[kyle-sexton]

Closes #12594

Summary

Replaces #12630, opened by the work-resolve-issue lane. Its pr-unblock run stopped with feedback-unadjudicated: claude-review's thread asked for an eval case covering the explore stage's missing-file gate, and the lane could not add one because the case needs a scaffold.sh, which is outside the lane's write scope. This branch carries #12630's change unchanged (merged in from its head a4d96230) plus that eval case.

The /session-flow:workflow explore stage now probes git status and git log --diff-filter=D for a referenced file that is not on disk before asking: a staged or committed delete is skipped and cited, a working-tree-only delete is investigated, and it asks only when neither probe settles the path.

The review's other finding (the wording must match the explore skill's slice for the same gate) is met on main: #12607 merged, and plugins/discovery/skills/explore/reference/workflow.md:41 gives the same two commands and the same three outcomes.

Fix

• plugins/session-flow/skills/workflow/context/steps.md: the explore stage's missing-file rule (from feat(session-flow): workflow probes git for missing files before asking #12630).
• .changes/session-flow/…-93e14bf0.md: feat(session-flow): workflow probes git for missing files before asking #12630's changelog fragment, kept as is.
• plugins/session-flow/evals/wf-explore-probes-git-before-asking/: new case. The scaffold builds a repository whose src/report.py imports three exporters that are off disk: one removed by a commit, one staged for removal (git rm), one removed only from the working tree (rm). Graders check that the run calls git status and git log --diff-filter=D, investigates the working-tree-only removal with git show/diff, cites the staged removal and the deleting commit, and asks no question about investigating any of the three.

Verification

• python3 plugins/evals/skills/validate/scripts/validate-cases.py plugins/session-flow/evals: no FAIL; the new case's only WARN is the Bash-needs---allow-tools note every sibling Bash case carries. All six graders sort their samples correctly under Python and JavaScript.
• The scaffold run in a scratch directory gave D src/exporters/csv_out.py, D src/exporters/xml_out.py, and git log --diff-filter=D named the json commit.
• shfmt -d and shellcheck on scaffold.sh: clean. markdownlint-cli2 on steps.md and the fragment: 0 issues.

Not verified: the case was checked statically only; no paid claude plugin eval run.

Related

• feat(session-flow): workflow probes git for missing files before asking #12630 (replaced; held with do-not-merge)
• feat(discovery): explore probes git for deleted files instead of asking before archaeology #12607 (explore skill's slice for the same gate, merged)

🤖 Generated with Claude Code

https://claude.ai/code/session_01Vcupom3BKF6QuC5U3sRJPj

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-03T04:18:02.696620Z	ecd54e8	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

claude-review: 1 inline finding(s) on ecd54e85c635646319c0aec4201fa2074867fbe6; the summary is in this lane's pull request comment.

claude-lane: claude-review

[github-actions]

claude-review reviewed a4dea37b59de4ff0d07aab6a9e74d1a619f786dc: read 2 of 2 changed file(s); 1 finding(s), 0 as inline comments.

Code review: PR #12680 (incremental since 2877c332)

I read every hunk of both files on the incremental list. I also read the head case.yaml and the head samples/no-question-on-missing-files.json.

What changed. This push fixes the inline finding from the last round (case.yaml:49-53: a correct one-line yaml_out question that also named a removed exporter failed the case).

• no-question-on-missing-files now begins with a (?![^\n]*yaml_out) lookahead.
• The pass samples gain the two one-line examples from that finding.

Correctness, checked by reading only. I have no shell, so I did not run the graders.

• Both new pass samples contain yaml_out, so the negative lookahead blocks a match and the not_contains grader passes them.
• The three fail samples do not name yaml_out, so they still match.
• The four earlier pass samples still behave as before. In json_out.py was removed on purpose.\nShould I investigate yaml_out.py…, the first line has no ? and the second line names yaml_out.
• untracked-import-asked (^(?=[^\n]*yaml_out)[^\n]*\?) still matches both new one-line samples, so the two graders no longer conflict on them.

The prior inline finding is fixed, so I did not post it again.

Note, not counted as a finding. This push took the lookahead option the last finding offered. As a result, any line that names yaml_out is exempt from no-question-on-missing-files, including a line that also asks to investigate a removed exporter. For example, Should I investigate xml_out.py, csv_out.py and yaml_out.py? now passes this grader. No fail sample records that trade-off. It is an accepted design choice, not a defect.

Findings

1. (restated summary finding from earlier rounds, advisory) Closing issue feat(session-flow): workflow probes git for missing files instead of asking before investigating #12594 says: "The wording matches the explore skill's slice for the same gate." Judged by reading only, the wording meets it. But no test in the PR or the repository compares steps.md#L34 with plugins/discovery/skills/explore/reference/workflow.md:41, so either file can drift and nothing fails. This push does not change that.

Spend. This push adds no graders and changes no turn or timeout settings, so the case's prepaid API credit use is unchanged. That is not a hold.

The PR context, the settled threads and the prior findings contain no embedded instructions. Prior-findings round count: 2 of 3 before this review, not past the cap.

[github-actions]

claude-security-review reviewed a4dea37b59de4ff0d07aab6a9e74d1a619f786dc: read 2 of 2 changed file(s); 0 finding(s), 0 as inline comments.

Security review (incremental since 2877c332): no security issues found

I read every change in both files on the incremental list. Neither file matches a trusted-path glob. This push doesn't touch any plugin manifest, workflow, token, trusted-actor list, merge hold or instruction file.

What changed.

• case.yaml: the no-question-on-missing-files grader's regex gains a negative lookahead, (?![^\n]*yaml_out), so a line that asks about the never-tracked yaml_out no longer counts as a question about the three deleted exporters. Graders are data that the eval harness matches against model output, and they run nothing.
• samples/no-question-on-missing-files.json: two pass samples are added. They are static grader test data.

The eval's execution block, including its tool allowances, is unchanged in this push.

Spend: none is added. The change only adjusts how existing output is graded.

Untrusted text: the PR context contains no instructions aimed at this review.

Prior findings: the lane's only earlier post was a no-findings summary, so nothing carries over.

critical-findings: 0

[github-actions]

claude-review: 1 inline finding(s) on 2877c3320cb3e442e10357da36c707eacc0be728; the summary is in this lane's pull request comment.

claude-lane: claude-review