Skip to content

docs: record the crewmate chat contract - #12923

Closed
CrysisDeu wants to merge 1 commit into
mainfrom
feat/crewmate-chat-bubbles
Closed

CrysisDeu wants to merge 1 commit into
mainfrom
feat/crewmate-chat-bubbles

Conversation

@CrysisDeu

@CrysisDeu CrysisDeu commented Sep 22, 2026 •

Copy link
Copy Markdown
Contributor

Problem / Motivation

A crewmate chat should show the conversation, not its internal work log.

Current scope

The implementation and tests from this stacked change are already on main through PR #12815. After rebasing onto current main, this PR now contains only the missing system-spec section in docs/system-specs/modules/crew-mode.md.

The section records the shipped contract:

  • member-mode chat filters machinery at render time without deleting history;
  • visible speech, user messages, approvals, errors, and stop cards remain;
  • a never-spoken crewmate points users to its Work log;
  • roster previews use the same speech rules as the chat;
  • messages form one bubble run per turn;
  • ordinary chats remain unchanged.

Backwards compatibility

Documentation only. No runtime behavior changes in this rebased diff.

Tests

  • bash scripts/docs-lint.sh
  • git diff --check origin/main..HEAD

Both pass. No local test suite was run.

Manual verification

Not needed for this rebased diff. The current branch changes documentation only.

Related work

Checklist

  • One commit
  • Documentation checks pass
  • No secrets, credentials, or internal references in the diff

@CrysisDeu
CrysisDeu requested review from a team and patrigao September 22, 2026 22:05
@CrysisDeu

Copy link
Copy Markdown
Contributor Author

Intent: A crewmate's chat on the Members page shows only what the crewmate says to the user, as Slack-style bubble runs under its avatar; the patrol/cron/sub-agent/tool machinery stays in history for the Work log but is not drawn. Ordinary chats are untouched.
Not a goal: the Work log surface itself, reply threads (the row stamp is only the hook for them), changing what the gateway persists, or fixing the pre-existing option-chip clearing by a later empty reply.

@github-actions github-actions Bot added the readiness: action required A blocking check or review needs attention label Sep 22, 2026
@github-actions

github-actions Bot commented Sep 22, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Fable 5) — ⏭️ skipped

Revision 2d3befb99a5ba98c2ad425556d9cb27233a23dbb ships no reviewable capability (docs, tests or generated files only), so there is nothing to inventory. Advisory — does not block merge.

@github-actions

github-actions Bot commented Sep 22, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Fable 5) — ✅ PASS

Design-level review of 2d3befb99a5ba98c2ad425556d9cb27233a23dbb — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

Verify the fine-grained claims (corner rules, footer overrides) still match #12815's shipped code — referenced symbols all exist, behavior details I could not execute.

[DESIGN-REVIEWED] 2d3befb

@github-actions

github-actions Bot commented Sep 22, 2026 •

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of 2d3befb99a5ba98c2ad425556d9cb27233a23dbb and found no blocking issues.

This comment is updated in place on each push.

Review details

FINDING -- docs/system-specs/modules/crew-mode.md:1137 -- duplicate ## A crewmate's chat repeats the contract while omitting the existing reply-thread clause -> Fix: delete the added duplicate section.
[GPT-REVIEWED] 2d3befb

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 2d3befb99a5ba98c2ad425556d9cb27233a23dbb: <one-sentence reason>

@github-actions

github-actions Bot commented Sep 22, 2026 •

Copy link
Copy Markdown
Contributor

UX Review (Fable 5) — ⏭️ skipped

Revision 2d3befb99a5ba98c2ad425556d9cb27233a23dbb touches no user-facing surface (no changes under website/ or committed screenshots), so the UX review was skipped. Advisory — does not block merge.

@github-actions

github-actions Bot commented Sep 22, 2026 •

Copy link
Copy Markdown
Contributor

Opus 5 Review — ✅ no blocking findings

Reviewed 2d3befb99a5ba98c2ad425556d9cb27233a23dbb — this comment is updated in place on each push.

