Repository navigation
fix: match self-nudge-loop steps to the popover and kanban-md (#16164) - #16208
Conversation
Design Review (Fable 5) — ✅ PASSDesign-level review of Design-Verdict: PASS Install lines still fetch kanban-md Suggestions
[DESIGN-REVIEWED] 4f0912a Description read[DESCRIPTION-READ] 10da996e8038976055e1fd467046c985b67558eccdb30183862988bb7500192d 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 (Fable 5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Verify the new test's kanban-md 0.37.0 flag tables against the real binary once — CI never runs kanban-md, so the pin is only as true as the author's capture. What this change shipsInventory (8 items) — 8 justifiedIntent: FIX — make the generated
Checked: [FIRST-PRINCIPLES-REVIEWED] 4f0912a Description read[DESCRIPTION-READ] 664794ae251b713420637d9fef17aabc4957f3db2dc414642a1bc1580e2f2fed That sha256 names the PR description this verdict read. Recompute it with |
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 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 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: |
0f38d1c to
ba69337
Compare
ba69337 to
156aed5
Compare
|
|
First Principles headline on 156aed5 asked to confirm the test's |
156aed5 to
a243244
Compare
|
a243244 to
db8b527
Compare
|
|
|
|
db8b527 to
82799e6
Compare
|
|
|
The LOOP.md that scaffold.sh writes told the user to fill four fields in
the Set a goal popover, the fourth a stop-sentinel path the popover does
not have. It now names the popover's three inputs by their labels and
says where the stop file comes from instead: the service gives a
popover-armed loop its own stop file, which {{STOP_FILE}} in the goal
text names; the nudge's first check also halts on the anchor's STOP
file, and REST arming makes that file the service's.
The nudge and SKILL.md's board commands used kanban-md flags kanban-md
0.37.0 rejects with "unknown flag": pick --assignee, move --status,
edit --add-body and handoff --notes. They now use pick --claim with
--status todo, a positional status for move, edit <id> --append-body
and handoff --note, each naming the claim: kanban-md requires one for
pick and handoff, and refuses an edit of a claimed card without it.
The claimant is one name per loop ("loop-$PROJECT" in the generated
nudge, quoted so a project name with a space stays one argument), each cycle resumes the card that name already holds before
picking a new one, handoff passes --release, and a held card the loop
walks away from as blocked is released with edit --block --release
(kanban-md refuses --claim with --release on edit). A card claimed in
one cycle stays workable in the next, and neither a handed-off nor a
blocked card is resumed. goal-loop's SKILL.md, which summarizes the
same cycle, gets the same claim, handoff and blocked-card commands;
its prerequisites now say the commands target kanban-md 0.37.0 and
which subcommands the cycle uses.
A new test reads the three files as text: the popover bullets must equal
the dashboard's English labels; every kanban-md command, including
one continued over two lines, must use a known subcommand and flags,
give a task ID where one is required, and name the claim where
kanban-md requires it (but never with edit --release); and each file
must use a single claimant (quoted in scaffold.sh), look for held work
before picking, and release the claim on handoff and on a blocked card.
Closes #16164
82799e6 to
4f0912a
Compare
|
|
iamwhatever
left a comment
There was a problem hiding this comment.
Reviewed: fixes both parts of #16164. The popover steps now name its three real inputs from the en locale and explain the {{STOP_FILE}} token, and every kanban-md command uses 0.37.0 flags (--append-body, pick --claim, positional move todo, handoff --note ... --release). The claim handling is needed once a claim name is required. Text-only tests fail on the old files. Caution for a maintainer: the flag table is pinned to 0.37.0 while the install lines fetch the latest release.
Problem / Motivation
Goal: A user who follows the self-nudge-loop skill's generated
LOOP.mdfills in the inputs the Set a goal popover actually has, and a loop that follows the ready-to-paste nudge runs every kanban-md command without anunknown flagerror.skills/self-nudge-loop/scaffold.shwrites aLOOP.mdthat tells the user to fill "all four fields" of the Set a goal popover, the fourth a stop-sentinel path. The popover has three inputs. The nudge andSKILL.mdalso give kanban-md commands with flags kanban-md 0.37.0 rejects:edit --add-body(the flag is--append-body),pick --assignee,move <id> --status todoandhandoff --notes.Why it matters
A repo-checkout user arming a loop from the popover looks for an input that does not exist. A loop that follows the nudge hits
unknown flagon its board steps every cycle, so it either spends tool calls finding the right flag or skips the claim and the progress record.Not a goal
stop_sentinel_pathor the{{STOP_FILE}}token.What changed (motivation → approach → change)
The skill text drifted from the two things it describes: the popover's inputs and kanban-md's CLI.
