Repository navigation
feat(slack): conductor work board in code channels - #21993
Conversation
UX Review (Opus 5.5) — ✅ human override accepted@kyleseaman overrode this lane for |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS No design blocker: the board is gated by the existing dispatch exemption, adds no store, and degrades when Notes
[DESIGN-REVIEWED] 71814c4 Description read[DESCRIPTION-READ] 98d6fd8414819554d66af266d85a42531311e05d9097af13170c63e605f22ecc 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) — ✅ human override accepted@kyleseaman overrode this lane for |
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: |
c9289e3 to
e6a8146
Compare
ef4956c to
106445c
Compare
Sensitive Change ReviewSensitive files changed:
|
e6a8146 to
408d621
Compare
757ccfc to
e550529
Compare
408d621 to
90250b1
Compare
e550529 to
a733fd1
Compare
d7a6ca4 to
0eaada3
Compare
cc01bfb to
a4dfdbe
Compare
Sensitive change review
|
Human judgment recorded@kyleseaman recorded a deferred release (class
This decision applies only to this commit. A new push requires a new judgment. |
|
/ai-review override first-principles 0b537cd product-shape: I approve removing the reaction ladder and queued-message hourglass in bound code channels. Show progress through Slack session status and the pinned work board instead. Leave DMs, threads and ordinary channels unchanged. |
|
AI-review override not recorded: |
Code change approver recommendation: Request changesWhy: the decider found 2 problems and estimates an 85% chance that this change introduces a defect, at or above the 80% level where it asks for changes. This PR adds a large conductor work board to Slack code channels. It also removes the reaction ladder there; the owner decided that, but no RFC records it. On the head commit, the new tests for the host-path scrub fail on Windows. The home path is either cut wrongly (%USERPROFILE%…/x, because the POSIX absolute-path rule runs after the home token is put in) or, per CI, left in the published approval detail. Either way it contradicts the PR's claim of Problems found:
An automated recommendation, not a review: a maintainer still decides. It approves on its own only when the estimated chance of a defect is at most 12% and the change has at most 400 changed lines. Decider v3c (217fe607) · commit |
A conductor-led code channel shows its helper agents' work in the channel: a pinned work board, edited in place, with the counts line, then one task per helper agent with its state, what it is doing now (the call in flight, host paths made home-relative, secrets removed) and for how long. A helper agent that needs the owner gets one threaded reply mentioning them, once per waiting spell and at most once a minute. The channel's status follows the helper agents between the conductor's turns, and code channels add no reaction ladder. A Stop stops the channel's own session only: the board's own passes are fenced by their admission alone, so the next helper-agent event still publishes, and an idle !stop says how many helper agents still run and where to stop one. Every board write (post, edit, pin, page) re-checks scope and Stop right before the raw send; a board counts as published only once it reached Slack, and an unsent one has one clean-context re-publish pending (backed off, and never before a failure's backoff or Slack's Retry-After). The Code tab's combined diff, each row's PR and CI, and the wrap-up's outcome lines follow in a separate PR.
Register slack/code_channels/board/model.py as the "Slack code-channel work board" redaction sink (it redacts the goal line and worker summaries posted to the pinned board) and allowlist board/render.py as non-egress, since its blocks ride that same board message.
Sensitive change review
|
|
/ai-review override first-principles 71814c4 owner-decision: I approve removing the reaction ladder and queued-message hourglass in bound code channels, as I decided in #21993 (comment). Progress shows in the Slack session status and the pinned work board. DMs, threads and ordinary channels are unchanged. |
Human judgment recorded@kyleseaman recorded a owner-decision release (class
This decision applies only to this commit. A new push requires a new judgment. |
|
/ai-review override ux 71814c4 owner-decision: I accept the idle !stop reply as it is for this PR, per my decision in DECISION_LINK. The reply will link each running helper agent directly in the next PR of this stack. |
Human judgment recorded@kyleseaman recorded a owner-decision release (class
This decision applies only to this commit. A new push requires a new judgment. |
Sensitive change review
|
Head commit: 71814c4 (base
main)Problem / Motivation
Goal: In a conductor-led Slack code channel, show each helper agent's state right in the channel, page the owner when one needs them, and keep the channel's status honest while helper agents run, so nobody has to open transcripts to know where the work stands.
