[kyle-sexton]

Closes #8333

Summary

In an engineering session /planning:interview now reads the consumer's standards before any bundled default, asks the intent question (is a second variant known, a category, or speculative; what the existing structure does; what upstream context says), and prompts quality requirements (NFRs) from a new prompt list, each scaled by the quality bar. The Brief records Quality bar: and Intent: lines under Captured assumptions and each quality requirement as a Quality requirement: line under Constraints.

Fix

• plugins/planning/skills/interview/context/loop.md gains an "Engineering defaults" section; context/nfr-prompts.md is the new prompt list; SKILL.md points to the section after the Step 1 survey.
• Q47 interpretation, for review:
  • A confirmed Brief is one whose Restatement: line reads confirmed by the user.
  • Text a bot or a teammate wrote informs the recommendation and never settles intent.
  • A refactor that extracts a new abstraction asks the full intent question.
  • Unattended, the bug-fix or refactor collapse applies only when the user's own words name a bug fix or a behavior-preserving refactor.
• lock rule: a quality prompt at the bar's depth that the request does not cover becomes a captured assumption with a revisit trigger (or is stated as not applying), never a gap that halts the run.
• Consumer standard wins: when a consumer standard contradicts a bundled default, the Brief template records - Overrides bundled default <principle> per <standards file> under Captured assumptions.
• Blocker last-line shape: unattended, an unstated intent is recorded blocked (arbiter USER-RESERVED), the Brief is still written, and the run's last line is Blocked: Q<N> <question> (arbiter: USER-RESERVED).
• Upstream context is recorded as a short summary; private titles and URLs are left out in public or unknown-visibility repositories.
• Five new eval cases under the planning plugin's evals; changelog fragment .changes/planning/feat-8333-interview-intent-and-quality-requirements-4c7d2a91.md (minor).
• Not done: no pointer was added inside SKILL.md Step 1. That section is digest-pinned by plugins/planning/tests/interview-defenses.test.sh, and SKILL.md:45 already points to "Engineering defaults".

Verification

• Paid evals, branch plugin vs main plugin, 3 runs per arm. The noise report says n is too small to call any of these.

  Case	Round	Branch	Main
  nfr-quality-requirement (unattended)	r4	1.00	0.00
  consumer-standard-override (unattended)	r4	1.00	0.50
  intent-unattended-blocked	r2	1.00	0.58
  intent-asked-recorded	r2	1.00	0.00
  intent-prototype-one-line	r1	0.78	0.67

• Two cases run unattended because an attended one-prompt lock correctly stops on open domain questions, so it cannot test what the Brief contains.

• bash scripts/check-affected.sh --tests --base origin/main: 8 passed, 1 skipped over 47 changed files.

• typos --format brief on the changed paths: clean.

Related

• feat(planning): carry universal, overridable design defaults through planning output #8326
• planning/prd: add a quality-requirements (NFR) section #8322
• feat(review): add /review:quality-gate code-design mode #8388
• feat(planning): add a design-principles reference cited to canonical sources #8393

🤖 Generated with Claude Code

[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 2026-10-19T07:25:15.500936Z	a3baffb	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 15c8314fa139c7eba0826259498791a3b317e7da: 0 finding(s), 0 as inline comments.

Security review (incremental, fa45f596..15c8314f): no security issues found.

The only change is one line in plugins/planning/skills/interview/context/loop.md:77. It splits the lock gap rule by mode. Attended runs still "STOP and surface it". Unattended runs follow the Unattended bullets.

Because the edit narrows a standing instruction, I read it as an instruction-surface change. I checked whether the Unattended path still enforces the same control, using the file's own text. I did not read docs/conventions/instruction-exception-register/README.md, so I did not check the change against its protected-class list.

• Unattended handling: the Unattended bullets (loop.md:108-119) and the ladder (loop.md:386-396) cover a missing or upstream-only part 1 answer. The row is recorded blocked with arbiter: USER-RESERVED. The Brief goes to ### Deferred questions and the run ends on a Blocked: Q<N> line. The run does not assume the answer or hold the lane. The contract is reported unconfirmed.
• No new assumption path: the Unattended bullets say only the user's own words, or a Brief or PRD the user confirmed, can settle part 1. Text from a bot or a teammate cannot, and conflicting sources count as unstated. That is stricter than "upstream only", so injected upstream content cannot settle a user-reserved decision.
• Auto-guard preserved: the ladder keeps the rule against silently folding a user decision into the Brief as an assumption (loop.md:398).

The reworded rule keeps its control, in both attended and unattended runs. No finding is warranted.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a3baffbb93

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".

[github-actions]

claude-review reviewed 15c8314fa139c7eba0826259498791a3b317e7da: 0 finding(s), 0 as inline comments.

Incremental review of plugins/planning/skills/interview/context/loop.md since fa45f59: no findings.

The change is one line, in the lock paragraph of "Engineering defaults". The gap sentence now reads "attended, STOP and surface it; unattended, the Unattended bullets below apply."

• It fixes a conflict. The old text said STOP on a gap, which contradicted the Unattended bullets. Those bullets record the intent row blocked and write the Brief anyway. Before this change, an unattended lock run hitting a missing or contradictory part 1 answer had two incompatible instructions.
• The pointer resolves. The **Unattended** bullets sit at lines 108–123, after the changed line. They cover the missing and contradictory cases: "Part 1 unstated" → blocked row, and "sources that conflict count as unstated".
• Both modes are covered. Attended runs keep the STOP behavior and unattended runs follow the bullets, so neither path is left undefined.
• Not a finding. The new line runs about 150 characters while the paragraph wraps near 100. Lines 56, 115 and 117 of the same file already run longer, and no linter catches it, so it falls below the bar.
• Closing issue feat(planning): interview asks intent and quality requirements, reads consumer standards #8333. The incremental change doesn't touch any of its acceptance criteria. I checked only this diff and the surrounding file. I did not re-review files reviewed at the earlier SHA, and I did not run the tests or evals. The PR body's verification numbers are author-claimed and unverified by me.