agentic-loop

Use AFTER multiple workers have returned envelopes, to normalize, deduplicate and merge their results into one distilled artifact before the orchestrator consumes it or escalates to review. Trigger with a list of envelope/artifact paths — NOT for single-worker results (the orchestrator reads those directly).

Use for frontier-quality drafting, hard reasoning, or a strong second opinion WHILE Fable is included in the subscription plan (check the window note in CLAUDE.md) — the subscription-covered alternative to scripts/call_fable.sh. NOT for mechanical work (cheap tiers) and NOT a replacement for Sol's cross-family review.

Use at the START of any multi-step task in the agentic loop to decompose it into delegable subtasks with tier assignments. Trigger when the orchestrator needs a work breakdown before fanning out — NOT for single-step tasks the orchestrator can do directly.

Use as the PRE-SOL escalation review WHILE Fable is included in the subscription plan (check the window note in CLAUDE.md): a fresh-context, frontier-capability blind reviewer, invoked after the routine loop-reviewer pass and before any metered Sol call. Trigger with a task spec + candidate artifact path ONLY — deliberately withhold the reasoning that produced the candidate. NOT a fixer: it reports findings, it never edits.

Use as the ROUTINE review tier before considering Sol: a fresh-context, subscription-covered blind reviewer for candidate artifacts. Trigger with a task spec + candidate artifact path ONLY — deliberately withhold the reasoning that produced the candidate. NOT a fixer: it reports findings, it never edits.

Factory build stage: claim the oldest reviewed spec from the queue, build it on an isolated branch with tier-routed workers under Red Gate discipline (tests must fail before implementation), and advance it for review. Designed to run unattended — one spec per invocation, loopable via /loop or the factory workflow.

Toggle the agentic-loop feature flags for the current project: observability (event log + tree reports), minimize (code-minimization ladder in build briefs), grill (pre-planning interview), guards (reviewer quality gates) and summarize (Ollama fallback summaries in reports). Use when the user wants to turn a loop feature on or off, check what is enabled, or render an observability report.

Scaffold the agentic-loop template into the current project directory: copies the worker shim scripts, routing-policy CLAUDE.md, .env.example, run-state directory and per-project README. Use when the user wants to set up the orchestrated multi-model agentic loop in a new or existing project.

Factory review stage: claim the oldest built spec, run a blind fresh-context review (security, optimization, test quality) with findings typed by layer, apply a bounded revision, verify in a real browser when the change has a UI, open a PR with an executive summary and test steps, and log the evening digest. Designed to run unattended — one item per invocation, loopable.

Turn an idea into a factory spec: triage its size, grill the user with one-question-at-a-time interrogation (adaptive depth), emit a single spec file with a machine-checkable done-condition, and gate it through a fresh-context spec review before it enters the build queue. Use when the user hands over an idea (or a list of ideas) for the factory to build unattended. Interactive — run it while the user is present.

4 hooks across 4 events