Fable as a skill thread - lets gather our knowledge together and refine

I published a small open-source repo for a workflow I’ve been using to coordinate coding agents on larger codebases:

https://github.com/sherlockholmesyes/fable-agent-orchestration

The basic idea is simple:

Don’t hand-code every change yourself, but also don’t let agents free-run and trust their summaries.

Instead, act as the conductor:

- Split the work into narrow slices.

- Launch build agents in isolated git worktrees.

- Require each agent to open a PR, not merge it.

- Validate each PR with two separate critics:

- - one checks whether the test/gate actually proves the task;

- - one reviews the code/change itself adversarially.

- Verify reviewer claims against the real diff, current code, and CI.

- Merge one PR at a time.

- Relaunch the next slice while other work is still running.

The repo includes a clean skill database under Apache-2.0:

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/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, rejects bad assumptions instruction-drift-control Keeping agent instructions and fix logs in sync Prevents stale duplicated guidance investigate-before-fix Before fixing a suspected root cause Prevents patches for unproven diagnoses long-run-continuity Long multi-PR runs or context resets Preserves queue, PRs, and residuals across breaks easy-vs-right-check When a step feels like progress Catches convenient work that dodges the real task periodic-retrospect When stalled or after repeated cycles Finds dropped threads and recurring failure patterns seal-both-types Designing typed contracts Prevents forged valid-by-construction states The main lesson:

The bottleneck is not only “make the generator smarter.” For large agent-driven work, the bigger win is often to strengthen the verifier: claim-to-diff validation, fail-under-broken tests, independent review, and serialized merge discipline.

I also included a machine-readable `catalog.json` and schema so the skill set can grow into a more organized agent-orchestration library.

I also try to make a community around open source AI where I'd like to share and discuss more , big ambitious projects and PoC feel free to join.

https://element.wearein.space/

think-work-try

credits : https://github.com/anmoln7/agent-standard-oss/ skill: instruction-drift-control

сredits : https://github.com/rennf93/opus-fable-playbook skill: behavior-contract-harness

credits: https://github.com/bjgreenberg/senior-engineering-partner phase-aware-engineering-ladder