dsh-skill-router

Intent-level skill auto-routing for DeepSeek Harness — the harness finally uses skills when it should.

The problem: DSH injects an <available_skills> catalog into the context, but whether a skill gets loaded depends entirely on the model's own judgment. With many skills, the model misses the right one. dsh-skill-router hooks agent/pre-step and routes every user message: on a hit, the full skill body is injected into the context automatically — the model never has to remember to call the skill tool.

How it works

user message → heuristic scoring (zero LLM cost)
                ├─ strong hit (≥ minScore) → inject skill body directly
                └─ weak signal → LLM judgment (12s timeout, falls back to the weak band)
                        ├─ judged hit → inject skill body (createUserMessage + skill-invocation source)
                        └─ fallback per weakForm ─ summary: compact hints block (default)
                                                ├ full: body fallback (v0.2 behavior)
                                                └ none: skip

• Chinese-aware heuristic: CJK bigram segmentation (stopword filtering), latin token overlap, skill-name and whenToUse trigger-phrase hits.
• Three-tier routing: strong hits inject the full body; weak signals go to LLM judgment; the unresolved weak band falls back to a compact skill-hints summary block by default (weakForm: summary) — borrowing the discovery/loading separation of dsh_cot_gw_dyn's skill_search (large-body injections perturb trajectories; summaries are cheap to discover from), while staying zero-tool-registration.
• Per-user personalization (v0.4): on first route the plugin scans this environment's actual skill catalog and MCP tool surface (ctx.tools.schemas enumerating mcp__<server>__<tool>) into a persisted baseline (refreshed periodically). The weak-signal hints block additionally lists message-matched MCP tools — every user discovers what their environment really has — and an incomplete live catalog falls back to baseline candidates.
• Compatible with 0.1.2-alpha.5: alpha.4 removed Session.events in favor of the on-demand snapshotEvents()/eventAt() API; first-message detection uses the new API and falls back to the events array on older kernels (v0.4.1, verified against alpha.1 / alpha.3 / alpha.5 kernels).
• Reuses the official injection pipeline (skill-invocation source) — the same mechanism as the /skill-name gesture.
• Never breaks the agent loop: any routing failure just falls back to the status quo.
• No tools registered, no tool-catalog pollution.
• Compatible with experimental presets: on strict on-demand compositions with NO skill channel at all (anchored-standard-style, no skill and no skill_load/skill_search) auto-routing steps aside by default; cot-family presets (cot-gw / cot-dyn) that expose skill_load/skill_search are auto-detected (onDemandLoaderTools) and routed normally — automatic injection coexists with on-demand discovery. The very first message of a fresh session is never injected (preserves the first-request Minimal anchor).

Install

Runtime injection via dsh-super-injector:

dev_inject_plugin <this repo dir>

Or add it to the profile bundles for persistent assembly.

Config

Test

node test-router.mjs    # functional test: standard / anchored scenarios
node test-integration.mjs  # upgrade verification: cot-family auto-detection + routing (12 assertions)
node debug-scoring.mjs  # scoring debug

Dependencies: @deepseek-ai/dsh-llm, @deepseek-ai/dsh-skill, @deepseek-ai/schemastery.

License

MIT