AI review gate non-functional: pass-2 extraction crashes (promptArgs + inline resume prompt)

[lsfera]

The AI review gate (ADR-0020) never produces a real verdict — its extraction pass crashes deterministically, so every review fail-safes to changes-requested and parks for a human. In practice /afk has been silently degrading to /hitl on every PR. Discovered dogfooding #80 (the reviewer ran on PR #81, pass-1 succeeded, then pass-2 crashed).

The bug

In reviewer-adapter.ts, review() runs a two-pass produce→extract flow. Pass-2 resumes the pass-1 session with an inline prompt string:

const extractResult = reviewResult.resume
  ? await reviewResult.resume(
      await loadPromptFile(join(_dir, "review-extraction.md")), // inline prompt STRING
      { output: Output.object({ tag: "output", schema: reviewOutputSchema }), cwd },
    )
  : await run({ /* fallback: promptFile, output, ... */ });

resume() carries the pass-1 run's promptArgs into this inline-prompt call, and sandcastle rejects the combination:

(FiberFailure) PromptError: promptArgs is only supported with promptFile.
Inline prompts (prompt: "...") are passed to the agent as-is — interpolate values
directly in JavaScript, or switch to promptFile to use {{KEY}} substitution.
  at @ai-hero/sandcastle/src/run.ts:661

review() catches it → returns null → reviewVerdict(null) → changes-requested → WaitForHuman. Deterministic on every review. (Note: review-extraction.md has no {{KEY}} placeholders, so the extraction pass needs no promptArgs at all.)

Goal (end-to-end value)

The reviewer's extraction pass succeeds and returns the schema-validated structured verdict, so a genuine pass actually enables auto-merge instead of every PR parking for a human. The gate becomes functional, not a disguised /hitl.

Acceptance criteria

• Extraction pass no longer crashes: the pass-2 resume invocation does not combine promptArgs with an inline prompt. Fix per the sandcastle error's guidance — pass extraction via promptFile (preferred: keeps review-extraction.md as the source of truth and {{KEY}}-substitution available), or pass the prompt inline with no promptArgs. The session-resume (so the agent recalls its pass-1 review) must be preserved — do not silently fall back to a fresh run() that re-reviews from scratch.
• Real verdict round-trips: a successful review returns the validated ReviewOutput ({verdict, summary, comments}); a genuine pass → EnableAutoMerge. The existing validate-then-filter (parseDiffLines/filterInlineComments) and fail-safe (null → changes-requested) are unchanged.
• Fail-safe still intact: a real extraction failure (timeout, malformed <output>, schema violation) still returns null → changes-requested. Don't widen the catch into a fail-open.
• Builds on Reviewer produce pass re-runs up to 8× — make it single-pass #80: the pass-1 single-iteration fix (buildReviewerPassOneConfig, maxIterations: 1) stays as-is. Mirror it with a pure config builder for pass-2 if it makes the fix testable.
• Don't touch reduce.ts (verdict mapping / gate, Review gate before auto-merge in /afk (AI Reviewer step) #74/ADR-0020) or pass-1.

Testing (highest seam — logic in TypeScript, plus one gated live check)

The bug lives in the live sandcastle resume interaction, which a pure unit test cannot fully catch — so cover it at two levels:

• Pure seam (default npm test, no Docker/network): extract a buildReviewerPassTwoConfig-style pure function (mirroring buildReviewerPassOneConfig from Reviewer produce pass re-runs up to 8× — make it single-pass #80) returning the extraction invocation shape, and assert it does not pass promptArgs alongside an inline prompt (e.g. it uses promptFile, or sets no promptArgs). Prior art: reviewer-adapter.test.ts (Reviewer produce pass re-runs up to 8× — make it single-pass #80).
• Gated integration (SANDCASTLE_INTEGRATION=1, integration.test.ts): run review() against a tiny real PR branch and assert it returns a non-null ReviewOutput (proves pass-2 actually round-trips). Keep this out of the default npm test / sandcastle CI job so they stay green without network.

Why the #80 tests didn't catch this

#80's unit tests assert the pass-1 config (maxIterations, promptArgs, name) — they never invoke the live resume(), where this bug lives. Hence the gated integration check above.

Constraints (per repo workflow)

Implement with /tdd, run shell via /exec, scope to reviewer-adapter.ts + its test (+ a gated integration case), no new dependencies, no pushing to main.

[lsfera]

AFK agent failed 3 times (retry limit: 2) — leaving this issue unlabelled for a human to inspect and re-label ready-for-agent if appropriate.