Review details

The entire diff is a byte-duplicate of the ## A crewmate's chat section already at crew-mode.md:1039, and the copy is already stale.

FINDING — docs/system-specs/modules/crew-mode.md:1137 — the added block re-adds ## A crewmate's chat verbatim (now two identical headings, 1039 and 1137; docs_lint.py has no duplicate-heading check, and #a-crewmates-chat resolves only to the first), and the second copy drops the two closing sentences naming pages/members/ThreadPanel as the reader, so the spec's final word is no DOM attribute is stamped until that reader exists while website/src/pages/members/ThreadPanel.tsx already consumes crewmateRunPosition/crewmateBubbleClass → Fix: delete the added 95 lines; the contract is already recorded at line 1039, and apply any genuinely new sentence there instead.

[OPUS-REVIEWED] 2d3befb

Verdict parsed from the review's SHA-scoped output markers for commit 2d3befb99a5ba98c2ad425556d9cb27233a23dbb.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 2d3befb99a5ba98c2ad425556d9cb27233a23dbb: <one-sentence reason>

@CrysisDeu
CrysisDeu force-pushed the feat/crewmate-chat-bubbles branch from 51e4cdc to aca7a8e Compare September 22, 2026 22:41
@github-actions github-actions Bot added readiness: checking Automated validation is still running and removed readiness: action required A blocking check or review needs attention labels Sep 22, 2026
@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • filterCrewmateChat removes policy-block evidence, mislabeling a system continuation as the user's steer — fixed (span=4eae54b776c7)

Fixed in aca7a8e: the row set takes the UNFILTERED transcript (crewmateTranscript, ChatPane passes paneMessages) and renderAssistantBubble reads turnHadPolicyBlock against it via policyBlockTranscript, locating the row by identity. The filtered list is used for layout only.
self-added: yes

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • renderAssistantBubble hides footer actions on run-ending messages — fixed (span=80c7587a1c52)

Fixed in aca7a8e: the crewmate assistant row passes forceFooter: pos === 'single' || pos === 'end'; the run end is the turn end on this surface, so the last bubble always carries its footer and hover actions. Streaming rows still draw none.
self-added: yes

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • data-crewmate-run is stamped for a reply-thread footer that does not exist yet — rebutted (span=e0f7355afe85)

The stamp is the seam the parallel reply-threads worker (screen 07, branch feat/reply-threads-api, in flight for the same launch) hangs its footer on; without it that PR must reopen CrewmateMessage.tsx to re-derive the run position it already exports. One data attribute, no runtime path, no user-visible surface.
Kept as the module boundary the launch plan assigned to this PR; it clears when the thread footer lands and reads it.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • A quiet patroller's chat filters to empty and shows the fresh-thread "Session ready" line — fixed (span=95c0c48cb17d)

Fixed in aca7a8e: when the filter leaves nothing although the history is not empty, ChatPane draws components.chatPane.crewmate_quiet — " hasn't said anything to you yet. What it has been doing is under Recent activity." — in all 12 catalogs; the fresh-thread line is kept for a truly empty thread. Screenshot (dark + light) added to the PR body.
self-added: yes

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Nothing on the filtered surface names where the machinery went — fixed (span=e8848baed3e4)

Fixed in aca7a8e for the state where it bites: the empty-chat hint names Recent activity as where the crewmate's work lives. A persistent pointer on a populated chat is not added — the RFC (screen 05/06) puts that material in the Work log tab, which is a sibling PR of this launch, and a dismissible banner on the conversation would be a second element to explain.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Option chips cleared by a quiet patrol — track before launch — accepted-and-deferred (span=5cc5f803c683)

Pre-existing in deriveFollowUpOptions (options.ts): the walk-back stops at any assistant row, so an invisible-only reply clears the previous question's chips. Outside this PR's render-time filter. Tracked in #12934 with the suggested fix (skip isHiddenInvisibleAssistantRow rows in the walk-back), routine on this surface, so filed before launch as asked.
self-added: no

