Repository navigation
Conversation
|
Intent: Let dashboard agents offer an explicit multi-turn objective as a distinct autonomous goal card, then open the existing Set-a-goal review flow with that objective prefilled. Keep ordinary suggested replies separate. |
Design Review (Opus 5.5, fork) — ✅ PASSDesign-level review of Design-Verdict: PASS No design-level blocker: the producer is gated to slots full chat can open, nothing runs without review, and the callers that measure offsets were moved to the tail-only grammar. Notes
[DESIGN-REVIEWED] d2d886c Description read[DESCRIPTION-READ] 16ed0546ee57aa39efb95e2782d499641957896fdfb15a44be420a7d7407d1f6 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, fork) — ✅ PASSUX-level review of UX-Verdict: PASS Nothing blocks: screenshots cover every card state (ready, review editor, goal active/running, monitor, stopped, storage failure, crew/member, started). No first-time reader has read the card cold; a same-repo branch run would add that. Notes
[UX-REVIEWED] eeafe0d Description read[DESCRIPTION-READ] 8e48fa88cedc5b970218a441286a08038ba5bcd01b5043eddf39f2c4d9a32b11 That sha256 names everything this verdict read: the PR description, plus the 23 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 |
GPT 6.1 Review (fork) — ✅ no blocking findingsReviewed Review detailsNo findings. |
First Principles Review (Opus 5.5, fork) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing blocks. The one open question is whether a rule that is on by default in every dashboard session counts as a product-shape change. That takes judgement, and nothing runs until the user reviews and starts the goal. Notes
What this change shipsInventory (10 items) — 9 justifiedIntent: let the agent offer a labeled multi-turn goal that the user reviews in the existing goal editor before it starts — ADDITION.
[FIRST-PRINCIPLES-REVIEWED] eeafe0d Description read[DESCRIPTION-READ] 72e0d3418cfabbcc4f1b5b66d1130cfa3b4eb9dcbf313dbfb18f6fff0d05a067 That sha256 names the PR description this verdict read. Recompute it with |
Opus 5.5 Review (fork) — ✅ no blocking findingsReviewed |
51ff5a5 to
54da932
Compare
|
Disposition for an earlier head, kept for the record (no longer a ledger record): self-added: yes
|
|
Disposition for an earlier head, kept for the record (no longer a ledger record): self-added: yes
|
|
Disposition for an earlier head, kept for the record (no longer a ledger record): self-added: yes
|
|
Disposition for an earlier head, kept for the record (no longer a ledger record):
|
|
Disposition for an earlier head, kept for the record (no longer a ledger record):
|
|
Disposition for an earlier head, kept for the record (no longer a ledger record): self-added: yes
|
|
UX evidence added on |
|
Design review note: replacing the remembered draft message is intentional prefill behavior only when the user chooses Review & start. The integration test proves cadence and cycle cap remain unchanged. In the existing-automation path, Open automation does not call |
|
Disposition for an earlier head, kept for the record (no longer a ledger record):
|
54da932 to
7fc8ba4
Compare
7fc8ba4 to
db6955c
Compare
|
Pass 11: rebased onto current main at |
cbf8e84 to
6264c44
Compare
|
Pass 12: rebased onto main |
6264c44 to
3260118
Compare
|
Pass 16: rebased onto main |
|
Pass 18: rebased onto main |
|
Pass 22: rebased onto main |
|
Pass 26: rebased onto main |
|
Head 37b9240 (one commit on main 4c9d96a) answers the review of 00514dd. The rebase follows the main changes this PR touches: Fixed:
Dispositioned:
On 3d6cda4, 2,305 scoped backend tests, 2,696 vitest tests, typecheck, zero-warning ESLint, |
|
Review of 37b9240: the First Principles block repeats the product-shape gate dispositioned above for 00514dd. Whether dashboard agents offer |
|
Pushed 7b043e2, rebased onto 6457624 (no conflict). GPT (upheld on a2413f1): fixed.
CI on a2413f1: Backend Tests shard 8 timed out in Tests
The First Principles RFC gate is unchanged. |
|
Pushed 413031e, rebased onto d2417a4.
|
|
Rebased onto main Resolution, keeping both sides:
The other 74 paths in the patch are identical to the previous head. Checks on
The previous head's only diff-independent reds were Backend shard 7 ( |
|
Fixed the upheld GPT finding on
Tests:
|
|
Fixed GPT's upheld finding on 4133fd4 (
|
|
Rebased onto aeb5c02 to clear a conflict with #23388 in
|
|
Rebased onto 9bf1d20 as 28c5383. The rebase resolved conflicts with #24766 (stale option chips) and #24776 (context pools).
|
|
Rebased onto main to resolve a conflict with #25042, which added a sub-agent critical-rules block next to this PR's goal block in
This PR's tests, the tests main changed since the last push, vitest, tsc, the composition contracts and the static gates pass. The only failures are ones main shows on this host too. |
|
First Principles gate (product shape). The First Principles lane's last verdict is BLOCK, on 28c5383: the default critical rules for full-chat dashboard slots now tell the agent to offer The same lane passed this shape on fa767f0 (2026-10-22, "Nothing blocks"). That head already had Product shape is a maintainer decision, so this PR does not add an RFC. If you accept the shape, please post The other reds on 5348de5 are GitHub API rate limits (the AI review lanes that could not run, PR Policy, Fork PR Description, PR Readiness) and fork-only checks (Sensitive Change Review, Dependency License Gate). They need a maintainer re-run; no code here changes them. |
|
Pushed ec5f8e1, rebased onto main On |
|
Rebased onto main
Two changes beyond the conflicts:
The description's file table now names both files; its stale test counts are gone. |
|
Pushed d2d886c, rebased onto main |
Let agents offer an explicit multi-turn goal action beside ordinary reply suggestions. Teach the marker only to dashboard slots full chat can open, and say in the rule that split panes offer the objective as an ordinary reply, since the same slot can also be shown in one. Reuse the per-session goal draft and review flow so users see the autonomy contract before starting it. A goal or monitor in the session disables the card, which says whether it is running, waiting on your reply or stopped, confirms a goal the card itself started, and offers Open goal, the composer's goal control. A failed browser save turns Review & start off and leaves Open goal as the one action. Strip goal lines at the shared strip_control_comments projection boundary, so mirror backfill gets the same treatment as previews, speech and channel egress.