Kyle's ask: "we need to show all the sub-session states for conductor, the diff, make the integration GREAT". A code channel's default agent is a
kirocrew-conductor(orkirocrew-team-lead) that splits the work and hands each piece to its own helper agent session. Until now the channel showed only the conductor. Test build 7 showed it plainly: the board said "3 idle" while all three helper agents streamed in the dashboard, and the channel showed no status.This PR is the first half of that work, split at Kyle's decision (board message first). The Code tab's combined diff, each row's PR and CI, and the wrap-up's outcome lines follow in a stacked PR. The per-helper-agent context-bar items are dropped from the stack: #22370's Workers item replaces them.
Why it matters
Conductor-led channels are the default Slack Code experience. Without this, the channel looks idle while helper agents run, and a helper agent stuck on a question waits silently.
Not a goal
What changed
The board. One pinned message, edited in place, drawn with Slack's own
planblock: a context line "Work board · updated HH:MM UTC" with the conductor's dashboard link, the summary as the plan's title ("2 working · 1 waiting on you · 1 of 4 done"), then one task per direct helper agent. Only a working row has Slack's spinner; idle and waiting rows get the empty circle, and a waiting row's title leads with "Waiting on you: " and ends with "Open in dashboard". Details show what the helper agent is doing now ("Runningpytest test/x.py -q· 4m", "Editingsrc/…/gitlab.py") with secrets removed and paths shown relative to its worktree or home-relative (~/…/<name>,%USERPROFILE%\…\<name>on Windows), anything else as…/<name>, or "Idle · 12m · ". At most 50 tasks (49 plus "+N more"). A Slack that refuses the plan block gets the same content as plain mrkdwn rows, logged once.Needs-you replies. A helper agent that waits on a person gets one threaded reply under the board that mentions the owner: once per waiting spell of the item (a helper agent repeating its report pages nobody again), once per pending tool approval, and never sooner than 60 s after that item's last page.
Workers by lineage. The watcher (
WorkspaceService.watch_tick, every 15 s, started at gateway boot) follows the conductor's helper agents wherever they run: the code-channel store's worker records plus every session the last board names, one level only. A helper agent's turn start or end, closed session, tool approval raised or answered, and every work-ledger write refresh the board; a live board is refreshed at least every 60 s.Channel status. Between the conductor's own turns the channel's status follows the helper agents:
suspendedwhile a row waits on you,processingwhile one runs, otherwise what the conductor's turn end showed. A helper agent whose queue holds only cards restored after a restart (is_held_queue_entry, waiting for a person) reads idle on the board, as in the session status. Code channels add no reaction ladder (owner's decision, see Premise).What a Stop covers. Stop stops the channel's own session and fences the work in flight at the press. The helper agents keep running, so the board's own passes are fenced by their admission alone and the next helper-agent event republishes the board. A
!stopwhile the session is idle but helper agents run answers "This channel's session isn't running, but N helper agents still are", pointing at the work board, where each one's chat has Stop generation.When a board write goes out. The board's four writes (post, edit, pin, needs-you reply) are its only writes of helper-agent content. Each re-checks the board's scope and the Stop right before its raw send, with no await between (
