orchestrator

Fable Orchestrator

Model-aware multi-agent orchestration for Claude Code. It detects the model in the session chair and injects the orchestration profile that fits it: token-frugal when you're on Fable, latency-lean when you're on Opus 4.8 (or any other model). Both profiles keep the same quality guarantees — a Requirements Ledger against silent detail loss and a verification pass — and both are backed by guard hooks that enforce the discipline mechanically.

The problem

Delegating bulk work to subagents is the right move when your session model is an ultra-scarce top tier (Fable): it preserves the usage limit and keeps the session context small. But that same heavy discipline — mandatory ledger before any delegation, sequential research relays, a fresh re-read verification pass on every close, disk hand-offs between every stage — is tuned to trade wall-clock latency for top-tier token frugality.

Run it under Opus 4.8, with its 1M-token context window and top-tier judgment+implementation in the chair, and the trade inverts. Opus tokens aren't the scarce resource you're protecting; the clock is. The ceremony that saves Fable's limit just makes Opus slow — for a benefit it no longer needs. The fix isn't "less discipline," it's "the right discipline for who's in the chair."

And the deeper issue stays the same across both: workflow instructions in CLAUDE.md are advice, not mechanism. The two rules that matter most — write the ledger, finish it before closing — are exactly the ones that get skipped under pressure. Hooks make them mechanical.

What this plugin does

Three layers:

1. Model-aware orchestration instructions — injected at every session start (no CLAUDE.md editing needed). The SessionStart hook reads the session model and injects one of two profiles:

   Profile	Injected when	Optimizes for	Behavior
   dynamic-workflow-fable.md	session model is Fable	usage limit	Delegate aggressively; bulk work to cheaper tiers; mandatory ledger + fresh-eyes verification. Accepts latency to keep the top-tier token share low.
   dynamic-workflow-opus.md	any other model (Opus 4.8, Sonnet…)	wall-clock latency	Do bounded/medium work inline; delegate only to buy parallelism or to keep bulk out of context; ledger and verification proportional to size and risk; parallel fan-out (Workflow pipeline()/parallel()) over sequential relays.

   Both route subagents by tier (haiku → mechanical, sonnet → implementation, opus → judgment) and both treat the filesystem as shared memory. The Opus profile just stops spending latency where it no longer buys quality.

2. A Requirements Ledger — every explicit requirement, implicit expectation, constraint, and edge case as a checkbox line. Files survive context compaction; conversation context does not. On Fable the ledger is written before any delegation. On Opus it's proportional: an inline checklist for small/medium work, a ./.workflow/LEDGER.md file for large/multi-session/high-risk work. Either way, detail loss becomes visible instead of silent.

3. Two guard hooks — they convert the two most-skipped rules from "should" into "must":

Guard	Trigger	Effect
Spawn guard (PreToolUse on Agent|Task)	A detailed delegation (spawn prompt over the threshold) while ./.workflow/LEDGER.md doesn't exist	The spawn is denied; the model is told to write the ledger first, or do small tasks directly. The threshold is model-aware: strict on Fable (1500), relaxed on Opus (4000) so ordinary fan-out isn't taxed. Short prompts always pass.
Close guard (Stop)	The turn ends while the ledger has open - [ ] items	The stop is blocked once and the open items are listed. - [x] (done + verified) and - [~] deferred: <reason> count as closed. A stop_hook_active check prevents infinite loops. On Opus, small tasks have no file ledger, so this simply never fires for them.

What it looks like in practice

A detailed delegation without a ledger gets bounced back to the model (threshold shown reflects the active profile):

LEDGER GUARD: this looks like a detailed delegation (spawn prompt > 4000 chars)
but ./.workflow/LEDGER.md does not exist. Per Dynamic Workflow Rule 1, first
write the numbered Requirements Ledger ...

A turn ending with unfinished ledger items gets held once:

LEDGER GUARD: ./.workflow/LEDGER.md still has 8 open item(s):
- [ ] A1. ...
- [ ] A2. ...
If you are CLOSING a workflow: address each item and mark it '- [x]' ...

Why the Opus profile is faster (at the same quality)

The Fable profile's latency sinks, and what the Opus profile does about each — without dropping a quality guarantee: