| name | model-orchestration |
| description | Session-model-led multi-model tiering. Use in ANY conversation (ambient) or in the pipeline to decide which model tier does which slice of work, and to delegate to the right model-pinned agent (strategist/deep-reasoner/fast-worker). The session model you selected IS the orchestrator; it self-selects its topology from its own model identity. |
Model Orchestration
The session model the user picked is the orchestrator. You do not detect or switch the session model — you are it. Subagents are pinned to other tiers via model: frontmatter using latest FULL model ids (claude-fable-5 / claude-opus-4-8 / claude-sonnet-5). Full ids, not aliases: the bare sonnet alias still resolves to claude-sonnet-4-6, so aliases would silently pin an older model. On Claude Code ≥ 2.1.172 subagents nest and honor their own model: frontmatter, so normal tiering needs no code — the pins carry it. Match your own identity to a profile below and delegate accordingly.
Tiers (latest full ids)
| Agent | Model id | For |
|---|
strategist | claude-fable-5 | Hardest, longest-horizon, highest-stakes / irreversible judgment. |
deep-reasoner | claude-opus-4-8 | Reasoning-heavy: complex debugging, architecture, algorithm design (fresh context). |
fast-worker | claude-sonnet-5 | Mechanical: boilerplate, tests-to-spec, formatting, simple edits. |
explorer | claude-haiku-4-5 | Cheapest, read-only: codebase sweeps, capability discovery, "find where X lives", broad file/grep reconnaissance. Never edits. |
Read-only research, codebase sweeps, and capability discovery go to soe:explorer (Haiku) — NEVER to Opus (deep-reasoner) or an unpinned general-purpose agent, both of which default to the expensive session model. Reserve deep-reasoner/strategist for slices that need actual reasoning or judgment, not information-gathering.
Per-slice routing — score the work, not the vibe
Rate each unit of work on three axes, then pick the tier:
- Stakes — how expensive is a wrong answer?
- Reversibility — how hard to undo? (low reversibility ⇒ raise the tier)
- Ambiguity — how much genuine judgment vs. mechanical follow-through?
High on all three → strategist (Fable). High reasoning/ambiguity, recoverable → deep-reasoner (Opus). Low/low/well-specified → fast-worker (Sonnet). When unsure, prefer the cheaper tier and escalate on signal — but never downscope work touching auth, secrets, payments, migrations/deletions, or prod: those always get full scrutiny regardless of size.
Topology profiles — self-select from your own model
You know which model you are running as. Apply the matching profile:
- On Fable → you orchestrate. Delegate reasoning to
deep-reasoner (opus), mechanical to fast-worker (sonnet). No strategist — you already are the top tier.
- On Opus → you orchestrate. Mechanical →
fast-worker (sonnet); reasoning stays with you or a fresh-context deep-reasoner; call strategist (fable) only for irreversible high-stakes calls, and only if the user has Fable.
- On Sonnet → you orchestrate. Reasoning →
deep-reasoner (opus); mechanical stays local with you.
Absent tiers (Fable, or Codex) are simply not used — soe-core never hard-depends on them. Pick a model, get the matching topology.
Context firewall (token discipline)
Delegated agents write full output to a shared scratch path (an absolute path outside any worktree) and return only path + 3-line summary + confidence. Trust the summary; pull the full file only when you must. This keeps your context clean across many delegations.
When a delegated agent can't predict what context it needs up front, have it apply soe:iterative-retrieval (dispatch → evaluate → refine) to gather context progressively instead of over-loading its window.
Two usage levels
- Ambient — no
/go, no track, no state. In any conversation (ad-hoc debugging, review, enhancement) you choose to spawn the right pinned agent per this skill. This is a convention you follow, not an enforced router.
- Pipeline — opt-in
/go. The Evaluate-Loop's own agents are already pinned to their tiers; do not duplicate the tiering here.
Explicitly out of scope
- No Advisor — no API-only side model; the user is on a subscription.
- No runtime auto-fallback — the user picks the model via
/model; you never silently switch the session model.
- Fable gate (config) — two independent switches govern the strategist tier: (1) availability — if the user is not on a Fable plan, the strategist is simply not invoked (topology degrades to
deep-reasoner/Opus); (2) config — .soe/config.json fable_enabled:false routes the strategist tier to the reasoner (Opus) even when Fable is available, to cap Fable spend. Resolve the strategist's model at dispatch via lib/model-resolve.js resolveModel(config, 'strategist') and pass it as the per-invocation model (per-invocation model outranks frontmatter). Default is fable_enabled:true — use strategist on merit (stakes/reversibility/ambiguity).
