Repository navigation
feat(slack): per-row PR and CI, the combined Code tab and wrap-up outcomes on the work board - #38181
Conversation
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS No design-level blockers. The new Notes
[DESIGN-REVIEWED] 5064b8f Description read[DESCRIPTION-READ] a0c666abd5d2a71ffdbf29b020e860ba6aa9ae04ae724c92c7a05500c1d66592 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: every item is one the base RFC What this change shipsInventory (9 items) — 9 justifiedIntent: in a conductor-led Slack code channel, see each helper agent's PR, CI, commits and outcome without opening the dashboard. This is an ADDITION.
[FIRST-PRINCIPLES-REVIEWED] 5064b8f Description read[DESCRIPTION-READ] 35966756e72fa4b2590193c5ecacb61d506bc328116801bce3850656c8e24d2e That sha256 names the PR description this verdict read. Recompute it with |
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: |
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: |
ea16bae to
2f366a3
Compare
2f366a3 to
1523fc3
Compare
The idle !stop reply sent people to the work board for a link only a waiting row carries. It now names each running helper agent by its board title, linked to its chat with the board's own link builder (no click token), up to five, the rest folded into one line pointed at the lead session. With no dashboard origin it names them and points at the lead session. The helpers are read under the board's scope check after the link builder's await, right before the send.
…comes on the work board The second half of the work board, split from #21993. Each board row now shows its helper agent's PR and CI: the number the worker reported, read from the channel's own repository (by number, else by branch), cached, concurrent and off the channel lock, with CI counts from the dashboard's check read. A worker whose session works in another repository shows no branch, PR or CI. The Code tab shows the combined change: the channel's own diff, then each worker's own commits under a header naming it, read from the channel checkout's object database only for a tip reachable from the worker's own tree, screened and redacted as the channel's diff is. setView takes a required send gate the client asks last, right before the raw send; a refused or failed worker tab is replaced by the channel's own or taken down, and leaves the pass unsent. The wrap-up lists one outcome line per helper agent, left out when the board's scope refuses the channel at the post or any re-send. Workers are not filed into folders, and the context bar carries no worker. Three board tests now fail under an identity redactor (they passed with redaction removed).
1523fc3 to
5064b8f
Compare
|
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/). Review result on head 5064b8f: no blocking findings. The review read the full source diff and checked every caller of the changed signatures ( Optional:
|
Head commit: 5064b8f (base
main)Problem / Motivation
Goal: In a conductor-led Slack code channel, show each helper agent's PR and CI on the work board, its commits in the channel's Code tab, and its outcome in the wrap-up, and make the idle
!stopreply link each running helper agent.#21993 put the work board in the channel; it was split at the owner's decision (board message first), and this PR is the second half: Kyle's ask was "we need to show all the sub-session states for conductor, the diff, make the integration GREAT". Today the board names each helper agent and what it is doing, but not its PR or CI; the Code tab shows only the channel's own diff, so a helper agent's committed work is invisible in Slack; and the wrap-up lists the channel's own PR but not the helper agents'. The UX review of #21993 also found that the idle
!stopreply sent people to the board for a link that only a waiting row carries.Why it matters
Without these, the owner still opens each helper agent's session in the dashboard to see whether its PR is open, green or merged, and what it changed. The
!stopreply promised a link the board does not have.Not a goal
session_createalone files them; the board writes nofolder_id).What changed (motivation → approach → change)
The idle
!stopreply (commit 1). It named a count and said "open it from the work board", where only a waiting row has a link. It now names each running helper agent by its board title, linked to its chat with the board's own link builder (no click token, a URL the redaction would change or that breaks a link is not linked, titles escaped), up to 5; the rest, and any the board has no row for, fold into "and N more", pointed at the lead session (linked when it can be). With no dashboard address it names them and says to open them from the lead session in the dashboard. The helpers are read under the board's scope check after the link builder's await, right before the send.Idle and waiting rows keep sharing the empty circle: Slack's
