Repository navigation
feat(followup): opt-in one-click send for follow-up Start, with a prompt preview - #25926
Conversation
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: the new mode is opt-in, stored only in the browser and never restored from the agent-writable backup, and the default still pre-fills as it does on main. Notes
[DESIGN-REVIEWED] 0cf21e1 Description read[DESCRIPTION-READ] c1ffb73738e65990df25d84a73f5727843f1458087dca825b86c0909213a3941 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 The relabel of "Start in new worktree" to "Start in a separate copy" changes the label for everyone, and the only reason given is a reader nobody names. Nothing blocks the merge. Notes
What this change shipsInventory (8 items) — 7 justifiedIntent: let a person who opts in start a follow-up suggestion with one click that sends its prompt, while Start keeps pre-filling by default. This is an ADDITION.
[FIRST-PRINCIPLES-REVIEWED] 0cf21e1 Description read[DESCRIPTION-READ] 6e7d515addc7b04b84dc261b90d251131b44ef1cca2ed80b6e6bd0564b6f75a6 That sha256 names the PR description this verdict read. Recompute it with |
UX Review (Opus 5.5) — ✅ PASSUX-level review of UX-Verdict: PASS Nothing blocks. The blind reader read both modes of the primary button correctly. They hesitated before "Run in a separate copy" only until they could read the whole prompt, and the PR's own Show prompt toggle removed that hesitation (shot 06). Notes
[UX-REVIEWED] 0cf21e1 Description read[DESCRIPTION-READ] cd771964ba227f0ff10d68096c6769e9cd190e9055774e733cba0edab8720f6b That sha256 names everything this verdict read: the PR description, plus the 18 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 — ✅ 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 Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
bolichen97
left a comment
There was a problem hiding this comment.
Thanks for this. The goal is clear and the send-before-switch ordering plus the Show prompt toggle are careful work. Requesting changes for these points at 27c3a32:
-
Product shape. Start now runs agent-authored text on one click, which replaces the pre-fill the base code documents as a deliberate product constraint ("a click can never start an unattended turn"). That is a consent change, so it needs an accepted RFC under docs/request-for-change/ or an explicit maintainer design sign-off before it can merge. The First Principles lane blocks on the same point.
-
Worktree retry after a refused send cannot succeed. When the send returns false on the worktree route, the scoped session is kept. The retry message tells the user to press Start again, but the reused worktree then hits the in-use check in the worktree create handler (a slot is already scoped to that path) and returns 409 every time. Either delete the just-created slot on a refused send, the way the scoping-failure path does, or retry the send into that same slot, and reword the message and the guide to match.
-
The owning spec is not updated. docs/system-specs/modules/learn-cron-dashboard.md still says both card actions pre-fill a composer and never send, and that the worktree route activates and then prefills. Please update that paragraph in the same commit: send, clear the card, then switch, plus the not-sent, not-opened and not-focused outcomes and the Show prompt toggle.
Also worth folding in: a transport error after the server has launched the turn makes send() return false, so the card says nothing was sent and invites a second run of the same prompt. Treating an indeterminate result separately from a refusal would avoid that duplicate run.
CI is red on this head (frontend, backend and coverage jobs), so those need a look too.
| // launch uses. Resolves false only when nothing was delivered; the card | ||
| // item owns that report, so the send posts no page notice of its own. | ||
| const delivered = await send(item.prompt, slotKey, undefined, true, true) | ||
| if (!delivered) throw new Error(fail.notSent) |
There was a problem hiding this comment.
A refused send on the worktree route leaves a retry that always fails.
When send returns false here, the new slot stays scoped to the worktree. The not-sent message asks the user to press Start again, the create call reuses the worktree, and the handler's in-use check finds that slot scoped to the same path and refuses with 409. The prompt can never be sent from the card after the first refusal.
Suggestion: on a refused send, delete the slot just created (as the scoping-failure path does), or retry the send into this same slot, then make the message and the guide say what actually happens.
Separately, a transport error after the server launched the turn also returns false here, so a retry can run the prompt twice. Consider telling "refused" apart from "unknown".
