Repository navigation
feat(dashboard): one-click send for Dashboard cards, behind an off-by-default preview - #33203
Conversation
7e7417e to
729ed00
Compare
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: |
UX Review (Opus 5.5) — ✅ PASSUX-level review of UX-Verdict: PASS Nothing blocks. The blind reader read every primary control correctly: the "Approve the plan" card button, the "Send replies in one click" switch and the Profile badge. Each new state and notice appears in shots 01–07. Notes
[UX-REVIEWED] 7525727 Description read[DESCRIPTION-READ] d7eab243a020e5ad8de31c904800870a73d3a7d5fb5765d742dea7ba48399d3c That sha256 names everything this verdict read: the PR description, plus the 10 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 |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS No design-level blocker: the feature is off by default, the gateway enforces it, the frame-tied click gate holds, and a card turn gets no auto-approval. Notes
[DESIGN-REVIEWED] 7525727 Description read[DESCRIPTION-READ] 6b1515628bddd175aff948a84a1161a69b73e1c04804ef848eb4c8311290cebf 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 |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing blocks. The one-click send ships off for every crewmate ( Notes
What this change shipsInventory (9 items) — 8 justifiedIntent: let a person's click on a Dashboard card's answer send that reply into the crewmate's own thread, instead of only filling the chat box. This is an ADDITION.
[FIRST-PRINCIPLES-REVIEWED] 7525727 Description read[DESCRIPTION-READ] aed7ffd647292787d0183f55797fd939e857a1dc679867ab21d2258f7e7f0baf That sha256 names the PR description this verdict read. Recompute it with |
Opus 5.5 Review — ✅ no blocking findingsReviewed Review detailsNo findings. Notes:
[OPUS-REVIEWED] 7525727 Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
729ed00 to
c0917d7
Compare
c0917d7 to
ef1186e
Compare
Disposition on head
|
ef1186e to
e1036df
Compare
e1036df to
6d9c72d
Compare
Disposition on head
|
6d9c72d to
b95ed3f
Compare
eff7076 to
c11a106
Compare
c11a106 to
7783eff
Compare
Disposition on head 70f5b40First Principles BLOCK (built-in cards still post Item 10 (zero consumers): taken. Items 5-7 (card-turn gating called oversized): kept. A card turn carries words a page chose, and the page is not the person: with the gating removed, a page's one click would inherit the thread's trust and YOLO and could start sub-agents and jobs that read that trust. The owner shipped this as an off-by-default preview with those limits on purpose; loosening them is a product decision for a maintainer, not this PR. GPT (late refusal overwrites a newer reply, head 7783eff): fixed in bb8bfa2. Every composer fill while a send is in flight is tracked, so a late refusal adds its text instead of replacing the newer one; covered for both the Static Ratchet Gate ( |
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Thanks for replacing the three regexes in dashboard/card_send.py with plain checks. That request is fixed on this head. Two things still block on 7166cc5:
-
The new UI states have no rendered evidence (
website/capture/dashboard-card-send.tsx:40). The UX lane on this head said "cannot evaluate: missing the one-click send on the Dashboard tab". A verdict a lane cannot reach for missing evidence is a block, not a pass. The head still commits onlytemp-screenshots/22367/1-4, all from a staged fixture. Shot 3 draws the send-failed notice above the Profile switch with no dismiss button and no composer. The real notice renders above the crewmate's page inCrewDynamicDashboard.tsx:1418with a dismiss control. No image showscrew-dashboard-send-unsure(CrewDynamicDashboard.tsx:1432), the Profile write-failed notice, or a card reply sent on the Dashboard tab. This was item 1 of my earlier review, and it still stands. Smallest fix: add capture states that mount the realCrewDynamicDashboardthe wayCrewDynamicDashboard.test.tsxdoes, withonSendresolvingrefusedand thenunknown. Add one forCrewProfileSettingswith the write rejected. Commit the shots, recaption shot 3 to say what it shows, and re-run the UX lane. -
With the preview off, a built-in card click no longer behaves like today's
act(website/src/pages/members/MembersPageSidePanel.tsx:571).onSendis passed wheneverconfirmedSlotis set, anddashboard_card_sendis never checked. The goal-board, office and project-report answer buttons now postsend, so every default-off click makes aPOST /api/chat. Suppose the connection drops, or the answer takes longer than the 10 s abort.sendCardReplythen returns'unknown', andCrewDynamicDashboard.tsx:546shows "The reply may not have been sent" without filling the composer, though the gateway would have refused withcard_send_off. A 401 or 500 shows "Couldn't send the reply" to someone who never turned sending on. This contradicts the PR's own Not a goal ("sendbehaves exactly as today'sactuntil a crewmate is opted in"). Smallest fix: passonSendonly when the active roster row hasdashboard_card_send === true. The gateway'scard_send_offcheck stays the enforcement point. Please add one test: switch off, plus anonSendthat would reject, means the composer is filled and no notice shows.