publisher._writable);test_slack_code_channel_commit_point.pylists every such writer, so a new one fails until gated. A board counts as published only once its writes reached Slack. A pass that read a board and did not get it out (fenced by a Stop, failed) arms exactly one re-publish (board.settle): one timer per channel, armed in a clean context, debounced, backed off to 5 minutes, kept when a re-armed pass cannot read its record, dropped when the board is sent or the channel no longer admits it. No pass edits the board before a failure's backoff or Slack'sRetry-Afteris over.Sidebar field
work_aggregate(for #21437): each board read writes the conductor slot'swork_aggregate(state,label,counts), pushed through the existingslot_patchframe only on change. The slot projection carries it (nullwhen no board), so three turn golden files gain that one key.Where helper agents are filed. Only
session_createfiles a code channel's helper agent, in the channel's own folder at creation (#22001); a channel whose session turned conductor mode on keeps its helper agents out of folders, as conductor mode does everywhere (decision 2026-10-27). The board writes nofolder_id: it follows its helper agents by lineage, wherever they are filed.slack/code_channels/board/source.py,dashboard/handlers/work_ledger_board.pyslack/code_channels/board/model.pyboard/publisher.py,board/render.py,board/links.pyworkspace/service.py,workspace/board_cache.py,workspace/worker_status.py,board/__init__.py,status.py,surface.pydashboard/state_parts/slot_turns.py,dashboard/interaction_coordinator.pybinding.py,slack/handler_runtime/turn.py,slack/renderer.py,slack/events.pyDocs:
docs/system-specs/modules/slack-gateway.md("Conductor work board"),docs/request-for-change/rfc-slack-code-channels.md(Stop),src/kiro_crew/docs/slack-integration.md.Published-fields rule (owner-sanctioned)
Only these reach the channel: each item's title, its derived state, the first line of its last report summary,
artifacts.branch(valid branch names only, in the plain-rows look), each working helper agent's current call and a waiting one's pending approval call. Every text field goes throughmodel.safe_line: the outbound redaction over the whole raw text first, then host paths and host names out, one line, thenclean_text(pings defused, redacted again, cut, redacted once more).Moved to the follow-up PR
The Code tab's combined diff (each helper agent's commits from its own worktree only), each row's PR and CI counts, the wrap-up's outcome lines, and their tests and docs.
Premise (for the maintainer override)
First Principles blocks the removal of the reaction ladder in code channels (eyes → thinking → done, and the queued-message hourglass) as a product-shape change without an accepted RFC. The removal is the owner's decision, from the build-7 test ("remove the lobster emoji, rethink that emoji bit in general"): drop the ladder in code channels, keep progress in Slack's session status and the pinned work board, and leave DMs, threads and ordinary channels unchanged. The gate is
binding.no_reactions(channel), true only for a bound code channel.Sensitive change
The Sensitive Change Review paths this PR touches are
src/kiro_crew/security_posture_sinks_delivery.py(+9) andsrc/kiro_crew/security_posture.py(+6), both in commitfix(security): classify the work board's redaction call sites. Both rows are byte-identical to the previously reviewed head3f775c2da68a0dd164a47e903e783f1415110383;exfil.pyis not touched. Filled reasoning for this head, for the owner to post (a worker never posts it):Backwards compatibility
Compatible: a channel whose conductor keeps no work ledger gets no board; the context bar and Code tab are unchanged from main; every new surface field has a default, so stored records load unchanged. In ordinary channels nothing changes. In code channels there are no per-message reactions any more (by design).
