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).
When a user switched profiles and created a new session, the session
was saved to the default profile directory instead of the active
profile directory — because get_hermes_home_for_profile() silently
fell back to _DEFAULT_HERMES_HOME when the profile directory didn't
exist yet on disk.
Root cause: api/profiles.py:156 had `if profile_dir.is_dir(): return
profile_dir; return _DEFAULT_HERMES_HOME`. New profiles (no session
yet, so no dir) routed every session back to default.
Fix: remove the is_dir() guard, return the profile path
unconditionally. The profile directory is created on first use by
the agent/session layer.
5 regression tests in tests/test_issue1195_session_profile_routing.py:
existing-profile, non-existent-profile (the core fix), None, empty-
string, 'default' all return the expected path.
Co-authored-by: bergeouss <bergeouss@users.noreply.github.com>
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>
Opus pre-release findings on #1370 applied:
SHOULD-FIX 1: Tightened parent_session_id exposure to only emit when
the parent's end_reason is in {compression, cli_close}. Without this,
two distinct WebUI sessions sharing a non-continuation parent (e.g.
'user_stop') would get clustered by frontend's _sessionLineageKey
(which falls through to parent_session_id when _lineage_root_id is
missing) and incorrectly collapsed into a single sidebar row.
Updated assertions in:
- tests/test_session_lineage_metadata_api.py::
test_non_compression_state_db_parent_does_not_create_sidebar_lineage
- tests/test_pr1370_lineage_metadata_perf_and_orphan.py::
test_non_compression_parent_does_not_extend_lineage
SHOULD-FIX 2: Chunked the IN-clause to 500 vars to stay under
SQLITE_MAX_VARIABLE_NUMBER. Python 3.9 ships sqlite 3.31 with the
default limit of 999. A power user with 2000+ sessions in the
sidebar would hit OperationalError, the silent except-wrapper would
swallow it, and lineage collapse would never work. Added
test_in_clause_chunked_for_large_session_set with SQL interception
to lock the invariant in source.
PR addition (per user directive — Opus + my review, no second
independent review round needed for combined batch):
#1372 from @NocGeek — fix: persist manual cron run results.
Self-contained 89 LOC fix split out from the held #1352. Mirrors the
scheduled-cron path (cron/scheduler.py:1334-1364) exactly: saves
output, marks job complete, treats empty response as soft failure
with matching error string. 2 behavioral tests using sys.modules
monkeypatch to mock cron.scheduler.run_job. CI not yet attached
because branch is brand-new; ran the new tests + adjacent suites
locally — all pass.
Final test count: 3471 passing, 0 failed.
Also adds 2 more regression tests for the perf-fix invariants:
- test_in_clause_chunked_for_large_session_set
- test_two_children_sharing_non_continuation_parent_not_collapsed
Manual WebUI cron runs previously called cron.scheduler.run_job(job)
and then only cleared the in-memory running flag. That meant output
could be dropped and job metadata like last_run_at / last_status was
not updated after a manual run.
This PR matches the scheduled cron path (cron/scheduler.py:1334-1364)
exactly:
- Save manual-run output via save_job_output
- Mark manual runs complete via mark_job_run
- Treat empty final_response as a soft failure with the same error
string as the scheduled path
- Record manual-run failures in job metadata via mark_job_run(False)
- Keep _run_cron_tracked self-contained for worker-thread execution
Includes 2 behavioral regression tests using monkeypatch.setitem on
sys.modules to mock cron.scheduler.run_job + cron.jobs helpers — the
right test pattern (exercises the real _run_cron_tracked code path).
Split out from #1352 (the larger profile-aware-cron-panel PR that's
on hold) per pre-release-review feedback. Self-contained, doesn't
touch the held PR's profile-filtering scope.
Co-authored-by: NocGeek <NocGeek@users.noreply.github.com>
Bundles:
- #1370 fix: expose session lineage metadata in API (@dso2ng)
Pre-release fixes applied:
1) Perf: replaced full table scan with parameterized WHERE id IN (...)
query. Original code did SELECT id, parent_session_id, end_reason
FROM sessions on every sidebar refresh. Measured 9ms cached scan at
1000 rows in production (up to ~450ms cold-cache); scales linearly.
New approach hits PRIMARY KEY + idx_sessions_parent — 50x faster
at 1000 rows, ~0.2ms regardless of total row count. Depth-bounded
to 20 hops to cap query count under pathological data.
2) Orphan-parent guard: suppress parent_session_id in API output when
the referenced parent row doesn't exist in state.db. The frontend's
#1358 _sessionLineageKey falls through to parent_session_id when
_lineage_root_id is missing — orphan references would create
never-collapsing single-row groups in the sidebar.
3) Regression suite (5 tests in
test_pr1370_lineage_metadata_perf_and_orphan.py):
- Pins the no-full-scan invariant by intercepting all SQL queries
and asserting no SELECT FROM sessions without a WHERE clause
- Pins orphan-parent suppression
- Pins cycle termination via threading.Event watchdog (2s timeout)
- End-to-end test for 4-segment compression chain root resolution
- Pins non-compression end_reason boundary stops walk
PR #1358 added the client-side lineage collapse helper, but
/api/sessions often did not include _lineage_root_id for the WebUI
JSON sessions visible in the sidebar. In that case the helper has no
grouping key and multiple same-title continuation rows remain visible.
