fix(shepherd): correct the settle procedure and stop restating its command - #430
Conversation
…mmand Closes #427. Both files told agents to answer a non-thread finding "(fix it, decline it, or file it), then settle" — and only two of those three can end there. `settle --disposition` accepts `declined|filed` and nothing else, and a fix means a push, which moves the head, invalidates the cycle the state file describes, and starts a fresh one. An agent that fixed such a finding and followed the prose reached a command that cannot succeed, having been told it completes the cycle. Both now separate the paths: fixing means push and let the new head be reviewed on its own merits; settle records the two answers that leave the code alone, with its required note, because the record is a human's adjudication and not a suppression. AGENTS.md additionally stops restating the invocation and routes to the recipe in the shepherd SKILL.md. That is not tidying: harmon-init#799 inlined the same command and three consecutive review rounds each found a different defect in it — missing required flags, this impossible fix path, and `--surface comment|review` parsing as a shell pipe when copied. Deleting the duplicate converged it immediately, and one authoritative copy in the file that ships the command cannot drift from itself. SKILL.md's fenced recipe was already correct and is unchanged apart from the fix-path sentence. Consumers already have the corrected text: harmon-init#799 fixed its own copy rather than waiting on this release, so a `task sync:skills` after this ships re-vendors matching guidance instead of overwriting it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JzKn4zxYJAZXX3zg9v4Nqy
|
@codex review |
|
Codex Review: Didn't find any major issues. Nice work! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
Ready for review — all gates cleanHead
One thing worth recording, because the guard caught me rather than the other way round: the first CI run failed the closing-keywords check. I had written acceptance criteria on #427, satisfied both, and never ticked them — then used a closing keyword anyway. That is precisely the lapse the guard exists to prevent, and it was mine. Both criteria were verified against the diff before ticking (one grep came back empty only because the phrase wraps across lines in AGENTS.md; the text is there and says what the criterion requires). Consumers already carry the corrected text: evanharmon1/harmon-init#799 fixed its own copy rather than waiting on this release, so a |
1 participant
What
Both
AGENTS.mdand the shepherdSKILL.mdtold agents to answer a non-threadfinding "(fix it, decline it, or file it), then
settle" — and only two ofthose three can end there.
settle --dispositionacceptsdeclined|filedandnothing else, and a fix means a push, which moves the head, invalidates the
cycle the state file describes, and starts a fresh one. An agent that fixed
such a finding and followed the prose reached a command that cannot succeed,
having been told it completes the cycle.
Both files now separate the paths: fixing means push and let the new head be
reviewed on its own merits;
settlerecords the two answers that leave thecode alone, with its required note, because the record is a human's
adjudication and not a suppression.
AGENTS.mdadditionally stops restating the invocation and routes to therecipe in
SKILL.md. That is not tidying — evanharmon1/harmon-init#799 inlinedthis same command and three consecutive review rounds each found a different
defect in it: missing required flags, this impossible fix path, and
--surface comment|reviewparsing as a shell pipe when copied. Deleting theduplicate converged it immediately. One authoritative copy, in the file that
ships the command, cannot drift from itself.
SKILL.md's fenced recipe was already correct and is unchanged apart from thefix-path sentence.
Consumers
evanharmon1/harmon-init#799 already corrected its own copy rather than waiting
on this release, so a
task sync:skillsafter this ships re-vendors matchingguidance instead of overwriting it with the broken text.
Verification
task verifygreen; localtask challengeandtask revieweach clean in oneround.
Closes #427