thermos

name	thermos
description	Launch both thermo-nuclear review subagents in parallel, then synthesize their findings. Use for thermos, double thermo review, or combined bug/security and code-quality branch audits.

Thermos

Run the two thermo review passes as async background subagents in parallel, then synthesize their results.

Model routing

Before each reviewer edge, invoke railyard:model-routing with exact contract railyard/model-routing/v1. It is the only public model, effort, budget, and transport router and internally applies ../../references/provider-task-routing.md; do not invoke that reference as a second router. Standalone Thermos admits its two reviewer actions. When Goal Driven Delivery invokes Thermos, consume its exact pre-reserved reviewer slots and never resolve or charge the same edge twice.

When a frozen decision selects a Claude Fable/Opus reviewer, override only the normal optional cross-model review executor through CE's existing attested read-only Claude adapter; Railyard never starts a parallel raw claude -p runner. Pass the frozen binding to that CE-owned seam, then feed its receipt-bound findings into the ordinary Thermos disposition. Until that seam attests the binding, the route is transport_unsupported. Preserve the reviewer concern, persona, input digest, output schema, and Thermos synthesis authority. Never modify Compound Engineering, pass Claude through a Codex selector, or claim a review without the fixed adapter receipt.

Workflow

1. Determine the review scope from the user request, PR, current branch, or relevant changed files.
2. Freeze one deterministic review packet: objective and stop condition, exact diff/file digests, relevant source excerpts, requirement map, changed runtime-artifact chain, simplification receipt when applicable, and reusable hash-bound validation receipts. Reviewers do not rerun unchanged broad suites.
3. Map concern coverage before launching. Require one correctness/security disposition and one code-quality disposition with distinct scopes. A matching independent CE or Sol review may satisfy a disposition only when independence, frozen input digest, concern scope, disposition schema, and authority all match. Launch only uncovered concerns or unresolved disagreements, normally both passes in parallel:
   • Cursor: launch both subagents in the same message with run_in_background: true:
     • subagent_type: "thermo-nuclear-review-subagent" for bugs, breakages, security, devex regressions, feature-flag leaks, and other branch-audit risks.
     • subagent_type: "thermo-nuclear-code-quality-review-subagent" for maintainability, structure, file-size growth, spaghetti, abstractions, and codebase-health risks.
   • Claude Code: launch both reviewers as two Agent tool calls in one message (fresh-context subagents). Give each the frozen packet plus the full instructions of its sibling skill — read ../thermo-nuclear-review/SKILL.md for the correctness/security agent and ../thermo-nuclear-code-quality-review/SKILL.md for the maintainability agent — since no predefined subagent types exist here.
   • Codex: spawn two explorer subagents in parallel.
     • For the correctness/security agent, attach or pass the thermo-nuclear-review skill. If structured skill attachments are unavailable, read ../thermo-nuclear-review/SKILL.md relative to this skill and include its instructions in the subagent prompt.
     • For the maintainability agent, attach or pass the thermo-nuclear-code-quality-review skill. If structured skill attachments are unavailable, read ../thermo-nuclear-code-quality-review/SKILL.md relative to this skill and include its instructions in the subagent prompt.
     • Codex spawn calls are background work; wait for both with the available wait tool, then synthesize.
4. Pass each reviewer the same frozen packet and ask for prioritized findings plus an explicit disposition for its assigned concern. A runtime artifact lacking producer/package/install/consumer proof and a material complexity increase lacking a simplification receipt are mandatory findings, not optional observations.
5. After the required dispositions finish or are validly reused, synthesize findings, deduplicate overlaps, and resolve disagreements. Add another model only for a unique unresolved question after routing preflight; review is a concern-coverage portfolio, not an additive swarm.

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