[melodic-automation-lanes]

Closes #12594

Summary

Opened by the work-resolve-issue lane. The run's own summary:

The workflow skill's explore stage used to ask before investigating a referenced file missing from disk. It now runs read-only git status --porcelain and git log --diff-filter=D on that path first. It skips a staged or committed delete and says so, investigates a working-tree-only delete, and asks only when neither probe names the path or a probe fails. To make the two skills give one rule (record row 28), I copied the rule and commands from the explore skill's slice for the same gate, issue ＃12590, whose wording sits in open PR ＃12607. That slice is not on main yet, so plugins/discovery/skills/explore/reference/workflow.md still says ask-first until ＃12607 merges.

Not verified

• The ai-slop deterministic detector (detect.sh) did not run, because the lane's Bash allows only the named wrappers. The rubric pass was a manual read and found no tells.
• Tidy skipped passes 1-3 (code passes): the change is markdown only. Passes 4 and 5 ran and made no edits.
• check-affected's CI-only checks (line-citations, purged-em-dashes, editorconfig and others) and its skipped validate-plugins steps are left to the PR's CI
• The explore skill's matching wording lands only with open PR ＃12607 (issue ＃12590); until it merges, the explore and workflow skills still disagree on main

Fix

• plugins/session-flow/skills/workflow/context/steps.md: the explore stage's missing-file bullet now runs the read-only git status and git log probe before any question. It classifies the result as intentional (skip it and cite the evidence), unexplained (investigate it) or not settled (ask).
• .changes/session-flow/work-resolve-issue-issue-12594-41563008213-1-93e14bf0.md: a new minor changelog fragment describing the behavior change.

Verification

The runner ran each check from the base copy against a4d9623:

• scripts/check-changelog-fragments.sh --check: passed
• scripts/check-changed-skills.sh --merge-base: passed

The run reports:

• read https: //github.com/melodic-software/claude-code-plugins/blob/7766ddefddf1e6d99842c105c27bf20a8a1d6a56/docs/audits/7005-confirm-gates/confirm-gates.md: row 139 (steps.md:34) and row 28 (explore) are both reversible-local, probe-first; probe-first means a read-only probe answers the question the
• read https: //github.com/melodic-software/claude-code-plugins/issues /7005＃issuecomment-6570157077: the owner approved probe-first for rows 21, 28 and 139
• read https: //github.com/melodic-software/claude-code-plugins/issues /12590: the explore slice of the same gate is open, with open PR ＃12607
• read https: //github.com/melodic-software/claude-code-plugins/pull /12607/files: the explore wording (the same two git commands, with intentional, unexplained and not-settled outcomes) that this change mirrors
• check-changelog-fragments: all 5 fragments valid (run before and after tidy and review)
• check-changed-skills: workflow PASS, 0 errors, 2 warnings (both about SKILL.md, which this change does not touch)
• check-affected: 11 passed, 7 skipped, the rest CI-only
• fresh-context review subagent: no defects; I applied two nits (limit the probe to investigation the task did not ask for; 'Skip investigating it')
• Grep: nothing else in plugins/session-flow states the old ask-first rule, and nothing cites steps.md by line number
• The lane run: https://github.com/melodic-software/claude-code-plugins/actions/runs/41563008213

Left out: the items under Not verified.

Related

#12594

[github-actions]

claude-security-review reviewed a4d96230ad99c65a5d998b140deeed9ea4448f05: read 2 of 2 changed file(s); 0 finding(s), 0 as inline comments.

Security review: no security issues found

I read every hunk of both changed files. Neither matches a trusted-path glob.

• plugins/session-flow/skills/workflow/context/steps.md: the explore stage used to ask before looking into a referenced file that is missing from disk. Now it first runs two read-only probes, git --no-optional-locks status --porcelain -- <path> and git log --oneline -1 --diff-filter=D -- <path>, and acts on the result: it skips an intentional delete, looks into a delete made only in the working tree, and asks when the probes don't settle it.
• .changes/session-flow/...93e14bf0.md: a changelog fragment, prose only.

Instruction-surface lens. This change narrows a standing instruction (an ask-first rule). I checked the removed rule against the six Gate 0 classes that docs/contracts/instruction-exception-register/README.md adopts: irreversible-action, secret-handling, data-integrity, external-publication, legal-compliance and agent-authority. Without the old question, the agent now reads git history and investigates a file that was deleted only in the working tree. Nothing is mutated, published or deleted, and no new authority is granted, so no protected class matches. Both probes are read-only. --no-optional-locks keeps status from refreshing the index, and the -- separator stops a hostile filename from being read as a git option. Nothing in this change widens a trust control, a token's scope, what untrusted text reaches a lane, or anything that can bill.

PR context. It contains no embedded directives aimed at this lane.

critical-findings: 0

[github-actions]

claude-review: 1 inline finding(s) on a4d96230ad99c65a5d998b140deeed9ea4448f05; the summary is in this lane's pull request comment.

claude-lane: claude-review

[github-actions]

No test covers the new rule. Closing issue #12594 sets two criteria:

• "The step runs the git probe before any question."
• "A question remains only for the case the probe cannot settle."

This bullet meets both. But no file under plugins/session-flow/evals/ covers the explore stage's missing-file gate. Searching for deleted, not on disk or diff-filter finds only the unrelated cs-env-worktree-not-pruned and re-first-split-hop cases.

That means a later edit could bring back ask-first wording, or drop one of the three branches, and nothing would fail. One example: an edit that makes a worktree-only delete ( D) count as intentional.

Suggested fix: add an eval case. A scaffold could build a repo with three files and reference all of them from code:

• one staged-deleted (git rm);
• one deleted only from the working tree (rm);
• one deleted by an earlier commit.

