deepseek-minimal-anchor

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 for deepseek-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 for deepseek-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 / react modes 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.