planblock has only four task statuses (pending, in progress, complete, error) and a waiting row is none of the other three (the spinner would say it is moving, the error mark that it failed), so the difference is carried in words a screen reader also announces: "Waiting on you: " leads the title, and the threaded reply pages the owner.Per-row PR and CI (commit 2,
workspace/worker_prs.py). A row's PR is the number its worker reported, then each refresh reads that PR (a URL built from the channel's own repository, so a number can only name a PR there) or, with no number, its branch's PR, for every row either look can show, concurrently, off the channel lock, cached 60 s (900 s once merged or closed). A worker whose session works in another repository (host and owner/name compared case-insensitively), or one whoseoriginthe channel cannot compare, gets no branch, PR or CI. The row showsPR #12(linked only when https and left whole by the redaction), its state when not open, andCI 4/6 passing · 2 running(or· 1 failing) for a live PR (from an internalciCountsmarker of the dashboard's check read, stripped before the chip cache). The summary line adds· N PRs open. A forge failure leaves the rows without PR and CI; the board still goes out.The combined Code tab (
workspace/worker_diff.py,service._build_diff). With a board, the tab is the channel's own diff, then one group per helper agent with commits of its own, underHelper agent 2 of 5: <title> · branch <b> · PR #12 open · working. Each group ismerge-base(channel HEAD, branch) .. branch, read from the channel checkout's object database, only for a branch name git accepts as a branch and only when its tip is reachable from that worker's own tree (a claimed branch shows nothing), through the channel diff's whole-file screen and the outbound redaction. One raw read budget is shared by the helper agents' reads (the channel's own read keeps its own); a worker left out adds the truncation note.setViewnow takes a required send gate thatCodeChannelsClient._callasks last, after every await (a gate that raises refuses): refused there, the channel's own diff goes instead. The workspace record keeps whether the tab up may carry helper agents' commits (diff_workers): written to disk before a send built with the board, kept across a restart, a rebind and a lost answer, and cleared only on Slack's word (a channel-only tab or a removal answered, or a send that never left). While it is set and the board is refused (shared, Off, Review), the tab is replaced by the channel's own or removed; one whose id was never learned is replaced by one fixed line (This Code tab was withdrawn.) and that id removed. A takedown Slack refuses stays owed and is retried at most every 60 s doubling to 30 minutes (the watch re-arms it after a restart), and the owner gets one dashboard notice; nothing is retried while Slack has the channel's views off. A channel Slack no longer lists us as the agent of has its owed tab taken down just after it is unbound (so no refresh's send gate, which reads the stored binding, admits helper agents' content past it), best effort: Slack may already refuse that write, and nothing retries it while the channel is unbound. When nothing can be built, such a tab is taken down once the board's scope refuses it, and kept while it admits it, the pass re-armed until it builds. A tab that carries helper agents, or must replace one, and did not land (a refusal, a failed send or removal, a build that raised) leaves the pass unsent and re-armed, with or without a board read.Wrap-up outcomes (
board.outcome_lines,wrapup.py). The wrap-up lists one line per helper agent (• Fix login: PR #12 merged,rejected, PR #9 closed,lost, no PR), from the board this process published for this conductor or a fresh scope-checked read filled as a refresh fills it, capped at 1,800 characters then+N more. The lines are left out when the board's scope refuses the channel at the post, and on every re-send of the message (the retire edit, the archive edit and its restore).Carried from #21993's review. Three board tests passed with redaction removed (M1-M3); each now fails under an identity redactor. M4: the dashboard Board says "stopped" for a run with open items and no active patrol, the Slack board does not; I recommend adding it in its own change, not here.
