dsh-wrapped

A DeepSeek Harness (dsh) plugin that turns a session into a shareable summary card — Spotify Wrapped, for your agent session.

Why

Every session already has real, interesting numbers behind it — how much time went to reasoning versus tool calls, how fast the model streamed, how many turns it took — but they only ever show up as a small stats strip you glance at and forget. dsh-wrapped turns those numbers into a dark, screenshot-shaped card built for actually posting somewhere.

Install

dsh plugin --profile <name> add dsh-wrapped

Add the package name to that profile's dsh.profile.bundles list (see
dsh-minimal-anchor's README
for why this second step is required — being a listed dependency alone
doesn't activate a plugin's dsh.bundle patch):

{
  "dsh": {
    "profile": {
      "bundles": [
        "@deepseek-ai/dsh-base",
        "@deepseek-ai/dsh-web-app",
        "dsh-wrapped"
      ]
    }
  }
}

Confirm it composed with dsh --profile <name> --dump-config — look for a
wrapped entry.

Usage

Type /wrapped in any session. It writes an SVG card to
.dsh-wrapped/wrapped-<sessionId>-<timestamp>.svg under the current working
directory and prints a plain-text summary alongside the file path. Open the
SVG in a browser and screenshot it, or convert it to PNG with any SVG tool.

Needs @deepseek-ai/dsh-session-stats mounted in the profile (it ships in
the dsh-web-app bundle by default) — without it, /wrapped returns a
clear error instead of a blank or wrong card.

What's on the card

• Turn and step counts
• Tool time vs. thinking (LLM) time
• Average time to first token
• Decode speed (tokens/sec), omitted below a 1-second sample floor to avoid
  a noisy/misleading rate from too few tokens
• A one-line "type" verdict derived from the tool-time/thinking-time ratio:
  The Explorer, The Deep Thinker, or The Balanced Operator

Configuration

# profiles/<name>/cordis.patch.yml
- insert:
    - id: wrapped
      name: 'dsh-wrapped'
      config:
        minDecodeSampleMs: 1000
        verdictRatioThreshold: 1.5

All of it comes from the session's own sessionStats projection — no
separate tracking, no extra token cost, nothing persisted beyond the SVG
file itself.

How it works

Registers a /wrapped command via @deepseek-ai/dsh-commands. The handler
reads ctx.sessionProjections.snapshot(invocation.agent.session) for the
sessionStats key (turns, steps, llmMs, toolMs, ttftMs/ttftSteps,
decodeMs/decodeTokens), transforms it through a pure buildCard()
function into display rows and a verdict, then renders that to a
self-contained SVG string and writes it with plain node:fs. Commands never
enter model history and cost no tokens — this doesn't touch a single
request.

buildCard and renderSvg are pure functions with no dependency on a
booted Context, so the interesting logic (verdict thresholds, decode-speed
sample floor, XML escaping) is unit tested directly — see
Development.

Verified against a real local dsh web boot, not just types: installed
into a scratch profile, drove a real browser session with Playwright, ran
/wrapped after a real multi-tool-call turn, and confirmed the written SVG
file's numbers matched the session's own stats strip.

Troubleshooting

/wrapped returns "sessionStats projection is not available". Your
profile doesn't have @deepseek-ai/dsh-session-stats mounted. It ships by
default in the dsh-web-app bundle; a custom or headless profile may need
it added explicitly. Check with dsh --profile <name> --dump-config.

"This session has no completed steps yet." /wrapped needs at least
one closed step (step/end) to have something to summarize — it can't
report on a session where nothing has happened.

Can't find the SVG file. It's written relative to the harness process's
current working directory, not your project directory or the harness
checkout — check the exact path in the command's success text, or look
under .dsh-wrapped/ from wherever you launched dsh.

Development

npm install
npm run typecheck
npm test

buildCard/renderSvg are plain functions with no dsh runtime
dependency — examples/render-standalone.ts renders a card from
hand-written stats, useful for iterating on the design without a live
session:

npx tsx examples/render-standalone.ts > card.svg

License

MIT