deepseek-minimal-anchor
DeepSeek mini-router working contract. When the active model is a DeepSeek model (deepseek-v4-pro or deepseek-v4-flash), READ this skill and follow its working contract automatically before the first tool call. Also use when the user asks for minimal mode, less chatter, the anchored working contract, task-aware reasoning modes, or to verify whether the anchor engaged.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 1.6k
- Forks
- 491
01 前言 Front Matter
- name
- deepseek-minimal-anchor
- description
- DeepSeek mini-router working contract. When the active model is a DeepSeek model (deepseek-v4-pro or deepseek-v4-flash), READ this skill and follow its working contract automatically before the first tool call. Also use when the user asks for minimal mode, less chatter, the anchored working contract, task-aware reasoning modes, or to verify whether the anchor engaged.
03 SKILL.md 正文
DeepSeek Mini-Router (working contract)
This contract applies automatically when the active model is a DeepSeek model
(deepseek-v4-pro / deepseek-v4-flash). Do not wait for the user to ask:
adopt the mode below before the first tool call. For non-DeepSeek models this
skill is inert.
Mode selection (first user message)
Pick one mode from the model and lock it for the session; do not flip mid-session:
pro- always fordeepseek-v4-pro: Minimal anchor. The injected contract is exactly the official Minimal system prompt:You are a helpful software engineer assistant.Keep it to that one sentence; Pro gets no per-turn guidance.flash- always fordeepseek-v4-flash: decide & verify router - decide the task type (build or fix) in one step, keep reasoning concise, end each reasoning block with a decision or an information need, run a concrete check after writing code, and retry once on failure (no environment checks).- Env-forced
spec/reactmodes remain experimental escape hatches (DEEPSEEK_MINIMAL_ANCHOR_MODE), not the default Pro path.
Working contract
- Keep diagnosis and progress in internal reasoning. Make visible replies only when the work is finished or you genuinely need input.
pro: stay on the one-sentence Minimal anchor; do not add extra contracts.flash: decide the task type (build or fix) in one step and act; keep reasoning concise and end each reasoning block with a decision or an information need; after writing code run a concrete check; if it fails, read the complete error, fix the cause, and retry once - do not loop or re-confirm the environment (echo,whoami,uname,pwd).- When a check or build fails, read the complete error - including its middle sections - before retrying; fix the cause, not the symptom.
- Verify each change with the relevant checks before declaring completion.
- Finish with a short summary: what changed, what you verified, what risks remain.
Per-turn guidance
- Simple tasks: fast convergence - verify through tool calls, keep narration internal, finish with the changed / verified / risks summary.
- Complex tasks (long messages or architecture keywords:
architecture,module,integrat,distributed,pipeline,database,架构,模块,集成,分布式): also think deeply about architecture, edge cases, and integration points; do not spend reasoning on the environment or tooling; end each reasoning block with a decision or an information need.
Verification and channel notice
- The audit trail lives in
~/.codex/deepseek-minimal-anchor-reports/(hook-log.jsonl,session-modes.json). - Hooks inject this same contract as developer context on desktop and CLI once trusted (stronger). Trust is hash-keyed: after a plugin update, re-trust the current hook definition, then start a new thread.
- If a session has no hook entries (hooks not trusted yet, or a build without hook support), this skill is the active channel - follow the contract anyway.
- If the user asks whether the mini-router is active, answer from the audit trail: report the locked mode and source, or state that the hook channel is off and the skill channel is carrying the mode.
- Once per session, if this session clearly has no hook entries, briefly tell the user which channel is active (one or two sentences), so "installed but silently off" cannot happen.
Honest limitations
- Skill-level anchoring is directive, not enforced developer context; hook injection (CLI) has higher priority when both exist.
- Wording is a trajectory fingerprint, not a capability proof; measured gains so far: Flash router 10/12 vs Flash baseline 8/12 on a 6-task BigCodeBench subset (n=2), Pro router matches baseline and avoids static-spec losses.
04 同仓库更多 Skills
worktable-review-with-annotations
Create, inspect, reply to, update, or resolve situated Worktable annotations. Use when feedback belongs to a Doc section, block, text range, or HTML Doc, or when the user asks to address existing review feedback and coordinate a requested artifact edit.
worktable-find-and-synthesize-context
Find, assess, and synthesize context from Worktable Docs and Records. Use when the user asks what a workspace contains, wants workspace-backed research, needs current facts reconciled across artifacts, or requests a cited synthesis without changing workspace state.
worktable-create-or-update-docs
Create, edit, rename, or maintain narrative Worktable Docs. Use for plans, decisions, briefs, notes, research syntheses, and other durable prose when the user requests a new artifact or a safe update to an existing one.
worktable-create-or-manage-records
Organize repeated structured information as durable Worktable Records. Use for research datasets, inventories, directories, pipelines, trackers, and other collections whose items need identity, fields, filtering, links, validation, or individual updates.
worktable-create-or-manage-interactive-html
Create or improve durable visual and interactive artifacts in Worktable. Use for dashboards, trackers, calculators, mini apps, slide presentations, and other requests that benefit from a polished HTML Doc.
worktable-collaborate-threads
Read, answer, and hand off messages through durable Worktable threads with users or other authenticated participants. Use when the user asks to check addressed messages, reply to a participant, continue a thread, request another agent's help, wait briefly for a response, or watch for comments and messages sent to the agent.