcrumb() function with clickable segments
- static/workspace.js: openFile() calls renderFileBreadcrumb(path)
12 new tests in tests/test_sprint35.py
Note: PR #302 branch contained several accidental regressions (removed app-dialog
system, onboarding CSS, _checkProviderMismatch, closeMobileFiles, etc.) that were
not part of its stated scope. This clean branch applies only the three intended
features on top of current master.
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: warn on provider/model mismatch, surface auth errors (#266)
Fixes#266 — WebUI silently ignores provider/model selection mismatch.
The problem: selecting an OpenRouter (or Anthropic/OpenAI) model while
Hermes is configured for a different provider (e.g. local Ollama) sends
the request to the wrong endpoint, which returns a 401 Unauthorized error
with no UI indication of why.
Three-layer fix:
1. api/streaming.py — detect 401/auth errors explicitly
Added is_auth_error detection covering '401', 'AuthenticationError',
'authentication', 'unauthorized', 'invalid api key', and the specific
Ollama error string 'no cookie auth credentials'. Auth errors emit
apperror with type='auth_mismatch' and a hint pointing to 'hermes model'.
2. static/ui.js — expose active_provider and warn on selection
- populateModelDropdown() stores data.active_provider from /api/models
as window._activeProvider (the field was already in the response but
the frontend never used it)
- New _checkProviderMismatch(modelId) helper: compares the selected
model's slash-prefix (e.g. 'openai/' from 'openai/gpt-4o') against
the active provider. Skips the check for 'openrouter' and 'custom'
to avoid false positives on configs that legitimately route any model.
3. static/boot.js — warn on model dropdown change
modelSelect.onchange calls _checkProviderMismatch() and shows a toast
when the selected model looks incompatible with the configured provider.
4. static/messages.js — distinct UI label for auth errors
apperror handler now distinguishes type='auth_mismatch' and shows
'Provider mismatch' as the error label instead of 'Error'.
5. static/i18n.js — provider_mismatch_warning and provider_mismatch_label
keys added to all 5 locales (en, es, de, zh-Hans, zh-Hant).
Tests: 21 new tests in tests/test_provider_mismatch.py covering all
five change areas. 679/679 total pass (658 baseline + 21 new).
* fix: t() call args spread + use i18n label for auth mismatch
1. ui.js: _checkProviderMismatch passed [modelId, ap] as a single
array arg to t(). Since t(key, ...args) spreads, the function
received the array as m and undefined as p. Fixed to pass as
separate args: t('provider_mismatch_warning', modelId, ap).
2. messages.js: 'Provider mismatch' label was hardcoded instead of
using t('provider_mismatch_label'). Now uses the i18n key with
fallback for when t() isn't available.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: decode HTML entities before markdown processing + zh/zh-Hant translations (#239)
Adds decode() helper in renderMd() to fix double-escaping of HTML entities
from LLM output (e.g. <code> becoming &lt;code&gt; instead
of rendering). XSS-safe: decode runs before esc(), only 5 entity patterns.
Also adds 40+ missing zh (Simplified Chinese) translation keys and a new
zh-Hant (Traditional Chinese) locale with 163 keys.
Fix applied: removed duplicate settings_label_notifications key in both
zh and zh-Hant locales.
Fixes#240
* fix: restore custom model list discovery with config api key (#238)
get_available_models() now reads api_key from config.yaml before env vars:
1. model.api_key
2. providers.<active>.api_key / providers.custom.api_key
3. env var fallbacks (HERMES_API_KEY, OPENAI_API_KEY, etc.)
Also adds OpenAI/Python User-Agent header and a regression test covering
authenticated /v1/models discovery.
Fixes users with LM Studio / Ollama custom endpoints configured in
config.yaml whose model picker silently collapsed to the default model.
* feat: Docker UID/GID matching to avoid root-owned .hermes files (#237)
Adds docker_init.bash with hermeswebuitoo/hermeswebui user pattern so
container files match the host user UID/GID. Prevents .hermes volume
mounts from being owned by root when using a non-root host user.
