[kyle-sexton]

Closes #9790
Closes #11187
Closes #11189
Closes #11191

Summary

Four /planning:interview changes that each stopped host-only in the issue lane only because the lane cannot re-pin the interview digests. They share the same pins, so one PR re-pins them once instead of four PRs conflicting on them. The lane wrote the rule text for each; this host run resolved conflicts with main, re-pinned and verified.

Fix

• docs(planning): reword interview phrases shared with mattpocock grilling, and re-pin the interview digests #9790: reword the phrases the interview shared with mattpocock/skills grilling, grill-me and grill-with-docs (at 24fe0ef). "The frontier is empty" becomes "no open question remains on the frontier" across the skill bodies, sweep, checklist template, evals and wayfind. The Stance wayfind trigger and the confirmation gate's opening clause are reworded too. One eval prompt main had added since the lane run is reworded as well.
• fix(planning): read back answers that depart from the recommendation before marking them answered #11187: the loop.md drift check reads back an answer that departs from the recommendation, or arrives dictated or unclear, and keeps its row open until the user confirms.
• fix(planning): restate only changed answers and unseen assumptions at final confirmation #11189: when each answer was read back as it arrived, the Step 3 restatement gives only changed answers, hedged rows and unseen assumptions.
• fix(planning): word the round's closing probe as 'something else' #11191: the round's closing probe is worded "something else".
• The Stance, Step 2, Step 3, assumption-sweep and loop.md register digests are re-pinned with tests/repin.sh, each with its justification. No rule or halt changed.

Verification

• scripts/check-affected.sh --tests: 14 passed, 2 skipped (includes interview-defenses.test.sh).
• fingerprint.mjs compare of SKILL.md and context/loop.md against each of the three upstream files: separation.fired is false for all 6 pairs.
• git grep -n -i 'frontier is empty' -- plugins/planning prints nothing.
• scripts/check-changelog-fragments.sh --check and --check-required origin/main both pass.
• Left out: the eval reruns of interview-readback-dictated-answer, interview-final-brief-changes-only and interview-closing-probe-something-else (3 runs, opus judge), which fix(planning): read back answers that depart from the recommendation before marking them answered #11187, fix(planning): restate only changed answers and unseen assumptions at final confirmation #11189 and fix(planning): word the round's closing probe as 'something else' #11191 ask for. This host lane runs them next and posts the results on each issue.

Related

• ci(pr-pipeline): let the issue lane re-pin interview digests without running checkout code #11576: let the issue lane re-pin these digests, so changes like these reach a PR without a host run.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VcupokZY9nMi5ZG7UdzT4q

[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 2026-10-31T15:14:23.696847Z	8bd94d5	Draft marked ready

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-security-review reviewed d26e74270e07f8fc3435e055a1e37122e7d20aa4: 0 finding(s), 0 as inline comments.

No security issues found in the incremental change (d26e742 over 8bd94d5). It touches a changelog fragment, the /planning:interview SKILL.md, and its defenses test.

Instruction-surface lens. The one SKILL.md edit rewrites the recap rule in the "Recap and procedure check" paragraph. The rule now applies "when every answer was accepted as recommended or confirmed after its read-back" instead of "when each answer was read back as it arrived". It tells the model to save a recap (export-ledger --recap) once each round's read-backs are confirmed, so the first confirmation can paste the --changed recap instead of the full one. The changed lines do not touch the checks that stop the model acting on unconfirmed or unresolved answers:

• The explicit-confirmation requirement and the "do not act on the interview's output until they confirm" line are untouched.
• The procedure check (check-open-questions.sh, with its exit code and procedure= field cited as evidence) is untouched.
• The register gate is untouched.
• The recap is still generated from the on-disk ledger.

The edit changes only how much is restated, so no protected-class action loses its check. I did not read docs/contracts/instruction-exception-register/README.md, so I judged this by consequence rather than against the register's class list.

Test file. plugins/planning/tests/interview-defenses.test.sh has a re-pinned digest for the "Step 3" section. This is the expected update after the section text changed. It also rewords some pin comments. No pin was removed and no defense was weakened.

