fix(dashboard): drop a slot's buffered stream text on slot_clear

[buluoray]

Problem

slot_clear (the /clear command) clears the transcript but leaves the slot's entry in the streaming chunk buffer (chunkBufRef). The buffer flushes once per animation frame, so a chat_chunk — or chat_thinking text, which shares the same buffer entry since #6666 — that arrived just before the /clear landed flushes on the NEXT frame and resurrects discarded text into the just-cleared pane.

This gap predates recent work but was widened by it:

• perf(dashboard): coalesce chat_thinking dispatches to one per frame #6666 routed chat_thinking into the same buffer, so buffered reasoning text now shares the leak.
• fix(chat): drive mobile panel slides on the compositor #6868 (open) defers flushes for up to ~1s while a panel slide holds the pipelines, widening the reproduction window from ~16ms to ~1s.

Flagged as a non-blocking finding while reviewing #6666 (see the review comment there); this PR is the promised follow-up fix.

What changed

One line in the slot_clear case of useWebSocket.ts: delete the cleared slot's chunk-buffer entry alongside clearing the transcript. The delete is keyed — other slots' in-flight buffers are untouched. This mirrors what chat_done already does for a finished turn.

Tests

New useWebSocket.slotClearBuffer.test.ts (4 cases), driving real WS envelopes through hand-driven rAF frames:

1. A chunk buffered before slot_clear does not flush back into the cleared transcript.
2. Buffered chat_thinking text is discarded too (shared buffer entry).
3. Clearing one slot leaves another slot's in-flight buffer intact.
4. Streaming resumes normally after a slot_clear (no stale prefix).

Mutation-verified:

• Remove the delete → tests 1, 2, 4 red (test 3 stays green, as it should).
• Replace the keyed delete with an over-broad .clear() → exactly test 3 red.

Full gates: tsc -b clean, eslint 0 errors on touched files, full vitest suite 26,049 passed on the rebased head.

Why no screenshot: no visual delta — the fix prevents a transient race (discarded text flashing back after /clear) that cannot be captured in a static frame; behavior is pinned by the four deterministic tests above.

No linked issue — the defect was found during code review of #6666 and is fixed directly here.

[github-actions]

UX Review (Fable 5) — ✅ PASS

UX-level review of e0ef425616e24bd513eee4926e4ef3b04f3355c7 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

This is a one-line race-condition fix in useWebSocket.ts plus tests — no new UI, strings, or visual surfaces. The change purely improves the existing /clear experience by preventing discarded stream text from flashing back into a just-cleared pane. Nothing for any UX lens to flag.

UX-Verdict: PASS

Invisible-by-design fix: /clear now actually clears — buffered stream text no longer resurrects into the emptied pane; no new UI or copy to review.

[UX-REVIEWED] e0ef425

[github-actions]

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of e0ef425616e24bd513eee4926e4ef3b04f3355c7 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] e0ef425

False positive or not applicable? A repository writer can comment:
/ai-review override gpt e0ef425616e24bd513eee4926e4ef3b04f3355c7: <one-sentence reason>

[github-actions]

Design Review (Fable 5) — ✅ PASS

Design-level review of e0ef425616e24bd513eee4926e4ef3b04f3355c7 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

The fix is a one-line keyed delete mirroring the existing chat_done cleanup at line 1612, with the same buffer lifecycle pattern. WebSocket message ordering guarantees the delete-at-slot_clear-receipt point is correct: any chunk the backend emitted before processing /clear arrives before slot_clear, so a keyed delete at that point discards exactly the pre-clear text and nothing after. Deleting the entry also resets lastSeq, which matches chat_done semantics (no phantom gap marker spanning a discarded transcript). The tests match the description's four claimed cases, and the description accounts for every hunk. No design-level concerns survive the kill-filter.

Design-Verdict: PASS

A narrow lifecycle fix in the right handler, mirroring the established chat_done cleanup, with ordering-correct semantics and mutation-verified tests.

[DESIGN-REVIEWED] e0ef425

[github-actions]

First Principles Review (Fable 5) — ✅ PASS

Premise-level review of e0ef425616e24bd513eee4926e4ef3b04f3355c7 — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

No shell was available, so the author's gate claims (vitest run, tsc) can't be re-executed — but everything this lane owns was verified by reading: the chat_done parity claim, the reconnect path, and the absence of unfixed siblings.

First-Principles-Verdict: PASS

A one-line keyed delete closing the last transcript-discard path that skipped the chunk buffer — every item traces to the reported defect.

What this change ships

Intent: text discarded by /clear must not flash back into the cleared pane a frame later. This is a FIX.

1. /clear now also drops the slot's buffered stream/thinking text — justified (defect reported in the perf(dashboard): coalesce chat_thinking dispatches to one per frame #6666 review; description matches the diff exactly).
2. Four deterministic tests pinning the drop, its keying, and post-clear streaming — justified (pins the fix and the over-broad .clear() regression).

Verification I ran, since the counts are the substance of a PASS here:

• Duplication: none. Grepped chunkBufRef (8 sites) — the only existing purges are chat_done's keyed delete (useWebSocket.ts:1612) and the reconnect whole-buffer clear (:752). Neither covers slot_clear; this is the missing third, in the same idiom.
• Siblings: grepped case 'slot_ (4 events) and the transcript-discarding reducers. slot_clear was the only discard path leaving a buffer entry; the parallel subagentChunkBufRef is not a sibling because clearMessages/clearSlotCache (chatSlice.ts:3293,3300) never discard subagent panel state, so nothing there resurrects.
• Depth: the fix sits at mechanism level — the discard handler now owns its buffer entry, matching the documented buffer lifecycle ("reconnect clear, chat_done delete, unmount cancel", useWebSocket.ts:246-249). The deeper "cause" would be abandoning the per-frame buffer, which exists for a measured per-token dispatch cost — not up for relitigation.
• New surface: zero — no key, flag, export, or event added.

[FIRST-PRINCIPLES-REVIEWED] e0ef425

[github-actions]

Opus 4.8 Review — ✅ no blocking findings

Reviewed e0ef425616e24bd513eee4926e4ef3b04f3355c7 — this comment is updated in place on each push.

Review details

The change is a single, well-scoped one-line delete plus tests. The discovery pass found no candidates, and my inspection of the buffer machinery confirms the fix is correct (keyed delete closes the buffer-then-flush resurrection window; subagentChunkBufRef writes to state that clearMessages/clearSlotCache don't touch, so no parallel fix is owed). Nothing survives to report.

No findings.

[OPUS-REVIEWED] e0ef425

Verdict parsed from the review's SHA-scoped output markers for commit e0ef425616e24bd513eee4926e4ef3b04f3355c7.

False positive or not applicable? A repository writer can comment:
/ai-review override fable e0ef425616e24bd513eee4926e4ef3b04f3355c7: <one-sentence reason>

[NicholasRBowers]

Tier 1 auto-approve: fix (2 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean, security checklist all-NO, AI reviewers green. Category: fix with clear root cause -- slot_clear now drops the slot's buffered stream text (keyed delete) so a chunk buffered before /clear cannot flush back into the cleared pane; frontend-only, plus a pinning test.