This PR:
- Reads parent_session_id and end_reason from state.db.sessions for
the WebUI sidebar's session ids
- Walks the parent chain when end_reason is 'compression' or
'cli_close', producing _lineage_root_id and _compression_segment_count
- Cycle-detects via a 'seen' set
- Preserves projected lineage metadata on imported/gateway session rows
- Allows sidebar collapse to group cross-surface continuation chains
(CLI-close → WebUI continuation) while keeping non-continuation
parent rows flat
Co-authored-by: Dennis Soong <dso2ng@gmail.com>
The new _handle_clarify_sse_stream handler in #1355 holds clarify._lock and
then calls clarify.get_pending(sid) under the lock. get_pending also acquires
_lock internally — and clarify._lock is a non-reentrant threading.Lock(),
so the second acquisition deadlocks the SSE handler thread the moment any
client connects to /api/clarify/stream.
Existing tests pass because they only exercise sse_subscribe, sse_unsubscribe,
_clarify_sse_notify, and submit_pending directly — none of them invoke the
route handler. The deadlock would only manifest when a real EventSource opens
the connection.
Reproduced with a tiny harness that holds _lock and calls get_pending: the
worker thread is still blocked after a 2s timeout. With the fix, both empty
and populated queue cases complete in <1ms.
Fix: read clarify._gateway_queues / clarify._pending inline under the same
_lock acquisition, mirroring the approval SSE handler's pattern at
api/routes.py:2785-2793. No recursive lock; head-of-queue snapshot is
identical to what get_pending would have returned.
Added tests/test_pr1355_sse_handler_no_deadlock.py with three tests:
- behavioural: empty queue snapshot completes within 2s
- behavioural: populated queue snapshot returns the head entry
- source-level invariant: routes.py must not call get_clarify_pending()
inside `with _clarify_lock:` block (locks the regression in)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the 1.5s HTTP polling loop for clarify with a Server-Sent Events endpoint at /api/clarify/stream that pushes clarify events to the browser instantly. Mirrors the approval SSE pattern from v0.50.248 (#1350) including all the correctness lessons:
- Atomic subscribe + initial snapshot under clarify._lock
- _clarify_sse_notify called inside _lock for ordering guarantees (no notify-out-of-order race)
- Notify passes head=q[0].data (head-fidelity, not the just-appended entry)
- resolve_clarify also calls notify after pop so trailing clarifies surface immediately (no stuck-clarify bug)
- Empty-state notify with None,0 after pop-empty so frontend hides the card
- 30s keepalive comments, _CLIENT_DISCONNECT_ERRORS handling
- Bounded queue (maxsize=16) with silent drop on full
- Frontend: EventSource with automatic 3s HTTP polling fallback on onerror
Co-authored-by: fxd-jason <wujiachen7@gmail.com>
Session.load_metadata_only().compact() was dropping is_cli_session, source_tag, session_source, and source_label, so imported CLI/gateway sessions lost their provenance in sidebar/API payloads. Adds these to METADATA_FIELDS and Session.compact().
Co-authored-by: Dennis Soong <dso2ng@gmail.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>
Pre-release Opus review caught three correctness bugs in the original
PR #1350 SSE wiring beyond the snapshot/subscribe race:
A) **Notify-ordering race (MUST-FIX A):** _approval_sse_notify took _lock
only for the subscriber-list snapshot, then released it before
put_nowait. With two parallel submit_pending calls, T2's notify
could fire before T1's, leaving the UI showing pending_count=1 while
the server actually had 2 queued.
C) **Trailing approval lost (MUST-FIX C):** _handle_approval_respond
never called _approval_sse_notify after popping. With parallel
tool-call approvals (#527), a second approval queued behind the one
being responded to was invisible until the next event ever fired —
in practice, the agent thread parked on it would appear hung.
D) **Payload showed tail not head (MUST-FIX D):** payload built from
the just-appended entry instead of queue[0]. /api/approval/pending
returns the head; SSE returned the tail. Diverging contracts.
Fix:
- Split into _approval_sse_notify_locked (caller holds _lock, no
internal locking) and _approval_sse_notify (convenience wrapper).
- submit_pending: call _locked variant inside the queue-mutation lock,
passing queue_list[0] as head.
- _handle_approval_respond: call _locked variant inside the pop lock,
passing the new head (or None/0 if queue is empty).
- Restore fallback poll to 1500ms (was bumped to 3000ms; degraded-mode
parity with v0.50.247 is more important than save 1.5s of polling).
New regression tests in tests/test_pr1350_sse_notify_correctness.py:
- test_second_submit_pending_sends_head_not_tail (D)
- test_respond_to_first_pushes_second_as_new_head (C)
- test_respond_to_only_pending_pushes_empty_state (C edge)
- test_pending_count_is_monotonic_under_contention (A)
Updated test_approval_sse.py to pin the new contract:
- _approval_sse_notify_locked(session_key, head, total)
- 1500ms fallback interval
Total: 3411 tests passing.