Configure via WANTED_UID and WANTED_GID env vars (default 1000/1000).
Readme updated with setup instructions.
Fix applied: removed duplicate WANTED_GID=1000 line in docker-compose.yml
that was overriding the ${GID:-1000} variable expansion.
* security: redact credentials from API responses and fix credential file permissions (#243)
Adds response-layer credential redaction to three endpoints:
- GET /api/session — messages[], tool_calls[], and title
- GET /api/session/export — download also redacted
- SSE done event — session payload in stream
- GET /api/memory — MEMORY.md and USER.md content
Adds api/startup.py with fix_credential_permissions() at server startup.
Adds 13 tests in tests/test_security_redaction.py.
Merged with #237 container detection changes in server.py.
* fix: cancel button now interrupts agent and cleans up UI state (#244)
Wires agent.interrupt() into cancel_stream() so the backend actually
stops tool execution when the user clicks Cancel, rather than only
stopping the SSE stream while the agent keeps running.
Changes:
- api/config.py: adds AGENT_INSTANCES dict (stream_id -> AIAgent)
- api/streaming.py: stores agent in AGENT_INSTANCES after creation,
checks CANCEL_FLAGS immediately after store (race condition fix),
calls agent.interrupt() in cancel_stream(), cleans up in finally block
- static/boot.js: removes stale setStatus(cancelling) call
- static/messages.js: setBusy(false)/setStatus('') unconditionally on cancel
Race condition fix: after storing agent in AGENT_INSTANCES, immediately
checks if CANCEL_FLAGS[stream_id] is already set (cancel arrived during
agent init) and interrupts before starting. Check is inside the same
STREAMS_LOCK acquisition, making it atomic.
New test file: tests/test_cancel_interrupt.py with 6 unit tests.
* docs: v0.46.0 release notes, bump version, update test counts
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: surface approval prompt in UI instead of getting stuck in Thinking
When a dangerous command was detected during streaming, the approval system
would call submit_pending() but no SSE 'approval' event would be emitted to
the frontend. The agent thread either blocked indefinitely (gateway path) or
returned an approval_required status the UI never saw (EXEC_ASK path). Either
way the chat UI stayed stuck in 'Thinking...' with no prompt shown.
Root cause: streaming.py used HERMES_EXEC_ASK=1 but never registered a
register_gateway_notify() callback. Without it, check_all_command_guards()
fell back to the legacy polling path (submit_pending only), which relies on
on_tool() polling -- but on_tool() fires *before* the tool runs, so by the
time the terminal tool detected the dangerous command and called submit_pending,
the approval event had already missed its window.
Fix (streaming.py):
- Register a gateway-style notify_cb via register_gateway_notify() before the
agent runs. The callback calls put('approval', ...) to emit the SSE event
the moment a dangerous command is detected, regardless of on_tool() timing.
- Unregister via unregister_gateway_notify() in the finally block to unblock
any threads still waiting if the stream ends or is cancelled mid-approval.
- Keep the on_tool() fallback poll for older approval module versions.
Fix (routes.py):
- Import and call resolve_gateway_approval() in _handle_approval_respond().
This unblocks the agent thread parked in entry.event.wait() when the user
clicks Allow or Deny in the UI. Without this call the thread would block
until the 5-minute gateway timeout.
Tests (tests/test_approval_unblock.py):
- 16 new tests covering: resolve_gateway_approval() event signalling, deny/
session/once choices, resolve_all, notify_cb registration/firing/cleanup,
unregister signals blocked entries, full end-to-end streaming simulation,
module symbol exports, and HTTP endpoint regressions.
515 tests pass (499 existing + 16 new).
