dev-team: effort-band model routing (task-size drives the model)

[outofrange-consulting]

Applies the upstream effort-band routing idea (v7.0) without breaking our tier model or copilot-preset: the model is derived from the task size and pipeline phase, floored by the agent's tier. Write-up: docs/effort-band-routing.md.

Model (phase-aware bump-from-floor)

The effort goes into spec/plan, not the build — a complex task needs deep reasoning to design, and a solid plan makes implementation routine. The agent's model: tier is a floor; the task size raises the band only while planning:

sizeBand   = { trivial: small, standard: balanced, complex: deep }
bumpStages = [needs-plan]                      # /scope -> /specs -> /plan
effective  = (stage in bumpStages) ? max(floor, sizeBand[size]) : floor

• Planning (needs-plan): a complex plan runs the architect + the 5 plan-review critics at deep.
• Build/review (plan-approved): no bump — implementers and reviewers run at their floor (don't spend opus on mechanical work, and don't route every lexical reviewer on opus just because the task is complex).
• Never below the floor → high-stakes deep agents (security/domain/arch-review, architect, security-engineer, codebase-recon) always hold at deep.
• No signal / trivial / unscoped → the floor (backward-compatible).

How

• model-routing.json → effortBand { ladder, sizeBand, bumpStages, enforcement }.
• lib/shared.ts → pure effectiveBand(floor, size, stage); 11 phase-aware unit tests.
• model-routing.ts → reads floor (agentModel) + size + stage (plan-gate); logs every dispatch; advisory warns (default), DEV_TEAM_EFFORT_ROUTING=enforce blocks, =off disables. /routing shows the effective band per floor for the current stage + size.
• plan-gate.ts → /scope records size (--trivial | --complex | standard); /plan-approve preserves it.

copilot-preset untouched (band picks a tier; modelRoles resolves the model).

Pre-existing routing debt cleaned up

• The orchestrator Resolution Procedure described a Claude-Code hook (agent-model-resolve.sh, .claude/model-overrides.json, updatedInput, deny, a non-existent docs/model-routing.md + ADR 0004) the OMP extension never implemented → rewritten to reality + the new behavior.
• 4 skills + agent-registry repointed off the stale hooks/agent-model-resolve.sh to the model-routing extension.
• Left for a separate pass: the /model-routing-check exec block's deeper .claude/ debt (/routing is the accurate diagnostic; the skill now points to it).

Verified

ci-validate-json 23/23 · extensions compile · unit suite green (incl. 11 phase-aware effectiveBand cases) · no agent-model-resolve.sh refs remain.

🤖 Generated with Claude Code