agent-canon

Use when agent-side working philosophy, interaction lessons, task retrospectives, repeated routing misses, missed skill invocation, or recurrence-prevention feedback should be logged without mixing them into user preferences.

Use when exploring, refactoring, or choosing an algorithm under proof obligations; builds JIT-canonical IR, lemma dependency graphs, algorithmic blocker frontiers, and algorithm-change guidance before handing terminal proof work to formal-proof-workflow.

Use when designing, implementing, reviewing, or diagnosing numerical optimization, solvers, preconditioners, convergence, gradients, Jacobians, Hessians, KKT conditions, tolerances, or optimization benchmarks; fixes the mathematical and validation contract before code or experiment changes.

Use when a change needs oracle/spec-risk classification or resilient, adversarial static test design, including behavior contracts, oracle choice, property/metamorphic candidates, mutation adequacy, or brittle-test diagnosis.

Use when updating AgentCanon itself, refreshing a vendored vendor/agent-canon submodule pin, repairing AgentCanon root runtime views, applying AgentCanon update TODOs, or routing local AgentCanon source commits through a proper AgentCanon branch and PR before parent pin updates.

Use when accumulated AgentCanon eval evidence is missing, stale, or failing; runs registered eval producers, validates eval family accumulation, and stores evidence through the log archive instead of hand-writing reports.

Use for code review, doc review, or AI-generated diff review when you need findings-first output focused on bugs, regressions, missing tests, and broken assumptions.

Use when Codex needs a context-independent execution path for a repository task, from intake and workflow selection through artifact placement, implementation, validation, and closeout.

Use when touching Docker, CI, dependencies, runtime compatibility, or repository-level development environment instructions.

Use when Markdown files changed, docs formatter/fixer output must be checked, or `agent-canon docs` formatting, heading, math, Mermaid, and link checks are in scope.

Use for owner-bounded repository edits after routing evidence shows a bounded owner, replaceable unit, targeted validation route, and no public behavior/schema expansion; also use for typo/link/format-only edits and Owner-Bounded Change work where Codex should run existing tools directly, record owner/tool/validation evidence, keep validation targeted, and avoid escalating to broad workflow prose.

Use when processing GitHub pull requests or issue queues: inventory open PRs, preserve PR Essence in bodies and run bundles, resolve conflicts, order merges, update branch protection evidence, merge only with authority, triage stale issues, and sync AgentCanon source PRs with parent pin PRs.

Use when a large refactor should run as a behavior-preserving refactor loop with explicit path mapping, semantic-delta controls, repair slices, and strong review gates.

Use when AgentCanon runtime dashboard evidence should be turned into owner-routed repair work, including dashboard next actions, repair failing hook evidence, hook entries status=fail, missing actual wave rows, workflow attribution gaps, consulted source URLs, reference missing URLs, AGENT_RUNTIME_DASHBOARD_WAVE_MISSING_ACTUAL, AGENT_RUNTIME_DASHBOARD_HOOK_WORKFLOW_MISSING, or AGENT_RUNTIME_DASHBOARD_REFERENCE_MISSING_URLS.

Use when a task needs specialist delegation, run-bundle bootstrap, explicit stage subagents, or Codex implementation routing.

Use when running tools, checkers, hooks, static analysis, or structural analyzers to find problems, preserve raw and structured full finding artifacts, mechanically rank every finding, and produce a complete finding report for implementation or refactor planning; before/after impact is optional when explicitly requested.

Use when the user explicitly asks to debug, repair, or refactor one issue at a time with visible problem statements before each edit and a next-issue prompt after each scoped fix.

Use this skill when preparing, running, or validating experiments.

Use when writing, exporting, saving, accumulating, or reporting tool/checker/hook/skill/eval/experiment results; creates durable raw and summary artifacts with unique IDs and no accidental overwrite.

Save and publish experiment run results with branch-safe retention. Use when Codex needs to preserve experiments/<topic>/result/<run_name>, create or verify experiment result manifests, write experiment reader reports, publish to experiment-results/<topic>, prevent overwrites, or keep failed/partial experiment runs as durable evidence.

Mandatory routing skill for repository tasks. Use before selecting workflow family, skills, review roles, subagents, model/team policy, runtime entrypoints, or run bundles for Codex routing.