Optional:
- An admitted card send claims its rate-limit slot and writes an "allowed" audit record before the late busy and pending-decision rechecks in
chat_send.py. A late refusal then leaves two records for one send, and the next click within 3 s gets a spurious 429. You could claim and audit only once the send commits. CrewDynamicDashboard.test.tsx:55mockstiedToFrameasdocument.activeElement === frame.hasGesturealready checks that, so removing thetiedToFramegate reds no test.useFrameLinkBridge.test.tsxshows how to drive the real one.- The draft RFC and its index row under
docs/request-for-change/still describe default on andactfor the built-in pages. Updating them here would let the reader see the design that ships.
The linked issue is still parked, and the RFC is still a draft. So once these are fixed, a maintainer still needs to sign off the design before this can merge.
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Thanks for the update. I reviewed head 3c7a859. Two earlier items still block on this head.
Status of earlier review items:
- Replace the three regexes in
dashboard/card_send.py: fixed. The head has noimport reand no compiled pattern. - Rendered evidence for the new UI states: still stands. See item 2.
- With the preview off, a card click should behave like today's
act: still stands. See item 1.
-
With the preview off (the default), a built-in card click still calls the gateway (
website/src/pages/members/MembersPageSidePanel.tsx:571).onSendis passed wheneverconfirmedSlotis set, and nothing on the client readsdashboard_card_send, even though this PR adds that field to the roster row. The goal-board, office and project-report answer buttons now postsend, so every default-off click makes aPOST /api/chat. Trigger: the connection drops or the answer takes longer than the 10 s abort. Outcome:sendCardReplyreturns'unknown', the tab says "The reply may not have been sent", and the composer stays empty. A 401 or 5xx response, or a refusal checked beforecard_send_off(card_origin_not_allowed,card_target_invalid), shows "Couldn't send the reply" to someone who never turned sending on. This contradicts the PR's own Not a goal ("sendbehaves exactly as today'sactuntil a crewmate is opted in"). The only test of the off path mocksonSendto resolve'off', so it never exercises this wiring. Smallest fix: passonSendonly whenactive.dashboard_card_send === true, and keep the gateway'scard_send_offcheck as the enforcement point. Add one test with the switch off and anonSendthat rejects or resolves'unknown'. It should show that the composer is filled, no notice appears, andapi.sendChatis never called. -
The new Dashboard-tab states have no rendered evidence (
website/capture/dashboard-card-send.tsx:39). On this head the UX lane said "cannot evaluate: missing the Dashboard tab, where the one-click send happens". A lane that cannot reach a verdict because evidence is missing counts as a block. The PR still commits onlytemp-screenshots/22367/1-4, and all four come from a staged Profile-switch fixture. Shot 3 draws the send-failed notice above the switch, with no dismiss control. The real notice renders over the crewmate's page inCrewDynamicDashboard.tsxand has a dismiss button. No image shows a card reply sent on the Dashboard tab,crew-dashboard-send-unsure, or the Profile write-failed notice. Smallest fix: add capture states that mount the realCrewDynamicDashboard, the wayCrewDynamicDashboard.test.tsxdoes, withonSendresolvingrefusedand thenunknown. Add one state forCrewProfileSettingswith the write rejected. Commit the shots at the new head, recaption shot 3 to say what it shows, and re-run the UX lane.
