The _normalized_message_timestamp_for_key helper was preserving
microsecond precision (%.6f). When the same message is persisted by
both the WebUI sidecar JSON writer and the Hermes agent state.db
writer, their timestamps can differ by a few microseconds, causing
_session_message_merge_key to produce different keys for the same
logical message and letting both copies survive the dedup pass in
merge_session_messages_append_only.
Truncating to second-level granularity collapses sub-second drift to
the same key, so the duplicate is suppressed correctly.
Fixes#2616
Four code-review comments from the automated Copilot reviewer on this PR:
1. `_journal_tool_already_present` dedupe was session-wide, so a
legitimately-repeated tool (e.g. a second `terminal: ls` in an
earlier turn) could cause the retry path to falsely skip
materializing the recovered tool card. The helper now takes a
keyword `stream_id` argument; when supplied, a tool card whose
`_recovered_stream_id` is set AND differs from the candidate is no
longer treated as a duplicate. Untagged tool cards (live tools, or
tool cards carried over from a pre-tagging core transcript) still
match, preserving the existing 'core transcript already has this
tool, don't duplicate' invariant. Two new tests in
`TestJournalToolDedupeScoping` cover both legs of the rule.
2./3. The troubleshooting FAQ pointed at `~/.hermes/webui/sessions/session_<sid>.json`
and `~/.hermes/_run_journal/...`. The actual sidecar filename has
no `session_` prefix and the run-journal lives under the WebUI
sessions dir (`~/.hermes/webui/sessions/_run_journal/<sid>/<stream>.jsonl`,
default). Both paths fixed and an explicit note added about
`HERMES_WEBUI_STATE_DIR` overriding the state root.
4. Drop unused `json` / `queue` / `Path` imports from
`tests/test_session_lost_response_regression.py` so the file stops
carrying noise that future linting would flag.
When the WebUI process restarts mid-stream and sidecar repair runs while
the run-journal for the dead stream is not yet visible on disk (WSL2 9p
/ DrvFs page-cache loss, un-fsynced journal tail on network FS, …),
`_append_journaled_partial_output()` returns False and the marker is
permanently baked with the "no agent output was recovered" wording even
though the journaled tokens appear on disk shortly afterwards.
This commit reframes the recovery contract so the read side can
self-heal:
* `_interrupted_recovery_marker` gains a `pending_retry=True` mode
that produces a third wording ("Recovering the partial output …
reload this session to retry.") and stamps a
`_pending_journal_recovery` flag.
* `_apply_core_sync_or_error_marker` now writes that pending-retry
marker (with `_journal_retry_stream_id`,
`_journal_retry_attempts`, `_journal_retry_first_seen_ts` meta)
whenever it cannot recover visible output AND the stream id is
known. The legacy "no output" wording is reserved for the
no-stream-id case. The core-sync branch leaves marker emission to
the existing visible-output check (the core transcript itself is the
canonical history in that branch).
* A new `_retry_journal_recovery_in_place(session)` helper re-runs
`_append_journaled_partial_output(…, dedupe_existing=True)` for the
latest pending marker. On success the marker is promoted in place to
the recovered-output wording, the journaled rows are reordered to
sit above the marker (preserving chronological order), and all
retry meta is stripped. On failure attempts is incremented; after
_JOURNAL_RETRY_MAX_ATTEMPTS (12) or _JOURNAL_RETRY_GIVEUP_SECONDS
(24h) the marker is demoted to a neutral "Partial output may have
been lost." wording.
* `get_session()` cheaply short-circuits via
`_session_has_pending_journal_retry()` and invokes the helper on
both cache-hit and cold-load paths when a pending marker is found.
`metadata_only=True` skips the helper to keep sidebar refresh
cheap. The retry call runs OUTSIDE the SESSIONS LOCK to avoid a
deadlock with `session.save()` write paths.
No streaming write path or run_journal fsync behaviour is changed — the
fix is read-side only.
When API server runs append messages directly to state.db, reconcile WebUI sidecar sessions with those canonical rows across API responses, model-facing streaming context, and active browser refresh.
Add append-only state.db merge helpers, metadata-only counts for refresh polling, and regression coverage for API visibility, context incorporation, and frontend refresh behavior.