Co-authored-by: jasonjcwu <jasonjcwu@users.noreply.github.com>
Bundles:
- #1349 fix(ui): show context indicator percentage without explicit context_length
- #1350 feat(approval): SSE long-connection for real-time approval notifications
Pre-release fixes applied:
- Inline subscribe + snapshot under a single _lock acquisition in
_handle_approval_sse_stream() to close the snapshot/subscribe race
flagged in pre-release review. A submit_pending() arriving between
the snapshot read and subscribe call would have been lost (appended
to _pending after our snapshot AND notified to subscribers before we
joined). Now atomic.
- Added tests/test_pr1350_sse_atomic_subscribe.py (4 source-level
invariants covering the atomic-lock-block guarantee).
Co-authored-by: jasonjcwu <jasonjcwu@users.noreply.github.com>
Replaces the 1.5s HTTP polling loop with a Server-Sent Events endpoint
at /api/approval/stream that pushes approval events to the browser
instantly. The backend uses a thread-safe subscriber registry
(_approval_sse_subscribers) with bounded queues to prevent memory
leaks from slow clients. Frontend uses EventSource with automatic
fallback to 3s HTTP polling on SSE error.
- Backend: subscribe/unsubscribe/notify lifecycle in api/routes.py
- New route: GET /api/approval/stream?session_id=
- submit_pending() now calls _approval_sse_notify() after queue append
- Frontend: EventSource with onerror -> _startApprovalFallbackPoll()
- 30s keepalive comments, _CLIENT_DISCONNECT_ERRORS handling
- 42 new tests (static analysis + unit + concurrency)
Co-authored-by: jasonjcwu <jasonjcwu@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 review on PR #1345 (Finding #3) flagged that
get_cli_sessions() was calling ensure_cron_project() once per cron
session in the loop — N lock acquires + N disk reads of projects.json
for N cron sessions per sidebar refresh.
Hoist a per-scan lazy memoizer (_cron_pid()) so we pay the resolution
cost at most once per get_cli_sessions() call. The memoizer is local
to the function (closure) so it's naturally scoped to a single scan
and doesn't leak across calls.
Could also have made ensure_cron_project() module-memoized, but that
would need invalidation on project deletion — the per-scan cache is
simpler and correct without coordination.
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>
release: v0.50.243
Batch release of 2 PRs.
- #1301 — fix: remove PRIMARY chip badge + add Claude Opus 4.7 label
Drops the chip-projected configured-model badge added in #1287 (chip
width 235px → 164px). Adds Claude Opus 4.7 label entries so the picker
no longer renders "Claude Opus 4 7" (missing dot).
Independently reviewed and approved by nesquena (commit c0bbd23).
- #1297 (@franksong2702) — fix: preserve cron output response snippets
Fixes#1295. /api/crons/output now preserves the ## Response section
when a large skill dump appears in the prompt section; falls back to
file tail when no marker exists.
Tests: 3254 passed, 2 skipped, 3 xpassed.
Independently reviewed and approved by nesquena (commit b262e4d).
Reverts the global assistant serif rule and removes the Calm theme that were shipped in v0.50.240 PR #1282. Pure deletion; 3252 tests passing. Override on independent review per Nathan.
When the user explicitly selects @provider:model from the picker,
_resolve_compatible_session_model() was stripping the prefix because
the hint matched the active provider (hint_matches_active=True → return bare_model, True).
This caused:
- The picker to snap back to the first duplicate entry on next render
- resolve_model_provider() to use the default provider instead of the
explicitly selected one, running the agent on the wrong backend
The hint_matches_active branch was intended for normalizing stale cross-
provider session models. But an @provider:model where the hint IS the
active provider is not stale — it is the user's deliberate selection.
Fix: return (model, False) so the full @provider:model survives to
resolve_model_provider() in config.py, which already handles it correctly.
Updates test_active_at_provider_session_model_preserved_with_hint and
adds test_issue1253_duplicate_model_id_active_provider_hint_preserved.
Closes#1253
All LRU cache operations (get, set, move_to_end, popitem) are already
protected by SESSION_AGENT_CACHE_LOCK. This addresses the reviewer's
concern about thread safety in multi-threaded ASGI servers.
Problem:
- GET /api/mcp/servers returned 404 error
- MCP servers management UI could not load server list
- Root cause: route was placed outside handle_get(), in unreachable code
Root Cause:
- The MCP servers GET route was incorrectly placed after handle_get() returned False (404)
- handle_get() function returns False at line ~1224, so any code after it won't execute
- The route was also in handle_post() area but without proper method checking
Solution:
- Moved GET /api/mcp/servers route inside handle_get() before the return False statement
- Removed the misplaced route from the old location (originally around line 1636)
- Also updated /api/profiles response format to include full profiles list
Testing:
- After restart: curl http://localhost:8787/api/mcp/servers returns {"servers": []}
- No more 404 errors
- WebUI can now properly load MCP servers list
starship-s