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Sep 22, 2026
@CrysisDeu
CrysisDeu force-pushed the feat/crewmate-chat-bubbles branch from aca7a8e to 9a20cb8 Compare September 22, 2026 23:17
@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • data-crewmate-run is stamped with zero consumers — fixed (span=2407024fea17)

Fixed in 9a20cb8: the CREWMATE_RUN_ATTR constant and the row stamp are deleted; crewmateRunPosition stays exported for the reply-thread footer to reuse, and the PR that reads a DOM attribute adds it. Spec and feature map updated to match.
self-added: yes

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Empty crewmate chat beside a roster row still quoting a last message — fixed (span=ee0b547815cf)

Fixed in 9a20cb8 on the server: the roster preview is speech-only. last_message_info(speech_only=True) (history_projection.py) skips every non user/assistant row for the TEXT while the recency epoch still reads the newest row; the live member/message event carries preview only for a speech row with visible text and the roster projection applies it only when present. A never-spoken patroller's row is blank, matching its chat. Pinned by test_members_preview_speech_only.py.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Hint points at Recent activity, whose counters the reader could not parse — fixed (span=05eea4b28eb8)

Fixed in 9a20cb8: the pointer is dropped; the hint is one sentence, " hasn't said anything to you yet." (12 catalogs). A link to the Work log is not added here — that surface is a sibling PR of this launch and does not exist on main yet; linking to it belongs to the PR that ships it.
self-added: yes

CrysisDeu added a commit that referenced this pull request Sep 22, 2026
Not this PR's change: the two commits it sits on, squashed into one so the
PR stays at two commits. #12797 = feat/reply-threads-api 32484ff,
#12923 = feat/crewmate-chat-bubbles 9a20cb8. Dropped as each lands.
@CrysisDeu
CrysisDeu force-pushed the feat/crewmate-chat-bubbles branch from 9a20cb8 to 87e3752 Compare September 22, 2026 23:43
@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Crewmate chat is incorrectly documented as client-only — fixed (span=2f6082500a24)

Fixed in 87e3752: the feature-map row names the backend half — handlers/members.py (speech-only roster preview), dashboard/state.py (member/message preview only for speech), dashboard/system_notices.py (is_speech_row) — and its endpoints (GET /api/members last_message, the member_projection WS frame).
self-added: yes

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • A crewmate with no history at all still shows "Session ready — type a message to start" — rebutted (span=cc262d07e293)

Out of this PR's shape on purpose. The truly-empty crewmate chat is the state #12924 (New crewmate, this launch) owns: it opens the new crewmate's chat with a greeting, so after that PR a crewmate chat is never empty, and the RFC already fixes the copy for a single crewmate's empty state ("Your AI teammate. Give it a job; it keeps working while you are away."). A third string here ("Say hello to {{name}}") would compete with both. This PR's hint covers the state only it creates — history that filters to nothing — and leaves the fresh-thread line to the PR that replaces it.
self-added: no

@github-actions

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ⚠️ review incomplete

GPT 5.6 did not produce a complete verdict for 5e75c2aa933161321f749615afa98b68ea0e22ec; inspect the workflow logs and re-run it.

This comment is updated in place on each push.

See the GPT 5.6 Review job logs; this commit has no completed GPT verdict.

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 5e75c2aa933161321f749615afa98b68ea0e22ec: <one-sentence reason>

1 similar comment
@github-actions

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ⚠️ review incomplete

GPT 5.6 did not produce a complete verdict for 5e75c2aa933161321f749615afa98b68ea0e22ec; inspect the workflow logs and re-run it.

This comment is updated in place on each push.

See the GPT 5.6 Review job logs; this commit has no completed GPT verdict.