Changelog fragment. It is prose only.

Pin-ceremony and friction-hold changes are not a finding in this phase. No manifests, token scopes, trusted-actor lists or merge holds changed.

critical-findings: 0

[github-actions]

claude-review: 2 inline finding(s) on 8bd94d56de5d8ae7f69a7194eb54cc24948e6930; the summary is in this lane's pull request comment.

claude-lane: claude-review

[github-actions]

claude-review reviewed d26e74270e07f8fc3435e055a1e37122e7d20aa4: 0 finding(s), 0 as inline comments.

Incremental review of d26e742 (since 8bd94d5): no findings.

What changed

• The Step 3 "Recap and procedure check" paragraph in plugins/planning/skills/interview/SKILL.md. It no longer says to replace the recap with one line naming the Q<N> ids that stand as read back. Instead, when every answer was accepted or confirmed after its read-back, the skill saves a recap after each round's read-backs are confirmed and does not paste it. The first confirmation then pastes the --changed recap, which lists only the answers changed since, plus one counting line. The restatement adds the intent, the hedged rows and the unseen assumptions.
• The .changes/planning/…11189….md fragment, reworded to match. It is still bump: patch, which fits a wording fix to a skill.
• interview-defenses.test.sh: the Step 3 pin digest and its justification comment were updated. The stale combined "Adds three rules" comments were replaced with comments for the rules each pin actually covers.

Checked

• The new instruction matches plugins/planning/surface/round.py (cmd_export_ledger, read_last_recap, around lines 2647–2700). A recap without --changed writes <ledger stem>.last-recap.json. A later --recap --changed run diffs against that file. If the file is missing or unreadable, the recap lists every row, so nothing is silently dropped.
• The context/loop.md "Drift check" section the new text links to exists (line 362).
• No other file in the repository uses the removed "read back as it arrived" or "one line names the register" phrasing, so no stale reference is left.
• The change adds no new row resolution. The register gate and procedure check still run unchanged, so the no-silent-resolve and confirmation-gate defenses are untouched.
• The fix(planning): restate only changed answers and unseen assumptions at final confirmation #11189 acceptance criteria are met in SKILL.md: the rule is stated once, names the interview-final-brief-changes-only eval as evidence, and points to the --changed recap. I did not run the test, and I did not recompute the new Step 3 digest (ed720477…). The PR's author-claimed check-affected.sh --tests result is unverified by this lane. The eval rerun is stated in the PR body as deferred to the host lane.

Not a hold: the change spends no billed resource. Eval reruns draw only on prepaid credit and are not a finding.

[kyle-sexton]

Host eval reruns for #11187, #11189 and #11191, on this PR's head d26e742 (which includes the review fix to the #11189 rule).

How they ran: claude plugin eval <out-of-repo copy of plugins/planning> --case <case> --runs 3 -j 2 --scaffold --trust-plugin --keep-temp --no-publish --allow-tools Bash Write Edit --judge-model opus, one case at a time, empty HOME and a short TMPDIR. Every run is VALID under run-validity.py --runs 3 --scaffolded. The only warning is "skill fired": the slash command loads the hub without a Skill call.

Case	Issue	With	Without	Bar (0.8)
interview-readback-dictated-answer	#11187	1.00	0.67	met
interview-closing-probe-something-else	#11191	1.00	0.50	met
interview-final-brief-changes-only	#11189	0.67	0.50	not met

On #11189's case, the llm grader changes-only-restatement passed 2 of 3 with-arm runs and 0 of 3 without. The regex grader omits-unchanged-items failed all 6 runs.

• In with-arm runs 1 and 3 it is a false fail. Those runs left the unchanged answers out and mentioned UTC or the 50,000 cap only while reasoning about Q5 and Q9.
• Run 2 re-listed them, a correct fail.

Filed as #11591. #11189's done-when makes the 0.8 bar the rerun's check, not this PR's gate.

This was generated by AI.