agent-loop-ext
Crash-resilient external agent loop with state persistence and CI/CD integration
AUTO-INVOKE for ANY AIWG capability question, framework/addon/extension query, steward routing, MCP profile, or when unsure where to start. ALWAYS consult before filesystem search on .claude/.factory/.codex/.warp/etc. AIWG core utilities quick reference — always-on framing,
$ npx -y skills add jmagly/aiwg --skill aiwg-utils-quickref --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/aiwg-utils-quickrefContext preview
The summary Claude sees to decide when to auto-load this skill.
AUTO-INVOKE for ANY AIWG capability question, framework/addon/extension query, steward routing, MCP profile, or when unsure where to start. ALWAYS consult before filesystem search on .claude/.factory/.codex/.warp/etc. AIWG core utilities quick reference — always-on framing,
name: aiwg-utils-quickref namespace: aiwg platforms: [all] kernel: true description: AUTO-INVOKE for ANY AIWG capability question, framework/addon/extension query, steward routing, MCP profile, or when unsure where to start. ALWAYS consult before filesystem search on .claude/.factory/.codex/.warp/etc. AIWG core utilities quick reference — always-on framing, steward, index, meta operations. triggers: - "i do not know what AIWG has" - "translate my goal into AIWG search terms" - "run two or three aiwg discover searches" - "capability discovery"
This is your always-loaded core directory for AIWG. It does **not** list every utility skill. Instead, it teaches the cross-cutting domains and gives you **curated search phrases** that map to `aiwg discover` lookups, plus a list of which framework-specific quickrefs are loaded.
Walk the action hierarchy in order. The agent always prefers a **skill** over a raw CLI invocation:
1. **Priority 1 — Local skill or agent**: check if a skill already in your context handles the need (kernel skills listed below, framework quickrefs, deployed agents like `aiwg-steward` / `aiwg-finder`). These cost nothing to invoke. 2. **Priority 2 — Discovered skill**: if no local match, pick a **curated phrase** from a capability domain (or paraphrase the user's words), run `aiwg discover "<phrase>"`, surface the top match, and fetch its body with `aiwg show <type> <name>` — never `find` / `ls` / `Read`. 3. **Priority 3 — Raw CLI**: only when no skill exists at priorities 1 or 2, OR the command is on the discovery surface (`aiwg discover`, `aiwg show`, `aiwg list`, `aiwg version`, `aiwg runtime-info`), OR you are inside a skill that's calling the CLI as its step.
**Do not enumerate skills from memory.** AIWG ships hundreds of skills; you only see the kernel set in your context. See the [`cli-secondary` rule](../../rules/cli-secondary.md) for the full hierarchy and per-command pairings.
Most AIWG skills (~460 of 480) are **not in your context** and are not in `<provider>/skills/` either — they stay at `$AIWG_ROOT` (no per-project copy by default, #1217). Two commands close the loop:
# 1. Find — returns ranked candidates with absolute paths aiwg discover "<the user's need, paraphrased>" # 2. Fetch — streams the SKILL.md body for the chosen candidate aiwg show skill <name> # or: agent | command | rule
`discover` and `show` are designed to compose. **You should never need to navigate the filesystem to read AIWG content.** That's the whole point of the CLI.
Most AIWG skills aren't kernel-listed, so the platform's Skill tool will reject them. This is expected behavior, not a bug. The fallback hierarchy:
1. **`aiwg show <type> <name>`** — primary fetch. Always works regardless of kernel status. 2. **`aiwg show <type> <name> --json`** — same fetch with `{ path, content }` envelope, useful for forwarding to a sub-agent. 3. **Read directly from `$AIWG_ROOT/agentic/code/`** — last-resort fallback. Only if the `aiwg` CLI itself is broken. The corpus at `$AIWG_ROOT/agentic/code/frameworks/<name>/skills/<skill>/SKILL.md` and `$AIWG_ROOT/agentic/code/addons/<name>/skills/<skill>/SKILL.md` is the canonical source — it is always present at the install root and survives any deploy-state corruption.
**Forbidden after `discover` returns a path**: running `find`, `ls`, `Glob`, or `Read` on a `<provider>/skills/` directory. Those reflect the kernel-pivot deploy state, not the full surface. Use `aiwg show`.
This is mandated by the `skill-discovery` HIGH rule. Surface the top match (or top-3) to the user — the search is auditable.
If a user asks about a specific framework's surface, the corresponding quickref is your first stop. These are kernel-resident, so they're already in your context:
If a quickref isn't loaded, the framework isn't installed in this project. Use `aiwg list` to confirm.
For everything **outside the framework set** — the ~214 skills across 29 addons (loops, voice, memory, color, testing, NLP, etc.) and 7 ops extensions (sys/net/sec/dev/it/stream/api-adapter) — load `aiwg-language-map`. It's also kernel-resident and acts as the orientation layer for the addon/extension surface, with curated discover phrases per capability domain.
aiwg-language-map → addon capability domains + extension domains + cheat sheet
These ops **skills** are kernel-resident — already in your context regardless of `aiwg discover`. They are your primary surface for self-maintenance. Each skill carries the priming the bare CLI cannot — pre-flight checks, dry-run preview, preservation logic, recovery patterns.
> **Skills first, CLI second.** Each row below shows what the skill is for and the CLI command it calls under the hood. Invoke the skill — let it call the CLI. Direct CLI invocation skips the priming and is acceptable only on the discovery surface (Rule 2 of `cli-secondary`). See [`cli-secondary` rule](../../rules/cli-secondary.md) for the full policy.
| Skill (preferred) | Purpose | CLI it wraps | |---|---|---| | `steward` | Provider capability awareness + command routing | (no single CLI — orchestrates several) | | `aiwg-doctor` | Installati
Reusable project context and specialist workflows for the AI tools you already use. Plan software, coordinate specialist reviews, prepare campaigns, investigate incidents, organize research, curate media, and maintain operational knowledge.
Repo: jmagly/aiwg
Crash-resilient external agent loop with state persistence and CI/CD integration
Detect requests for iterative autonomous agent loops and route to the appropriate loop executor
Automatically execute tests when code-generating agents modify source files, enforcing the execute-before-return pattern
Enable agent loops to learn from similar past tasks and share patterns across loops
Query and manage the executable feedback debug memory
Execute tests on generated code and iterate until passing