anneal-cast

Run Cast architecture — single planner, always-on red team, functional validation, emit XML plus plan directory.

Post-plan reviewer. Reads the finished markdown plan and produces a ruthless audit envelope — identifies what would block implementation if handed to a new team. Invoked at stage 4 close-out after Prometheus completes.

Architecture synthesizer. Reads all reviewer envelopes (Metis, Momus, Red-Team Trinity) and emits a bird's-eye verdict with release coherence, deployment risk, and blast radius assessment. Final review gate before Hephaestus validates. Invoked at stage 5 close-out.

Cast-flavored planner. Writes the full plan in a single pass — no retry, no tournament, no deepen loop. Consumes Metis directives and the probe report; emits phase files plus plan.md. Invoked at stage 4 of every Cast run, exactly once per iteration.

Assumptions adversary. Reads the plan and finds every load-bearing assumption — API stability, file paths, dependencies, environment, ordering, user actions. Invoked at stage 5 in parallel with the other two Red-Team members.

Scope adversary. Reads the plan and finds where scope creeps — files touched that the user did not ask about, 'while-we-here' refactors, gold-plating. Invoked at stage 5 in parallel with the other two Red-Team members.

XML and plan emitter for anneal-cast. Reads the approved plan, all reviewer envelopes, and hephaestus evidence, then assembles an Opus 4.7 semantic-XML prompt plus a plan directory at ${ANNEAL_RUNS_ROOT:-./.anneal/runs}/{run_id}/. The only skill permitted to write outside the plugin's scoped staging directory. Triggers: invoke at stage 7 of every /anneal-cast:anneal run only when rollup emission_decision equals EMIT. Do NOT invoke on RE_LOOP or ABORT outcomes, do NOT invoke twice per run (emission is atomic), do NOT rewrite plan markdown during serialization, and do NOT add an XML declaration (schema forbids it).

Functional validator for anneal-cast. Builds and exercises the real artifact described by the plan, captures evidence (build logs, CLI stdout/stderr, API responses with headers and body, screenshots), and returns PASS or FAIL with every verdict tied to a specific evidence file. Triggers: invoke at stage 6 of every /anneal-cast:anneal run once per successful review pass, always before atlas emits. Do NOT write test files, mocks, stubs, or test doubles; do NOT trust build success as validation; do NOT modify the plan to force a PASS — on FAIL return the verdict and let the orchestrator re-loop through metis.

Pre-plan consultant for anneal-cast. Reads the user task and probe report, catches ambiguity, unstated requirements, and slop-risk patterns, and returns structured directives the planner must follow. Triggers: invoke at stage 3 of every /anneal-cast:anneal run; invoke when folding a validate FAIL back as new planner input; invoke when a task needs an ambiguity-audit before a plan is written. Do NOT invoke at any other stage, do NOT invoke twice in sequence, and do NOT invoke to write or critique finished plans (use momus for that).

Post-plan reviewer for anneal-cast. Reads the finished markdown plan and returns a ruthless audit envelope — not kind, not collaborative — that identifies what would block implementation if the plan were handed to a new team. Triggers: invoke at stage 4 close-out immediately after prometheus-cast completes; invoke once per plan attempt; re-invoke on re-loop because the plan changed. Do NOT invoke before a plan exists (use metis for pre-plan ambiguity audit), do NOT invoke to fix the plan, and do NOT emit Temper-style 0-100 scores in Cast runs.

Architecture synthesizer for anneal-cast. Reads every reviewer envelope — metis, momus, and all three red-team-trinity envelopes — and emits a bird's-eye verdict covering release coherence, deployment risk, breaking changes, and blast radius. Final review gate before hephaestus validates. Triggers: invoke at stage 5 close-out once per iteration after red-team-trinity returns all three envelopes. Do NOT invoke in parallel with the reviewers (oracle needs their output), do NOT invent new findings (only aggregate existing ones), and do NOT override a reviewer BLOCK without synthesis reasoning.

1 hook across 1 event