Skip to content

[Draft] feat(prompt-preset): add delegation + skill gates to deepseek-v4 and kimi-k3 (canary measuring) - #912

Draft
grim-susemi wants to merge 1 commit into
code-yeongyu:mainfrom
grim-susemi:feat/add-deepseek-k3-delegation-presets
Draft

[Draft] feat(prompt-preset): add delegation + skill gates to deepseek-v4 and kimi-k3 (canary measuring)#912
grim-susemi wants to merge 1 commit into
code-yeongyu:mainfrom
grim-susemi:feat/add-deepseek-k3-delegation-presets

Conversation

@grim-susemi

@grim-susemi grim-susemi commented Aug 17, 2026

Copy link
Copy Markdown
Contributor

Draft — canary measuring, ready on M1-M4

본 PR은 **카나리아(개인 확장 omo-subagent-skill-correction)**가 행동 게이트를 통과하면 ready로 전환됩니다.

  • 배경: GPT 프리셋에는 DELEGATION 등 위임 규칙이 있으나 DeepSeek V4(3종)와 K3에는 없어, K3 위임률 0.21%(1/470), DeepSeek 첫 위임 62.3메시지, SKILL.md 0회 read로 관측됨 (.omo/plans/omo-model-presets-deepseek-k3.md Findings).
  • 변경:
    • deepseek-v4.ts 5→7 rules: delegate-separable-work + load-matching-skills (ALL_PRESETS, 4규칙 중 2)
    • kimi-k3.ts core Working the Task에 위임 1회 결정 + 스킬 terminal gate 2문단 추가 (canary Appendix A verbatim)
  • 테스트: prompt-presets-deepseek-v4 66/66, prompt-presets-kimi-k3 포함 — 기존 canary 규칙 id와 동일(deepseek-v4.delegate-separable-work 등), 구조 파싱만 (no pinned prose).
  • 게이트: Appendix B — ≥20 자격 턴 또는 2주, M1 +10%p, M2 +10%p, M3 ≤+5%p, M4 ≤+2%p 통과 후 ready. 통과 전 머지 금지(G13).

Canary: ~/.omo/agent/extensions/omo-subagent-skill-correction.ts (DISABLED=false) — 같은 4규칙이 이 머신에서 이미 주입 중. 측정·리포트는 매일 09:00 Discord 수집수세미 채널.

Related: .omo/plans/omo-model-presets-deepseek-k3.md, .omo/scripts/omo-subagent-skill-correction.mjs


Summary by cubic

Adds delegation and visible-skill gates to DeepSeek V4 and Kimi K3 presets to address under-delegation and missed SKILL.md usage. Previously these presets kept work local and skipped the skill scan; now they partition work at intake, batch separable task calls, and enforce a terminal visible-skill scan that loads matching skills and passes load_skills when delegating.

Review and rollout

  • Changes are limited to packages/coding-agent/src/core/extensions/builtin/prompt-preset/deepseek-v4.ts, packages/coding-agent/src/core/extensions/builtin/prompt-preset/kimi-k3.ts; packages/coding-agent/CHANGELOG.md adds a fix note.
  • DeepSeek V4 adds rule IDs: delegate-separable-work, load-matching-skills (ALL_PRESETS). Kimi K3 adds two paragraphs under “Working the Task” for delegation-at-intake and the terminal visible-skill gate.
  • Tests pass (66/66). Verify each preset renders the new rules exactly once and that no non-rule prose is pinned.
  • Canary measuring only. Hold promotion until gates pass: ≥20 qualified turns or 2 weeks, with M1 +10pp, M2 +10pp, M3 ≤+5pp, M4 ≤+2pp vs baseline.

Written for commit 50b882e. Summary will update on new commits.

Review in cubic

…kimi-k3 [draft]

DeepSeek V4 (flash/flash-0731/pro) and Kimi K3 under-delegate and
under-load skills vs GPT (K3 0.21% task rate, DeepSeek first delegate
avg 62.3 msgs, 0 SKILL.md reads). This draft ports the canary's 4
rules into builtins for review while the canary measures M1-M4 at
~20 qualified turns (Appendix B gates).

- deepseek-v4.ts 5→7 rules: delegate-separable-work + load-matching-skills
  (subagent-delegation / skill-utilization, ALL_PRESETS)
- kimi-k3.ts core: two new paragraphs under Working the Task — delegation
  once at intake + visible-skill terminal gate (verbatim from canary)

Tests: 66/66 pass. Prompt-behavior-coverage (no pinned prose) holds.
Evidence: existing preset tests render once per rule.
@grim-susemi
grim-susemi force-pushed the feat/add-deepseek-k3-delegation-presets branch from 0c25619 to 50b882e Compare August 22, 2026 06:18
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant