[cleak]

Summary

This is S3a of docs/interview-challenge-research.md. It's stacked on #57 (eval harness), which is stacked on #55 (skill) and #54 (Rust). It relies on #54's graph_traverse {include_body}, interview_show {full}, and journal_search tag/reference filters.

Research and five rounds of critic review pointed the same way. Critics work best when they are independent, run in a fresh context, define their target narrowly, and get filtered deterministically. A multi-agent debate panel would not have done as well. Critics also belong in an on-demand command that can challenge any node, including ones no interview ever touched. They should not be wired into every turn.

tempyr-critic (subagent)

• One critic. The lens is passed in:
  • kill-pass: the single assumption that kills the idea, plus the cheapest test for it. It has a mandate-focus variant.
  • pre-mortem
  • assumptions
  • outside-view
• It sees a node packet and the settled items, never the conversation.
• Every objection must quote at least 15 characters verbatim from its target node, so it can't get away with generic "consider security". It must not argue settled items. "No objection, here's what's well-reasoned" is a valid answer.

/tempyr-spar <node> [hops N] [lens …|panel]

1. Builds the packet with graph_traverse. Outside-view and kill-pass packets drop solution nodes.
2. Collects prior rebuttals and dismissals from the journal by tag and reference.
3. Validates quotes after normalization (CRLF, whitespace, curly quotes, case). It drops anything targeting a settled item, a needs: line, or a generated block.
4. Shows the top 3. The user answers in one line, e.g. 1a 2r launchd handles it 3d.
5. Applies each choice:
   • Accept previews and then writes a risk or open question. If the node type has no valid edge for that, it appends to the body instead.
   • Change edits the node.
   • Rebut and dismiss only record.
6. Journals every choice with references and tags: ["disposition:…"], so a rebutted objection never comes back.

Interview kill-pass

• When the interview enters Technical, the skill launches the critic in the background with an outside-view packet and keeps interviewing.
• The result becomes the next push: "An independent red-team pass flagged one thing: …".
• If it's still pending at Review, the skill waits up to ~30s and then lists it.
• At hard level, an announced pre-Review sweep also runs.

Also

• The extractor returns a weakest_claim (a push candidate written from a fresh context) and marks weak answers with needs:.
• The managed files gain SparSkill and CriticAgent. They're implied by the existing Skill and Agent selections, so existing projects get them on update. The manifest still records only the user-facing artifacts, so older binaries can read it.
• Codex gets mirrored interview and spar skills under .agents/skills/.
• Eval harness: for arms whose skill has a kill-pass, the harness runs the critic itself at the Technical transition, because workflow subagents can't launch their own. It passes the result to the next interviewer turn and scores whether the interviewer surfaced it. The target is ≥ 8/9 runs.

Test plan

• cargo test -p tempyr-cli passes. New test: Skill and Agent imply the spar skill and critic, and the manifest records only Hooks, Skill, and Agent. Updated two tests that assumed three managed files.
• cargo clippy -p tempyr-cli --all-targets and cargo fmt --check.
• Ran workflow.js with mocks. The kill-pass runs once per kill-pass arm after Technical, is delivered on the next turn, and the baseline reports none.
• setup.sh copies the critic into the sandbox.
• Live /tempyr-spar on a real graph, and kill-pass numbers from a full /tempyr-eval run.

🤖 Generated with Claude Code

[coderabbitai]

Warning

Review limit reached

• Run on-demand review

_{This review includes 12 billable files and costs up to $3.00.}

• Ask an admin to make reviews automatic

Open in CodeRabbit

Reviews can continue after your included limit without a manual trigger. An admin must approve usage-based billing.

Or wait 43 minutes for your next included review.

Check out review usage here.

View limit details

Limit details: You’ve used the included review currently available. Your 74 included PR review attempts over the past 7 days set your current allowance at 1 review per hour.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: ASSERTIVE
• Plan: Essentials
• Run ID: 239aadf6-b26a-4f3c-b509-2b9365d231fb

📥 Commits

Reviewing files that changed from the base of the PR and between 1725952 and 0827503.

📒 Files selected for processing (12)

• .claude/agents/tempyr-critic.md
• .claude/agents/tempyr-extractor.md
• .claude/skills/tempyr-interview/SKILL.md
• .claude/skills/tempyr-spar/SKILL.md
• crates/tempyr-cli/assets/tempyr-interview.codex.SKILL.md
• crates/tempyr-cli/assets/tempyr-spar.codex.SKILL.md
• crates/tempyr-cli/src/commands/init.rs
• crates/tempyr-cli/src/commands/managed.rs
• evals/interview/rubric.md
• evals/interview/setup.sh
• evals/interview/template/.claude/skills/tempyr-eval/SKILL.md
• evals/interview/workflow.js

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}