🧩 Skill Trail

August 23, 2026 · View on GitHub

每條回覆如實標出「觸發了什麼、成功用了什麼」。 基於引擎日誌的逐回覆技能/工具徽章,零模型自報。

⚠️ ** superseded by @bryan-cmf/dsh-skill-ledger ** — this plugin's per-reply badges live on there (dual-seat, deliverables-safe), with the alternative-skill hint and the Skills tab. This repo is archived as the design-history record.

English · 繁體中文

A DeepSeek Harness (DSH) plugin that renders a compact trail under every assistant reply: which skills were triggered (🧠) and which tools were actually used (✓) or failed (✗).

$ 🧩 技能軌跡 [🧠 \text{ak}-\text{hk}-\text{stock}-\text{dd} ✓] [\text{read} \times 3 ✓] [\text{bash} \times 2 ✓] [\text{web\_search} ✗] $

Why it is truthful

The badges are not produced by the model. A skillTrail session-projection unit (Host) folds each session's committed log — every tool/call event becomes a triggered entry; the paired tool/result (matched by callId) settles it as ok / err. The Client reads the finished value through useProjection('skillTrail') and renders it in the zero-risk conversation.chat.turnTail slot.

✅ Structural — the platform records, nobody self-reports
✅ Replayable — projection folds are pure; persisted cache + cold-start backfill come from the DSH projection subsystem for free
✅ Per-turn — each reply shows exactly its own turn's trail
✅ Skill-aware — skill tool calls render as 🧠 with the resolved skill name

Install

The package ships a Host half (lib/index.js) and a browser Client half (lib/client.js, window.__ModuleLoader__ closure-factory format).

Path A — Host composition (always-on, every session & preset)

Add to the profile's package.json:

"dsh": { "profile": { "bundles": ["@bryan-cmf/dsh-skill-trail"] } }

then install and restart DSH. The projection unit is registered process-wide; the client half renders in every conversation.

Path B — Agent preset

Add one row to your preset's agent.cordis.yml (a consumer-only row — it publishes no service, so it sits loose; no isolate realm needed):

- id: skill-trail
  name: '@bryan-cmf/dsh-skill-trail'

Path C — Dynamic experiment

Inside a session, @bryan-cmf/dsh-skill-trail can also be run through the dynamic Cordis flow; prefer A or B for anything standing.

Requirements

DSH >= 0.1.0-rc.6 (@deepseek-ai/dsh-session-projection mounted in the host composition — it is part of the shipped dsh-base bundle)
Browser UI needs the conversation.chat.turnTail slot (shipped chat view)

Configuration

Key	Default	Meaning
`maxTurns`	`500`	Turns retained per session; oldest pruned first
`maxCallsPerTurn`	`50`	Triggered-entry cap per turn (wire-payload guard)

Projection value

useProjection('skillTrail') → Record<turn, { triggered, used, failed }>

{
  "12": {
    "triggered": [{ "name": "skill", "skill": "ak-hk-stock-dd", "status": "ok", "callId": "call_1" }],
    "used": { "skill": 1, "read": 3, "bash": 2 },
    "failed": ["web_search"]
  }
}

Development

pnpm install
pnpm build        # tsc declarations → lib/types + tsdown → lib/{index,client}.js

License

MIT