fix: sanitize sync chat history with config

[Michaelyklam]

Thinking Path

• Follow-up: complete #2297 coverage in _handle_chat_sync (POST /api/chat) #2398 identifies a leftover fallback path from the v0.51.65-if you use deepseek v4 pro, after use vision or sth related to picture, the session will fail #2297 text-mode image-history fix.
• The streaming chat path already passes active config into _sanitize_messages_for_api(...) so historical native image_url parts are stripped for text-mode providers.
• The legacy synchronous POST /api/chat path still sanitized history without config, so direct API callers could replay the same rejected image parts.
• This PR keeps the fix narrow: pass the active WebUI config into that one fallback callsite and pin it with focused regression coverage.

What Changed

• Updated _handle_chat_sync() to call _sanitize_messages_for_api(_previous_context_messages, cfg=get_config()) before agent.run_conversation(...).
• Added a regression assertion in tests/test_native_image_attachments.py so the sync route stays aligned with the text-mode sanitizer contract.
• Added an Unreleased changelog note for Follow-up: complete #2297 coverage in _handle_chat_sync (POST /api/chat) #2398.

Why It Matters

Direct users of the legacy synchronous chat endpoint now get the same text-only provider protection as the main streaming chat path: old multimodal history will not be replayed as native image_url parts when agent.image_input_mode resolves to text.

Closes #2398

Verification

• env -u HERMES_CONFIG_PATH -u HERMES_WEBUI_HOST /home/michael/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_native_image_attachments.py -q — 39 passed
• /home/michael/.hermes/hermes-agent/venv/bin/python -m py_compile api/routes.py api/streaming.py
• git diff --check

Risks / Follow-ups

• Low risk: this only supplies config to an existing sanitizer callsite.
• The compression precondition/idempotency sanitizer calls remain display/session comparisons rather than provider-facing history construction, matching Follow-up: complete #2297 coverage in _handle_chat_sync (POST /api/chat) #2398's scoped report.

Model Used

AI-assisted change with repository inspection, targeted editing, and shell-based test verification.

[nesquena-hermes]

Ready for independent review (small, scoped, low-risk fix).

This PR pairs with nesquena-hermes/hermes-webui-workspace-tracked issue #2398 which was filed during the v0.51.75 release as a deferred follow-up: the legacy synchronous POST /api/chat route was missing the cfg= argument that the streaming path uses to strip historical image_url content parts for text-mode providers.

This 1-line fix brings the sync route in line with the streaming path. The diff is tiny and the regression test directly pins the line. Should batch cleanly with #2407 (docs-only) and #2408 (Activity-group split) into a follow-on stage release once all three have reviews.

CI green on 3.11/3.12/3.13.

[nesquena]

Review — end-to-end ✅ (clean APPROVE, no fix pushed)

What this ships

1-line fix at api/routes.py:7905 to pass the active config into the conversation-history sanitizer on the synchronous /api/chat fallback route. Brings the sync path into parity with the streaming chat handler (which already passes cfg=_cfg at lines 3733, 3944, 4723). Closes #2398; companion to #2297 which fixed the streaming path.

Touches: api/routes.py (+1/-1), tests/test_native_image_attachments.py (+10), CHANGELOG.md (+4). 3.11/3.12/3.13 CI green.

Traced against upstream hermes-agent

Pulled a fresh tarball. No cross-tool surface: _sanitize_messages_for_api lives at api/streaming.py:1878 (webui-only). It mutates the message list passed to agent.run_conversation(conversation_history=...). The agent sees the sanitized list. No config.yaml write, no session schema change, no AIAgent param touched.

End-to-end trace

• _sanitize_messages_for_api(messages, *, cfg: dict = None) at api/streaming.py:1878-1897. Line 1897: strip_native_images = cfg is not None and _resolve_image_input_mode(cfg) == "text". With cfg=None (the pre-fix sync path), strip_native_images=False — historical image_url parts pass through unchanged → 400 from text-mode providers.
• _resolve_image_input_mode(cfg) at api/streaming.py:757-781 reads cfg["agent"]["image_input_mode"]. Default "auto"; explicit "text" → returns "text"; "native" → "native"; auto + auxiliary vision configured → "text".
• _handle_chat_sync(handler, body) at api/routes.py:7801-7908. The fixed callsite at api/routes.py:7905:

  conversation_history=_sanitize_messages_for_api(_previous_context_messages, cfg=get_config()),

• Streaming counterparts at api/routes.py:3733, 3944, 4723 already use the cfg=_cfg pattern. Sync now matches.

Other audit — things that are correct already

• get_config() is already imported and used 4+ times in the same file (lines 8421, 8451, etc.). Safe to call here.
• cfg= is a keyword-only argument (*, cfg) — caller has to be explicit. Test pins the exact callsite shape.
• No regression for image-mode providers: the strip is gated on _resolve_image_input_mode(cfg) == "text". Native/auto-without-vision paths are unchanged.
• No thread safety question — get_config() reads a snapshot, and the sync handler is request-thread-only.

Behavioural harness

no cfg:           image_url retained = True  (pre-fix behavior)
text cfg:         image_url retained = False (post-fix — strip)
native cfg:       image_url retained = True  (unchanged)
auto+no vision:   image_url retained = True  (unchanged)

All 4 cases match expectations. The strip is correctly gated.

Edge-case trace

Scenario	Expected	Actual
Text-mode provider, historical native image	strip → no 400	✅ harness
Native-mode provider, historical native image	keep	✅ harness
Auto-mode, no auxiliary.vision configured	keep	✅ harness
Auto-mode, auxiliary.vision configured	strip	✅ resolver logic at line 776+
No historical images	no-op	✅ no image_url parts to touch
cfg shape malformed	safe default (no strip)	✅ resolver defaults to "auto" → "native"
Cross-tool: CLI reads same config	unaffected (no shared state changes)	✅

Tests

• PR-targeted: 1/1 pass (test_sync_chat_history_sanitizer_receives_config).
• CI: 3.11/3.12/3.13 all green.

Minor observations (non-blocking)

1. String-grep regression test. The new test pins the exact textual form of the call. A defensive refactor that wrapped get_config() in a helper would break the test for cosmetic reasons. The grep test is fine as long as future maintainers know to update it alongside the call shape.
2. No matching /api/chat integration test. The streaming path has integration tests that exercise the full flow; the sync path's regression coverage stays at the string-grep level. Behavioural harness above covers the sanitizer side; the route wiring side has CI 3-version pass.

Recommendation

Approved. Tight 1-line fix that closes a real parity gap. Sanitizer logic correctly gated; no regression risk for non-text-mode providers. Behavioural harness confirms all 4 mode permutations.

✅ Parked at approval — ready for the release agent's merge/tag pipeline.

[nesquena-hermes]

Shipped via stage-372 / PR #2414 / v0.51.79 (Release BC). Attribution preserved in the squashed stage commit and the v0.51.79 CHANGELOG entry. Thank you!