feat(planning): route the ungrillable question to a prototype detour, and leave plan mode off while interviewing

[kyle-sexton]

Closes #2998

Summary

Course lane 5's /planning:audit-answers pass — two fresh-context validators reading the upstream
primaries independently — overturned the lane's provisional zero-work-items answer and surfaced two
grilling-family techniques absent from planning:interview. This adopts both, and reconciles the
planning:plan sentence that sat in tension with the second.

Documentation only; no behavior change. planning 0.32.0 → 0.33.0.

Fix

The ungrillable question → prototype detour. Some questions cannot be answered by asking,
however well the round is framed: the user has to see a thing and react to it before they know what
they think. Upstream fires the escape mid-grilling ("stop grilling… build the throwaway version…
come back and answer in one line"). Our route existed only downstream — wayfind, plan,
brainstorm — while the interview's mid-session composition named research, exploration, and
handoff only, and its categorization taxonomy had no needs-an-artifact arm.

• interview/SKILL.md "Mid-interview composition" now carries the detour, routing a look-and-feel
  question to /prototype:explore-directions and a logic/state/data-shape question to
  /prototype:pressure-test (invoked via the Skill tool, if installed), then returning to answer in
  one line. A matching composition-table row keeps the two surfaces consistent.
• interview/context/loop.md gains a Needs-an-artifact arm beside resolvable / blocked /
  defer-with-assumption / defer-fully, marked explicitly as a route, not a deferral.

The prototype is the instrument that produces the answer, not a deliverable — it is thrown away. The
question is registered open before the detour so it cannot be lost across it.

Plan mode off while interviewing. Beyond upstream's taste point, the audit surfaced a mechanical
edge: the ask-time open-question register is a disk write (the ledger's ## Open-question register section — verified against context/loop.md "Write at ask-time, not at answer-time", not
inferred), and plan mode's read-only enforcement blocks it. A round asked under plan mode therefore
leaves nothing on disk holding it — precisely the failure the register exists to prevent,
reintroduced by the permission mode.

• New interview/context/gotchas.md entry under the Q&A surface, beside the other register gotchas.
• plan/SKILL.md "Plan Mode Integration" reconciled with lane 4's asset-rush doctrine: its
  plan-mode round is now scoped to a scoping confirm, and substantive what are we building
  questions route to /planning:interview outside plan mode, on both the register-write mechanics
  and the doctrine that plan mode primes the run toward producing the asset while the job is still
  reaching shared understanding.

Merge with main (2026-08-20)

main shipped planning 0.32.0 for #2936 while this branch had already claimed that number for
#2998, so the branch merged main and re-numbered:

• plugins/planning/CHANGELOG.md — the planning: interview prototype detour + plan-mode-off guidance (aihero lane 5 audit findings) #2998 entries move up under a new ## [0.33.0] heading;
  main's ## [0.32.0] section and its Lane C: TDD wiring + zero-assembly default chain #2936 entry are preserved verbatim beneath it (all 55
  headings confirmed preserved by check-changelog-parity.sh --check-preserved origin/main).
• plugins/planning/.claude-plugin/plugin.json — 0.32.0 → 0.33.0 to match.
• plan/SKILL.md auto-merged cleanly; both Lane C: TDD wiring + zero-assembly default chain #2936's test-boundary naming and this branch's
  plan-mode routing are present in the merged file.

Verification

Full local gate re-run on the merge commit, all green:

Gate	Result
scripts/validate-plugins.sh	all manifests + catalog validated
generate-catalog.mjs --check / generate-cheatsheet.mjs --check	both in sync
check-changelog-parity.sh --check / --check-order / --check-bump origin/main / --check-preserved origin/main	all pass; 83 changelogs newest-first, 55 headings preserved
check-changed-skills.sh origin/main	2 skills checked, 0 failed
markdownlint-cli2 (6 touched files)	0 issues
typos	clean
planning test suites (standards-binding, goal-condition-length, check-open-questions)	PASS=9 FAIL=0, plus both script suites

The three check-changed-skills warnings are pre-existing on origin/main, not introduced here:
the SKILL.md soft line-count target (interview and plan were both already over the 200 target before
this change) and the numbered-procedure completion-criteria warning on the unchanged
Plan Review Mode list. The REcompile() - panic lines in that run are the known mawk/ERE-interval
artifact tracked as #3005, not a failure — the script exits 0.