|
Pushed 6a6db15, rebased onto main
Main's |
Problem / Motivation
Goal: Let an agent offer a clearly labeled multi-turn goal that the user reviews in the existing goal editor before autonomous work starts.
Agents can suggest one-turn replies, but users must copy a longer objective into Set a goal and infer that it continues across turns.
Why it matters
The choice between one reply and autonomous work should be explicit before work starts. A goal card lets the user review the objective, cadence, and cycle cap instead of accepting an ambiguous suggestion.
Not a goal
What changed (motivation → approach → change)
A standalone
[GOAL: objective]line in an assistant reply is a typed suggestion, not a command to start work. The dashboard critical rules of any slot full chat can open teach it: offer at most one goal, only for multi-turn work, on its own line before any final[OPTIONS:]line. One slot can be open in full chat, a split pane and the companion chat at once, and the server cannot tell which host shows it, so the rule says so directly: full chat and the companion chat (an embedded full chat) show an Autonomous goal card, and split panes offer the same objective as an ordinary one-turn reply. Crew, member, and App Kit slots, Side Chat, cron, sub-agents, and every messaging runtime keep a rules block that never mentions the marker.Full chat shows the objective in an Autonomous goal card. Review & start stores it in the existing Set-a-goal draft and opens that editor. A goal or monitor in the session, running, paused or stopped, or a crew or member session, disables the card. The status line says which: a goal or monitor waiting on the reply options reads as active and waiting on the reply, as the goal panel's "Goal active" does, and the copy names the goal panel rather than the button. When the session's active goal is the suggested objective, the card confirms "Started. Running in this session." instead of asking the user to stop it. Each blocked or started state offers Open goal, a labeled button that opens the composer's goal control, so the user never has to find the icon in the toolbar. That status line is the accessible description of both buttons. If the browser refuses to store the objective, the error notice says so, Review & start turns off, and Open goal becomes the primary action. The objective stays visible in the card to select and paste.
The bounded parser removes complete marker lines without backtracking, including CRLF-terminated lines. Split panes, Side Chat, ChatEmbed, and composers without goal review offer the objective as an ordinary editable reply. Goal lines are removed inside
strip_control_comments, the projection boundary previews, text-to-speech, channel renderers, channel egress and both mirror backfills already call. Slot projections and the Slack options extractor, which do not call it, strip them too. Callers that measure a tail offset or ask whether a tag ended the reply usestrip_trailing_control_tags, the unchanged tail-only grammar. Marker-free output is returned unchanged.context.pyand itscontext_assembly/ownersoutput_contract.py,session_start.pyandturn_message.pyconstants.py,dashboard/state_parts/slot_transcript.py,slack/format.py,slack/handler.py,slack/handler_runtime/stream.py,messaging/egress.pyapp-sdk/protocol/,ChatEmbed.tsx,SideChat.tsx,AssistantMessage.tsxChatInput.tsx,chat-input/props.tsandsuggestedGoal.ts,ChatPane.tsx,FollowUpBar.tsx,ChatPage.tsx,ComposerDockView.tsx,followUpChips.ts,chat/crewmateChips.tsgoalDrafts.tsfeature-map/README.md,website/src/i18n/locales/,website/src/test/,test/Backwards compatibility
Compatible: replies without a standalone
[GOAL: ...]line keep existing follow-up behavior, and unsupported hosts keep the objective as an editable reply. Only dashboard sessions full chat can open receive the new rule text. Other runtimes receive byte-identical critical rules.Removes nothing that has shipped.
Tests
goalSuggestionvitest andtest_preview_text.pyscanner tests fail on the previous head and pass here:[GOAL: Fix dict[str, Any] parsing]and[GOAL: check arr[0] until it is set]yield the full objective and are stripped from projections, while[GOAL: a] b]and a closer followed by text stay prose; while such a line streams, its marker stays hidden until its own closer.ChatInput.goalSuggestiontests pin the card's started/blocked states, andsearchableTexttests keep goal lines out of message search. On this revision 3,180 scoped backend tests and 99 PR-scoped vitest tests (888 with the ChatEmbed suites) passed.Manual verification
Images 01 and 02 come from the real built SPA with fixture APIs; the sidebar preview there is the backend
_ChatSlot.to_dict()projection of a reply that carries the marker, and every capture asserted that no[GOAL:text node is visible. Images 03 to 10 were rendered from the real composer, card and goal panel components with synthetic data, so they show the transcript and composer without the sidebars; each capture checked the status text, card state and buttons first. The split-pane and Side Chat reply fallback is covered by theChatPaneand protocol tests.No first-time reader has looked at the card cold yet. The evidence is committed because a read-only fork PR cannot use
gh pr edit --attach.Screenshots / video
Related Issues
Closes #12835
Checklist
Contribution License Agreement
Contributions are accepted under the repository's Apache 2.0 license.