Optional:
CrewDynamicDashboard.test.tsx:53-56mockstiedToFrameasdocument.activeElement === frame, whichhasGesturealready checks. Removing|| !tiedToFrame(frame)therefore fails no test.useFrameLinkBridge.test.tsxon main shows how to drive the real check.- No card-turn test covers the native-crew, own-worker and own-child auto-approve exclusion in
tool_approval.py:1114. Deletingnot _card_turn andfails no test. - The draft RFC under
docs/request-for-change/still describes default on andactfor the built-in pages. Updating it here would show readers the design that ships.
Once these are fixed, a maintainer will still sign off the design before merge.
|
@KiroCrewCaptain thanks. Point by point on head be8a930: 1. Preview off, a built-in card click still called the gateway ( 2. No rendered evidence for the Dashboard-tab states: added. The live shots turned up a real layout bug: the floating Download control sat on top of the send notices' dismiss button. The notices now keep clear of it ( Optional items:
|
Disposition on head 0d8efbbGPT 6.1 (security-fenced, upheld): stopping a card turn let its in-flight spawn through. Fixed, no override. The card mark is released when the turn stops, but a request the turn sent can still be inside the handler's awaits. That covers the spawn handler's agent warm-up as well as the auth and memory reads.
|
Disposition on head ca880c3UX (BLOCK on 0d8efbb): the send mode was invisible where the person clicks. Fixed. With one-click send on, the Dashboard tab now shows a "One-click send on - change on Profile" badge, styled like the layout-lock badge. It opens the Profile tab where the switch is. It shows only when the host offers a send, which is only when the switch is on, so the badge and the click's behaviour cannot disagree. Shot 4 is the switch off: no badge, and the click fills the chat box. Shots 5-8 and 10 show the badge. Notes, all taken:
Also on this head: Backend Lint's black failure in |
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Thanks for the update. On ca880c3 all three items from my earlier review are fixed. Two new CI reds come from this diff and block the merge.
Status of earlier review items:
- Replace the three regexes in
dashboard/card_send.py: fixed. - With the preview off, a card click should behave like today's
act: fixed.MembersPageSidePanel.tsx:575now passesonSendonly whenactive.dashboard_card_send === true. - Rendered evidence for the new UI states: fixed. Shots 4 to 10 cover switch off, send, refused, unsure and the Profile write failure, and the UX lane passes on this head.
-
The
spawnsignature pin was not updated (test/test_subagent_parts_composition_contract.py:969). The PR adds_card_turn: bool = Falsetospawninsrc/kiro_crew/subagent_parts/admission_surface.py:314. The pinned string still ends at_root_session_key: 'str | None' = None) -> 'SubagentInfo | None'. As a result,test_the_manager_keeps_every_base_member_with_its_kind_and_signaturefails withAssertionError: spawnand Backend Tests (3.12, 8) is red. Smallest fix: add, _card_turn: 'bool' = Falseafter_root_session_key: 'str | None' = Nonein that string, then runpython -m pytest -n0 test/test_subagent_parts_composition_contract.py. -
Two new locale strings fail the fr and ko style tests (
website/src/i18n/locales/fr.json:22109,website/src/i18n/locales/ko.json:21045). Indashboard_card_send_badge_label, the fr value has a plain space before:("activé : un"), butfrStyle.test.tsrequires U+202F there. Indashboard_send_unsure, the ko value quotes{{reply}}with ASCII double quotes, butkoStyle.test.tsrequires ‘ ’. Frontend Tests (1) and (2) each fail one test, and Coverage Gate then fails closed. Smallest fix: put U+202F before the:in the fr value and check the other new fr keys for;:?!. Change the ko value to‘{{reply}}’. Then runcd website && npx vitest run src/i18n/style.
Optional:
CrewDynamicDashboard.test.tsx:55still mockstiedToFrameas the same focus checkhasGesturemakes. So deleting|| !tiedToFrame(frame)atCrewDynamicDashboard.tsx:517fails no test. A controllable mock plus one case with activation on, the frame focused andtiedToFramefalse would pin it.- The "Dashboard card forged send" row in
docs/system-specs/modules/security.md:1024describes only thehasGesturegate and calls the self-focus case open.dashboard-iframe-hosts.mdand the PR body saytiedToFramenarrows it. Making them agree would help security readers. - The draft RFC under
docs/request-for-change/still describes default on,actfor the built-in pages, and dropping a second reply in the window. A dated amendment listing the departures would match what ships.