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 5e75c2aa933161321f749615afa98b68ea0e22ec: <one-sentence reason>

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • system_notices.py re-declares three sub-agent envelope constants — fixed (span=af62da9f48bd)

Fixed in 7c01908: the three literals are gone; system_notices.py imports SUBAGENT_COMPLETION_PREFIX, SUBAGENT_BATCH_COMPLETION_PREFIX and SUBAGENT_COMPLETION_META_KEY from kiro_crew.constants (no cycle — the module already imported kiro_crew.preview_text).
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Envelope classifiers are pinned to each other, not to the writers — fixed (span=9cdbde56538e)

Fixed in 7c01908 the first way you named: TestEnvelopeClassifiersFollowTheRealWriters imports workflow_inject._summarize (the builder, not a copied string) and asserts is_speech_row rejects its output for finished / failed / cancelled snapshots — and that the same text with a broken header IS speech, so the pin fails loudly on either drift. The sub-agent envelope is composed inline in the Slack gateway with no importable builder, so its row is pinned on the part the gateway guarantees: the subagentCompletion meta stamp (kind / agentId / outcome from kiro_crew.constants), rejected under both roles even with a reworded header line.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Stamp a meta.kind on the workflow-completion envelope at write time — rebutted (span=2a1b8be548d0)

Not in this PR: it changes the writer (workflow_inject.py) and the frontend card's own predicate (isWorkflowCompletionMessage, which this PR only calls), i.e. the wire contract of a surface this PR does not otherwise touch, to shield a drift the new writer-built test now catches at CI time instead. Worth doing where the workflow card is owned; here it would widen a render-time-filter PR into the envelope's format.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Working turns render nothing: with tool rows hidden, a mid-turn crewmate chat sits still until it speaks — needs-a-decision (span=d8835fe58d8e)

Real, and a product-shape call for the maintainer, not a fix I should slip in: the accepted RFC (screen 05) defines what a crewmate's chat DRAWS and says nothing about a working indicator, so adding a typing bubble is a new element outside the RFC text. What I checked: the pane keeps every existing mid-turn signal an ordinary chat has before its first token — the header's running dot, the busy composer — and the stubbed-/api harness cannot render a live turn (busy state comes from the SSE stream), so I have no honest mid-turn frame to attach. @CrysisDeu, one of two: (a) accept as-is for this launch — nothing for you to do, the filter ships and a typing bubble can follow as its own small PR with a recording; (b) reply "add the working bubble" and I add a Slack-style " is working…" bubble under the author line while the turn runs with no streamed words yet (one new string ×12 catalogs, a recording as evidence), ~one more round. My recommendation: (a), so this PR stays the RFC's shape.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Malformed sub-agent headers corrupt the roster preview (a pasted glyph-less header is read as machinery) — fixed (span=55fa97098b0c)

Fixed in de49e15 as you named: the backend's legacy header path now requires an outcome glyph (✅❌⏹⚠, the card's own OUTCOME_GLYPH_RE) beside the agent id, mirroring parseSubagentCompletion's "no glyph means this is not a shape this card understands; degrade to normal rendering". One correction to the trace: the frontend twin already refused the glyph-less header (it returns null without a glyph), so the pasted text WAS drawn — only the roster quote lagged; now both twins agree. Two fixture rows pin it: a user pasting [Subagent completion event]\nAgent \demo`` and an assistant quoting a glyph-less header in prose are both speech; the real envelope with its glyph stays machinery.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Crewmate chat feature-map row contradicts visible boundary behaviour — fixed (span=2f6082500a24)

Fixed in de49e15: the row now names the retained rows the user must see or act on (errors, notices, files, OAuth prompts, a pending approval, the stop card, the live stream), lists resolved approvals and workflow envelopes among the hidden rows, and states that a run breaks on a user message, on any other visible row (error, pending approval, stop card) or on five minutes of silence — the same rule crew-mode.md already records.
self-added: no

@github-actions

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ⚠️ review incomplete

GPT 5.6 did not produce a complete verdict for de49e15ef67e702066a3064534dc4865402f89fa; inspect the workflow logs and re-run it.

This comment is updated in place on each push.

See the GPT 5.6 Review job logs; this commit has no completed GPT verdict.

