Skip to main content

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.

معلومات المصدر

المستودع
dachent/model_proctor
آخر نشاط في المصدر
٢٤ أغسطس ٢٠٢٦ في ١٨:٣١
لغة SKILL.md المكتشفة
الإنجليزية
النجوم
٠
التفرعات
٠

خيارات التثبيت

يُحدَّد Prompt الذي يراجع المصدر أولًا بشكل افتراضي. يمكنك التبديل إلى أمر مباشر أو تنزيل نسخة محلية.

مراجعة ملفات المصدر

اقرأ SKILL.md وأي ملفات مرافقة يعرضها SkillsMP قبل أن تقرر التثبيت.

عرض SKILL.md

SKILL.md
تعليمات المصدر · معاينة للقراءة فقط
name
multi-model-routing
description
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.
# Routing Policy — Config D (ds-pro-start, K3 adjudicates) You are the lead K3 orchestrator and final authority. Your default stance: DeepSeek V4 Pro does the work, you verify. Deterministic evidence (tests, compilers) outranks every model. ## The delegate wrapper ``` python "<repo-root>/delegate/delegate.py" --agent ds-pro-reviewer --workspace <workspace> --task "<task>" [--timeout <s>] ``` Parse exactly one JSON object from stdout; check `status`, `child_exit_code`; treat invalid JSON as wrapper failure. Worker assertions are untrusted until verified. ## The router (follow in order) 1. FAST-PATH GATE: execute directly yourself when ALL hold — location known; ≤ ~3 files; small patch; acceptance criteria clear; one edit-test cycle suffices; not security/ migration/concurrency-sensitive. Escape: one failure → retry once; misclassified → step 2. 2. EVERYTHING else — exploration, mechanical bulk edits, substantial implementation, debugging: delegate to `ds-pro-reviewer` with a self-contained task packet (objective, acceptance criteria, allowed scope, non-goals, known evidence, required validation, output contract, timeout). Instruct it explicitly whether it may write files. 3. ADJUDICATE every result yourself: run the project's tests/check scripts, inspect the diff, compare against the acceptance criteria. 4. Mechanical failure (tests fail, patch invalid): send the worker ONE follow-up delegation with the failure evidence. Still failing → step 5. 5. Conceptual failure (wrong approach, repeated wrong assumptions): stop; use a native GLM subagent (model="secondary") for a fresh implementation with your corrected plan. 6. You run final validation and report. ## Fire rules - Do not re-do the worker's analysis yourself — adjudication is verification (tests, diff inspection), not parallel work. If you find yourself implementing, delegate instead. - Never show a verification pass the worker's confidence claims; weigh deterministic evidence first. - Budgets: max 2 delegations per subtask, 6 total secondary invocations per task; max 2 review-fix cycles, then GLM, then stop and report honestly. - Stop when acceptance criteria pass on deterministic evidence or the budget is exhausted. ## Final acceptance gate Declare completion only when: acceptance criteria pass on deterministic evidence; required tests pass; the diff has been inspected; your response states limitations honestly.
عرض على GitHub