Tests
test/test_slack_code_channel_board.py,test_slack_code_channel_board_live.pyandtest_slack_code_channel_board_review_rounds.py, with Slack, git and the ledger faked at their seams. Covered: the build-7 repro (folderless channel, helper agents in their own worktrees), one level of lineage, conductor and team-lead alike, the 60 s backstop and restart re-arm, the publisher (edit gap, backoff,Retry-Afterhonoured by every pass, plan refusal fallback, re-pin, a re-arm taken even when a publish raises), paging once per waiting spell with the 60 s gap, safe text (keys, URL credentials, paths on POSIX and Windows, whole path segments only, a project above home never cut, host names, a huge command), the Stop rules (a Stop with the conductor idle still lets helper-agent events publish; a Stop mid-pass on the last helper-agent event still ends with a published board; the stopped turn's own refresh is skipped; the idle!stopreply), the send gate for each board write, the structural writer list, andwork_aggregate. Every new test was shown to fail without its fix; tests edited during the split were each shown to fail under a mutation of the behaviour they name. Determinism: waits usekiro_crew.testing.wait; the flake-pattern ratchet passes. Also covered: a failing channel PR read, bar or diff step never keeps the board from going out; a write the client's native-Stop fence would drop counts as unsent; Retry-After is timed from Slack's answer; a long ledger history keeps every open row plus the newest 100 settled ones, with totals, both looks andwork_aggregateidentical to the unbounded result; the per-item page clock is pruned past its gap.Host paths and names, after two fresh-context redaction reviews:
board/model.pyreads every token that holds/or\as ONE path (quoted, it may hold spaces and the other quote;\stays inside), normalises it lexically withposixpath/ntpath(nothing read from disk or expanded), and shows it as project-relative (inside the project root, no root, drive or..left), the home marker alone or~/…/<last>(%USERPROFILE%\…\<last>) below home, a plain relative path or public http(s) URL as written, else…/<last>, or…when the last name is missing or is the account or this machine's name. An unterminated quote, or a path the scan bound cuts, is…whole. A read or edit call's structured path is redacted, then scrubbed as one whole value beforeReading/Editing. A name-shaped run holding this machine's name becomeshostwhole, so no name is rebuilt from what is left. The tests (test_slack_code_channel_board_review_rounds.py, published code elements and steps) carry every reproduction from both reviews: Windows homes in any case and separator, projects above home in any spelling or by.., project-name siblings (repo$other,repo=other), flagged drives, UNC, device, rooted and drive-relative forms, quoted, opposite-quote, unterminated and past-the-bound paths, legal punctuation and escaped spaces, structured paths with spaces, literal%prefixes, raw home markers with.., host-name reassembly, public URLs and relative paths kept, and every case published twice unchanged. The board writes no folder: a structural test fails if any module underslack/code_channels/board/assigns afolder_idor writes a session's metadata record, and a stored surface row carrying an earlier build'sboard_filedlist still reads (this build ignores the key and keeps it on disk). The board suites pin the scrubber's home, account and host names in a shared fixture (cases about a specific home or host install their own), so they pass on an account namedtesttoo.Repo-wide gates this change meets: the forced-full-save census stays at 26 (the board writes no session record);
slack/code_channels/board/publisher.pygets onepost_messagebaseline row (the needs-you reply: text already escaped and redacted by the board,_writablere-checked just before the send).Local checks. Earlier heads of this change passed broader runs that this head reuses only where their inputs are unchanged: c9289d525ec319e63cc429de9a2871b58fbe0897 (on main 65dd048) passed the full local prepush with
vitest related, and a339ce5974533b93c9f1a1952e458904f99fd167 (on main e81acb9) passed 9911 tests in 230 repo-wide gate files. d81370cdfe0452b06e45223ea4f1b604c7c2d851 is that change rebased onto main 6ef9d6b, and 14273669df6b280a752fb85bcf4ee5b445453805 is d81370cdfe rebased onto main 39fd360 with no conflicts and both commits identical ingit range-diff; 29c337d2b5b1c725ca2d8be8b702d12968606022 is 14273669df rebased onto main 8579038, again with no conflicts and both commits identical. One conflict, indashboard/handlers/work_ledger_board.py, was two sets of new functions added at the same place; both are kept, and the PR's lines in that file are unchanged. Two changes follow main: a helper agent whose queue holds only restored cards reads idle (main #33214's rule), and the board's refresh test covers main's newaskledger action. Every other file's change is identical to the previous head. On d81370cdfe: the Python gates (source, docs, static and flake ratchets, flake8, isort, black, mypy, line cap), the changed-plus-importers tests (87 files), 106 focused vitest files (i18n catalog and style, settings registry and search, the Slack Code settings panel, code-channel sidebar, slot list, API client surface) and the i18n style tests pass; the settings registry, pseudolocale and config baseline generators reproduce the committed files byte for byte