* feat: full approval UI — i18n buttons, keyboard shortcut, loading state, scoping fix
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
The v0.39.0 security sprint introduced _ENV_LOCK to protect env var
mutations in the streaming path. The implementation held the lock for
the entire agent run (potentially minutes), then tried to re-acquire
it in the finally block — a guaranteed deadlock on any non-reentrant
threading.Lock().
Result: first message completes (done event fires before finally hits),
but the lock is never released. Every subsequent chat/start POST blocks
forever waiting for that lock.
Fix: narrow the lock scope to just the env mutation. Set the vars inside
the with block, then let the lock release before the agent starts. The
finally block re-acquires cleanly since it no longer re-enters an
already-held lock.
No logic change — only the critical section boundary moves.
* fix: sync message_count to state.db for /insights (#163)
sync_session_usage() didn't write message_count to state.db, so
/insights showed 0 messages for all WebUI sessions even with
sync_to_insights enabled.
Added message_count parameter to sync_session_usage() and pass
len(s.messages) from both the streaming and non-streaming chat paths.
Fixes#163
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use callable pattern for _execute_write in sync_session_usage
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The previous implementation read SOUL.md files from a filesystem directory.
The Hermes agent uses config.yaml agent.personalities section with string
or dict format (system_prompt, tone, style), resolved via
_resolve_personality_prompt() and passed to AIAgent via
ephemeral_system_prompt.
Changes:
- /api/personalities: reads from config.yaml agent.personalities, not
filesystem SOUL.md directories. Calls reload_config() to pick up
config changes without restart.
- /api/personality/set: resolves prompt from config.yaml using the same
logic as hermes-agent cli.py (string or dict with system_prompt/tone/style)
- streaming.py: passes personality via agent.ephemeral_system_prompt
(agent's own mechanism) instead of prepending to system_message
- Removed unused 're' import from streaming.py
- Updated tests to match config-based approach
Fixes#139
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The previous fix (#142) prefixed non-default provider models with
'provider/model' which then hit the cross-provider guard and routed
to OpenRouter — worse than before for users without an OpenRouter key.
New approach: non-default provider models use '@provider:model' format
(e.g. @minimax:MiniMax-M2.7). resolve_model_provider() parses this
hint and returns (bare_model, provider, None). streaming.py and
routes.py then pass the resolved provider to
resolve_runtime_provider(requested=provider) which gets the correct
per-provider API key and base_url from hermes-agent.
This uses the agent's own credential resolution instead of reinventing
routing logic in the webui.
Fixes#138
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: /personality slash command with backend integration
Add /personality command to switch the agent's system prompt personality.
Hermes CLI supports personalities stored at ~/.hermes/personalities/<name>/SOUL.md.
Backend:
- GET /api/personalities: lists available personalities from the active
profile's personalities directory (reads first line of SOUL.md for desc)
- POST /api/personality/set: sets active personality on the session, reads
and validates the SOUL.md file exists, returns the prompt text
- streaming.py: injects personality prompt (SOUL.md content) as prefix to
the system_message when run_conversation is called
Frontend (commands.js):
- /personality with no args: lists available personalities as a local message
- /personality <name>: sets the personality with a toast confirmation
- /personality none|default|clear: removes the active personality
Session model: new 'personality' field (backward-compatible, defaults to None)
Closes#139
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: path traversal in personality name + case sensitivity
Security: personality name is now validated with regex ^[a-zA-Z0-9][a-zA-Z0-9_-]{0,63}$
in both routes.py (POST /api/personality/set) and streaming.py (system
prompt injection). Defense-in-depth: resolve().relative_to() check ensures
the path stays inside the personalities directory even if regex is bypassed.
Also: removed toLowerCase() from frontend command handler so personality
names are case-preserved (filesystem may be case-sensitive).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: /personality command — hardened, compact() fix, tests
Fixes on top of original PR:
- compact() was missing 'personality' field — UI couldn't know active
personality after page load. Added to Session.compact().
