Repository navigation
feat(chat): draw the crew window with the shared transcript and composer - #18095
Conversation
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Check one thing before merge: the Approve button in the shared transcript should reach the peer through the real [DESIGN-REVIEWED] 91a262b Description read[DESCRIPTION-READ] a4f19ae3570f12c08200dd24477a1ba30ed5464dc7d5926607ff08f5a2305269 That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
UX Review (Opus 5.5) — 🟡 CONCERNSUX-level review of UX-Verdict: CONCERNS The composer's unlabeled red-square Stop replaces the labeled "Stop", and the blind reader would not use it because they couldn't tell what it ends. Watch
Suggestions
[UX-REVIEWED] 91a262b Description read[DESCRIPTION-READ] 23c0d1a99a1210d640073eebe106dbb6bcf1e8555af49106dfa69e63f6521378 That sha256 names everything this verdict read: the PR description, plus the 8 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Before merge, check on a real peer (not the stubbed proxy) that Approve in the shared transcript row reaches the peer with the What this change shipsInventory (10 items) — 10 justifiedIntent: People driving a crew session should see the same chat rows and composer as in a local chat, so every chat feature does not have to be built twice. ADDITION (wave 3 of the
[FIRST-PRINCIPLES-REVIEWED] 91a262b Description read[DESCRIPTION-READ] fb4e6c2298a3f139105b304da38339320297fc20e5f63c1bfe6017d922062df4 That sha256 names the PR description this verdict read. Recompute it with |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Opus 5.5 Review — ✅ no blocking findingsReviewed Review detailsNo findings that block the merge. FINDING — website/src/pages/chat/crew-window/crewWindowRenderers.tsx:35 — [OPUS-REVIEWED] 91a262b Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
a28ea4c to
a774688
Compare
Disposition: fixed. Fixed in a774688: each permission row hands back its own |
Disposition: fixed. Fixed in a774688: a disabled fieldset turns the whole composer off offline, so the textbox is disabled again; the test is back to |
Disposition: fixed. Fixed in a774688: same change as d82d2f36b2b5; the RFC's "input disabled" holds again and the body no longer claims otherwise. |
Disposition: fixed. Fixed in a774688: the filter is now an allowlist ( |
Disposition: fixed. Fixed in a774688: the existing |
Disposition: fixed. Fixed in a774688: the row now reads "The agent on sent . The file stays on that crew." (new |
Disposition: rebutted. This is the shared composer's Stop control, the same one every local chat uses, with |
a774688 to
8a5f6c7
Compare
Disposition: fixed. Fixed in 8a5f6c7: the crew window registers its own |
Disposition: fixed. Fixed in 8a5f6c7: the same |
Disposition: deferred. Same item answered on 8713de0: it opens the hub's own Workflows tab, writes nothing and reaches no peer. Every check is green, so it rides the follow-up (drop workflow_completion from PEER_SAFE_ROWS). |
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765
…#18801) The crew chat window passed the shared composer none of its model, effort or agent props, and drew its own reply row plus loose "Rewind to here" / "Regenerate" buttons. In a two-machine test it showed a bare message box, misplaced actions and a reply row unlike the local chat. - The composer's model chip and model + effort picker (the shared ModelEffortDropdown) now read the peer's roster and write the PEER slot's model and effort through the hub proxy. ReasoningEffortDropdown gains an optional writeEffort seam for that write. - Attach, the agent picker and the approval-mode picker would write this machine's state, not the peer's, so the window does not draw them. - User and reply rows are the local UserMessage / AssistantMessage bubbles with the same hover rows: Quote stages a message for the next send, Edit & Resend rewinds on the peer, Regenerate sits on the newest reply's row. Peer rows keep the #18095 fences: decision records are stripped before drawing (no verdict thumbs) and the reply draws no file-change chips; code stays copy-only. Closes #18765 Co-authored-by: Joe Guo <zejiangg@amazon.com>
Problem / Motivation
Goal: A crew session's window shows the same chat rows and composer as a local chat, so tool calls, cards, approvals and future chat features reach crew sessions too.
The crew chat window (
CrewChatWindow.tsx, wave 3 of the remote-crew sidebar RFC) kept its own copy of the transcript, the approval card, the running label and a bare textarea. A tool call on the peer showed as a raw🔧 …line, a sent file as raw JSON, and no new row type ever reached it.Why it matters
Every chat feature had to be built twice, and the crew copy fell behind. People driving a crew session saw less than in a local chat for the same work.