The case would then assert that:

• both probes run before any question;
• the staged and committed deletes are skipped, with the status line or commit cited;
• the worktree-only delete is investigated;
• no question is asked.

claude-lane: claude-review

[melodic-automation-lanes]

pr-unblock / address-feedback: UNCERTAIN, left open

Evidence: plugins/session-flow/skills/workflow/context/steps.md:34, plugins/session-flow/evals/workflow-stage-5-settings-screen/case.yaml:8

The finding holds as far as the code shows. The new explore-stage rule (steps.md lines 34-44) sets three outcomes: intentional, unexplained and not settled. No case under plugins/session-flow/evals/ exercises it. No fix was made in this round, though. The suggested eval would need a git repo holding a staged delete, a worktree-only delete and a committed delete. Every eval case in this repository builds its fixture through context.scaffold_script, which points to a scaffold.sh file next to case.yaml, and the review-thread tool's check-paths refuses plugins/session-flow/evals/<case>/scaffold.sh because script files are outside this lane's scope. A case.yaml with no scaffold could not reach the gate. A person should add the case, with case.yaml, scaffold.sh and samples, or file it as a follow-up. The graders should check that both probes run before any question, that the staged and committed deletes are skipped with the evidence cited, that the worktree-only delete is investigated, and that no question is asked.

[github-actions]

claude-review reviewed a4d96230ad99c65a5d998b140deeed9ea4448f05: read 2 of 2 changed file(s); 2 finding(s), 1 as inline comments.

Code review: PR #12630, feat(session-flow): workflow probes git for missing files before asking

I read both changed files in full: plugins/session-flow/skills/workflow/context/steps.md and the new changelog fragment .changes/session-flow/work-resolve-issue-issue-12594-…md. The new rule is consistent. Status is checked first, and the delete commit counts only when status prints nothing. A staged D means skip, a worktree-only D means investigate, and anything else means ask. Both commands are read-only, and --no-optional-locks keeps git status from writing the index. The fragment's bump: minor fits a behaviour change in a skill, and its wording matches the new rule.

Closing issue #12594: acceptance criteria

• "The step runs the git probe before any question." Met (steps.md:34-38). No eval covers it (see the inline comment).

• "A question remains only for the case the probe cannot settle." Met (steps.md:44). No eval covers it (same inline comment).

• "The wording matches the explore skill's slice for the same gate (record row 28), so the two skills give one rule." Not met at this head (blocking). The explore skill's matching gate at plugins/discovery/skills/explore/reference/workflow.md:41 still says "Inline, ask first", and its dispatched branch records an open_questions entry. If this PR merges alone, the two skills give opposite rules for the same gate. The PR body says the matching wording is in open PR feat(discovery): explore probes git for deleted files instead of asking before archaeology #12607. I could not read feat(discovery): explore probes git for deleted files instead of asking before archaeology #12607 from this lane, so it is author-claimed and unverified that the two texts match. Either change, or both, would fix it:

  • merge this PR only after feat(discovery): explore probes git for deleted files instead of asking before archaeology #12607 lands, then confirm the two texts agree;
  • bring the explore slice into this PR.

  Choosing the merge order needs a person.

Other notes

• Path-scoped rules: .claude/rules/skill-bodies-state-current-rules.md and .claude/rules/cost-claims.md match steps.md. Neither is broken: the change restates no volatile upstream specific and makes no cost claim.
• No plugin manifest changed, so no manifest lines.
• No embedded instructions found in the PR context.
• Verification in the PR body (check-changelog-fragments, check-changed-skills) is author-claimed. I did not run it.

[melodic-automation-lanes]

The unattended merge path armed native auto-merge at head a4d96230ad99c65a5d998b140deeed9ea4448f05.

[melodic-automation-lanes]

pr-unblock / address-feedback finished round 1 of max 4 on a4d9623.

[melodic-automation-lanes]

The pr-unblock lane stopped address-feedback on this pull request: the lane left a review finding open and no adjudicator verdict settled it: none was recorded, it found the finding does not hold where the lane had not disputed it (an UNCERTAIN or VALID-defer reply), which only an INCORRECT reply lets a verdict resolve, it found the finding asks only for tests whose covering check runs had not all passed when the round began, or it raised a security question on a finding the security-review lane rated below CRITICAL; so a session builder takes it. The lane labels it needs-session for a session builder, who makes the change this lane cannot, and the run ends in a failing pr-unblock/address-feedback status on the head the lane gated.

• Reason: feedback-unadjudicated
• Head: a4d96230ad99c65a5d998b140deeed9ea4448f05
• Run: https://github.com/melodic-software/claude-code-plugins/actions/runs/41571055692

When a retry on this head can change the outcome, the lane runs one on its own, up to its per-head cap, and edits this comment if it stops again; a run that does not stop clears the status. When no retry can (a review thread the lane does not take or settle, or a failing check after the retries are spent), a person resolves the threads or gets the checks green, then dispatches the workflow below: that run finds the cause gone and clears the status. On a pull request the lanes App opened, the dispatchers start that run once per stop; on any other, a person does. A new commit to the branch starts clean.

To run it again on this head now, dispatch the pr-unblock workflow:

gh workflow run pr-unblock.yml -R melodic-software/claude-code-plugins -f pr-number=12630 -f activity=address-feedback

[kyle-sexton]

Hold: replaced by #12680 (owner-login replacement after pr-unblock escalated). Releases when: #12680 merges, then this PR closes.

[kyle-sexton]

Superseded by #12680, which carries this branch's head plus the eval scaffold the lane's path scope refused. Closed so the open-PR count reflects live work.