fix: strip reasoning blocks from AI outputs

[Vassista]

This PR hardens the output sanitization layer to ensure that reasoning blocks (e.g., <think>...</think>) from models never leak into user-facing views or exports. It expands support for additional model tags, preserves original Markdown formatting more cleanly, and applies the sanitizer universally across all UI components and features.

Changes:

• Sanitizer Improvements: stripReasoning now supports <thinking>, <thought>, and <reasoning> tags. It preserves legitimate Markdown indentation and cleanly removes blank-line residue after a block is stripped.
• Backend Parity: Updated the Markdown normalizer (simple_recorder.py) to handle the expanded set of reasoning tags, keeping frontend and backend parsing logic in sync.
• Universal Application: Reasoning blocks are now consistently stripped from all user-facing paths. This resolves previous leaks in custom report rendering, shareable note uploads, clipboard copying, search indexing, and list previews.
• Testing: Expanded unit tests to cover the newly supported tags, formatting preservation, blank-line cleanup, and relevant streaming edge cases.

[cubic-dev-ai]

No issues found across 3 files

_{Re-trigger cubic}

[Optic00]

Review

Good PR - the two closed gaps are real (reports rendered activeReport.content raw, and composeShareBody uploaded the raw summary), and I checked the new regex against the tricky cases (case-insensitive backreference, the $(?![\s\S]) end-of-string anchor under /m, the unclosed streaming block, backtracking behaviour on large unmatched input) - all sound. Two remaining gaps of the same class the PR is fixing, plus two smaller notes:

1. copyNotes still copies the raw summary to the clipboard

MeetingDetail.tsx:483 - const summary = meeting.summary?.trim() goes into navigator.clipboard.writeText unfiltered. Same pattern this PR fixes in composeShareBody, one screen up. With meeting.summary = "<think>private chain</think>\nPublic summary", the rendered note and the share upload are clean, but "Copy notes" puts the reasoning block on the clipboard. One-line fix, ideally in the same PR.

2. List preview and search snippets render the raw summary

• PreviousRow.tsx (previewText, ~line 169) shows meeting.summary?.trim() as the row preview on Home.
• CommandPalette.tsx:184 passes raw m.summary into snippet() (noteSearch.ts:33), which renders at line 206. This one is double: the search itself also matches text inside reasoning blocks, so a query hit inside <thinking>...</thinking> surfaces reasoning text as the result snippet.

A summary that starts with a reasoning block shows that text in the meeting list and palette even though the detail view is now clean. Wrapping both sources in stripReasoning fixes display; whether search should also match stripped text only is a judgement call (I'd say yes, same input for index and snippet).

3. Backend tag list now lags the renderer

simple_recorder.py:79 - _REASONING_TAG_HEADER_PATTERN normalizes headers after think|thought only. The renderer now also knows thinking|reasoning, so <thinking>...</thinking>### Action Items inline would skip the parser normalization. Not a blocker, but worth syncing the lists (or extracting one source of truth) so they don't drift further.

4. Correction on the PR description (doesn't change the value of the PR)

The regex on main has no /m flag - /<(think|thought)>[\s\S]*?(?:<\/\1>|$)/gi - so $ already matched end-of-string, and it strips closed multi-paragraph blocks and unclosed streaming blocks correctly (I ran both cases against the old code to be sure). What the new version actually improves: the two extra tag names, no more blank-line residue between stripped blocks, and preserved indentation instead of trimStart(). Worth adjusting the description so the changelog doesn't record a bug that wasn't there.

Nits

• The "handles nested tags" test covers <think>outer<thinking>inner</thinking></think> (different tags) but not same-tag nesting: <think>a<think>b</think>c</think> leaks c</think> with old and new regex alike. Pre-existing limitation and probably fine for real model output - just don't let the test name suggest it's covered.
• Dropping trimStart() is a deliberate contract change (the tests pin it), looks right to me - flagging it only so it's a conscious decision.

Happy to take 1-3 as a follow-up if you'd rather keep this PR small.

[Vassista]

Review

Good PR - the two closed gaps are real (reports rendered activeReport.content raw, and composeShareBody uploaded the raw summary), and I checked the new regex against the tricky cases (case-insensitive backreference, the $(?![\s\S]) end-of-string anchor under /m, the unclosed streaming block, backtracking behaviour on large unmatched input) - all sound. Two remaining gaps of the same class the PR is fixing, plus two smaller notes:

1. copyNotes still copies the raw summary to the clipboard

MeetingDetail.tsx:483 - const summary = meeting.summary?.trim() goes into navigator.clipboard.writeText unfiltered. Same pattern this PR fixes in composeShareBody, one screen up. With meeting.summary = "<think>private chain</think>\nPublic summary", the rendered note and the share upload are clean, but "Copy notes" puts the reasoning block on the clipboard. One-line fix, ideally in the same PR.