Related

• Refs planning/interview: aihero course lane 5: grilling parity check #2903 (course lane 5, the grilling-interview parity lane that filed this)
• Refs planning: aihero course lane 4: plan-mode asset-rush critique vs interview-first sequencing #2902 (lane 4, the asset-rush doctrine the plan reconciliation answers to)
• Refs docs(upstream): AI Hero course journey roadmap — remaining lanes and filed follow-ons, in order #2994 (the durable AI Hero course roadmap this item sits in)
• Refs Lane C: TDD wiring + zero-assembly default chain #2936 (the planning 0.32.0 release this branch re-numbered around)
• Refs skill-quality: check-skill.sh fresh-eyes scanner breaks under mawk — ERE interval expressions unsupported #3005 (the mawk/ERE-interval scanner artifact visible in the gate output above)

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01QbfCrj3X9FfGL7VRZYmrn4

---

Generated by Claude Code

[claude]

Claude finished @kyle-sexton's task in 4m 36s —— View job

---

Code review complete

• Gather context (PR diff, CLAUDE.md, changed files)
• Run skip gate — not applicable, PR is open and has meaningful review surface
• Review changed files for correctness/maintainability
• Post findings (inline + summary)

This is a documentation-only change to the planning plugin's interview and plan skills (SKILL.md prompt files, which double as this repo's executable spec, so I reviewed them for instructional consistency, not just prose). Verified the two factual claims the PR body makes against the actual files: the ## Open-question register heading name (interview/templates/checklist.md:14) and the "write at ask-time" rule (interview/context/loop.md:171-173) both check out — the register mechanism described is accurate. The /prototype:explore-directions and /prototype:pressure-test skills referenced from the new mid-interview detour both exist in the repo. The version bump (0.31.0 → 0.32.0) and CHANGELOG entry match the diff.

One finding posted inline:

• plugins/planning/skills/plan/SKILL.md line 253 — the new "Substantive rounds do not belong in plan mode" text says to route to /planning:interview "run with plan mode off," but doesn't say how, and this conflicts with the section's own established doctrine two lines above (line 249: "the skill does not automatically enter plan mode — the user controls permission modes") — which handles the symmetric entering-plan-mode case by suggesting it to the user, not doing it directly. Without an equivalent "ask the user to exit" step here, a session mid-formulation in plan mode that hits a substantive question has no scripted way to satisfy "plan mode off," and the likely fallback is invoking /planning:interview while still in plan mode — reproducing the exact register-write failure this same paragraph exists to prevent. Full detail and a suggested fix in the inline comment.

Everything else — the new Needs-an-artifact categorization arm in interview/context/loop.md, the gotchas.md addition, and the composition-table row in interview/SKILL.md — is internally consistent and cross-references correctly.

[claude]

Claude finished @kyle-sexton's task in 48s —— View job

---

Security Review

• Gather context (PR diff, changed files)
• Run skip gate / assess security-relevant surface
• Review changed files for security issues
• Post final findings

Skip gate triggered — no security-relevant surface.

All 6 changed files are documentation-only (Markdown skill instructions + CHANGELOG.md + a plugin.json version bump, 0.31.0 → 0.32.0):

• plugins/planning/.claude-plugin/plugin.json — version bump only
• plugins/planning/CHANGELOG.md — changelog prose
• plugins/planning/skills/interview/SKILL.md — new "ungrillable question" mid-interview detour row/paragraph
• plugins/planning/skills/interview/context/gotchas.md — new gotcha about plan mode blocking the ask-time register write
• plugins/planning/skills/interview/context/loop.md — new "Needs-an-artifact" categorization arm
• plugins/planning/skills/plan/SKILL.md — reconciles plan-mode question scope with the interview skill

There is no code, no script, no GitHub Actions workflow, no permission/config change, no dependency/action pin, and no executable logic anywhere in the diff — just prose describing when/how an agent should invoke other skills (/prototype:explore-directions, /prototype:pressure-test, /planning:interview) via the Skill tool. No injection surfaces, no secret handling, no trust-boundary changes, no authorization logic.

No security issues found. Per this lane's high-signal bar, nothing further to report.

[github-actions]

Last security-reviewed head: 9b58bfbf754fb6ec353c1f7ebbd5b90ce0dc56f7. On the next push, the relevance gate compares only the commits since this SHA; delete this comment to force a full re-review.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 91cbd014d6

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".

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".

[claude]

Claude finished @kyle-sexton's task in 43s —— View job

---

Security Review

• Gather context (PR diff, changed files)
• Run skip gate / assess security-relevant surface
• Review changed files for security issues
• Post final findings

Skip gate triggered — no security-relevant surface.

