fable-agent-orchestration

Apache-2.0 skill database for Fable-style agent orchestration.

The repo provides a public-clean set of reusable AI-agent workflow skills: coordinate many coding agents, keep work isolated in git worktrees, validate agent pull requests against real diffs and CI, and prevent fake-green tests.

This is intentionally generic. It does not include private project formulas, identity material, private continuity-system details, third-party source, keys, or closed operational notes.

Contents

• SKILL.md - single-file Fable orchestration skill.
• PUBLIC_GOAL_PLAYBOOK.md - standalone public goal prompt/playbook for applying the orchestration pattern in other agent workspaces.
• catalog.json - machine-readable index of the clean skill database.
• schemas/skill-record.schema.json - schema for catalog entries.
• skills/*/SKILL.md - individual public-clean skills.

Current skill records:

Skill	When to use	Why it matters
fable-orchestrator	Running many PRs with several agents	Keeps parallel work coordinated and merge-safe
autonomous-finish-loop	When reversible work remains	Prevents stopping on plans, promises, or tool noise
think-work-try	One risky implementation slice	Forces investigate -> build -> prove
one-slice-worker-cycle	Giving one agent a narrow task	Prevents vague broad PRs
two-critic-review-loop	Reviewing non-trivial PRs	Splits test review from code review
agent-pr-validator	Checking an agent-made PR	Compares claims to real diff and CI
adversarial-reviewer	Before trusting a change	Finds the strongest real objection
task-relative-test-gate	Verifying tests themselves	Stops fake-green tests
review-verifier	After a reviewer gives a verdict	Catches stale or wrong review feedback
orphaned-wip-adopter	Salvaging abandoned agent work	Reuses good WIP instead of rebuilding
agent-dispatch-packet	Delegating work to an agent	Turns vague goals into scoped, testable packets
peer-review-packet	Asking another model or person	Sends only clean, relevant context
fable-session-skill-miner	Mining agent sessions for reusable skills	Extracts procedures without leaking raw logs
external-workflow-adapter	Importing outside workflows	Keeps useful ideas and rejects bad assumptions
contributor-evidence-gate	Reviewing public contributions	Requires narrow scope and current proof
instruction-drift-control	Keeping agent guides in sync	Prevents stale canonical instructions
behavior-contract-harness	Testing agent behavior changes	Turns style claims into measurable contracts
phase-aware-engineering-ladder	Matching rigor to project phase	Keeps the security floor while scaling gates
investigate-before-fix	Before fixing a delicate diagnosis	Avoids patching the wrong cause
long-run-continuity	Work crosses session boundaries	Persists queue, PRs, and residuals
easy-vs-right-check	Before choosing the convenient step	Keeps work aimed at the real task
periodic-retrospect	On stalls or periodic review	Recovers dropped threads and repeated patterns
seal-both-types	Designing typed acceptance contracts	Prevents unchecked values from bypassing validation

Install

For the compact version, copy the root SKILL.md into your local skills folder:

~/.claude/skills/fable/SKILL.md

For modular use, copy any folder from skills/ into your agent's skills directory.

Core idea

The user or lead agent acts as a conductor. Agents do implementation slices. The conductor owns scope, evidence, review, and merge decisions.

The reliable loop:

1. Split work into independent slices.
2. Launch one build agent per slice in an isolated git worktree.
3. Give every agent a packet with role, work scope, invariant, non-scope, proof gate, output contract, and rules.
4. Track every live worker in a parent registry so completion events are routed to the right task and verified against real artifacts.
5. Require each agent to open a pull request, not merge it.
6. Run two separate critics: one critic reviews the test gate, and one critic reviews the code change.
7. Verify the critics' claims against the current code and CI.
8. Merge one PR at a time on green evidence.
9. Relaunch the next slice while reviews and CI run.
10. Do not stop on a plan, promise, or background notification while a reversible next action remains.

The rule of thumb for delegation:

Spawn an agent only when you will immediately go do something else.

If you would sit and wait, the agent adds overhead instead of throughput.

Clean-public boundary

This repo includes only generic engineering workflow primitives:

• worktree isolation;
• fail-under-broken tests;
• sabotage or negative-control checks;
• two separate critic roles for tests and code review;
• PR claim-to-diff validation;
• CI and runtime evidence review;
• stale review detection;
• recovery of stalled work;
• autonomous finish-loop discipline for reversible in-scope work;
• label-stripping session mining for reusable skills;
• instruction-drift control for canonical agent guides, fix logs, and keep-in-sync contracts;
• behavior-contract harnesses for prompts, hooks, eval probes, telemetry, and real-work regression gates;
• phase-aware engineering ladders that scale verification depth without weakening the security and data-loss floor.

It excludes:

• private formulas or methodology names;
• private peer identities or memory systems;
• third-party source dumps or reverse-engineering artifacts;
• raw session transcripts or bulky logs;
• project secrets or credentials;
• private business, product, or architecture notes.

Credits

This repository contains public-clean workflow skills and may adapt general workflow ideas from permissively licensed projects. It does not vendor external source code unless explicitly stated here.

agent-standard-oss

• Source: https://github.com/anmoln7/agent-standard-oss
• License: MIT License
• Inspected commit: 263c7e8bbd2f337fcb1e0c352eaa2508192ce757
• Used for: inspiration for the instruction-drift-control skill's public treatment of canonical agent instruction files, fix logs, keep-in-sync contracts, verifier decay, and autonomous-loop contracts.
• Vendored code: none.

The resulting skill is a rewritten synthesis for this repository's public bounded-agent orchestration model, not a copy of the upstream scripts, installer, hooks, or project instructions.

opus-fable-playbook

• Source: https://github.com/rennf93/opus-fable-playbook
• License: MIT License
• Inspected commit: 160acb6a59f83d80e45ef8e1148bac28bd46c143
• Used for: inspiration for the behavior-contract-harness skill's public treatment of behavior doctrine, stop gates, prompt nudges, honesty hooks, telemetry, critic agents, golden transcript evals, and measured variance.
• Vendored code: none.

The resulting skill is a rewritten synthesis for generic agent behavior contracts. It does not copy upstream hooks, scripts, transcripts, prompts, or project instructions.

senior-engineering-partner

• Source: https://github.com/bjgreenberg/senior-engineering-partner
• License: Apache License 2.0
• Inspected commit: 8b4bb74600a3b49edcb9bb4a7f7522580a883040
• Used for: inspiration for the phase-aware-engineering-ladder skill's public treatment of prototype/MVP/production rigor, fixed security floor, promotion triggers, and deferred gates.
• Vendored code: none.

The resulting skill is a rewritten synthesis for this repository's bounded agent orchestration model. It does not copy upstream scripts, references, eval scenarios, or project instructions.

License

Apache License 2.0.