Thermos

Two diff-scoped review lenses run in parallel on the same scope, then one synthesized verdict.

• agent thermo-review (skill thermo-nuclear-review): bugs, breakage, security, devex regressions, feature-gate leaks. Only added or modified code.
• agent thermo-quality (skill thermo-nuclear-code-quality-review): maintainability, code-judo, 1k-line rule, spaghetti, boundaries, module depth, test value.

Thermos is the default review gate. interrogate is the escalation: use it when design is contested, stakes are high, or the user asks for multi-model or adversarial review. Thermos is read-only. It never edits. Fixing is the owner's job.

Who calls this

• opening-a-pr: pre-PR gate.
• babysit: every new patch, scoped to the delta since the last verdict.
• shipping: independent per-PR verdict, one scope per stack layer.
• autopilot: verifier rounds, as mandatory lanes next to the swarm lanes.
• orchestrate and multi-phase plans: per-PR review step.
• feature, bug-fix, refactoring, and perf playbooks: the "review before PR" step.
• review-bot-triage: thermos is the first-party counterweight to review bots. A bot finding thermos also raised is high signal.

Run no-comments and deslop before thermos, not after.

Workflow

1. Determine the scope from the user request, PR, current branch, or local uncommitted work. Resolve base (default main) and head. For a PR, get the number, title, and body with gh pr view <n>.
2. Write the scope spec. Do not paste the diff. The spec holds:
   • base ref and head ref (or "working tree" for uncommitted work)
   • PR number if any
   • an intent paragraph: what the change is for and what breakage is intended
   • for a stack layer: the one layer under review (git diff <parent-branch>...<branch>) and the names of the layers below, for context only
   • optionally a pointer to a saved diff at local://review/<id>.diff when the diff is big
3. Read model roles. Call pstack_models with role: "thermos-review" and role: "thermos-quality". If the tool is absent, read rule://pstack-models and fall back to the defaults. The extension also applies these roles automatically when the agents spawn, so passing model is optional. Pass it when the role resolves to something other than the agent's alias.
4. Spawn both lenses in ONE task call. tasks[] holds 2 x scopes items, two per scope (one scope for a normal review, one per layer for a stack):
   • agent: thermo-review with the scope spec and a solutionSpace like "open-ended audit, breakage paths unknown".
   • agent: thermo-quality with the same scope spec and a solutionSpace like "structure may admit a much simpler reframing". Use one task call with heterogeneous agent items, not a workpool. A workpool binds one agent and replaces its structured output with a free-form schema. The task tool keeps each agent's own schema, so you get the verdict and priority 0-3 findings back structured. Name items by scope and lens (for example Layer2Review, Layer2Quality).
5. Wait for all items. Synthesize.

Synthesis

Findings first, then the verdict. No praise.

Tag each finding with the lens from the item that produced it (thermo-review or thermo-quality). Do not trust a finding's own lens value for that. The agents' lens field is review-bot or human only when the finding was sourced from PR discussion.

1. Dedup findings across the two lenses and across scopes. Merge same-root findings into one entry with both lenses credited.
2. Overlap raises weight. A finding both lenses raised, or one a review bot raised that a lens also raised, ranks higher.
3. Resolve disagreements with your own judgment. Read the code. Do not average.
4. Bucket every finding:
   • act-on: real, in scope, fix before the step proceeds.
   • consider: plausible or stylistic, owner decides.
   • noted: true but out of scope or intended, recorded only.
   • dismissed: wrong or already handled, with a concrete reason.
5. Verdict is approve or changes_requested. Any P0 or P1 in act-on means changes_requested. It blocks the step until fixed or dismissed with a concrete reason.
6. If show-me-your-work is active, add a trail entry: scope, lenses run, buckets, verdict, dismissals with reasons.

If the agents' summaries are already visible to the user, do not restate them. Surface the unified verdict, the highest-signal findings, and remaining uncertainty.

Review a stack

One scope per unmerged layer, bottom to top.

1. Get the layers from gh stack view --json (or the stacks REST endpoint).
2. Each layer's diff is git diff <base-parent-branch>...<layer-branch>, that PR's base to its head. The bottom layer's parent is trunk.
3. Lower layers are context only. A layer is not blamed for code that lower layers added.
4. Spawn a thermo-review and thermo-quality pair per layer in the same single task call (step 4 above), then synthesize per layer. Report one verdict per PR.
5. Re-review a layer when its head SHA changes. Do not re-review unchanged layers. Scope the re-review to the delta: base = the SHA of the last verdict, head = the layer's current head. Name the layer's parent branch in the spec as context so the reviewer can tell what lower layers own.