Per the lane's skip gate, this run compares only the commits since the last-reviewed head (91cbd014d6843f2af62a7ceecf539b47293a7432) through the current HEAD (5d4212c3c7dee340e4d5a2af063a8a6efca2883c). That delta is three files, all prose:

• docs/upstream/aihero-course.md — marks the two ADOPT rows "(delivered)" and links this PR
• plugins/planning/CHANGELOG.md — changelog prose addition describing the prototype-gate fix
• plugins/planning/skills/interview/SKILL.md — reworks the "ungrillable question" mid-interview
  detour paragraph to route through a scope-confirm + checkpoint step before invoking
  /prototype:explore-directions or /prototype:pressure-test, addressing the earlier Codex
  finding that the detour skipped the prototype skills' own model-initiated entry gate
  (plugins/prototype/context/discipline.md)

No code, no script, no GitHub Actions workflow, no permission/config change, no dependency/action
pin — just Markdown instructing an agent when/how to invoke another skill via the Skill tool. No
injection surfaces, no secret handling, no trust-boundary changes, no authorization logic.

Combined with the earlier pass on 91cbd01 (also a full skip — the whole PR is documentation-only:
skill instructions, a CHANGELOG.md, and a plugin.json version bump), the entire PR head has now
been covered. No security issues found. Per this lane's high-signal bar, nothing further to report.

[github-actions]

Claude has reviewed this PR 3 times. The lane skips further automatic reviews after 5; deleting this comment resets the count.

[claude]

Claude finished @kyle-sexton's task in 46s —— View job

---

Security Review

• Gather context (PR diff, changed files)
• Run skip gate / assess security-relevant surface
• Review changed files for security issues
• Post final findings

Skip gate triggered — no security-relevant surface.

Per the lane's skip gate, this run compares only the commits since the last-reviewed head
(5d4212c3c7dee340e4d5a2af063a8a6efca2883c) through the current HEAD
(19c7cbee29001add9ce5c6cab955ea6af34b916c). That delta is one commit, two files, both prose:

• plugins/planning/CHANGELOG.md — changelog clause describing the fix below
• plugins/planning/skills/plan/SKILL.md#L253 — adds the missing mechanism for exiting plan mode before invoking /planning:interview (the correctness gap flagged in the earlier code review): the skill now explicitly asks the user to exit plan mode (shift+tab) rather than asserting "plan mode off" with no scripted way to get there.

No code, no script, no GitHub Actions workflow, no permission/config change, no dependency/action
pin — just Markdown clarifying when/how an agent should ask the user to change permission mode
before invoking another skill via the Skill tool. No injection surfaces, no secret handling, no
trust-boundary changes, no authorization logic.

Combined with the earlier passes on 91cbd01 and 5d4212c (both full skips — the whole PR is
documentation-only: skill instructions, a CHANGELOG.md, and a plugin.json version bump), the
entire PR head has now been covered. No security issues found. Per this lane's high-signal bar,
nothing further to report.

[claude]

Claude finished @kyle-sexton's task in 2m 51s —— View job

---

Code review complete

• Gather context (PR diff, CLAUDE.md, changed files, prior review rounds)
• Run skip gate — not applicable; this sync commit specifically targets the one open finding from the prior round
• Review the delta since the last-reviewed head, and re-check the full diff for consistency
• Post findings

This lane already reviewed this PR twice (at 91cbd01 and again after 5d4212c). The only open finding was mine, from the 91cbd01 pass: plugins/planning/skills/plan/SKILL.md#L253 said to route to /planning:interview "run with plan mode off" without saying how plan mode gets turned off, given the skill can't toggle permission modes itself — and it broke symmetry with the entering-plan-mode case two lines above, which does script a "suggest to the user" step.

This commit (19c7cbe) fixes it, and the fix is correct. The new sentence at line 253 now reads:

Getting there is the user's move, not yours — symmetric to entering plan mode above: you do not toggle permission modes, so when plan mode is active and a substantive round comes due, say why and ask the user to exit it (shift+tab), then invoke the interview once they have. Do not invoke it from inside plan mode on the assumption the register write will survive — it will not.

This closes the exact gap flagged: it's now symmetric with the entering-plan-mode doctrine at line 249 (suggest, don't act), it names the concrete mechanism (shift+tab), and it makes the failure mode explicit rather than implicit. The matching CHANGELOG entry (plugins/planning/CHANGELOG.md [0.32.0] → Changed) describes the same fix accurately.

No new issues found. I re-checked the full diff against origin/main (all 7 files), not just the incremental commit:

• The Needs-an-artifact categorization arm added to plugins/planning/skills/interview/context/loop.md:80 isn't hardcoded anywhere else in scripts (confirmed no script greps the category label set), so it can't silently break a gate.
• The claims in the new plugins/planning/skills/interview/context/gotchas.md:19 gotcha and the SKILL.md:300 "ungrillable question" paragraph — the ## Open-question register heading name and the "write at ask-time" rule — check out against interview/context/loop.md:162 and :171.
• The earlier Codex/security-lane finding about honoring the prototype skills' own model-initiated entry gate (plugins/prototype/context/discipline.md:14-16) was addressed in 5d4212c and remains intact here — the detour text confirms scope and checkpoints before handing off, and the ledger/Brief-per-lock-in mechanism it names is a legitimate (if non-literal) substitute for the /session-flow:handoff example in that gate, since the interview's own on-disk state already survives a crash/clear per its own design.
• docs/upstream/aihero-course.md and plugins/planning/.claude-plugin/plugin.json (0.31.0 → 0.32.0) diffs match what the PR body and CHANGELOG describe.