Once these are fixed, a maintainer still needs to record the design decision before this can merge.
|
Pushed Both earlier blockers are addressed on the current head, and three CI reds that were red on the prior head are fixed:
Touched backend contract tests, the three touched frontend component test files, and all locale style/consistency gates pass locally. The code-scanning lane's earlier failure was an installation API rate-limit during result upload, not a finding in this diff. |
|
This branch is owned by local session chat-3354-1794188640 under lead chat-3338; other agents please do not push here. Held head: a412c33. |
|
@raymondchen625 You force-pushed c616448 here, so this branch is yours now. There is a newer review on that head (a change request from the automated reviewer at 14:57Z). Do you want us to address those findings on top of c616448, or will you take them? We will not push to this branch unless you say yes. |
|
Ownership correction: the 16:09Z note above is wrong. c616448 and the new head 0daac97 were both pushed by the owning worker session (lead chat-3338); the Raymond Chen committer name comes from the shared clone git identity, not a takeover. @raymondchen625 no action needed from you. Held head is now 0daac97; please do not push to this branch. UX BLOCK on c616448 fixed in 0daac97: the visible badge now states the outcome, "Reply buttons send right away - change on Profile", in all 13 locales (en-XA regenerated); the aria-label already said the same. Evidence shots 5-8 and 10 re-captured with the new text. Notes not taken here: the [Dashboard card] tag is the protocol marker named in the RFC and the injected-message spec, so renaming it is a separate change; the quoted reply in dashboard_send_unsure was asked for by the previous UX round. |
|
GPT finding on 525f260 (card_send.py:127, cron reads and stops refused) fixed in c9ebde4. POST /api/crons/tools is no longer matched by path. Its handler now checks the validated tool name for a card turn: cron_list, cron_pause, cron_remove and cron_remove_all pass; every other cron tool (add, update, resume, trigger, secret_request, and any added later) gets the same 403 as other delegation routes. Tests: a per-tool table in test_card_turn_subagent_policy.py and a wiring test through the real handler in test_cron_tools_forwards_session_token.py. RFC route list updated. Held head is now c9ebde4. |
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Thanks. Almost everything from my earlier review is in place on this head. One test problem still blocks.
- The send-interval tests depend on the real clock.
website/src/test/CrewDynamicDashboard.test.tsx:314and:331asserttoHaveBeenCalledTimes(1)after a third post. Oncerelease('sent')clears the in-flight flag, the only thing that holds back that third post is the interval check inwebsite/src/pages/members/CrewDynamicDashboard.tsx:523-524, which reads an unpinnedDate.now()againstSEND_MIN_INTERVAL_MS(3000). ThewaitForat:312and:329passes on its first check, because the count is already 1, so it does not wait for anything. Trigger: a loaded coverage runner spends more than 3 s between the first post and the third. Outcome: the component correctly sends again,onSendis called twice, and both tests fail even though the product works. Smallest fix: in both tests, addvi.spyOn(Date, 'now').mockReturnValue(T)with a fixed, nonzeroT(the existingafterEachvi.restoreAllMocks()undoes it). Then wait on the outcome itself, not on a count that is already 1. Could you also add one healthy-side case? After a'sent', move the stub toT + SEND_MIN_INTERVAL_MSand assert that the next post callsonSendagain. Today no test goes red if the interval never expires.
Follow-up on my earlier review:
- Rendered evidence for the new states: fixed. Shots 5 to 10 now come from the real
CrewDynamicDashboardandCrewProfileSettings, and shot 3 is gone. - Preview off still called the gateway: fixed.
onSendis passed only whendashboard_card_send === true, and a side-panel test pins it. POST /api/crons/tools: fixed. It is now named in the route test, and the handler gates it per tool, with a wiring test.tiedToFramemocked to the same focus check thathasGesturemakes: still open. See the optional notes.