- GET /api/personalities: add symlink guard (is_symlink() skip) and
resolve() check — prevents reading SOUL.md from symlink targets
outside personalities dir.
- POST /api/personality/set: require() only checks session_id (not name)
so clearing with name='' works correctly instead of 400.
- POST /api/personality/set: add MAX_FILE_BYTES size cap on SOUL.md to
prevent unbounded context window consumption.
- POST /api/personality/set: return personality:null (not '') when cleared.
- streaming.py: same MAX_FILE_BYTES guard before prepending to system msg.
Added tests/test_sprint28.py: 11 tests for API round-trip, listing,
symlink guard, path traversal rejection, clear, size cap, persistence.
Tests pass in isolation; full-suite run has a test-isolation interaction
with shared server state across sprint tests (tracked as follow-up).
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* security: fix four audit findings -- env race, signing key, upload traversal, password hash
1. Race condition in os.environ (HIGH): Per-session _agent_lock didn't
prevent cross-session env writes from interleaving. Added global
_ENV_LOCK in streaming.py that serializes the entire env save/restore
block across all sessions.
2. Predictable signing key (MEDIUM): sha256(STATE_DIR) was deterministic.
Now generates a random 32-byte key on first startup and persists it to
STATE_DIR/.signing_key (chmod 600). Existing sessions invalidated on
first restart (acceptable for a security fix).
3. Upload path traversal (MEDIUM): Filename '..' survived the regex
sanitization (dots are allowed chars). Added explicit rejection of
dot-only names and safe_resolve_ws() check to verify the resolved
path stays within the workspace.
4. Weak password hashing (MEDIUM): Replaced bare SHA-256 with PBKDF2-
SHA256 (600k iterations per OWASP). Uses stdlib hashlib.pbkdf2_hmac,
no new dependencies. Note: existing passwords must be re-set after
this change (hash format changed).
Closes#106
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use random signing key as PBKDF2 salt (replaces predictable STATE_DIR salt)
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
WebUI sessions were invisible to 'hermes /insights' because the WebUI
bypasses the gateway and calls AIAgent.run_conversation() directly,
never writing to state.db.
New 'Sync usage to /insights' setting (default: off) that mirrors
WebUI session metadata (tokens, cost, model, title) into state.db
after each turn. Uses absolute token counts to avoid double-counting.
Components:
- api/state_sync.py: bridge module with sync_session_start() and
sync_session_usage(). Uses ensure_session() (idempotent) and
update_token_counts(absolute=True). All wrapped in try/except.
- api/config.py: new 'sync_to_insights' boolean setting
- api/streaming.py: calls sync_session_usage() after s.save()
- api/routes.py: same for the non-streaming chat path
- Settings UI: checkbox toggle with description
Default off because:
- Writing to state.db while CLI/gateway also writes could cause
WAL lock contention on busy systems
- Some users may not want WebUI sessions in /insights stats
Closes#92
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The context indicator in the composer footer now shows real data from
the agent's context compressor instead of hardcoded estimates:
- last_prompt_tokens / context_length (e.g. '12.4k / 200k (6%)')
- Bar color: blue <50%, yellow 50-75%, red >75%
- Hover tooltip shows exact numbers + compression threshold
- Cost appended when available
Backend: streaming.py now reads context_length, threshold_tokens, and
last_prompt_tokens from agent.context_compressor after run_conversation()
and includes them in the usage dict sent with the 'done' SSE event.
This matches the CLI's context window display (the bar that shows
current context vs total window).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The agent's run_conversation() already triggers context compression
internally, but the WebUI was unaware of the side effects:
1. Session ID rotation: compression creates a new session_id inside
the agent. The WebUI kept writing to the old session file, causing
silent data loss. Fix: detect agent.session_id mismatch after
run_conversation(), rename the session file, and update in-memory
caches.
2. No user notification: compression was invisible. Fix: emit a
'compressed' SSE event when compression is detected. Frontend shows
a system message and toast.
