interrogate

Files (5)

• SKILL.md
• code-quality-review.md · references/
• lead-judgment.md · references/
• reviewer-prompt.md · references/
• rubric.md · references/

SKILL.md (first lines)

# Interrogate

Read [the workflow contract](../../references/workflow-contract.md) and
[review routing](../../references/review-routing.md). Model diversity—not personas
on one model—is the independent signal here. It uses `reviewer`, not the thermos
lens agents; no model pins in definitions, edits or automatic fixes. Models come
from per-call `model` on Rho's `agent`, or from host model-selecting delegation.

1. **Scope and intent.** Resolve the requested files/diff/PR and actual base/head,
   not an assumed main. Pin requested working-tree evidence if applicable. State
   the intended outcome and repository constraints; ask only if missing intent
   materially changes correctness. Parent gathers all Git/PR evidence and pinned
   context. A moving snapshot invalidates affected conclusions.
2. **Check capability.** Prefer `reviewer` calls with per-call `model` (see review
   routing), using models the user named or configured aliases; ask once if none
   are known. Otherwise discover the host's model-selecting delegation API (in T3,
   `orchestrator_capabilities`, `delegate_task` and the live catalog). Choose at
   least two genuinely different model families; do not infer diversity from
   different IDs. Without either, state **cross-model interrogation unavailable**,
   and offer an ordinary same-model `reviewer` pass only if the user accepts. Do
   not install definitions or change global configuration.
3. **Launch.** Fill `references/reviewer-prompt.md` with the same intent, pinned
   scope/evidence, `references/rubric.md`, `references/code-quality-review.md` and
   workflow contract contents for every reviewer. Supply actual text, not unresolved
   paths. Require read-only operation, no builds/tests, no delegation, no external
   actions, and explicit missing-context/status reports. Use immutable evidence
   or supplied snapshots, never mutable-worktree substitutes. Launch independent
   child runs in one batch (host tasks follow host instructions). Record run IDs,
   actual models and any errors/substitutions; no top-level conversations or
   homemade shell model runner. Await every required completion, without polling
   loops or duplicate work. Failed/unavailable routes are not clean reviews.
4. **Synthesize.** Deduplicate by actual issue, keep model attribution and disagreements.
   Apply `references/lead-judgment.md`; verify real paths/counterevidence before
   accepting or dismissing. Consensus is signal, never proof; a lone security or
   correctness finding still warrants investigation.
5. **Judge.** Sort into Act on (concrete material issue), Consider (valid tradeoff),
   Noted (not actionable now), Dismissed (wrong/context/style, with rationale).

Return Intent; Reviewers (actual provider/model, status, finding count); Act on /
Consider / Noted / Dismissed with evidence/remedies; Agreement map; failures and
unreviewed areas. Never describe a user-accepted same-model alternative as
cross-model interrogation. For PR work, use `thermos` as the one combined review
unless a distinct interrogation was requested; do not stack redundant review runs.

From matthewyjiang/rho-pstack, MIT.

More from matthewyjiang/rho-pstack