devflow-adaptive-routing

Adaptive Routing

Pick exactly one mode, then assign model tiers. This file is the ONLY source of the routing rules and the model matrix — commands reference it, never restate it.

Decision rules — evaluate in this exact order

1. debug — user describes a problem that already exists: quoted error/stack trace, "X is broken", "test fails", "crash", "regression", "wrong output". NOT debug: "add error handling", "improve auth", "harden security" — that's feature work (standard/deep) unless an actual failure is also described.
2. tiny — exact scope is one sentence AND no behavior risk: typos, comments, log/message wording, single-line config, format-only changes, missing import, local rename. Applies regardless of file location (auth.js, payment.go) as long as the change cannot alter runtime behavior.
3. deep — ANY of: new architecture/module/boundary; data model/schema/migration; auth/security/payments/PII/billing behavior; breaking public API contract; multiple modules or services touched; unclear or contradictory requirements; high regression risk on a core path; multi-session work.
4. standard — everything else (multi-file but architecture clear, behavior change with tests, no architectural decision pending).

Conflicts: debug + deep → start debug (root cause first), escalate to deep after root cause. Unsure tiny vs standard → standard. Unsure standard vs deep → deep. Unfamiliar code → never tiny.

Common shapes

Task	Default
SQL extraction / cleanup / formatting	tiny (standard if restructuring the query)
New endpoint / service / repo method	standard
Bug, failing test, 500, crash, container exit	debug
Review only (no implementation)	/devflow-review
Deployment / config troubleshooting (Docker, env, secrets)	debug
Architecture / data model / migration / payments / auth refactor	deep

Model + effort tiers

One model everywhere: Fable (model: fable). Tiers differ by effort, not model. Both are pinned in each agent's frontmatter (agents/*.md) — the Agent tool has no per-call effort param, and a per-call model= would override frontmatter. Commands therefore never pass model= on dispatch; picking the agent picks the tier.

Tier	Effort	Agents	Use for
max-reasoning	xhigh	devflow-planner	planning, specs
working	medium	devflow-plan-reviewer, devflow-implementer, devflow-reviewer, devflow-debugger, devflow-scout	plan review, implementation, review, verification, debugging, discovery

Inline work (tiny mode, orchestration) runs at the session model + session effort. Update this table and the agent frontmatter together when models or effort levels change.

Matrix (canonical)

Mode	Phase	Agent (effort)
tiny	all (inline — no subagent)	session model @ session effort
standard	discovery + design (one dispatch)	devflow-planner (xhigh)
standard	plan review	devflow-plan-reviewer (medium)
standard	implementation	devflow-implementer (medium)
standard	review + verify	devflow-reviewer (medium)
deep	discovery (standalone pass)	devflow-scout (medium)
deep	spec + plan	devflow-planner (xhigh)
deep	plan review	devflow-plan-reviewer (medium)
deep	implementation	devflow-implementer (medium)
deep	review + verify	devflow-reviewer (medium)
debug	all	devflow-debugger (medium)

Planning always runs max-reasoning — never downgrade it. A wrong plan executed perfectly is wasted work. Everything else — plan review, implementation, review, debugging, discovery — runs medium: reasoning budget is reserved for producing the plan, and a good plan makes medium-effort execution sufficient.

Escalation: tiny-mode ESCALATE → re-classify as standard.

Dispatch announcement — MANDATORY

Immediately before every Agent dispatch, print one line in the conversation (hooks cannot render this; only your own text is guaranteed visible):

▸ devflow-planner · fable @ xhigh

Agent → effort: planner = xhigh, plan-reviewer/implementer/reviewer/debugger/scout = medium. Same values as the agents' frontmatter — frontmatter is the source of truth.

Output format

Mode: <mode> — <one sentence naming the deciding rule>
Route: <phase>=<agent>, ...

Two lines. No filler. No alternatives listed.