3. No manual control: Fix: add /compact slash command that sends a
message to the agent requesting context compression. Shows in the
autocomplete dropdown.
Detection works two ways:
- agent.session_id != original session_id (ID rotation)
- agent.context_compressor.compression_count > 0 (compressor state)
Closes#90
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When the user's config uses a non-Anthropic provider with an
Anthropic-compatible endpoint (e.g. MiniMax at
https://api.minimax.io/anthropic), chat in the WebUI fails silently
with APIConnectionError on every request, while the hermes CLI and
messaging gateway work fine with the same config.
Root cause: both api/routes.py and api/streaming.py constructed
AIAgent using only (model, provider, base_url) from
resolve_model_provider() and never passed api_key. When the base URL
ends in /anthropic, AIAgent uses the anthropic_messages adapter, but
only falls back to ANTHROPIC_TOKEN when provider == "anthropic" (a
safety check to avoid leaking Anthropic credentials to third parties).
For MiniMax and similar providers the effective key becomes "", and
the auth failure surfaces as a generic "Connection error" after three
retries.
The CLI and gateway resolve the key via
hermes_cli.runtime_provider.resolve_runtime_provider(), which reads
MINIMAX_API_KEY (and similar) from ~/.hermes/.env. This patch does the
same before creating the AIAgent in both chat paths.
Fixes#77
The webui stores display-only fields on messages (attachments, timestamp,
_ts) for UI rendering. These leaked into the conversation_history passed
to AIAgent.run_conversation(). Most providers ignore unknown fields, but
Z.AI/GLM tries to deserialize 'attachments' as its native ChatAttachments
type, causing HTTP 400 on every subsequent message after an image upload.
Fix: _sanitize_messages_for_api() creates a clean copy with only
API-standard keys (role, content, tool_calls, tool_call_id, name,
refusal) before passing to run_conversation(). Applied to both the
streaming path (streaming.py) and non-streaming path (routes.py).
Closes#66
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Token usage display:
- Add 'show_token_usage' boolean to settings (default: false, off by default)
- Settings panel: checkbox 'Show token usage after responses'
- /usage slash command: instant toggle with toast feedback, persists to
server, updates checkbox if settings panel is open, re-renders messages
- Boot: load show_token_usage alongside send_key on startup
- ui.js: gate usage badge on window._showTokenUsage flag
Timestamps:
- streaming.py: stamp 'timestamp' on every message that lacks one at
conversation completion; old messages (no timestamp field) now get a
wall-clock time the first time they're touched by a new turn
- messages.js: stamp _ts on the last assistant message at done-event time
so the time shows immediately on the current turn before next reload
- Timestamps already render in the UI (Sprint 14): faint time on each
role header line, full opacity on hover, full date in title tooltip
Track A: Token/cost display
- Read agent usage attrs (session_prompt_tokens, session_completion_tokens,
session_estimated_cost_usd) after run_conversation in streaming.py
- Add input_tokens, output_tokens, estimated_cost fields to Session model
- Include usage in done SSE event payload
- Store usage on S.lastUsage in messages.js done handler
- Render usage badge below last assistant message (input/output/cost)
Track B: Subagent delegation cards
- Add subagent_progress to toolIcon map with shuffle emoji
- Special-case subagent_progress in buildToolCard: "Subagent" label,
strip double emoji from preview, add tool-card-subagent CSS class
- Indented border-left styling for subagent cards
- Clean delegate_task display name
Track C: Skill picker in cron create form
- Add skill search input + tag chips to cron create form HTML
- Skill picker JS in panels.js: search/filter, click-to-add tags,
remove tag chips, pre-fetch skill list on form open
- submitCronCreate sends skills array in POST body
- Skill picker dropdown + tag CSS
Track D: Skill linked files viewer
- Add file query param to /api/skills/content endpoint
- Serve linked files from skill directory with path traversal protection
- Ensure linked_files key always present in skill content response
- Render linked files section below SKILL.md content in preview panel
- openSkillFile function for viewing individual linked files
Track E: Bug fixes and code quality
- Expand Session.__init__ and compact() to readable multi-line format
- Remove inline import json as _j2 inside loop in streaming.py
- Fix tool_calls: capture args from assistant messages, skip unresolved names
- Store args snapshot in persisted tool_calls for reload display
6 new tests. Total: 421 (409 passing).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two fixes for Camanji rate limit UX:
1. api/streaming.py — pass fallback_model from profile config to AIAgent
The agent already supports fallback_model (a dict with provider/model/base_url)
for automatic rate-limit recovery, but streaming.py never read it from config
or passed it to AIAgent. Now reads get_config().get('fallback_model') at
call time (not module-level snapshot) and passes it through.