27c3a32 to
e90a02e
Compare
|
|
|
|
|
|
|
question (blocking): product shape. Agreed this reverses a recorded consent rule. Pearce asked for Start to run the prompt, and I have asked him to get a maintainer decision or open an RFC. Would a design sign-off here from you be enough, or do you want an RFC under docs/request-for-change/ first? issue: worktree retry after a refused send. Fixed in e90a02e. A send that never reached the server now closes the session it created, so pressing Start again reuses the worktree. The retry message and the guide say so. issue: owning spec. Fixed in e90a02e. The learn-cron-dashboard.md paragraph now covers send, clear the card, then switch, the not-sent, unconfirmed, not-opened and not-focused outcomes, and Show prompt. suggestion: lost receipt. Done in e90a02e. A transport error is now unconfirmed, not refused: the card clears, the session is kept, its first message shows the pending-delivery mark, and a notice says to check that session, so no second run is offered. note: the red CI on 27c3a32 came from tests this PR does not touch; they pass after the rebase onto current main. |
|
e90a02e to
f6353bb
Compare
|
|
bolichen97
left a comment
There was a problem hiding this comment.
Re-reviewed at f6353bb against our earlier request at 27c3a32.
Fixed:
- Worktree retry after a refused send. A refused send now deletes the just-created slot before reporting not-sent. The worktree create handler builds its in-use list from live slot projects, so the retry no longer hits the 409 "a session is already open in it" refusal. The new tests assert the delete on refusal and no delete on an unconfirmed send.
- Owning spec. docs/system-specs/modules/learn-cron-dashboard.md now says Add to this session pre-fills and Start sends, and lists the refused, unconfirmed, not-opened and not-focused outcomes plus the Show prompt toggle.
Also addressed: a lost receipt is now treated as unconfirmed rather than refused, so the card no longer invites a second run of a prompt that may already be running.
Minor, non-blocking: the delete on a refused send swallows its own failure. If that delete fails, the card still says to press Start again and the worktree retry returns 409 as before. Consider a different message in that case, or leave as is if you judge it rare enough.
Still open: the product-shape question. Start now runs agent-authored text on one click instead of pre-filling the composer, and that needs an accepted RFC or an explicit maintainer design sign-off. That decision is still pending with the maintainers, so this is a comment rather than an approval. The GPT lane finding on the duplicate-start guard resetting on route navigation also still stands on this head.
f6353bb to
aef533c
Compare
76c05e1 to
e578b40
Compare
|
|
@bolichen97 the Run-mode footer change from your 06:12Z review is in on e578b40. note: With one-click send on, the footer now opens with "Sends right away because Send follow-ups with one click is on. Turn it off in Settings." The second sentence is a note: The branch is rebased onto current main again. Main changed |
|
The branch now conflicts with main. Could you rebase it onto current main, keeping one or two commits for PR Hygiene? In the same push, please take the open point from my last review. With one-click send on, the Run-mode footer still does not say that this behavior comes from a setting, or where to turn it off. At 5715df6, I'll approve with auto once the rebased head has that and is green. |
|
note (non-blocking): rebased onto main as one commit, and each Run-mode footer string now opens with the reason and the way back. Head 1900001 sits on main eb0ecb9. In one-click mode the three footer strings ( At 5715df6 the card drew that lead from two separate keys placed before the string you quoted, so the string alone did not say it. Those keys ( |
resolution: narrowed -- the worktree copy wording stays as the maintainer set it; further footer copy changes are out of scope for this PR |
|
|
|
|
@bolichen97 question (non-blocking): could you record a UX override for 69457a0? The UX lane has now flipped three times on copy that did not change. It passed f18601d, blocked 47ecf5a and d6cde3b, passed 9970f77, and blocked 69457a0. The labels, footers and send glyph are byte-identical across all five heads, and the rebases between them touched only generated files, the bundle budget and test assertions. Every other lane passes on 69457a0: Design, First Principles, GPT 6.1 and Opus 5.5. The footer opens with "Sends right away because Send follow-ups with one click is on. Turn it off in Settings → Chat → Composer.", as you asked on Nov 8, and the setting is off by default. If you would rather change the wording first, tell me which line and I will push it. The command is |
|
|
Update (08:40Z): UX passed on 0cf21e1, so no override or change is needed. Every lane is green on that head (Design, First Principles, GPT 6.1, Opus 5.5 and UX), and all 97 checks pass. If you approve, please leave auto-merge to me. Main has moved 36 commits since this head's base, so I will rebase onto it first and arm auto-merge once that head is green. Earlier question, kept for the record and no longer needed: @bolichen97 question (non-blocking): the rebase onto main moved the head to 8ec1bd4, so the UX override I asked for on 69457a0 no longer applies. UX now blocks for a different reason: in one-click mode, Run sends a prompt whose fourth line onward is behind Show prompt. Design, First Principles, GPT 6.1 and Opus 5.5 all pass on this head. The three-line preview is Decision 3 of the RFC, settled in your review, so I have not changed it. Which do you want?
