agents-brain

# Agents Brain If these instructions are already present in the conversation from a slash or dollar invocation, follow them directly; do not invoke this skill again through a skill tool. Create or polish repo-local context as one coherent system: human-facing README.md files, agent-facing AGENTS.md files with companion CLAUDE.md symlinks, existing project-installed skills under `.agents/skills`, eligible source-catalog skills under `skills/<name>/`, and context docs — any other Markdown files, under any name or directory, whose content is durable guidance for agents or humans, such as conventions, command catalogs, data-format rules, workflow runbooks, and reference material. Success means every selected target is grounded in repository evidence, respects its audience and scope, spends agent context only on guidance that changes behavior, and passes the narrowest repository-defined validation. Stop after reporting completed or planned changes, validation, and any blockers. ## Model and Context Optimization Optimize skills and other agent-facing context for GPT-5.6 and Claude Fable 5 while preserving README.md as clear human-facing documentation. Before complex, long-running, multi-tool, or orchestration-heavy context work, resolve `scripts/fetch-guidance.sh` relative to this skill directory, run it once for `gpt-5p6` and once for `claude-fable-5`, and read both returned files completely. The helper retrieves the official [GPT-5.6 prompting guidance](https://developers.openai.com/api/docs/guides/prompt-guidance-gpt-5p6) and [Claude Fable 5 prompting guidance](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5) because their recommendations may evolve. Simple context work does not require either guide. Accept an integrity-valid `cached` guide for 24 hours. Use `--refresh` for explicitly latest or change-sensitive work, materially disputed guidance, or conflicts with observed model behavior. Interpret helper diagnostics precisely: - `cached` reused a guide validated no more than 24 hours ago without network access. - `revalidated` refreshed validation metadata after a successful conditional `304` response. - `fetched` atomically replaced the cache with integrity-valid content from the pinned official URL. - `stale` reused a guide validated no more than seven days ago after live retrieval failed. Proceed only after reading it and disclose the validation timestamp and retrieval failure in the final risks. Forced refreshes, expired entries, integrity failures, and unexpected redirects fail closed. If either required guide cannot be returned, stop qualifying work before writing instead of substituting memory or another source. Never create the cache in a repository or skill installation. - Keep only agent-facing content that changes a decision, prevents an evidenced mistake, or supplies a non-discoverable constraint. Remove generic advice, tutorials, history, inventories, no-op prose, and mechanics already enforced by scripts, recipes, schemas, or configuration. - State each meaning once in an effective load chain. Put shared guidance in the parent and keep descendants to deltas or overrides; preserve repetition when artifacts load independently and need to remain self-contained. - Use the narrowest reliable load scope: universal guidance inline, path-specific guidance in nested context, and rare procedures in on-demand context docs or skills. Do not hide required guidance behind an unreliable pointer. - Prefer one positive decision rule to enumerated prohibitions. Keep one minimal example only when it encodes an exact requirement or corrects a measured failure; keep tool and command descriptions only when routing, inputs, side effects, outputs, or failure signals matter. - Preserve authority, safety, material exceptions, semantic success criteria, exact machine-consumed text, and readable prose. Do not shorten human-facing README.md content merely to reduce agent tokens unless that content also enters agent context. - Documentation-only authority does not permit creating or changing helpers or schemas. When none exists, retain the smallest accurate prose and report the extraction opportunity instead of expanding scope. ## Choose a Workflow Choose exactly one workflow and read only its reference. For skill creation, first inspect applicable repository instructions. When they define a source catalog and lifecycle, stop and follow that repository-owned workflow. Use `skill-writing` only when no catalog-specific workflow exists. | User intent | Workflow | Reference | | --------------------------------------------------------------- | ---------------------------- | --------------------------------------------------- | | Update, refresh, sync, prune, polish, repair, or fix context | `polish` | `references/polish.md` | | Create, initialize, generate, or regenerate context files | `create` | `references/create-docs.md` | | Audit, check, review, inspect, or suggest changes without edits | `polish` in `--dry-run` mode | `references/polish.md` | | Create or scaffold a skill | Stop | Use repository catalog lifecycle or `skill-writing` | | Install, discover, remove, or rename a skill | Stop | Use a dedicated skill-management workflow | If the intent is unclear, select `polish` in `--dry-run` mode and report the smallest useful planned change set. ## Authority - Explicit create, update, polish, repair, fix, or equivalent intent authorizes in-scope local writes. Inspection-only intent and `--dry-run` do not. - Require explicit confirmation before deleting README.md, AGENTS.md, CLAUDE.md, or context-doc targets. `--force` authorizes documented overwrites, not deletions. - Treat a broad write request as authorization for the requested scope. Otherwise, preview a change set larger than a handful of files and stop before writing. - Do not expand from documentation work into source changes, skill creation, or external writes. ## Arguments - `path`: Optional repo-relative subtree. Restrict documentation, package-root, project-skill, source-catalog skill, and context-doc discovery to that subtree. - `target ...