seiya
@seiya
Public Skills
atmofab-review-loop
by seiya
Use when implementation in this repository reaches a pause and review begins, when running subagent or Codex review rounds, when fixing findings and moving to the next round, and when judging whether the loop has converged. Required reading for 「レビューして」「レビュー回して」「指摘を直して」「codex review」「この PR merge していい？」「まだ見るべきところある？」 and immediately after finishing the implementation of an audit finding or an issue. The subject is review of changes you made; do not use it for reading existing spec, docs, or implementation to judge whether they are sound (review without a change).
atmofab-enforcement-change
by seiya
Use when changing this repository's enforcement machinery — the deterministic gates, the validators, the build-runtime server's argument validation, the pure-leaf launch refusals and the read-only sandbox profile, the DEV hook layer — and whenever you are about to classify a review finding as 「残余」「到達不能」「対象外」 (residual / unreachable / out of scope). Required reading for fixing a fail-open, adding to an allowlist or denylist, touching the gate functions in mcp_servers/build_runtime_server.py / tools/orchestration_runtime.py, tools/hooks/ (the DEV layer; the leaf hook layer, the capability gate and write_root derivation went in Z4, issue #171, and the episodes about them stay in references/), or the gates in validate_pipeline_semantics.py, for fixing an audit finding, and for triaging a subagent's or Codex's review findings.
workflow-promote
by seiya
Use this when running the Promote stage and promoting an artifact whose verdict and aggregate_verdict passed to releases/ and updating the official_releases of spec_catalog.yaml. It applies to the work of satisfying the release_id invariant and the tracking-information registration.
spec-input-check
by seiya
Use this before starting the workflow to check whether a spec's input set (controlled_spec.md / deps.yaml / tests.md) is sufficient and self-consistent to run Spec -> Compile -> Generate -> Build -> Validate. It detects missing required items, cross-file contradictions, ambiguous statements, unresolvable dependencies, and declared tests that cannot reach their expected_outcome as written, and reports them as proposals. It never modifies the spec.
workflow-tune-generate
by seiya
Use this when running the generate of the Tune stage and generating lowering-plan (target_lowering_plan) override candidates while keeping the spec.ir.yaml finalized in the core workflow and the target profile invariant. As an optional flow, it applies to the work of expanding trial candidates from the performance-exploration tuning.spec.
workflow-tune-verify
by seiya
Use this when running the verify of the Tune stage and evaluating the physics passing, quality conditions, and performance objective function of a candidate lowering-plan (target_lowering_plan) variant to judge whether to adopt it. It applies to fixing the best impl and judging the move to regression.
workflow-timing-audit
by seiya
Use this when investigating where a workflow orchestration spent wall-clock time and output tokens — breaking a run down per leaf (step.substep), separating LLM leaves from the conductor's in-process deterministic steps, and attributing each LLM leaf's time to model generation vs. tool execution vs. dominant turns. Also flags the waste/failure signals a time-and-token table hides: turns cut off by max_tokens (thinking that emitted nothing) and leaves that died of an API/transport error. Handles the transcript multiple-counting traps. Also the rule and the cross-run report for comparing cost between two sets of runs (first launch vs retry, per node and step.substep). The target orchestration_id is auto-detected. Claude Code-only.
workflow-audit
by seiya
Use this when investigating why a workflow orchestration failed or fail-closed, which substeps it attempted more than once, whether sandbox enforcement recorded a violation, or what the run cost. Runs tools/audit_orchestration.py over the in-repo records of one orchestration; backend-neutral (claude_cli / codex_cli / HTTP providers).