|
|
The design is signed off: the RFC What the review checked on 0cf21e1:
It found no blocking issue. The failure table is turned into tests one row at a time, and the consent fence is tested from the attacker's side as well. For the security review:
Optional:
|
bolichen97
left a comment
There was a problem hiding this comment.
Approving on 0cf21e1. All 97 checks are green, including UX Review and PR Readiness.
The last point from my review at 76c05e1 is fixed. Each one-click footer (footer_worktree_start_sends, footer_start_sends_add_prefills, ..._project_directory) now opens with "Sends right away because Send follow-ups with one click is on. Turn it off in Settings → Chat → Composer.", and the setting's own name is a SettingsLink to that control. The path is built from the tabs' own labels, so it matches the nav, and the spec says the same.
Also in this head, compared against 76c05e1: a refused send now discards the empty session through deleteSlot({ discard: true }) and offers the retry only when that discard really closed it. A 404 or a refusal counts as kept, so a retry never meets a session that comes back holding the copy. The sidebar "could not close" notice stays quiet there, because the kept-session notice already says what to do.
The design is the one agreed in review and in rfc-followup-start-sends.md. Pre-fill stays the default for both buttons. One-click send is opt-in under the browser-local mc-followup-start-sends, outside the host backup. No regex in product code, no new restrictions.
Problem / Motivation
Goal: A person who turns it on can start a follow-up suggestion with one click, so its prompt goes to the new session as the first message. By default, Start keeps pre-filling the new session as it does today.
Today "Start in new worktree" and "Start in new chat" open the new session with the prompt in its message box, so every start needs a second press of Send. Some people want the click on Start to be the decision to run it. A click that sends agent-written text changes consent, so it is an opt-in and the default does not change.
The card also has two problems in both modes. The button says "worktree", which a first-time reader does not understand. The card shows only the agent's title and description, not the text that runs.
Why it matters
A follow-up exists to hand work to a new session in one step. With the pre-fill, a person who switches away before pressing Send leaves an empty session behind with the prompt in its draft. The people who want one step can now have it, and nobody else gets a click that runs agent text.
Not a goal
What changed (motivation → approach → change)
The setting is the browser-local key
mc-followup-start-sends(website/src/pages/chat/followupStartSends.ts), shown as "Send follow-ups with one click" under Settings, Chat, Composer (website/src/pages/settings/ChatPanel.tsx). It is read throughusePersistedBool, so only the stored string'1'is on. It is not aChatConfigfield and not inDURABLE_PREF_KEYS, because the uiPrefs host backup is agent-writable (themc-yolo-ackrule).ComposerDockView.tsxreads it andwebsite/src/components/FollowUpCard.tsxtakes it asstartSends. That one value picks the label (Start or Run), the send glyph and the footer, and goes to Start asmode: 'send' | 'prefill'. Each one-click footer opens with why it sends and a link to turn that off. InopenFollowupSession(website/src/pages/useChatPageSlotSwitches.ts) only an explicit'send'sends.The pre-fill path is main's code and keeps main's worktree messages. The one-click path uses new
*_sent_*keys that point at the sidebar.The one-click path sends
item.promptverbatim through the composer'ssend()to the new slot by key, before the switch. A newsend()callback incomposerSend.tsreports a refusal or a lost answer. The RFC's failure table applies:deleteSlot({ key, discard: true }), which refuses unless the session is still empty.recordSendAttemptkeeps the prompt on ↑.captureForkOrigin,forkMayOpen).After the person leaves the Chat route, failures go to the shell banner.