Use as the general explanatory-doc DSL-to-prose adapter for README, workflow, guide, migration, or specification documents whose file responsibility is reader-facing explanation; do not select this skill by text length alone.

Use when a report, experiment plan, Eval output, presentation storyboard, PPT/deck plan, document, paper, HTML view, or refactor needs a structure contract before prose, rendering, interpretation, follow-up runs, or edits.

Use when natural-language mathematical claims, JIT-canonical implementation claims, proof sketches, or theory assumptions should be converted into formal-proof obligations, generated Lean evidence, theorem-graph targets, and checker-gated evidence.

Use when an algorithm should be designed and checked in Lean before production implementation; models candidate algorithms independently of existing code paths, proves or refutes convergence, stopping, certificate, filter/restoration, and inner-solver contracts, then hands a checked design contract to implementation or implementation-derived proof workflows.

Use when reviewing experiment topics, run.py files, experiment registries, GPU/JAX environment ownership, notebook artifacts, or experiment README/report readiness.

Use when planning, running, validating, or diagnosing GPU/CUDA/JAX/XLA/IREE backend execution, GPU validation blockers, nvidia-smi evidence, CUDA_VISIBLE_DEVICES handling, ExperimentRunner-based Python runs, or JAX/XLA preallocation-disabled execution.

Use when analyzing accumulated AgentCanon skill/tool/workflow/hook/eval logs, missed or late skill invocation, routing misses, weak skills, over-constrained related-skill coverage, or selection gaps; first convert raw logs into a structured dashboard summary with AgentCanon source generate_agent_runtime_dashboard.py before reading or interpreting evidence.

Use when checking, validating, or diagnosing repository dependency manifests, expanding code/header/search dependencies into a change-impact packet, or preparing repair-planning and subagent handoff context before editing, review, or closeout.

Use when producing a browser-readable HTML experiment or Eval report; first decide the primary figure, then plan and run an evidence-backed report renderer while keeping domain authority in the original tool.

Use when converting accumulated prompt history, run bundles, hook logs, skill/tool/workflow routing evidence, eval summaries, or agent reports into durable AgentCanon skill issues; groups repeated evidence by abstract cause, shards multi-agent review by evidence partition, and writes issue candidates from structured dashboard artifacts.

Use when a task needs paper search, prior-art mapping, contradictory-source hunting, or a reusable bibliography.

Use when creating, scaffolding, planning, or implementing an MVP, prototype, runnable vertical slice, product skeleton, v0, or thin vertical slice and the agent must prevent overbuilding. Trigger for MVP作成, プロトタイプ, 骨格だけ, core runnable path, thin vertical slice, scope creep, over-polish, and cases where early implementation is getting unnecessary UI, architecture, features, or tests.

Use when choosing short AgentCanon tool, skill, profile, check, runtime, closeout, or evidence routes from long candidate names, broad workflow text, routing misses, over-constrained related-skill candidates, public/system skill delegation, skill splitting, or skill/tool routing refactors.

Use when a request asks to visualize code, repository structure, runtime behavior, state, data movement, dependencies, types, proof state, an interactive graph, or a diagram embedded in a document; infer the user's context question, embedding context, and precision need, select the diagram family and evidence source, then delegate rendering to the owning skill or tool.

Use when the user asks to run the OOP readability checker, SOLID check, OOP check, readability check, produce a mechanical OOP report table, or interpret/prioritize OOP readability results; keep mechanical tool output separate from agent analysis.

Python 差分を pyright、pytest、ruff、型境界、API 挙動、OOP 可読性根拠で厳密に確認する。

Use this skill to review current checkout authority, run-bundle drift, legacy worktree cleanup evidence, and cleanup readiness.

Use when organizing repository documents, finding non-canonical docs, separating source canon from generated reports, eval results, closed issues, duplicate headings, or stale document paths.

Use when repository structure review, repo-refactor requests, expected AgentCanon layout, directory responsibilities, canonical README ownership, path layout, root views, project .codex/.agents views, personal ~/.codex runtime boundaries, or responsibility-scope maps must be reviewed, repaired, or refactored using structure contracts, recursive directory README analysis, source/view ownership checks, stale-surface sweeps, dependency manifests, and behavior-preserving move/rename gates.