2. List preview and search snippets render the raw summary

• PreviousRow.tsx (previewText, ~line 169) shows meeting.summary?.trim() as the row preview on Home.
• CommandPalette.tsx:184 passes raw m.summary into snippet() (noteSearch.ts:33), which renders at line 206. This one is double: the search itself also matches text inside reasoning blocks, so a query hit inside <thinking>...</thinking> surfaces reasoning text as the result snippet.

A summary that starts with a reasoning block shows that text in the meeting list and palette even though the detail view is now clean. Wrapping both sources in stripReasoning fixes display; whether search should also match stripped text only is a judgement call (I'd say yes, same input for index and snippet).

3. Backend tag list now lags the renderer

simple_recorder.py:79 - _REASONING_TAG_HEADER_PATTERN normalizes headers after think|thought only. The renderer now also knows thinking|reasoning, so <thinking>...</thinking>### Action Items inline would skip the parser normalization. Not a blocker, but worth syncing the lists (or extracting one source of truth) so they don't drift further.

4. Correction on the PR description (doesn't change the value of the PR)

The regex on main has no /m flag - /<(think|thought)>[\s\S]*?(?:<\/\1>|$)/gi - so $ already matched end-of-string, and it strips closed multi-paragraph blocks and unclosed streaming blocks correctly (I ran both cases against the old code to be sure). What the new version actually improves: the two extra tag names, no more blank-line residue between stripped blocks, and preserved indentation instead of trimStart(). Worth adjusting the description so the changelog doesn't record a bug that wasn't there.

Nits

• The "handles nested tags" test covers <think>outer<thinking>inner</thinking></think> (different tags) but not same-tag nesting: <think>a<think>b</think>c</think> leaks c</think> with old and new regex alike. Pre-existing limitation and probably fine for real model output - just don't let the test name suggest it's covered.
• Dropping trimStart() is a deliberate contract change (the tests pin it), looks right to me - flagging it only so it's a conscious decision.

Happy to take 1-3 as a follow-up if you'd rather keep this PR small.

Thanks for the thorough review and for checking the regex edge cases.

I'm working on fixes for the remaining issues you called out:

• copyNotes copying the raw summary
• list preview / command palette using the raw summary (including search/snippet behavior)
• syncing the backend reasoning tag list with the renderer

I'll also update the PR description to accurately reflect what the regex change improves, and I'll clarify the nested-tag test name so it doesn't imply same-tag nesting is supported.

[Vassista]

• copyNotes now strips reasoning blocks before copying.
• Meeting list previews and search indexing/snippets now use the sanitized summary, so reasoning content is no longer surfaced there.
• Synced the backend reasoning tag list with the frontend to keep them aligned.
• I also updated the PR description to accurately reflect the final behavior and renamed the nested-tags test to clarify that it covers different-tag nesting rather than same-tag nesting.

Let me know if you spot something else.

[Optic00]

Re-checked the follow-up commits - all four points and the nits are addressed:

1. copyNotes strips the summary before it hits the clipboard.
2. PreviousRow.previewText and noteSearch (both searchNotes and snippet) now work on stripped text, so search matching and result snippets agree - the judgement call landed where I'd hoped.
3. Backend _REASONING_TAG_HEADER_PATTERN is synced to the four-tag list.
4. Description + test name corrected.

I also re-ran the new regex shape ((?:^[ \t]*)?...\n? with the startsWithReasoning guard) against the edge cases from the first pass - indentation preservation, blank-line residue, unclosed streaming block, uppercase tags, cross-tag nesting - all behave as the tests pin them. LGTM.

Coordination note: #318 (just opened) rewrites the same copyNotes block - it extracts the payload builder to lib/notesCopy.ts and fixes "Copy notes" to copy the open template report instead of the Standard note. It already applies stripReasoning to both copy sources (summary + report content), matching this PR's direction. So whichever merges second has a small conflict in MeetingDetail.tsx with a safe resolution: take #318's copyNotes (the strip behaviour from this PR is preserved there). Everything else in the two PRs is disjoint.