cheap-tier-recon-dispatch

name	cheap-tier-recon-dispatch
description	When dispatching cheap recon / spec-drafting work in a MURE workflow, fan across canary-proven cheap-tier routes by sub-task rather than running it inline in Sol. Trigger: any R0 read-only recon or R1 spec drafting that can be split across cheap-frontier models.

Cheap-Tier Recon Dispatch — MURE Roster Rule

Owner directive (Marcel, propagated via Griffin fleet message 2026-07-16 msg-8ae7c9b0): when cheap recon or spec drafting is part of a MURE workflow, fan across the canary-proven cheap-tier roster — do NOT run it inline in Sol. Match model to sub-task.

Canary-proven cheap routes

Route	Backend	Sub-task fit
deepseek-flash	ollama-cloud	scan / extract / classify (R0)
mimo (opencode-go/mimo-v2.5)	opencode	scan / extract / classify (R0); supports R1 with a config flip
composer-2.5	cursor	code + spec drafting (R1)
gemini-3.5-flash	cursor	code + spec drafting (R1); broader scope fallback
kimi-k2.7	ollama-cloud	code + spec drafting (R1)
nemotron-3	ollama-cloud	code + spec drafting (R1); second-opinion lane

Fail-closed carve-outs

• deepseek DIRECT → fail-closed. Use ollama-cloud ONLY. Direct request shape not currently canary-proven.
• Any route not on the canary-proven list → fail-closed (Terra / Luna / Sol also fail-closed per OMP default; Cline provider unavailable).

Sub-task routing

• R0 read-only (scan / extract / classify / evidence-gathering): deepseek-flash or mimo — in parallel, fanned out by independent slice.
• R1 spec / code drafting: composer / kimi / nemotron — parallel drafts of independent slices.
• R2+ verification or architecture: escalate to the heavy frontier (Opus / Sonnet thinking), not Sol-inline.

Hard rules

1. R0 recon MUST be fanned — never inline in Sol unless it's the trivial one-shot (single file, ≤50 lines, no multi-stage bash, per fleet-economy Iron Rule 5).
2. R0 may also be R0 + evidence (cheap produces only R0 + evidence, never R1+ semantic — promote the producer, not the cheap lane).
3. Sub-task slices dispatched with the cheapest tier that fits — never escalate to a heavier tier when a cheaper one canary-proven for the slice.
4. All fan-out respects the OMP dispatch rule for the current session (see provider route eligibility in OMP RULES).

Operator protocol

Before dispatching recon:

1. Slice the work into independent R0 tasks (scan / extract / classify / code-draft).
2. For each slice, pick the canary-proven route from the table above.
3. Confirm depth (R0 vs R1 vs R2+) matches the lane's allowed tier.
4. Dispatch in parallel via task tool; collect results; verify locally before integrating.

Provenance

• Origin: Griffin fleet message (worker-term-mrnswa2u-k) via Marcel directive, 2026-07-16.
• One known violation this session: the gap-closure spec was produced inline in Sol instead of fanned to deepseek-flash (R0 scan/extract) + composer (R1 spec drafting). Next dispatch on this thread MUST route per the table.