Optional:
CrewDynamicDashboard.test.tsx:53-56still mockstiedToFrameasdocument.activeElement === frame, so deleting|| !tiedToFrame(frame)atCrewDynamicDashboard.tsx:517turns no test red. A mock with its own flag, plus one case wherehasGestureis true andtiedToFrameis false, would pin it. The Proof line also still names an older head.docs/request-for-change/rfc-dashboard-card-send.mdstill says the switch is on by default (line 168), that built-in pages keep postingact(line 398), thathasGesturealone is the gate, and that #22379 is the implementation. Bringing it in line with the four departures in the description would make it accurate for whoever accepts it.CrewDynamicDashboard.tsxgrows from 1,534 to 1,713 lines. Moving the send state into a sibling hook would keep the file under 1,600 lines and give the interval tests one place to inject a clock.
…-default preview A button on a crewmate's Dashboard tab can send its reply into that crewmate's own thread in one click (kirocrew-dashboard:send), instead of only filling the chat box. It is a per-crewmate preview, off by default: until a crewmate opts in on its Profile tab, send behaves exactly as act. The send crosses a trust boundary, so the person must be the one who sends: the host sends only when the click is tied to the page's own frame (tiedToFrame, which a page's self-focus() cannot pass), and the gateway re-checks the payload and refuses what a card may not do (busy thread, open approval or question, rate limit). A card turn is tagged [Dashboard card], never counts as an approval, carries no owner directive authority, and inherits no auto-approval even in a trust thread. The sandboxed frame's CSP is unchanged. Fixes #22367
Per review: the repo takes no new regex in product code. _CONTROL_CHARS becomes a per-character ord() test. _DELEGATION_ROUTES and _TASKRUNNER_STOPS become sets of segment shapes after /api/, where None is any one non-empty segment and one trailing slash is dropped. Segments come from yarl.URL.build(path=..., encoded=True).parts (no decoding, empty segments kept). Same accept/refuse result as the old patterns on every route; tests pin trailing-slash, empty-segment and extra-segment cases and more control characters. Also: a different reply posted inside the send interval (or while a send is in flight) now fills the composer instead of being dropped; only a repeat of the same text is dropped as a double click. A send refused after a newer reply was filled while it was in flight is added to the composer (onActKeep -> mergePaneDraft), so it cannot swap out that newer reply.
|
Thanks for the steady follow-through on this PR. I reviewed head 7525727 and found no blocking defect. This changes behaviour users already rely on, so it needs a design sign-off from a maintainer before it can merge. A maintainer will post the decision here. To propose a design, follow the request-for-change process in CONTRIBUTING.md (docs/request-for-change/). The draft RFC for this feature already exists, so acceptance would be recorded in that RFC. Status of my earlier review items:
What this review checked on this head:
Optional:
|
bolichen97
left a comment
There was a problem hiding this comment.
Approved at 7525727. The direction is signed by the maintainer: an ordinary feature behind a per-crewmate preview, off by default, where off behaves exactly like main.
Our earlier point is fixed. card_send.py no longer uses re at all, and a search of the added product lines found no new regex in Python or the frontend. The control-character check in CrewDynamicDashboard.tsx is the existing actText pattern moved unchanged into replyText.
What I checked:
- The gate:
card_send_refusalruns before the slot is created. It admits only the owner's dashboard (no app token), and only an existing crewmate DM whose binding matches. The crewmate'sdashboard_card_sendmust be a realTrue. The text is checked (trimmed, at most 600 chars, no control characters), the thread must be idle, nothing may be open (no approval, permission or question), and there is a 3 s per-thread limit. Every decision is written to the SEL. - After admission: the body is reduced to the text and the thread. The turn carries the
[Dashboard card]prefix and no directive or owner authority. Every auto-approve rung is skipped for it (hook list, native crew, trusted patterns, trust-reads, trust and YOLO), while a hook deny still denies. - Delegation from a card turn: delegation routes, the non-read cron tools and sub-agent spawns are refused, using the arrival-time mark so stopping the turn mid-request can't unmark it.
- Re-runs: continue, regenerate, recovery requeue, queue drain and the shared-crew restart note all carry the card mark, so a re-run stays a card turn.
