🧚 Fable Token-Saving Skills Orchestrator

Fable is excellent at judgment: strategy, architecture, decomposition, review, and synthesis. It is also expensive enough that letting it type every routine line of code, reread giant logs, or wake from a cold prompt every few minutes is usually the wrong shape.

This repo packages the practices we have found useful so far into an additive kit for Claude Code style setups. It gives you a CLAUDE.md addendum, hooks, skills, and docs you can layer onto your existing configuration without replacing it. It is additive, not a replacement.

The point is not "never use Fable." The point is to use Fable where it is uniquely valuable, then route mechanical work, large-output digestion, parallel verification, and long-running side lanes elsewhere.

flowchart LR
    fable["Fable<br/>judge, spec, synthesize"]
    packet["Dispatch packet<br/>bounded work"]
    lanes["Cheaper lanes<br/>implement, test, digest"]
    proof["Bounded digest<br/>and proof"]
    review["Fable review<br/>accept or reroute"]
    wait{"Waiting?"}
    keepalive["Short wait<br/>self-extinguishing keepalive"]
    cold["Long idle<br/>let cache die once"]

    fable --> packet --> lanes --> proof --> review --> wait
    wait -->|under cache horizon| keepalive --> review
    wait -->|long idle window| cold

Quick Start

Dry-run the install first:

python3 scripts/install.py --dry-run --claude-home "$HOME/.claude"

Apply everything:

python3 scripts/install.py --apply --claude-home "$HOME/.claude"

Install only one surface:

python3 scripts/install.py --apply --install-hooks
python3 scripts/install.py --apply --install-skills
python3 scripts/install.py --apply --append-addendum

The installer is additive:

• it appends a marked block to CLAUDE.md
• it copies hooks into ~/.claude/hooks
• it copies skills into ~/.claude/skills
• it registers Stop hooks in settings.json
• it writes timestamped backups before changing existing files
• it does nothing unless --apply is present

Why This Works

Anthropic prompt caching is the economic backdrop. Official docs say the default prompt cache lifetime is 5 minutes and that cache use refreshes the cache. Official pricing says 5-minute cache writes cost 1.25x base input and cache reads cost 0.1x base input.

Sources:

• Anthropic prompt caching docs
• Anthropic pricing docs

That makes long orchestration loops weirdly sensitive to idle gaps. If Fable dispatches lanes, waits 6 to 8 minutes, processes one wake, and waits again, you can repay cold context repeatedly. A short keepalive read is much cheaper than a cold rewrite when a lane is expected back soon.

The practical rule:

• Do dense turns while lanes run.
• Keep Fable doing judgment and integration, not rote typing.
• Digest large outputs before Fable reads them.
• Use a self-extinguishing keepalive only for short waits.
• Let the cache die once for genuinely long idle windows.

What To Read

• Install guide
• Lane routing
• Dense turns
• Keepalive economics
• Run economics
• Adapting model names
• Related tools and inspirations
• Troubleshooting

Repo Layout

templates/       Copy-paste or installer-managed Claude config fragments
hooks/           Sanitized Stop hooks
scripts/         Installer, checks, and keepalive helper scripts
skills/          Claude skill templates
docs/            Public explanations and setup notes
tests/           Focused tests for installer, hooks, and docs safety

Safety Boundaries

This repo does not promise universal savings. It documents patterns that worked for us, ties the cache math to official pricing, and gives you tools to adapt the routing to your own models, prices, and agent harness.

Do not paste credentials into scripts. Use environment variables or your own secret manager. The example GLM/off-budget lane docs intentionally use placeholder env vars, not tokens.

Related projects such as openai/codex-plugin-cc and blader/arbitrage are useful references for adjacent Codex delegation workflows and token-arbitrage framing. See Related tools and inspirations for the boundary.

Contributing

Please read CONTRIBUTING.md and SECURITY.md. Public examples must not include raw transcripts, tokens, private paths, customer data, or machine-local credentials.