fix(entry-skills): close the handoff fences — six files render their arms wrong#550
Open
leeovery wants to merge 6 commits into
Open
fix(entry-skills): close the handoff fences — six files render their arms wrong#550leeovery wants to merge 6 commits into
leeovery wants to merge 6 commits into
Conversation
…arms wrong Found by a prose-test walker, which had to guess whether a fenced block ends at the next fence (what markdown does) or at the next arm heading (what the author meant), and recorded the ambiguity rather than picking silently. Markdown pairs fences in order, so a missing closing fence makes the next arm's heading render *inside* the previous arm's code block, and that arm's own arguments render as plain prose. In workflow-specification-entry every one of the four arms was mispaired. Since each arm says "invoke the skill with the next fenced block as its arguments", a session following it literally is handed the wrong text. Six of eight entry skills were affected; discussion and research close their blocks correctly, which is what shows this is an omission rather than a house style. Additions only — eleven closing fences, no wording touched. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ates it Lee's call, and the right one: a log the agent writes is the same trust problem relocated. A PostToolUse hook declared in the walker's own agent frontmatter — so it fires for walkers and nothing else — appends every tool call to .walk-actions.log in the world, excluded from the diff. Proven necessary, not assumed. A diagnostic walker's world held the agent row at `incorporated` plus the report file, so it had run dispatch, stub, scan and incorporate; its transcript began at the fifth step. Four cases came back INVALID today for that reason alone, none of them a prose fault. Two attempts to fix it with wording — a stricter log format, then moving to Opus — both failed, because a narrative composed at the end is a recollection however it is instructed. Evidence now splits by kind: recorded actions are the authority on what was done, the transcript on what was reasoned (which arm, which guard, what was emitted), and the recording wins any disagreement. INVALID WALK keys off the actions, so a walker that did the work and described it poorly is judged rather than discarded. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
smoke-boot-in-world asked for what it forbade: the act stopped right after boot returned, while the assertions were about renderings that only happen in the arms after that point. The act now covers the whole of initialisation and the claims match what it reaches. implementation-picks-first-task scripted one answer against a flow that asks three — environment setup, then project skills, then linters — so both runs halted at an unscripted question, in different places. Rather than script the extra gates, the case is narrowed to stop once the environment is settled: answering them writes project_skills and discovered linters whose values depend on what the fake project happens to contain, which no expected world can pin. The task pickup wants its own case, against a fixture where those are already settled. Both were found by the UNSCRIPTED QUESTION marker and by the world diff doing exactly what they exist for. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…udge without a record The PostToolUse hook never fired: its command was a relative path, and the walker's cwd is the world directory, so it resolved to <world>/tests/prose/lib/… and failed silently. Now resolved via $CLAUDE_PROJECT_DIR, with a throwaway shell probe alongside it so one run distinguishes "hooks don't fire here" from "my path was wrong". Hardening from the same round: an absent action log is the harness failing, not the walk. run.cjs assert now refuses to emit a prompt at all in that case, and the orchestrator reports HARNESS ERROR — five runs had reported INVALID instead, which reads like a test result and buried the actual cause. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Agent-frontmatter hooks do not fire in this environment — proven with a bare `echo` probe that produced nothing across a full run, so it was never the relative-path theory. Hooks in settings.json would fire, but for every tool call in the project rather than just a walker's. The harness already writes each subagent's session to a JSONL transcript, which is the same evidence and better scoped: lib/actions.cjs extracts the ordered tool calls, and `assert` now requires --transcript. The orchestrator passes the output_file the walker's dispatch reports. The hooks block, record-action.cjs and the action-log plumbing are gone rather than left dormant. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…trust Lee was right to push back on abandoning them. Frontmatter hooks in a project subagent are gated on hasTrustDialogAccepted (v2.1.218+) and fail silently when it is false — which it was for this repo, despite onboarding having completed long ago. Nothing was wrong with the config: the shape matched the documented example exactly, and a bare echo probe proved the hooks simply never ran. With trust accepted, the walker's own PostToolUse hook fires and records every tool call to .walk-actions.log in the world. The project-level settings.json hook is gone — it worked, but fired for every tool call in the repo rather than just a walker's — and so is the transcript-parsing fallback it was standing in for. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
workflow-specification-entry, it had to guess whether a fenced block ends at the next fence (what markdown actually does) or at the next arm heading (what the author meant) — and recordedAMBIGUOUSrather than choosing silently.discussionandresearchclose their blocks correctly — which is what shows this is an omission rather than a house style.Enumerated across every
invoke-skill.mdbefore fixing any of them; the first sweep only matched bare```and missed```bashopeners, which under-reported the damage — the corrected pass found six files rather than four.Test plan
npm testgreen: 1699 tests, 0 fail.🤖 Generated with Claude Code