The local Qwen driver does everything. Cloud help is rare, optional, and stateless.

The conductor adds cloud help as TOOLS, never a backend swap — the Hermes model stays the orchestrator and integrator the whole time. It is fully presence-gated: with zero cloud keys set, every rung is OFF and you proceed local-only. Reach up ONLY when the difficulty classifier says HARD/novel, the watchdog reports genuine-stuck after local recovery already failed, or there's a major architectural fork. Routine work stays local at $0 — do NOT reach up for it.

Step 1 — ask the policy which rung (don't guess)

Call conductor_advise(signals, verifiable) with the cheap signals you already have (file_count, novelty, prior_failures, lines_changed, cross_module) and whether the subtask has an objective TEST ORACLE (verifiable=true iff you can write tests that decide success). It returns the ladder rung, presence-gated — it never fires a call. The ladder by subtask type:

• routine (easy/medium) → tier=local. Stay on the local model. Stop here.
• verifiable + hard → tier=parallel_draft → if none pass, synthesize.
• ambiguous + hard → tier=steer (cheap nudge) → if it doesn't unblock, synthesize.
• frontier-novel / synth-failed → tier=escalate (Opus) ONLY when synth failed verify TWICE or two synth opinions DISAGREE on a high-blast-radius change. Otherwise the gate stays shut.

If a role is OFF (no key) the plan degrades automatically (pool off → synthesize → local; steer off → synthesize; synth off → local; Opus off → surface to human/local). Honor what tier it returns.

Step 2 — assemble the brief deterministically (never hand-write it)

You (the weak local model) write ONLY two fields: current_blocker and decision_needed. Then call brief_assemble(task_id, current_blocker, decision_needed, profile) — it pulls goal/done/constraints/ success from PLAN.md, architecture + failed_approaches from the KG + watchdog, and token-budgeted code from codebase-rag. Profiles: compact for steer, full for synthesize, draft for parallel_draft (pass acceptance_tests — the oracle). Use brief_request_more if the cloud asks for more.

Step 3 — fire the chosen rung (stateless; never raises)

• parallel_draft: parallel_draft(task_spec, tests=<oracle>, …) on mcp-search — fans best-of-N across the free pool and the VERIFIER (not a model) selects the green winner. None pass → it returns route_to=synthesize. Integrate the winning selected_files yourself; the slop models never touch the repo.
• steer / synthesize: conductor_steer(brief) / conductor_synthesize(brief) — first present rung wins, silent fall-with-log on failure, proceed_local if the role is OFF. The cloud returns a STRUCTURED DIRECTIVE; it is ADVISORY.
• escalate: only when the Opus gate is met — the existing escalate tool (capped).

Step 4 — GATE the directive before you execute it (the cloud is smart but BLIND)

Call directive_verify(directive, repo) before touching the tree. It checks every assumptions entry against ACTUAL repo state (a false one is rejected + recorded as a failed_approach — re-brief), confirms apis_to_use exist, requires concrete tests_to_write, and on low-confidence + high-blast-radius asks for a second synth opinion (compare_directives decides agree→execute vs disagree→escalate/human). Execute + checkpoint ONLY when execute is true. Write the prescribed tests FIRST.

Step 5 — record the outcome (the compounding flywheel)

Call conductor_record_outcome(subtask, tier, outcome) at subtask end so the difficulty classifier learns which subtasks needed which tier and repeated blockers reuse prior directives. Periodically check conductor_frequency_report — synth ≤ ~15/project and Opus ≤ ~3 are the honest targets. If Opus calls exceed 3, the brief-assembler quality is the bottleneck — fix the assembler, don't spend more on Opus.

The discipline (non-negotiable)

Slop-draft the verifiable, synthesize the ambiguous, escalate only the frontier-novel. Never run parallel_draft on an ambiguous subtask (no oracle → the verifier can't select → wasted slop). Never swap the Hermes backend model. No secrets in any brief. With all keys unset this whole skill is a no-op and the bare local harness runs unchanged.