authorize_and_add_nudgedefaults a blankstop_sentinel_path,src/kiro_crew/autonudge_authz.py), and{{STOP_FILE}}in the goal text names it at fire time. The line also says the nudge's first check halts on the anchor's STOP file either way, and REST arming makes that file the service's. The stop-methods table row fortouch $STOP_PATHnow says the service honors that file only for a loop armed over REST with it.SKILL.md:pick --claim … --status todo, positionalmove <id> todo,edit <id> --claim … --append-body,handoff <id> --claim … --note … --release.pickandhandoffrequire a claimant, and kanban-md refuses an edit of a claimed card that does not name the same claimant, so the claimant is one name per loop ("loop-$PROJECT"in the generated nudge, quoted so a project name with a space stays one argument;loop-<project>in the templates), not one per cycle. Becausepickalways takes a new card, the claim step first lists the cards that name holds (list --claimed-by) and resumes one. Every way a held card leaves the loop releases its claim, so the resume step never returns it again:--releaseon handoff when it is done, andedit <id> --block "<reason>" --releasewhen the loop walks away from it as blocked (kanban-md refuses--claimwith--releaseonedit, andpickskips blocked cards).skills/goal-loop/SKILL.mdsummarizes the same cycle (goal-loop runs this scaffold), so its claim step gets the same resume-or-pick command, its move-to-Review step names the samehandoff … --release, its blocked-card rule names the sameedit … --block … --release, and its prerequisites now say the commands were written for and checked against kanban-md 0.37.0 and list the subcommands the cycle uses, without changing which version the install lines fetch.Backwards compatibility
Compatible: only skill prose and a new test change. No code path, API or stored format is touched.
Removes nothing: the stop-sentinel bullet named a popover input that never existed, and every replaced kanban-md flag is one kanban-md 0.37.0 already rejects.
Tests
test/test_self_nudge_loop_skill.py(new) readsscaffold.sh,SKILL.mdand goal-loop'sSKILL.mdas text, so it needs neither bash nor kanban-md:test_popover_steps_name_exactly_the_popover_inputs: the popover bullets equal the dashboard's English labels fromwebsite/src/i18n/locales, with no extra field.test_every_kanban_md_command_is_one_kanban_md_accepts(per file, including goal-loop'sSKILL.md): every kanban-md command uses a known subcommand and its 0.37.0 flags, gives a task ID where one is required, and names a claim where kanban-md requires one.test_a_command_continued_on_the_next_line_is_scanned_whole: a command split with a trailing\is checked as one, so the flags onscaffold.sh's two-linekanban-md initare checked too.test_claimant_is_stable_across_commands_and_cycles(per file): every--claimin the file names the same claimant with no per-cycle token, alist --claimed-byfor that claimant comes before any new pick, everyhandoffpasses--release, and the file has anedit … --block … --releasepath for a blocked held card.test_edit_release_and_claim_are_mutually_exclusive: aneditthat passes--releasemust not pass--claim(kanban-md refuses the pair), while any otheredit, andhandoff, still need the claimant.test_scaffold_claimants_are_double_quoted: every--claimand--claimed-byvalue inscaffold.shis double-quoted, since$PROJECTexpands into the generated text.The popover test and all per-file command and claimant tests fail on the files before this change and pass after it.
Manual verification
bash -n skills/self-nudge-loop/scaffold.shpasses.scaffold.sh --with-boardinto a scratch anchor writesLOOP.md,README.mdand a board initialized by the real kanban-md 0.37.0.list --status todo --json,pick --claim loop-1 --status todo,show 1,edit 1 --claim loop-1 --append-body …,move 2 todo,handoff 1 --claim loop-1 --note …. Each old form exits 1 withunknown flag(--assignee,--statuson move,--add-body,--notes). Barepickexits 1 withrequired flag(s) "claim" not set,handoffwithout a claim exits 1 withclaim name is required, and an edit of a claimed card without--claimis refused.LOOP.md(--project demo): cycle 1 finds no held card and picks card 1 asloop-demo; cycle 2 finds card 1 held, edits it, and hands it off with--release; cycle 3 finds nothing held and picks card 2. Every command exits 0, and card 1 ends in review with no claim. With the old per-cycle claimant (loop-1thenloop-2), cycle 2's edit of card 1 exits 1.edit 1 --block "waiting" --releaseexits 0 and leaveslist --claimed-by loop-demoempty, and the nextpicktakes card 2.edit 1 --claim loop-demo --releaseexits 1 withcannot use --claim and --release together.scaffold.sh --project "my proj"; the generated step-4 pick, run throughbash -con the real binary, claims the card asloop-my proj. Unquoted, the same pick claims it asloop-myand the handoff exits 1 (accepts 1 arg(s), received 2).kanban-md 0.37.0
--helpfor the rewritten subcommands (captured from the release binary; global flags and-homitted)Related Issues
Fixes #16164
Pattern harvest
Rule candidate: agents-md
Pattern: skill prose that quotes another surface's UI labels or a third-party CLI's flags drifts from it silently; pin such text with a test that reads both sides.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)