- Non-card sends: every new branch is behind
card_originorcard_turn_seen, so typed sends behave as before. - Host: the host sends only on a click tied to the frame, to the tab's own confirmed slot. Any refusal fills the composer, and an indeterminate result shows a notice instead of resending.
CI: all 100 checks are green on this head. Sensitive Change Review's earlier rate-limit red passed on rerun.
One doc fix, please: line 174 of docs/request-for-change/rfc-dashboard-card-send.md says default on, but the code ships off and off is what is signed. Please correct it, and record the maintainer's acceptance in that RFC (status and section 7).
|
Demo video: a button on a Bot's Dashboard page sends its message to the Bot, and the Bot answers in its chat. Recorded on main 01fcf6d. pr33203.mp4 |
Problem / Motivation
Goal: A button on a crewmate's Dashboard tab can send its reply into that crewmate's own thread in one click, behind a preview that is off by default - instead of only filling the chat composer.
Today a page posts
kirocrew-dashboard:act {text}, the host validates it and puts the text in the composer, and the person must then switch to the chat and press send. For a card that offers a reply ("Approve the plan", "Show me the diff"), that is two steps for a choice the person already made with one click. The page is agent-written HTML, so the host never lets it send on its own - a page that could speak as the person would let a crewmate answer for them.Why it matters
The built-in cards (a needs-you row with offered answers, a review row with two buttons) are the one place the second step adds nothing: the person chose the reply by clicking it, then has to find the chat and press send again. Issue #22367 asks for that click to send. The gap is small per card but constant, and it is exactly the interaction the Dashboard exists to make direct.
Not a goal
sendbehaves exactly as today'sactuntil a crewmate is opted in.focus()residual the RFC names; this reduces it withtiedToFrameand leaves the host-confirmation / trusted-click-bridge follow-ups to a later change.What changed (motivation -> approach -> change)
Goal: let a person's one click on a card send the reply. Approach: add a second page-to-host message rather than widen
act, gate the send so only a person's real in-frame click can trigger it, and ship it off by default so it changes nothing until opted in. It followsdocs/request-for-change/rfc-dashboard-card-send.md(on main,status: draft), with four deliberate departures noted below.kirocrew-dashboard:send {text}(same text rules asact: trimmed, <=600 chars, no control char except newline). The host sends it as a user turn into the Dashboard tab's own crewmate DM thread, nowhere else.tiedToFrame, the stronger of the gates the PR-link bridge uses: it holds only when focus came into this frame straight from the dashboard and no click or key outside it is still live, so a page script that callsfocus()on its own frame after a click elsewhere does not pass.dashboard/card_send.py), which refuses what a card may not do: not the owner's own dashboard, not an existing crewmate DM thread, the preview off, the thread busy, an open approval / question / permission prompt, or inside the 3-second rate limit. Every refusal falls back to filling the composer, so a reply is never lost.[Dashboard card], never an approval, no owner directive authority, and no inherited auto-approval - its tool calls reach the approval card even in atrustthread (hook allowlist, trusted patterns, trust-reads, session trust, scoped grant, YOLO and native-crew all skipped and audited). It cannot spawn sub-agents or reach any delegation route.dashboard_card_sendconfig field defaultsfalse; a Send replies in one click switch on the Profile tab turns it on. With it off the host offers no send at all (MembersPageSidePanelpassesonSendonly when the roster row hasdashboard_card_send === true), so a page'ssendfills the composer exactly asactdoes and nothing reaches the gateway; the gateway'scard_send_offcheck stays the enforcement point. With it on, the Dashboard tab shows a "One-click send on - change on Profile" badge (styled like the layout-lock badge, opens the Profile tab), so the person sees what a card click will do where they click. The sandboxed frame's CSP is unchanged.send- goal-board's offered options, office's two review-row buttons, and project-report's needs-you choices. A starter the person finishes keeps postingact: goal-board's and the conductor board's Reply ("About ...: ") and project-report's "Write my own". With the switch off (the default), or on any refusal, the host fills the composer exactly as today.card_send.py: the control-character rule is a per-characterord()test, and the delegation and task-runner-stop routes are sets of path-segment shapes after/api/(None= any one non-empty segment), matching exactly the routes the earlier patterns matched.Departures from the RFC: (1) default off, not on (RFC section 3.5) - shipped as a preview so it is not a product-shape change for every crewmate; (2) the send gate is
tiedToFrame, nothasGesturealone - RFC section 4.1 accepts the weaker check with the page-self-focus()case as a residual;tiedToFrame(added to the repo after the RFC) does not pass that case; (3) the built-in pages' finished answers postsend(the RFC's compatibility section kept them onact), so the cards its own motivation names get the one click; (4) nodashboard_frameconstants foract/send(RFC section 3.1): nothing on the server read them, so the literals are pinned between the tab and the built-in pages by a test instead.Backwards compatibility
Compatible:
dashboard_card_sendis absent on every existing record and reads as off, and with it off asendfrom a page fills the composer exactly asactdoes, so the built-in pages' switch tosendchanges nothing anyone sees until they turn the preview on. Custom pages that postactare untouched. The host and gateway ship in one build. No field is rejected, renamed or narrowed.Removes nothing: the diff deletes no public name and moves no helper. The crew-steer split this PR once carried is now on main (#33545), and this PR keeps the layout from main.