Each item shows the first three lines of its prompt, with "Show prompt" for the rest. The worktree button reads "Start in a separate copy", its tooltip names the git worktree, and both of its footers say what the copy is. Every footer says "message box", and every worktree message says "copy".
The tool description in
schemas.py, the model note insession_directive_apply_slot.py,followup-suggestions.md, the spec paragraph inlearn-cron-dashboard.mdand the trust-model doc describe both modes.rfc-followup-start-sends.mdand its index row record the 2026-11-05 amendment. New strings are translated in all 12 catalogs, anden-XA.json, the settings registry and the UI index are regenerated.Security note. With the setting off, nothing runs until the person presses send, as on main. With it on, one click sends agent-written text as a user turn in a new session. Before that click the person sees the title, the description and the first three lines of the exact prompt, as plain text, with the rest one toggle away. The prompt is still redacted server-side when the card is posted. The setting is off by default, is never restored from the host backup, and is not writable by a change card.
Backwards compatibility
Compatible. The default behaviour of Start is unchanged, and the
suggest_followuptool's arguments and result are unchanged.The two footer keys
both_actions_pre_fill_the_composer_nothing_is_seandnew_chat_footer_same_project_directoryare replaced byfooter_both_prefillandfooter_both_prefill_project_directory, which say "message box".FollowUpCard.tsxwas their only reader.Tests
ChatPageFollowup.test.tsx, with the real card and composer: by default Start creates the worktree, scopes the session and puts the prompt in its message box with no send, and the new-chat route does the same. With the setting on, Start sends the prompt once to the new slot and leaves every message box empty. A double click sends once. A refused send shows on the item with no second notice.ChatPageW3Coverage.test.tsx: the pre-fill block covers the prefill channel, a route called with no mode, no key, a switch that fails with the item kept, and focus that never arrives. The one-click block covers send before switch, the refused, kept and close-failed rows, the withdrawn close notice, both unconfirmed wordings, ↑ recall, and a person who moves during the start.FollowUpCard.test.tsx: the default footer and no send glyph, the one-click footers and glyph, the mode passed to Start, the plain label and the tooltip, the three-line preview and Show prompt.SettingsChatPanelCoverage.test.tsx: the toggle is off by default, writes'1'and'0'under its own key, never touchesmc-chat-config, and the key is not inDURABLE_PREF_KEYS.ChatPageFollowup.test.tsxalso checks that a stored'true', or afollowupStartSendsfield in the chat-config blob, keeps Start pre-filling.switchSlotCallsiteClassification.test.tsclassifies the secondswitchSlotsite.Proof: not provable -- frontend test
Manual verification
I rendered the built SPA with every API call stubbed and dummy data. With the setting off, Start in a separate copy opened the new session with the prompt in its message box and sent no request to
/api/chat. With the setting on, onePOST /api/chatcarried the prompt to the new session.Screenshots / video
The images are in
temp-screenshots/followup-start-sends/.Before, the card footer says both actions pre-fill:
Before, Start opens the new session with the prompt pre-filled:
After, by default: the plain label, no send glyph, the first three lines of each prompt, and a footer that says nothing is sent:
After, the first item's prompt expanded:
After, by default, Start opens the new session with the prompt in its message box:
After, by default, in a project that is not a git repository:
The setting, off and on:
With the setting on, Start carries the send glyph, and the footer opens with why it sends and where to turn that off:
With the setting on, Start sends the prompt and the new session runs it:
With the setting on, in a project that is not a git repository:
With the setting on, a refused send shows the server's reason and the retry on the item:
With the setting on, a refused send whose session will not close removes the item and shows one notice that points at the sidebar:
With the setting on, a lost answer says the dashboard could not confirm the prompt reached this session:
Related Issues
Design: #26380 (
rfc-followup-start-sends.md), amended 2026-11-05 in review of this PR.Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)