phrazzld/critique

/critique

Single-lens signal. Not a merge gate, not /code-review, not a parallel bench.

Contract

• --lenses lists available lens names by reading harnesses/shared/references/lenses.md; do not hardcode the list here.
• --lens <name> resolves the matching ## <name> block from that rubric. Unknown lens means stop and print the available names.
• --target <path> is required for a focused critique. Resolve it against the invoking repo and read only the files needed to answer.
• Never reference agents/<lens>.md. Lenses are rubrics, not static persona files.
• The critic is read-only. Do not edit, stage, commit, or emit a Ship verdict.

Dispatch

Spawn one fresh read-only critic subagent with:

• role: critic
• objective: apply <lens> to <target> and find production-relevant design, correctness, test, or operability risks
• scope: target path, relevant nearby contracts, and the exact lens excerpt
• boundary: no edits, no broad review, no author reasoning trail
• output: structured findings only

If the harness cannot spawn a fresh subagent, the primary adopts the exact lens text for one read-only pass and marks the header lens-adopted-by-primary. This is a fallback, not fresh-context separation.

Output

## Critique
- Lens: <name>
- Target: <path>
- Evidence read: <files/commands inspected>
- Fresh-context mode: <subagent|lens-adopted-by-primary>
- Findings:
  - Finding: <concise claim>
    Evidence: <file:line>
    Impact: <high|medium|low>
    Suggested next step: <smallest useful action>
- Residual uncertainty: <what was not read or could change the verdict>

Findings require file:line evidence. If no finding survives, say so and cite the files that were read. Do not return Ship, Conditional, or Don't Ship; that verdict belongs to /code-review.

Examples

• /critique --lenses
• /critique --lens ousterhout --target src/auth
• /critique --lens cooper --target tests/auth.spec.ts