Note: I did not independently re-run the gate scripts (validate-plugins.sh, check-changelog-parity.sh, markdownlint-cli2, etc.) listed in the PR's Verification table — that table is the author's own claimed output, not something I confirmed myself.

[claude]

Claude finished @kyle-sexton's task in 1m 29s —— View job

---

Code review complete

• Gather context (PR diff, CLAUDE.md, changed files, prior review rounds)
• Run skip gate — skip gate triggered (see below)
• Review the delta since the last-reviewed head
• Post findings

Skip gate triggered — condition 4 (this head already has a successful review from this lane that still applies).

This lane has already reviewed this PR three times, at 91cbd01, 5d4212c, and 19c7cbe — the last of which fixed the only open finding (the plan/SKILL.md "run with plan mode off" mechanism gap) and was confirmed correct in the follow-up pass.

The only commit since 19c7cbe is 9b58bfb, a merge of origin/main into this branch. I diffed origin/main...HEAD (the actual net PR diff, matching the task's reported 7 files / +46/-5) and separately diffed 19c7cbe..HEAD on the PR's own files to isolate what the merge changed:

• plugins/planning/skills/interview/SKILL.md, context/gotchas.md, context/loop.md, docs/upstream/aihero-course.md — byte-identical to the already-reviewed 19c7cbe versions.
• plugins/planning/skills/plan/SKILL.md — the PR's own content (the line-253 "Getting there is the user's move" fix) is unchanged; the only diff since 19c7cbe is an unrelated "Name the test boundaries" paragraph at line 93 that arrived from origin/main (PR Lane C: TDD wiring + zero-assembly default chain #2936), already merged and reviewed independently on main.
• plugins/planning/.claude-plugin/plugin.json / CHANGELOG.md — version bump shifted 0.32.0 → 0.33.0. This is a correct, mechanical rebase: origin/main had already claimed 0.32.0 via the unrelated Lane C: TDD wiring + zero-assembly default chain #2936 merge (confirmed via git show origin/main:plugins/planning/CHANGELOG.md), so this branch's own [0.32.0] heading was renumbered to [0.33.0] with its original content preserved verbatim, and the prior [0.32.0] (from main) sits below it intact.

No new PR-owned content was introduced by the merge, so there is nothing new to review. Per this lane's skip gate, posting nothing further.

[claude]

Claude finished @kyle-sexton's task in 1m 39s —— View job

---

Security Review

• Gather context (PR diff, changed files, prior review rounds)
• Run skip gate / assess security-relevant surface
• Review changed files for security issues
• Post final findings

Skip gate triggered — no security-relevant surface, and this head's diff already fully reviewed.

HEAD 9b58bfb is a merge of origin/main into this branch (bringing in unrelated main-line progress, e.g. plugins/planning/skills/plan/context/plan-template.md and the audit-skill-visibility skill) plus one unrelated main commit (d1c0935). Neither touches this PR's own change set.

Comparing this PR's actual contribution against the current origin/main (git diff origin/main...HEAD, merge-base d1c0935) shows the same 7 files as before, and the added content is byte-for-byte the same prose already covered by the three prior full passes on this PR (91cbd01, 5d4212c, 19c7cbe):

• plugins/planning/skills/interview/SKILL.md — the ungrillable-question detour paragraph and composition-table row
• plugins/planning/skills/interview/context/gotchas.md
• plugins/planning/skills/interview/context/loop.md
• plugins/planning/skills/plan/SKILL.md
• plugins/planning/CHANGELOG.md, plugins/planning/.claude-plugin/plugin.json, docs/upstream/aihero-course.md

The only mechanical difference from the previously reviewed state is that the merge rebased this PR's version bump onto a newer origin/main baseline (main independently landed its own 0.32.0 in the meantime for an unrelated change, so this PR's bump now reads 0.32.0 → 0.33.0 in the diff instead of 0.31.0 → 0.32.0) — a renumbering artifact of the merge, not new content, and not security-relevant.

As established across the earlier passes: this is a documentation-only change (Markdown skill instructions, a CHANGELOG.md, a plugin.json version bump, and one upstream-tracking doc). No code, no script, no GitHub Actions workflow, no permission/config change, no dependency/action pin, no injection surface, no secret handling, no trust-boundary or authorization logic anywhere in the diff.

No security issues found. Per this lane's high-signal bar, nothing further to report.