mrcha033/skills

Advisor Review

Use the bundled scripts from this skill directory. An explicit invocation is a latch: run the reviewer in the same parent task before a substantive final answer or an action that follows or conflicts with the review. Read-only evidence collection may continue while constructing the packet.

Do not search the global tool catalog for an advisor. Do not open a second user-visible task. The only child execution is the isolated codex exec process started by scripts/run_advisor.py.

Requirements

Require shell execution, an authenticated local codex CLI, access to the requested gpt-5.6-sol model route, and outbound Codex API access from the parent shell. A parent sandbox that blocks subprocess networking also blocks the nested reviewer. If any requirement is missing, report that the independent review is blocked; do not weaken the sandbox without user or task authority.

The runner is not a confidentiality boundary. Sanitize evidence and artifacts before building the packet.

Workflow

1. Announce that Advisor Review is being used and why.

2. Select one phase:

   • plan: challenge a proposed approach.
   • stuck: diagnose failed or looping work.
   • pivot: compare a materially different approach.
   • final: audit completed work and its endpoint evidence.

3. Read references/context-contract.md.

4. Choose the context mode:

   • packet: short source-anchored facts are sufficient.
   • bundle: sanitized logs, timelines, diffs, or other bounded text artifacts materially affect the diagnosis.

5. Build advisor-context-2.0 with scripts/build_context_packet.py.

   • For stuck, include actions and exact results in chronological order.
   • Give every observed fact a concrete source.
   • Include negative results, conflicting state, and context limitations.
   • Never silently truncate decision-critical text. A builder failure is a blocked review.

6. Read references/review-rubric.md.

7. Select effort:

   1. Honor an explicit request for high, xhigh, or max.
   2. Use max only for an explicitly requested deepest review or a critical, hard-to-reverse decision with major unresolved ambiguity.
   3. Use xhigh for cross-component failures, conflicting verified evidence, two or more failed approaches, major pivots, or high-impact completion claims.
   4. Otherwise use high.

8. Run exactly one isolated review by default and save its receipt:

   python3 scripts/run_advisor.py \
     --input context-packet.json \
     --effort xhigh \
     --output advisor-receipt.json
   

   Keep the returned receipt in the invoking task. A receipt produced in another user-visible task is not a handoff.

9. Check the receipt:

   • request.requested_model must be gpt-5.6-sol.
   • request.requested_effort must match the selected effort.
   • observed_model and observed_effort may be null; the current CLI does not authoritatively expose serving-side identity. Never describe requested identity as verified actual identity.
   • Any runner error, schema failure, timeout, or context-hash mismatch blocks the review.
   • The isolated child uses a temporary CODEX_HOME with only the host authentication file linked when present. It also disables plugins, remote-plugin loading, skill search, and multi-agent execution so unrelated skill catalogs cannot consume the review context.

10. Read references/decision-contract.md. Resolve every R# recommendation as adopt, reject, or defer, then run:

    python3 scripts/validate_decision.py \
      --receipt advisor-receipt.json \
      --decision advisor-decision.json
    

11. Do not act on advisor recommendations until the decision record validates. A destructive recommendation additionally requires confirmed user authority.

12. Continue the parent task using the validated next action and stop condition.

Stuck-work rules

• Diagnose before changing more state.
• Distinguish the execution boundary: parent task, isolated reviewer process, and any other user-visible task are separate.
• Do not let a separate task's receipt stand in for a same-task handoff.
• Prefer one read-only or reversible experiment that separates competing hypotheses.
• After a repeated user correction, do not repeat the same prescription without new evidence.
• Do not broaden scope from one failing subsystem to plugins, services, repositories, or packages unless evidence connects them.

Review budget

Use one review by default. A second is allowed only when new evidence materially changes the decision or a high-risk final re-check is justified. A third is allowed only to reconcile a concrete conflict. Never create an open-ended review loop.

Skill-change behavior gate

When modifying Advisor Review itself, run the deterministic tests and at least one sanitized behavior fixture:

python3 scripts/evaluate_advisor_behavior.py \
  --fixture behavior-fixture.json \
  --receipt advisor-receipt.json

Use paired old/new fixtures where practical. Judge detection, evidence linkage, bounded experiments, and risky-action rate—not prose similarity.

Reporting

Report:

• phase, context mode, requested Sol model, requested effort, and identity verification status;
• advisor verdict and evidence-linked diagnosis;
• each consequential recommendation's adopt/reject/defer disposition;
• the selected next action and stop condition;
• unresolved risks or context limitations.

Call the result an advisor-style independent review, not Claude Advisor parity.