fix: stop replaying reasoning-only history

[Michaelyklam]

Thinking Path

• Long WebUI sessions rely on persisted transcript/context reconstruction to keep reloads and follow-on turns stable.
• Issue 思考过程错误显示 #2514 includes a redacted session where old reasoning-only assistant entries have multiplied into thousands of blank assistant turns with stale Thinking metadata.
• Those entries are useful as display metadata for interrupted/canceled turns, but they are not valid provider-facing conversation history.
• The narrow fix is to keep visible Thinking cards in the transcript while preventing reasoning-only display entries from being replayed to the model or reinserted into restored context.

What Changed

• Added a backend guard that identifies reasoning-only assistant display entries and excludes them from _sanitize_messages_for_api() / API-safe context alignment.
• Stopped _restore_reasoning_metadata() from reinserting prior reasoning-only display entries into model-facing context while still restoring them for the visible transcript when needed.
• Kept visible assistant replies that also carry reasoning metadata intact.
• Cleared stale settled .agent-activity-thinking rows before compact Activity rerenders rebuild the transcript, avoiding duplicate Thinking rows on repeated renders/session switches.
• Added regression coverage for provider-history sanitization, reasoning metadata restoration, and settled Activity rerender cleanup.
• Added a release-note entry.

Why It Matters

This prevents long sessions from feeding blank assistant Thinking artifacts back into the next turn and compounding stale reasoning/tool metadata. The UI can still show legitimate Thinking entries, but old display-only internals no longer metastasize into conversation history or inflated Activity counts.

Verification

• env -u HERMES_CONFIG_PATH -u HERMES_WEBUI_HOST /home/michael/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_sprint49.py tests/test_ui_tool_call_cleanup.py tests/test_issue1361_cancel_data_loss.py tests/test_pr1375_partial_tool_calls_sanitize.py tests/test_orphaned_tool_messages.py -q — 63 passed
• env -u HERMES_CONFIG_PATH -u HERMES_WEBUI_HOST /home/michael/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_sprint49.py::test_restore_reasoning_metadata_preserves_existing_timestamps tests/test_sprint42.py::test_streaming_restores_prior_reasoning_metadata_after_followup tests/test_sprint42.py::test_routes_restores_prior_reasoning_metadata_after_followup -q — 3 passed
• python3 -m py_compile api/streaming.py
• /home/michael/.hermes/hermes-agent/venv/bin/python -m py_compile api/streaming.py api/routes.py
• node --check static/ui.js
• git diff --check
• Manual redacted-session probe from the issue attachment: _sanitize_messages_for_api() reduced the attached 2,202-message sample to 440 provider-facing messages with 0 blank assistant turns.

UI media: not attached because this is a history/context sanitation and rerender idempotence fix rather than a new visual surface; the visible outcome depends on a large reporter session attachment and is covered by source/regression checks.

Risks / Follow-ups

• Existing already-bloated saved sessions may still contain old reasoning-only display entries until a future repair/migration cleans persisted session files. This PR prevents them from being replayed into provider history and stops settled rerenders from stacking duplicates.
• Tool-call duration semantics mentioned in the issue thread may deserve a separate focused follow-up if reporters still see per-tool timing confusion after the history bloat stops.

Closes #2514
Refs #2535

Model Used

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

[nesquena-hermes]

CI failure — test (3.11) red on this head

Heads up @Michaelyklam — the most recent push has test (3.11) failing with test (3.12) and test (3.13) cancelled. Won't be able to fold this into a batch release while CI is red.

Could you check tests/test_sprint49.py and tests/test_ui_tool_call_cleanup.py against the new guard? The fix shape itself (transcript display vs provider-replay separation) looks right per #2514, but the regression coverage needs to land green before we can route this through the merge queue.

Will revisit on the next sweep tick once CI clears.

[Michaelyklam]

Rebased this on the latest master after the release batch moved the changelog, keeping the PR note under [Unreleased] and preserving the current release notes.

Verification on the rebased head:

• pytest tests/test_sprint49.py::test_restore_reasoning_metadata_preserves_existing_timestamps tests/test_sprint42.py::test_streaming_restores_prior_reasoning_metadata_after_followup tests/test_sprint42.py::test_routes_restores_prior_reasoning_metadata_after_followup tests/test_ui_tool_call_cleanup.py -q — 24 passed
• python -m py_compile api/streaming.py api/routes.py
• node --check static/ui.js
• git diff --check origin/master...HEAD
• git merge-tree --write-tree origin/master HEAD
• GitHub Actions test (3.11), test (3.12), and test (3.13) — passed on the rebased head

The PR is mergeable again.