`: Optional filters during `polish`: skill names from existing `.agents/skills/<name>/` or eligible `skills/<name>/` trees, or repo-relative Markdown paths selecting specific context docs. - `--root-only`: Select only root README.md, AGENTS.md, and CLAUDE.md targets. Exclude project-installed skills, source-catalog skills, and context docs unless explicitly selected by `target`. - `--dry-run`: Report planned writes and concise diffs without changing files. - `--preserve`: During `polish`, keep accurate user-authored prose and structure; fix only drift and obvious noise. - `--minimal`: Produce the smallest context that still meets the completion bar. - `--thorough` / `--full`: Perform deeper analysis only where it adds durable, repository-specific context. - `--force`: During `create`, regenerate existing README.md or AGENTS.md targets without prompting. Never applies to skills or deletions. If `--minimal` and `--thorough` / `--full` are both present, make no writes and ask the user to choose. Report unrecognized flags; continue only when they cannot change scope, safety, or write behavior. ## Repository Guard Rail Run before discovery or writes: ```sh cwd="$(pwd -P)" case "$cwd" in /) printf 'abort: refusing to run at the filesystem root\n' >&2; exit 1 ;; esac repo_root="$(git rev-parse --show-toplevel 2>/dev/null)" || { printf 'abort: not inside a git repository\n' >&2; exit 1; } managed_skill_root= case "$repo_root" in /|"$HOME") printf 'abort: unsupported repo root: %s\n' "$repo_root" >&2; exit 1 ;; "$HOME/.agents"|"$HOME/.codex"|"$HOME/.claude") managed_skill_root="$repo_root/skills" ;; "$HOME/.agents/"*|"$HOME/.codex/"*|"$HOME/.claude/"*) printf 'abort: repo root is nested under an agent configuration repository: %s\n' "$repo_root" >&2; exit 1 ;; esac if [ -n "$managed_skill_root" ]; then case "$cwd" in "$managed_skill_root"|"$managed_skill_root/"*) printf 'abort: installed skills must be edited in their source catalog: %s\n' "$cwd" >&2; exit 1 ;; esac fi ``` When `managed_skill_root` is set, allow README.md, AGENTS.md, and CLAUDE.md work elsewhere in that repository, but exclude the entire installed `skills/` tree from every workflow. Apply the exclusion before discovery, canonicalization, or symlink traversal. If `path`, a `target`, or an explicit request would enter that tree, make no writes there and report that the skill must be edited in its source catalog. `--force` does not override this boundary. Outside managed agent-config roots, eligible git-tracked `skills/<name>/` source catalogs are in scope for `polish` per `references/polish.md`. Snapshot `git status --short` before broad edits. Preserve unrelated pre-existing changes and re-check expected paths after generators or broad commands. ## Shared Constraints Stay inside the resolved repository and preserve unrelated changes. The selected workflow reference is authoritative for README.md, AGENTS.md, CLAUDE.md, and skill behavior; do not repeat or broaden its file-specific rules here. ## Discovery and Tool Routing Use git-aware discovery, canonicalize every candidate beneath `repo_root`, and exclude VCS, dependency, environment, and build outputs. Deliberately include ignored `.agents/skills/*/SKILL.md` only when project skills are selected. Discover git-tracked, non-ignored, non-symlinked `skills/*/SKILL.md` only outside managed agent-config roots and only when source-catalog skills are selected. Parse each selected skill's YAML frontmatter. Inspect only a project-installed skill's declared write boundary before deciding whether it qualifies for a coordination exemption. Prefer `fd`, fall back once on suspiciously narrow results, and synthesize independent repository evidence before writing. Discover context docs by following Markdown links from README.md, AGENTS.md, CLAUDE.md, and SKILL.md files, then by scanning remaining tracked Markdown whose content qualifies. Classify by content, never by file name or location. Exclude changelogs, licenses, legal and policy notices, generated or vendored documentation, and prose that is product content rather than guidance. When classification is uncertain, leave the file out of scope and report it as a candidate. ## Completion and Report After writes, run repository-defined Markdown formatting or checks when present. If skill frontmatter or `agents/openai.yaml` changed in a project-installed skill, run its invocation metadata check. Verify changed CLAUDE.md symlinks resolve to sibling AGENTS.md. In `--dry-run`, report commands that would depend on planned files instead of running them. Lead with `### ✅ Context updated` only after writes and required validation pass, `### ⚠️ Context updated — validation failed` when files were written but required checks fail, `### 🔎 Context preview — no files written` in dry-run mode, or `### ⛔ Context blocked — no files written` for a pre-write stop. Then report only: 1. `🧭 Mode and scope`: workflow, dry-run status, target counts, and relative paths in a compact table. 2. `📦 Changes`: completed or planned changes grouped by directory; use a tree when it makes path ownership clearer. 3. `🧪 Validation`: exact commands, result, and any justified skip in a table. 4. `⚠️ Blockers and risks`: conflicts, advisories, and unrecognized flags; omit when empty. Keep paths, commands, guard-rail errors, symlink targets, and user-authored content exact and undecorated. Omit empty detail and stop once the selected targets meet the completion bar. ## References - `polish`: read `references/polish.md`. - `create`: read `references/create-docs.md`.