Also reads platform_toolsets.cli from the active profile's config at call
time so profiles with custom toolset lists use the right tools.
Camanji has fallback_model: {provider: openrouter, model: anthropic/claude-sonnet-4.6}
so hitting the direct-Anthropic rate limit will now automatically retry via
OpenRouter before giving up.
2. api/streaming.py + static/messages.js — show error inline, not 'Connection lost'
Previously: agent threw -> put('error', msg) -> SSE connection closed ->
browser's network-level 'error' event fired -> generic 'Connection lost'.
The actual error message was invisible to the user.
Fix: renamed server-side error event to 'apperror' (distinct from the SSE
spec's network error event). Added source.addEventListener('apperror', ...)
in messages.js that renders the error as a styled assistant message:
⏱️ Rate limit reached: <full message>
*Rate limit reached. Fallback model exhausted. Try again in a moment.*
Also added source.addEventListener('warning', ...) for non-fatal notices
(future use: fallback-activated status bar update).
Tests: 426 passed, 0 failed.
Add full profile management to the web UI, matching the hermes-agent CLI
profile system. Profiles are isolated HERMES_HOME instances with their own
config, skills, memory, cron, and API keys.
Backend: new api/profiles.py wrapping hermes_cli.profiles, dynamic config
reloading, 5 new API endpoints, profile-aware path resolution, HERMES_HOME
env save/restore in streaming, module-level cache patching for skills_tool
and cron/jobs.
Frontend: profile chip in topbar with dropdown, Profiles sidebar panel with
CRUD UI, boot-time profile fetch, cascade refresh on switch.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Tracebacks exposed file paths, module names, and potentially secret
values from local variables. Now logged server-side only; clients
receive a generic error message.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Update _PROVIDER_MODELS['minimax'] from stale ABAB 6.5 models to
current MiniMax-M2.7/M2.5/M2.1 lineup (matching hermes-agent upstream)
- Update _PROVIDER_MODELS['zai'] from GLM-4 to current GLM-5/4.7/4.5
lineup (matching hermes-agent upstream)
- Extend resolve_model_provider() to also return base_url from config.yaml,
so providers with custom endpoints (MiniMax, Z.AI) are routed correctly
- Pass base_url to AIAgent in both streaming and sync chat paths
Fixes#6
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace duplicated inline provider resolution in routes.py and streaming.py
with a shared resolve_model_provider() helper in config.py.
Improvements over original:
- If model ID has a prefix matching any known direct-API provider
(not just the config provider), strip it and route correctly.
This handles edge cases like localStorage restoring a model from
a different provider group.
- Single source of truth for the resolution logic.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When the model dropdown sends a prefixed ID like "anthropic/claude-xxx",
AIAgent interprets the provider/model format as an OpenRouter path and
routes through OpenRouter instead of the direct Anthropic API.
Fix: read the configured provider from config.yaml model section. If
the model ID starts with the configured provider name followed by "/",
strip that prefix and pass the provider explicitly to AIAgent. This
ensures direct API providers (Anthropic, OpenAI, etc.) are used when
configured, regardless of the model ID format from the dropdown.