Not a goal
dashboard/peer_redaction.py,handlers_instances.py).api/chatandapi/stream, so a peer file cannot be played here./commands, pickers). The composer is the side-panel preset.What changed (motivation → approach → change)
The window now feeds the peer's slot data into the shared parts instead of drawing rows itself.
messages.mapChatMessageList+ dashboard row set (reply row kept copy-only)ChatFooter+ the composer's StopChatInput(side-panel preset)Approve answers only the row the peer says is pending: same request id AND same row id (
meta.mid). The peer's slot detail used to drop that id:_prepare_messages_scoped(chat_utils.py) replacesmetawith the parsedclsdict, andclshas nomid. It now keeps the stored row's non-emptymidwhenclsreplacesmeta, forpermissionrows only, so every other row's wire meta is unchanged (amidinsideclsstill wins). Without this, every Approve against a real peer was refused. The peer reuses request ids, so a stale card for a reused id makes no peer call. The "Approve runs this one command…" hint still shows under a pending approval.crewWindowRenderers.tsxlets in only shared rows on an allowlist (PEER_SAFE_ROWS). A new shared row falls back to the store-free default until someone adds it. Left out: the launch cards (they read the hub's run by id) fall back to the tool line, and a sent file shows as "The agent on sent . The file stays on ." The user row adds the peer's rewind. The reply row stays plain markdown (it comes after the shared notice and completion cards, so those still win): the shared reply's verdict thumbs would write the hub's own decision log.Every code fence in the window is copy-only. A new
ReadOnlyCodeCtxincomponents/markdown/contexts.tsdoes for a subtree what thereadOnlyCodeprop does for one renderer; the window wraps its transcript in it, so no peer fence (reply, user, skill or inject row) offers Edit or Run in THIS machine's terminal.Offline, a disabled fieldset turns the whole composer off, as before and as the RFC says.
The shared rows and composer mount under a slot key no local session has, so they never pick up a hub session's state.
Five
crewWindowstrings only the old card and textarea used are removed from every locale; one new string (file_on_crew) is added. The feature map row names the new file.Backwards compatibility
Compatible: the window reads and posts the same peer routes with the same bodies; only how rows are drawn changed.
Reader: website/src/i18n/locales/*.json:pages.chat.crewWindow.{approval_title,approve,reject,stop,send} -- crew page -- no reader left;
git grepon origin/main finds them only inCrewChatWindow.tsx, which no longer uses them (npm run i18n:checkkey-refs passes).Reader: website/src/test/CrewChatWindow.test.tsx:crew-window-approval / crew-window-running test ids -- crew page -- no other reader on origin/main; the tests now find the shared Approve button and the footer's
Thinking…status.Tests
test/test_prepare_messages_permission_mid.py: a runner-shaped permission row (_ChatSlot.append("permission", ..., json.dumps({...}))) through_prepare_messageskeeps itsmeta.mid; fails on origin/main. Three more cases: amidinclsis not overwritten, a row with no storedmidgains none, and a non-permission row's wire meta gains nomid(red when the role check is removed). The 14 pytest files that call_prepare_messagesor read slot detail pass (1492 tests).The crew window fixture now sends the real wire shape: JSON
cls,metaparsed from it,midcarried by the server. A new case pins that a row with nomid(an older peer) gets no peer call.website/src/test/CrewChatWindow.test.tsx, updated for the shared parts, plus eight new cases:draws a peer tool call as the shared tool line, not its raw textdraws a peer file as its name, never as the hub outbox file of that nameanswers only the approval the peer itself reports pendinganswers no stale card whose request id the peer has reused: red when the mid check is removed (mutation run)says what Approve and Reject do while an approval is pendinglets in only shared rows decided safe for a peer (a new row falls back): fails on any new shared row id nobody decideddraws a peer code fence copy-only, with no Run into this machine: red whenreadOnlyCodeis removed (mutation run)draws a code fence in a peer USER row copy-only too: red when the context is off (mutation run)still draws a peer system notice as its card, not as a reply: red when the reply row is moved ahead of the shared rows (mutation run)The existing cases still pin approve, stop, continue, regenerate, rewind, send, draft recovery, the event feed, the version gate and the offline disabled composer. Determinism: 38/38 on three repeats and with
--sequence.shuffle.Manual verification
Rendered the real
CrewChatWindowin Chromium against a stubbed proxy (user row, reply, tool call, sent file, pending native approval). Checked in the DOM: tool line present, raw🔧text gone, file row shows its name, the old approval card is gone, Approve is in the transcript, composer present.Local AI review: WARNING, the extracted gpt and opus lanes (and a sonnet fallback) all failed on Bedrock throttling with no report, so this head had only a manual self-review against both contracts before push. CI's lanes are the first model review.
Screenshots / video
All "after" shots were taken on head
8713de07c2; later heads change no pixel, the realCrewChatWindowagainst a stubbed proxy.Before (origin/main): raw tool text, raw file JSON, separate approval card, bare textarea.
After, idle: Rewind on the user row, a typed draft, Send.
After, running: thinking indicator and the composer's Stop.
After, approval pending: the sent-file row, the approval in the transcript, and its hint.
After, peer offline: "devbox is not connected" and the composer turned off.
Related Issues
Closes #16774
Part of #6180
RFC: docs/request-for-change/rfc-remote-crew-sidebar.md
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)