False positive or not applicable? A repository writer can comment:
/ai-review override gpt de49e15ef67e702066a3064534dc4865402f89fa: <one-sentence reason>

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Unflushed speech is overwritten by stale transcript reconciliation — fixed (span=96b17a80a753)

Fixed in 6e2630e as you named, mirroring _reconcile_slot_window's gate: _slot_has_unflushed_rows(slot) (module level in handlers/members.py) is True when the live slot holds rows past _disk_window_len, is mid-rewind, or is dirty; api_members evaluates it per row BEFORE the roster observation and the disk read, and a slot flagged there is excluded from the correction whatever its disk read said — the row still carries the read (the client falls back to the folded quote), nothing is written, and the next read after the flush finds roster and transcript agreeing. Pinned: unit cases for the three gates; an API test where the member log already holds "Newer speech." from the live emit while disk holds "Older speech." and the slot has one unflushed row — the fold keeps ("Newer speech.", 2.0) — and after the flush the read agrees and appends nothing.
self-added: no

@github-actions

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ⚠️ review incomplete

GPT 5.6 did not produce a complete verdict for 6e2630e90a69184f4516d48d4ce337d16d3d029d; inspect the workflow logs and re-run it.

This comment is updated in place on each push.

See the GPT 5.6 Review job logs; this commit has no completed GPT verdict.

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 6e2630e90a69184f4516d48d4ce337d16d3d029d: <one-sentence reason>

1 similar comment
@github-actions

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ⚠️ review incomplete

GPT 5.6 did not produce a complete verdict for 6e2630e90a69184f4516d48d4ce337d16d3d029d; inspect the workflow logs and re-run it.

This comment is updated in place on each push.

See the GPT 5.6 Review job logs; this commit has no completed GPT verdict.

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 6e2630e90a69184f4516d48d4ce337d16d3d029d: <one-sentence reason>

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • The quiet chat's "Work log" pointer reads as a statement, not an action — fixed (span=1ac9e6106356)

Fixed in 2ea62ab with your copy: crewmate_quiet_where is now verb-first — "See what it has been doing in its Work log." — in 12 catalogs + en-XA (each locale in its own imperative, informal register per the style rules), spec and body updated, and the two "at HEAD" quiet frames re-shot with the new line (the capture asserts the exact text, that it is an underlined <button>, and clicks it to open the Work log).
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • "2 runs" may be read as "chats renamed" — confirm the unit is used consistently — rebutted (span=fadb9a27e790)

Confirmed consistent, no change needed beyond what ships: the unit appears in exactly two user-visible places on this page — the day rows (activity_chat_count_*, "N run(s)") and each entry's hover title (activity_chat, which defines a run: "a time this crewmate woke to work — a patrol, a schedule, or a chat you opened") — and nowhere else in the drawer or panel says "chat" for this count (the activity_routed_* sibling says "auto-picked"). There is no separate activity drill-in surface on main. The translator-context examples were the one stale spelling left ("1 chat · 1 auto-picked") and are fixed in the same head (your suggestion below).
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • en.context.json activity_routed_count_* example still says "1 chat · 1 auto-picked" — fixed (span=834a1dde0f06)

Fixed in 2ea62ab: both activity_routed_count_* context entries now read "1 run · 1 auto-picked", matching the unit beside them.
self-added: no

@github-actions

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ⚠️ review incomplete

GPT 5.6 did not produce a complete verdict for 2ea62ab28c875c3ec1fa843ac322e3d3abfa1769; inspect the workflow logs and re-run it.

This comment is updated in place on each push.

See the GPT 5.6 Review job logs; this commit has no completed GPT verdict.

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 2ea62ab28c875c3ec1fa843ac322e3d3abfa1769: <one-sentence reason>

1 similar comment
@github-actions

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ⚠️ review incomplete

GPT 5.6 did not produce a complete verdict for 2ea62ab28c875c3ec1fa843ac322e3d3abfa1769; inspect the workflow logs and re-run it.

This comment is updated in place on each push.

See the GPT 5.6 Review job logs; this commit has no completed GPT verdict.