andi18n-checkpasses; 78 work-ledger, approval, session-status, queue, KAS, prompt, config-generator and turn-golden files pass (2776 tests); and 107 repo-wide gate files that scan the test tree or touch code channels and queues pass (4333 tests), the other 123 of the earlier 230 reused from a339ce5974. On 14273669df: the Python gates, 44 focused test files at the seams main changed (restore order, slot detail, model validation, process identity, history, cron turns) with the PR's own tests, 113 focused vitest files (on a rerun: the first run's one failure, insrc/i18n/navLabels.test.tsx, which this PR does not touch, passed alone and in the rerun), the generators andi18n-check, the 9 gate files that import a module main changed and the PR's inventory gates (887 tests) pass. On 29c337d2b5b1c725ca2d8be8b702d12968606022: the Python gates, 23 focused test files (config defaults and patching, the Slack Code settings, the board's turn-end, liveness, publication and filing suites, session status, facades, turn goldens, config baseline, settings registry, security posture; 1499 tests), 97 focused vitest files (i18n, Slack Code settings panel, settings registry and search, WebSocket refresh kinds, slot list; 1632 tests) and the i18n style tests pass on their first run; the generators reproduce the committed files;i18n-checkpasses with its diff-scoped checks against the base; and the 3 earlier gate files that import a module main changed pass (234 tests).An earlier pushed head, 96182e6 is 29c337d2b5b1c725ca2d8be8b702d12968606022 rebased onto main ac4062d. Main had added two session-row fields (
participants,participants_dropped) beside this PR'swork_aggregateindashboard/slot_projection.py,test_chat_slot_facade_contract.pyand three turn goldens; both sides are kept, main's first. Apart from one line, named at the end of this paragraph, every added and removed line of the PR is byte-identical to 29c337d2b5b1c725ca2d8be8b702d12968606022: only context moved, the second commit is identical ingit range-diff, and the two Sensitive Change Review rows are unchanged. On the pushed head: 27 focused test files (the 23 above plus shared-crew participants, Slack upload, channel upload field types and thepost_messagecall-site census; 1584 tests) pass; the settings registry, pseudolocale and UI index generators reproduce the committed files;i18n-checkpasses with its diff-scoped checks against the base;tscand the 11 i18n style test files (114 tests) pass. The 97 focused vitest files and the Python lint and type gates were not rerun on the pushed head; they ran on 29c337d2b5b1c725ca2d8be8b702d12968606022, and of their inputs only the locale catalogs changed, which the i18n checks cover. One line differs from 29c337d2b5b1c725ca2d8be8b702d12968606022:board/model.pysplits a normalised POSIX path withposixpath.sepinstead of the literal"/"(same value; the PR Policy portability scan flags the literal), and the board, worker-status and workspace suites (621 tests) pass with it.An earlier pushed head, d9feb74 is 96182e6 with three changes, rebased onto main 0df116f (one conflict: two imports added at the same place in
slack/events.py, both kept).board/source.pyscope_refusalis the one place that classifies a refusal. New tests fail by assertion on 96182e6.board/links.pygets direct tests (Coverage Gate measured it at 30% on 96182e6; 100% locally now).SAMPLE_SLACK_BOT_TOKENfromkiro_crew.testing.synthetic_credentials, because main's feat(security): turn on the structured credential detectors, with owner engine and prompt-and-logs switches #29423 makes the platform redactor keep AWS access key IDs visible as identifiers, and the board uses the platform redactor unchanged;store_channel_bindingmoved totest/slack_code_channels_helpers.pyto keeptest_slack_code_channel_workspace.pyunder the 2,000-line cap.Local checks for that head. On 64f86198a5602312c19c2fb313a5085bac26d44a (same patch, one main commit earlier, before a two-line test import change): the full pre-push gate (source gates, static and flake ratchets, flake8, isort, black, mypy, line cap, 90 changed-plus-importer test files, 97 focused vitest files, the i18n style tests); the settings registry, pseudolocale and UI index generators reproduce the committed files;
i18n-checkpasses with its diff-scoped checks against the base;tscpasses. On d9feb74: the board, dispatch-exemption, worker-status, workspace, commit-point, stop-fence, security posture, facade, turn-golden andpost_messagecensus suites (889 tests). Per-file coverage from the board suites alone is 80% or more for every board and workspace file this PR adds or changes. Reviews: the product code is line-for-line the code a local Opus-lane review and a fresh-context delta review passed on a118b47c4223b8ab71d39af081720cba3464e483; a further fresh-context review of the rebase (name and behaviour drift across main's 687 commits, a 58-shape redaction corpus before and after, the changed tests under a no-redaction mutant) found nothing blocking. Its notes: three older tests assert on a credential that their own shortening step would remove anyway, and the dashboard Board now says "stopped" for a run whose patrol ended with open items, a word the Slack board does not have.The pushed head 71814c4 is d9feb74 with three changes, rebased onto main 7d1ba38 with no conflicts. Every other added and removed line is identical to d9feb74, and the two Sensitive Change Review rows are unchanged.
