dachent/model_proctor

model-proctor

Task-owning worker policy with deterministic acceptance and evidence-driven escalation (model_proctor) — run substantial work through the gated path (runner.py init/dispatch/verify/accept; flash-first with escalation to glm-5p3 then kimi-k3 on recorded…

Idioma del texto original: inglés

model-proctor

Use when substantial coding or proctor tasks require runner-governed delegation, or when the user explicitly invokes model-proctor. Do not use for trivial tasks — execute those directly. Requires the model-proctor install.

Idioma del texto original: inglés

multi-model-routing

Multi-model routing policy for Kimi K3 — delegate coding subtasks to fast/cheap models (k2p7 scout, DeepSeek V4 flash/pro, GLM-5.2) via the delegate wrapper and native GLM subagents, with read-only Codex/Claude advisors. Use ONLY for nontrivial coding tasks…

Idioma del texto original: inglés

multi-model-routing

Full multi-model routing policy — delegation to fast/cheap models via the delegate wrapper plus native GLM subagents (eval config C). Use ONLY for nontrivial coding tasks (unfamiliar codebases, multi-file changes, debugging, migrations, security-sensitive…

Idioma del texto original: inglés

multi-model-routing

ds-pro-start routing policy — DeepSeek V4 Pro is the default worker and you adjudicate (eval config D). Use ONLY for nontrivial coding tasks. Do NOT use for trivial tasks (known location, small patch, one edit-test cycle) — execute those directly.

Idioma del texto original: inglés

multi-model-routing

Routing policy for delegating coding work to native GLM secondary subagents only (eval config B). Use ONLY for nontrivial coding tasks (unfamiliar codebases, multi-file changes, debugging). Do NOT use for trivial tasks (known location, small patch, one…

Idioma del texto original: inglés