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 2ea62ab28c875c3ec1fa843ac322e3d3abfa1769: <one-sentence reason>

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • The one-time flush snapshot allows stale preview corruption — fixed (span=96b17a80a753)

Fixed in 665b68d as you named: _slot_flush_generation(slot) samples (len(messages), _disk_window_len, _dirty_gen) per bound slot BEFORE the roster observation; after the transcript read the correction is authorised only when the slot is still flushed AND that tuple is unchanged, so a reply that lands anywhere inside the window (after the pre-await sample, before or during the disk read) refuses the stale disk quote. Pinned by an API test that appends the reply from inside the patched last_speech_info (slot clean at the sample, unflushed at the re-check): the fold keeps "Newer speech." and no stale-preview event is appended; plus unit cases that each of the three counters moves the generation.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • The unflushed check runs only before the awaits; a slot that appends during the window is still admitted — fixed (span=cd9678020c04)

Fixed in 665b68d exactly as suggested: the preview_authoritative loop re-asks _slot_has_unflushed_rows(state._slots.get(row["slot_key"])) after tails resolves (the pre-await sample is kept), and additionally requires the slot's persistence generation to match the pre-observation sample. Same test as the GPT record: a reply landing during the read is not corrected from disk.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Run grouping chains bubbles across turns on a five-minute gap; the RFC decided runs are one turn's bubbles — fixed (span=c683c675d0f5)

Fixed in 665b68d, the subtraction you named: CREWMATE_RUN_GAP_MS and the time arm of chained are deleted. crewmateRunPosition(messages, index, transcript) now reads the turn boundary off the unfiltered crewmateTranscript the pane already passes: two drawn replies are one run only when every row between them in the transcript is the turn's own machinery (tool / thinking / done / state rows); a user message, a patrol wake (nudge), an inject or subagent envelope between them ends the turn. Tests: a filtered nudge between two replies splits the run, tool/thinking/done rows do not, inject/subagent envelopes do, and a 30-minute silence inside one turn no longer breaks it. The evidence frames are unchanged: the fixture's six replies stamp start,end,single,start,cont,end under the turn rule exactly as before (re-shot and asserted), so no re-attach.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • The run definition lives only on each entry's tiny timestamp hover; put it on the day count too — fixed (span=cfad57950772)

Fixed in 665b68d: the day row's count ("2 runs") is wrapped in a <span title={t('pages.membersPage.activity_chat')}>, so hovering the count itself shows the same one-sentence definition; no new string.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Shots 03/07 show pre-HEAD copy under an "After" label — fixed (span=0521f137060c)

Fixed in the body for 665b68d: the two superseded quiet frames ("Crew summary", "chats", plain-text hint) are removed; the never-spoken state is shown only by the two at-HEAD frames (dark + light), relabelled as the "After" quiet frames. Six attachments remain: before/after conversation in both themes, and the quiet state at HEAD in both themes.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Bounded history is mistaken for a never-spoken crewmate — fixed (span=42cfccb1d7fb)

Fixed in 2b10cd5 along the line the adjudication drew: the never-spoken hint is gated on the completeness signal the store already holds. crewmateQuietUnproven = crewmate pane AND slotPaneHasMore is true AND the window has rows AND the filter drew none of them; while it holds, the hint is not rendered and the pane's hydration latch flips to the unbounded read (the same latch the streaming upgrade uses), so the older speech is drawn instead of a false claim. An unbounded read that is still all machinery (has_more false) is the one case that says "hasn't said anything to you yet". Pinned in ChatPane.crewmate.test.tsx: speech behind a bounded has_more window is drawn and no hint appears; a bounded window whose full read is also machinery shows the hint only after the unbounded call; a complete first read (has_more false) triggers no second read.
self-added: no

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • Workflow completion fails to break message runs — fixed (span=1f8db16bf0e0)

Fixed in 2b10cd5 as you named: chained checks isTurnEnvelope(between) (= isWorkflowCompletionMessage || isSubagentCompletionMessage, the two predicates the module already imports) BEFORE the assistant pass-through, so a workflow or sub-agent completion written under the assistant role ends the turn and the reply after it opens a new run. Pinned in crewmateBubbles.test.ts for the workflow envelope, the single sub-agent envelope and the batch digest (both neighbours single), and the negative: an assistant row that only quotes the prefix is speech, drawn, and chains as an ordinary row.
self-added: no

@github-actions

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of 2b10cd5921e7adda12291712dca6b2d43e592f62 and found no blocking issues.