- Session.composer_draft field: {text, files} stored in session JSON
- POST+GET /api/session/draft endpoint for save/load
- loadSession: save draft before switch, restore from S.session.composer_draft
- textarea input: debounced 400ms auto-save to server
- send(): clear draft after message is sent
- lockComposerForClarify(): save draft before card locks composer
- _restoreComposerDraft: clears textarea when target has no draft, guards
against stale responses racing new session loads, exact text comparison
- Session.compact(): includes composer_draft in response
- Fix: use handler.command instead of parsed.method (ParseResult has no .method)
Co-authored-by: Minimax <noreply@minimax.io>
SHOULD-FIX: rate-limit _repair_stale_pending repair-firing telemetry. Switch
from unconditional logger.warning to age-keyed: WARNING when pending_age <
5min (the diagnostically valuable race window — actual leak-path candidates
that slipped past the grace guard) and DEBUG for the long-tail (orphaned
sidecars from prior process lifetimes). Prevents reconnect loops on stuck
sessions from flooding the log while preserving the diagnostic signal we
want for tuning _REPAIR_STALE_PENDING_GRACE_SECONDS empirically.
NIT: _LOCAL_SERVER_PROVIDERS expanded with lm-studio (hyphenated alias used
in some custom_providers configs and already recognized at api/config.py:2189
for SSRF host trust) and localai (LocalAI project). Test parametrize expanded
from 7 to 11 names, also covering pre-existing koboldcpp and textgen for
symmetry. +4 regression tests.
NIT (docs): CHANGELOG callout for the RFC1918 behavior change. Internal-
network OpenAI-compatible proxies now preserve the model prefix on private-IP
base_urls. Documented the migration path: configure as a custom_providers
entry to bypass the local-server detection.
NIT (deferred, optional): narrowing the heuristic to is_loopback only is
left as future work; the broader scope was an explicit goal in the bug
body and Opus flagged it as SHOULD-DISCUSS-but-not-block.
4184 -> 4188 passing. 0 regressions. ~10 LOC absorbed total.
Closes#1623 — Lower SSE app heartbeat from 30s to 5s at every long-lived
handler (main agent, terminal, gateway-watcher, approval-poller, clarify-poller).
Kernel TCP keepalive declares peer dead at 25s worst-case (10s KEEPIDLE +
5s KEEPINTVL * 3 KEEPCNT, added v0.50.289 #1581). 30s app heartbeat let the
kernel tear sockets down on flaky networks before the app sent its first
keepalive byte — drops at ~10s during long thinking phases. New named
constant _SSE_HEARTBEAT_INTERVAL_SECONDS=5; regression test pins the
inequality (app_heartbeat * 2 <= kernel_window) so future tuning can't
re-introduce the misalignment.
Closes#1624 — Add 30s grace period to _repair_stale_pending() trigger.
Without it, any narrow race between the streaming thread clearing
pending_user_message and STREAMS.pop(stream_id) produces a false-positive
'Previous turn did not complete.' marker on a turn that finished correctly
(reproducible after every command-approval turn). Defense-in-depth, not
the root-cause fix — the actual streaming-thread leak path is tracked
separately. Falsy pending_started_at (legacy sidecars) treated as
'old enough' so legitimate legacy-data recovery still works. Plus
logger.warning telemetry on every legitimate repair so the next batch of
user reports tells us whether the underlying race still fires.
Closes#1625 — Local model servers (LM Studio, Ollama, llama.cpp, vLLM,
TabbyAPI, koboldcpp, textgen-webui) now keep the full HuggingFace-style
model id (e.g. 'qwen/qwen3.6-27b' instead of stripped 'qwen3.6-27b'). New
_LOCAL_SERVER_PROVIDERS set + _base_url_points_at_local_server() loopback/
RFC1918 heuristic — either signal triggers no-strip. Backward compat
preserved for OpenAI-compatible proxies on public hosts (LiteLLM at
litellm.example.com still strips openai/gpt-5.4 -> gpt-5.4). Updated the
existing #230/#433 test to reflect that #1625 supersedes the strip-on-custom
rule for loopback hosts (see api/config.py and test_model_resolver.py
docstring update). Reported by @akarichan8231 in Discord on 2026-05-04.
42 regression tests across:
tests/test_issue1623_sse_heartbeat_alignment.py (3)
tests/test_issue1624_repair_stale_pending_grace.py (9)
tests/test_issue1625_local_server_model_id_preservation.py (30)
4142 -> 4184 passing. 0 regressions.
nesquena-hermes