Backwards compatibility
Compatible: a channel with no board keeps the plain single diff and its bar; the new
WorkerRowfields andFacts.workershave defaults;set_diff_view's newgateis keyword-only and both its callers pass it (the withdrawal's fixed line passesgate=None).Removes nothing: only the idle
!stopreply's wording changes; no key, record, route or caller is removed.Tests
test/test_slack_code_channel_board_diff.py(new): the combined diff (headers, order, a key edited inside a PEM, whole-file reads whatever the repo's diff config, the shared raw budget, a claimed branch, the plan's rows, settled rows past the view), the Code tab's scope (refused after the build and at the send, a tab taken down when no channel-only diff can be built, a takedown owed after a refused read retried until it lands, and still owed after a restart, a build that raised or a failed removal re-armed, a transient build failure keeping an admitted tab, git listing no tree, the client's gate asked last and failing closed), per-row PR and CI (by number over a reused branch, by branch, cache per host, off the lock, another repository, an origin the channel cannot compare, one differing only in case, a failing rollup, both looks, theciCountsmarker kept out of the chip cache, a failing forge read), and the wrap-up (outcome lines, the cap, the scope at the post and at every re-send, a Stop in the retire task, the fallback read filled like a refresh, a replaced conductor).test_slack_code_channel_commit_point.py:set_diff_viewrequires a gate, the client asks it last with no await before the send, and every caller passes one; no board module nor the new worker readers assigns afolder_id. The idle!stoptests sit with the board and routing suites (including a scope change during the link builder's await), and the routing suite takes an owed Code tab down once a revoked channel is unbound, and unbinds when that takedown raises. The needs-you reply and its tests are main's, unchanged. Every test added after the first review round failed on the head before its fix. Mutation checks: removing the send gate, the PR fill, the unsent-on-failed-tab rule or the take-down-on-refusal rule each fails a named test.This head is de08c6c69af7b0aad4d22771f8c74da2aabc53e0 rebased onto main 537166e, identical in
git range-diff. On it everytest_slack_code_channel*suite, the chip and source-provider files, everytest_*ratchet*.pygate with the PR base set and the config-baseline, content-scan and facade-mirror gates pass (2,879 tests). Two main commits earlier, on the same two commits: everytest/test_slack*.pywith those gates and 38 more test files that name a module this PR changes (9,559 tests), three repeat passes of the 6 files CI's stability check selects (677 each), docs lint and the source gates all pass. Local checks before that rebase, on de08c6c69af7b0aad4d22771f8c74da2aabc53e0 (two commits on main 8cc64c0, identical ingit range-diffto the previous local head on 084b4c1): the full pre-push gate (source gates, docs lint, static and flake ratchets, flake8, isort, black, mypy, line cap, the changed-plus-importer test files) passes; everytest/test_slack*.pyfile plus the chip files and the other changed test file (196 files, 7,392 tests) pass; three repeat passes of the 9 changed test files (902 tests each), as CI's stability check runs them, pass; CI's new-test stability replay (the 6 filesscripts/new_test_stability_scope.pylists, three passes with-n auto) passes five rounds in a row with eight busy loops loading the 16-CPU host; every ratchet and tree-scan test file with the PR base set as CI sets it (224 files) gives 12,612 passed and 3 failed, and those three fail the same on a clean main 8cc64c0 on this host (test_client_bindings.pyandtest_runtime_bindings.py, from the local Python 3.13'spathlib._local, and thetest_sandbox_launcher_program_parts.pytree scan). Every test changed in this round failed first: the Windows fixes under a simulated CRLF checkout, a tied clock or a slow host, the restored needs-you pins on the previous head, and the revocation takedown on the previous head. Local Opus-lane and GPT-lane reviews of the change since the last pushed head pass (GPT blocked once on a revocation race, fixed by taking the tab down after the unbind). No Sensitive Change Review path is touched and no website file changes.Manual verification
None on a live Slack workspace for this PR. Two points are unverified locally because no harness speaks to Slack:
setView(type=diff, contentThis Code tab was withdrawn.) that replaces a tab whose id was never learned.setViewreplaces the channel's one diff tab, which the takedown of that replacement relies on.Screenshots / video
No new screenshots: no local harness renders Slack's
planblock, the Code tab or a posted reply, and no live workspace was used. The rows' PR output, the Code tab's groups and the wrap-up lines are pinned as rendered text intest_slack_code_channel_board_diff.py.Related Issues
Follows #21993.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)