pi-fairy-tales-chains

Six durable skill chains for the pi coding agent — multi-phase workflows where each phase is its own skill and state flows between them through a file on disk, so a long job survives across sessions and context windows and resumes exactly where it stopped.

Each chain is built to leverage pi-fairy-tales capabilities — subagents, /ultraplan, the artifact tool, guard-rail hooks, memory, and the status line. Authored with pi-skill-system-creator and validated with its chain validator.

Why chains? pi-fairy-tales is powerful but mostly in-session — a /ultraplan run or a subagent fan-out is gone when the session ends. A chain writes state between phases, turning that horsepower into a multi-day, resumable pipeline with human review between steps.

How the two packages fit together — repos, runtime footprint, config precedence: Ecosystem Map.

Install

Requires pi-fairy-tales — the chains call its tools (agent, artifact, /ultraplan, fetch, hooks). Install both as top-level packages:

npm i -g @earendil-works/pi-coding-agent    # pi itself
pi install pi-fairy-tales                     # the capabilities the chains call (declared as a peer dependency)
pi install pi-fairy-tales-chains              # the chains

Both load independently in the same pi and compose at runtime. They are deliberately not bundled into one install: pi loads each package's tools once, and bundling a second copy of pi-fairy-tales would conflict with a standalone install of it. Two installs, zero conflicts.

Or from git:

pi install git:github.com/KcAnom/pi-fairy-tales
pi install git:github.com/KcAnom/pi-fairy-tales-chains

The chains

Each is a pipeline (human-in-the-loop): every phase reads the shared chain state, does its work, updates state, and hands off with Run /<chain>-<next>. You review between phases; the durable state means you can stop and resume days later.

Since 0.2.0, all six chains track state through the bundled chain tool: authoritative JSON at .pi/fairy-tales/chains/<chain>/state.json with a human-readable state.md projection, TTL'd cross-session locking, and automatic import of legacy .<chain>-state.md files (see docs/STATE-CONTRACT.md). Heavy phase work runs as quests (pi-fairy-tales ≥ 0.15) with idempotent dedupe keys, so a crashed phase resumes the same work instead of duplicating it.

Chain	Phases	What it does	Fairy-tales muscle
release	changelog → bump → tag → publish → verify	Cut an npm/git release end to end	/commit conventions, ship skill, status line
feature-ship	spec → build → review → ship	Take a feature request from spec to merged PR	plan agent + artifact, /ultraplan worktree, deep-review
onboard	map → deepen → digest	Understand an unfamiliar repo over multiple sessions	parallel explore subagents, artifact HTML, memory
bughunt	repro → diagnose → fix → verify	Reproduce, root-cause, fix, and confirm a bug	explore/review subagents, post-edit self-fix hook, checkpoints
migrate	inventory → transform → verify	Large codebase migration that spans sessions	explore subagents, /ultraplan, per-site durable checkoff
research	gather → synthesize → artifact	Research a topic and produce a polished report	SSRF-safe fetch, artifact HTML

Using a chain

Start the first phase, then follow each handoff:

/release-changelog            # phase 1 — creates .release-state.md
# … review, then it tells you:
/release-bump                 # phase 2
/release-tag                  # phase 3
/release-publish              # phase 4
/release-verify               # phase 5 — "Chain complete!"

Interrupted? Just run the next phase whenever you come back — the durable state remembers where you were (/chains shows every run in the project). Each phase refuses to run if the previous one didn't complete, so you can't skip a step or act on a half-written state.

How a chain is built

Every phase skill follows the pi-skill-system-creator contract:

• Clean YAML front matter so pi can discover the skill
• State validation on entry — the chain tool verifies the run is active at exactly this phase and refuses out-of-order or missing-state runs (legacy chains parse frontmatter + chain_version instead)
• Atomic state writes — the tool writes JSON + projection via temp file + rename, so an interrupted write never corrupts the state
• Cross-session locking — a TTL'd lock (renewed on every write, auto-broken when stale, force-clearable with the tool's unlock action) catches two sessions running the same chain
• An explicit handoff to the next phase (the last phase terminates with "Chain complete")

Validate any chain yourself:

python3 ~/.pi/agent/skills/skill-system-creator/validate-chain.py <chain-name>

License

MIT © KcAnom. Built for the pi coding agent; pairs with pi-fairy-tales and pi-skill-system-creator.