This comment is updated in place on each push.

Review details

FINDING -- website/src/i18n/locales/en.json:7200 -- "One run" claims patrols and schedules are counted, but activity records only track chats and routing decisions, so those runs leave the count unchanged -> Fix: restore copy describing chats across the changed catalogs. (origin: validation)
[GPT-REVIEWED] 2b10cd5

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 2b10cd5921e7adda12291712dca6b2d43e592f62: <one-sentence reason>

@CrysisDeu

Copy link
Copy Markdown
Contributor Author
  • User workflow-envelope messages contradict the speech contract — fixed (span=2f6082500a24)

Fixed in 955d1f5 as you named: is_speech_row now excludes a well-formed workflow-completion envelope only under the assistant role (the gateway's injection path); the same text pasted by the user is the user speaking, so the roster quotes what the chat draws. Pinned three ways: a new user-role row in test/fixtures/crewmate_speech_rows.json (speech: true, read by both suites), the writer-built _summarize test asserts the intact envelope is speech under user, and crew-mode.md's speech paragraph says "under the assistant role only".
self-added: no

iamwhatever
iamwhatever previously approved these changes Sep 24, 2026
@bolichen97

Copy link
Copy Markdown
Contributor

Closing as superseded: PR #12815 (e4b5f09) already put this section on main

The description says the branch now carries "only the missing system-spec section". That section is not missing. PR #12815 (squash e4b5f09, merged 2026-09-24) added it. Its message reads "chore: carry #12923 (bubbles) as this stack's base", and the section sits on origin/main at docs/system-specs/modules/crew-mode.md:1039-1136. The section was also already in this PR's merge base, which is why the diff's context lines (@@ -1134,6 +1134,101 @@) are the end of main's copy.

How this PR's change lines up with main:

This PR (head 2d3befb) origin/main c66f95b
crew-mode.md:1137: a second ## A crewmate's chat H2 crew-mode.md:1039: the same heading
crew-mode.md:1137-1229 byte-identical at crew-mode.md:1039-1131
crew-mode.md:1230 ends on ...no DOM attribute is stamped until that reader exists. crew-mode.md:1132-1135 goes on: The reply thread's own panel (pages/members/ThreadPanel, see history.md) is that reader...

So merging would add a stale duplicate and no new content. The file would have two ## A crewmate's chat headings, and the #a-crewmates-chat anchor only resolves to the first. The second copy would also say the reader does not exist yet, when it does: website/src/pages/members/ThreadPanel.tsx:38 imports crewmateRunPosition / crewmateBubbleClass and uses them at :61 and :167. On this head, the GPT 5.6 and Opus 5 review lanes both flagged the same duplicate at crew-mode.md:1137.

The behaviour the section describes is on main in code as well (spot-checked):

  • isCrewmateSpeech / filterCrewmateChat / crewmateRunPosition at website/src/components/chat/crewmateBubbles.ts:86/112/166
  • is_speech_row at src/kiro_crew/dashboard/system_notices.py:111
  • speech_preview at src/kiro_crew/preview_text.py:99
  • test/fixtures/crewmate_speech_rows.json

Nothing left to harvest: every added line is already on main. The option-chip follow-up is tracked in issue #12934, and closing this PR doesn't affect it.

If you meant to change a sentence in main's section, the cleanest route is a small PR that edits crew-mode.md:1039-1136 in place. If I've missed something this branch carries, reply here or reopen.

From a full open-PR audit against origin/main c66f95b (2026-09-25); findings were checked by independent reviewers reading the code. Reply here if this is wrong.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants