Ansonfishing/dsh-cap-profile

Languages: English | 中文

Per-model capability profiling — a plugin for DSH (DeepSeek Harness).

Why

The more sessions you run, the harder it is to answer "which model is reliable at what, and where does it tend to fail". Reading raw session logs is slow, and manual tallying misses things.

This plugin turns the session history in ~/.dsh/sessions into per-model capability profiles, directly answering: per-model session count, tool-call volume, error count, error rate, most-used tools, and top error signatures.

Quick start

Prerequisites: DSH (with web) + pnpm; Node ≥ 24 (zstd decoding uses the built-in zstdDecompressSync).

cd ~/.dsh/profiles/web                      # your DSH web profile directory
pnpm add github:Ansonfishing/dsh-cap-profile

Then add "dsh-cap-profile" to the dsh.profile.bundles array in package.json and restart dsh. A "Capability Profile" tab appears in the conversation view; the first open triggers a background initial scan (~20s for large histories), during which mock data is shown.

Features

• Model comparison table — per model (provider / model): sessions, tool calls, errors, error rate.
• Top-5 tools / Top-5 error signatures — per-model most-used tools (with a per-tool error-rate column: 0 gray / <5% amber / ≥5% red); normalized high-frequency error patterns (e.g. bash: [exit code: 137] OOM) reveal where a model tends to fail.
• Multi-model compare (≤4) — tick several models to open a compare view: one card per model with side-by-side metrics (best/worst/top tones), a Top-10 tools matrix, Top-5 error signatures, and a mini daily calls/errors bar chart (today's errors marked red); fewer than 2 valid selections falls back to the single-model detail.
• Time-range filter — All / Last 7 / Last 30 / Last 90 days / Today / Yesterday (anchored to the max date in the data, resilient to system clock drift).
• Read-only — analysis and display only, never mutates session data; the route validates a client header plus Origin/Referer and returns 403 on cross-origin requests.
• Incremental caching — per-file mtime baseline with background scanning (60s incremental / 24h full re-scan); while the initial scan is still running the route falls back to the last cache or mock data, so it never blocks.
• Zero runtime dependencies — pure Node.js.

No DSH? Take a look anyway

Clone this repo and open test/harness/index.html in a browser — a zero-dependency render harness (mock data; ?scenario=mock|live|empty|error switches scenarios, ?chrome=0 hides the harness bar).

Development

npm test                     # node --test test/*.test.mjs

Local development: after cloning, use pnpm add link:../path/to/dsh-cap-profile in your profile. Client-only changes need a browser refresh; Node-side changes (index.js / lib/*.js) need a dsh restart.

License

MIT © Ansonfishing