board/filing.py, theboard_filedfield of the surface record, its unsaved-list cache and theslack.code_channel.file_workerSEL action are removed;session_create.pyis not changed. A stored surface row that carries an earlier build'sboard_filedlist still reads.dashboard/session_control.pylistsAUDIENCE_UNREAD_REFUSALin the facade's forwarded names.test_facade_mirror_guards.pyfailed on d9feb74 (Backend Tests (3.12, 7) and (Windows) (2), and Coverage Gate after them) because the constant was forwarded but not listed.Local checks. On f63287f4ae4539ffe3fd0d79b54bfebd08fe3c64 (the same patch, four main commits earlier; both commits identical in
git range-diff): the full pre-push gate (source gates, docs lint, static and flake ratchets, flake8, isort, black, mypy, line cap, 104 changed-plus-importer test files, 97 focused vitest files, the i18n style tests) and the 230 repo-wide gate files that scan the source tree, the facades, the registries and the censuses (the set that holdstest_facade_mirror_guards.py, which the changed-plus-importer run had missed): 9960 pass, and the same four host-only tests named above fail as they do on a clean base. On 71814c4: docs lint, the feature map check, and the board, dispatch-exemption, worker-status, workspace, commit-point, stop-fence, facade mirror, security posture, slot facade, turn-golden,post_messagecensus and Slack render suites (1146 tests). Reviews: a local Opus-lane review and a local GPT-lane review of the delta from d9feb74 both passed on 8b830ad18f926ae52d685fa94dcac9a72326e742; this head is that commit rebased, plus three prose lines those reviews asked for (one code comment, one test docstring, one sentence in the Slack user doc).Four gate tests fail on d81370cdfe and on its clean base 6ef9d6b alike, so they are not this change:
test_sandbox_launcher_program_parts.py(tree scan), two intest_codex_session_mcp.py(the installed Codex adapter) andtest_hooks_edit_gate_diff_path.py(absolute edit path).Manual verification
A harness ran the real
WorkspaceServicewith its 15 s watcher, the real board reader, publisher, debounce, backstop and edit gap for four minutes, with a conductor and two helper agents each in its own worktree and no channel folder (build 7's shape); Slack and the slot table were faked and recorded. This ran on the full branch before the split, whose board code is this PR's plus the per-row PR. The board posted and pinned within 4 s of the dispatch, then followed each start, approval, report and end; while anything was live, consecutive board writes were 23–57 s apart (never closer than the 20 s gap, never staler than 60 s), the status wentprocessing→suspended(an approval) →processing→active, and the board stopped writing once everything was done.Screenshots / video
Before (build 7, three helper agents streaming, board idle):
The threaded reply that pages the owner for a waiting row:
The empty board, before any helper agent starts:
The earlier board, context-bar and Code-tab shots showed the per-row PR and the follow-up's surfaces, so they move to that PR.
Related Issues
Fixes #30538
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)