dsh-flash-godmode

V4 Flash "god mode" for DeepSeek Harness: w7 persona anchoring + first-turn tool anchoring + complexity-dispatched guidance — turn Flash's unstable thinking into reproducible capability.

!MIT !DSH 

> 简体中文版见 README.zh.md

V4 Flash on reasoning: max is powerful but unstable: it over-thinks simple tasks, under-thinks complex ones, and drifts mid-session. This plugin turns that instability into a routed capability, using the measured persona bands from dsh-router-standard:

• w7 persona (author-measured Flash optimum) — neutral identity + classify-then-act instruction + recall/anti-runaway anchors. No deep-thinking sentence baked in (author's correction: per-message guidance is negative on related-task chains, P21).
• Complexity-dispatched guidance — complex tasks get the decision-closure deep guide, simple tasks get the fast-convergence guide. Dispatched on the very first request, thanks to a capture path that works on headless (see below).
• First-turn tool anchoring — the first request sees only the core tool set (read/write/edit + shell); after the first durable tool call the full catalog is exposed.

Dependencies: DSH 0.1.x (headless profile) + any openai-completions provider (built and tested with the opencode-go channel, deepseek-v4-flash, reasoning: max).

---

Why the capture path matters

The author's own router reads the first user message via user/message events. On headless, that event lands after the first prompt assembly — too late for the path-committing first request. This plugin listens for agent/inbox/spliced instead, which carries the task text before the first assembly (verified on rc.6 headless). Result: the guidance decision is present on request #1, where it counts.

Quick start

# bundle install (recommended)
dsh plugin --profile headless add github:Cavan-Ou/dsh-flash-godmode

# verify: run any task; stderr shows the dispatch decision
dsh --profile headless "简单任务：运行 python3 -c 'print(1+1)' 并报告输出" 2>&1 | grep godmode
# → [godmode] taskLen=NN complex=false guide=fast mode=weak

Flash models are routed to weak mode automatically (model-id detection). Non-Flash models fall back to keyword task classification (react/spec/weak).

What's inside

lib/
├── router-core.mjs        # pure routing logic: personas, bands, complexity heuristic, task classification (zero deps)
└── router-bootstrap.mjs   # Cordis plugin: task-text capture, guidance dispatch, tool anchoring
tests/
└── router.test.mjs        # node:test unit tests (11 cases)

Observed results

Known edges

• The complexity heuristic (text.length > 120 or architecture keywords) is the author's original — a filename like analyze.py counts as complex. False positives lean toward the safe direction (deep guide, not no guide).
• If the agent/inbox/spliced capture fails on a future DSH version, the plugin falls back to the bare w7 persona (no guidance) — safe per P23, and the task still runs.
• All author-measured numbers above come from the official DeepSeek API; this plugin was verified on the opencode-go channel (same weights, different serving) — see the local runs in the table.

Relationship to dsh-routing-suite

dsh-routing-suite is the author's full kit (runtime injector + router presets) aimed at the web profile (PowerShell installer, preset selection per session). This plugin is the headless-native single-plugin form: installable via dsh plugin add, no installer script, first-request guidance on headless (which the suite's own issue tracker flags as not taking effect on rc.6).

Contributing

This plugin grows from real use. Hit an edge not listed here? Open an issue with the incident (symptom → root cause → fix). PRs welcome in Chinese or English.

License

MIT — with attribution to dsh-router-standard (MIT), whose measured persona bands and guidance texts this plugin builds on.

---

Built by running real tasks on headless, not by reading docs.