Tests
New and updated tests, each of which fails with the fix removed (mutation-verified): the accepted send and every refusal (
test_dashboard_card_send.py); the untrusted-turn gate across every auto-approve tier (test_card_turn_untrusted.py- forcing the card flag off fails 7); no sub-agent / no delegation (test_card_turn_subagent_policy.py); the preview defaults off, round-trips, non-bool reads as off (test_crew_card_send_setting.py); preview off = fill the composer (today's behaviour, pinned) and a send only on a frame-tied click, a page script cannot send (website/src/test/CrewDynamicDashboard.test.tsx- removing the gate fails the "a script cannot send" cases); the Profile switch default-off and write (CrewProfileSettings.test.tsx); the wire carriesorigin: dashboard_cardand nothing a typed send does not (sendCardReply.test.ts,crewDashboardAct.test.ts). Each built-in page pins which buttons postsendand which stayact, and fails if a finished answer goes back toact(test_dashboard_template_goal-board.py,test_dashboard_template_office.py,test_dashboard_template_project-report.py); the switch-off fallback to the composer is pinned inCrewDynamicDashboard.test.tsx("fills the composer without an error when the person turned one-click sends off"). The route shapes are pinned one case per route, with trailing-slash, empty-id, extra-segment and lookalike-prefix cases (test_card_turn_subagent_policy.py), and a distinct reply inside the send interval fills the composer (CrewDynamicDashboard.test.tsx).Proof: PROVEN - backend: forcing
_card_turn=Falsefails 7 untrusted-gate tests; frontend: removing the frame-tie gate fails the "a script cannot send" / "click elsewhere" cases (head 729ed00).Local gates green on the changed surface:
pytest -n 0(card-send + contract + steer suites),vitest(frontend) + i18n catalog parity,tsc,eslint,isort,flake8,mypy,black,check_file_line_cap.py(0 failures),config-baseline.jsonregenerated. CI runs the full cross-surface suite.Manual verification
Real-browser run of the real components (
website/scripts/capture-22367-card-send-live.mjsoverwebsite/capture/dashboard-card-send-live.html, Vite dev server, network stubbed in the page, no gateway). For each Dashboard state the script clicks the page's own button inside the sandboxed frame, so the host's real gates (hasGesture+tiedToFrame) decide, then asserts the branch before the shot:[Dashboard card]row, no notice, the chat box stays empty;The wiring with the switch OFF is pinned in
MembersPage.sidePanel.test.tsxon the real side panel: a page'ssendfills the composer andapi.sendChatis never called (fails if thedashboard_card_sendgate is removed).Screenshots / video
Every shot mounts the real
CrewDynamicDashboardorCrewProfileSettings. The "Thread" and "Chat box" strip under the frame is the fixture's stand-in for the Members page's transcript and composer: it shows whatonSendsent and whereonActfilled the reply. Committed undertemp-screenshots/22367/(the review lanes read committed media there):Related Issues
Fixes #22367
Checklist