dsh-session-eval

dsh-session-eval

English | 中文

Retrospective evaluation for DeepSeek Harness agent sessions — deterministic grade cards computed from sessions that already happened, plus cross-session regression diffs that answer "did my agent get better or worse after that plugin / prompt / model change?"

No benchmark suites to author. No LLM judges. No new runs. Every grade is a fixed, documented threshold over persisted session logs — fully reproducible.

Install

Requirements: Node.js 20 or 22 · a DeepSeek Harness profile that mounts the commands and sessionQuery services (the shipped web / agent profiles qualify).

dsh plugin --profile web add dsh-session-eval

Or from GitHub:

dsh plugin --profile web add github:kittimzhe/dsh-session-eval

The bundle overlay mounts /eval, /eval-diff, and /eval-history. If you maintain your own cordis.patch.yml, keep this row:

- insert:
    - id: session-eval
      name: 'dsh-session-eval'

Try it once

In DeepSeek Harness:

/eval

Session eval — a1b2c3d4… (dsh-session-eval v0.2.2)
Overall: A
  Reliability   A    2/87 tool results errored (2.3%)
  Re-ask        A    1 re-ask signal(s) over 23 turns (0.4/10 turns)
  Tool load     B    3.8 tool calls per turn
Notes:
  • Wall time 41.2 min over 23 turns.

The measurement layer of the session toolchain

Plugin	Layer	Answers
dsh-session-export	Evidence	"What exactly happened in this session?"
dsh-session-recall	Memory	"What did I do before, and where is it?"
dsh-session-eval	Measurement	"Was that session good? Is the trend improving?"

All three read through the same trusted ctx.sessionQuery seam, so any persistence backend (JSONL or SQLite) works without touching raw artifacts.

Contributing

• Local dev: npm ci && npm run typecheck && npm test && npm run bundle (Node 20 or 22).
• Start in the source: src/grade.ts (thresholds, grade cards), src/metrics.ts (metric extraction), src/trend.ts (/eval-history trend logic). See CONTRIBUTING.md for the source map and first-PR suggestions.
• Open gaps: #1 (configurable thresholds) and #2 (/eval-history --since) shipped in v0.3.0 — browse issues labeled good first issue.
• Rules: behavior changes need tests; doc changes must update README.md and README.zh.md in sync; version releases belong to the maintainer. Details: CONTRIBUTING.md.

How it differs from benchmark-style evaluation

	dsh-eval (benchmark)	dsh-session-eval (this)
Evaluates	Runs you launch ahead of time	Sessions that already happened
Setup	Author benchmark YAML + cases	None — read existing logs
Judge	Metric folding over trials	Fixed thresholds, deterministic
Best at	Pre-deploy benchmarks	Regression tracking over real work

Both are useful; they don't compete. Benchmarks test what you predicted; retrospective evaluation watches what actually happened.

Commands

Command	Effect
/eval	Grade card for the current session
/eval --id <sessionId>	Grade a specific session
/eval --json	Machine-readable metrics + card
/eval-diff <beforeId> <afterId>	Regression comparison, noise-tolerant
/eval-diff … --json	Machine-readable trend report
/eval-history [N]	Grade the last N sessions in this workspace and show the trend
/eval-history … --json	Machine-readable history report
/eval-history --since 7d	Grade only sessions with activity inside the window (24h, 30d also parse)

--since filters by each session's latest activity, not its creation time: a session opened long ago but active an hour ago stays in the window; a freshly created but silent one drops out. With --since 7d the trend compares the first and last of the in-window set, and an empty window prints a clear message instead of a crash. Without the flag, behavior is exactly as before (last N sessions).

Use /eval-history --out report.md to save the terminal card as Markdown, or add --json to save the JSON payload. Relative paths resolve against the current session workspace (the process working directory when unavailable). Parent directories must already exist; file errors are reported without creating directories. Existing output files are overwritten.

/eval-history finds the sessions for you via the same sessionQuery seam /archive uses (scoped to the current workspace's cwd), grades each, and diffs the first against the last. /eval-diff remains the tool when you already know the two ids that matter.

Configuring grade bands

The thresholds are defaults, not laws. A gradeBands block under the plugin's config node (naturally per-profile) overrides any dimension — values are the A–D upper bounds, ascending:

session-eval:
  gradeBands:
    reliability: [4, 8, 15, 30]   # tool error-rate percent
    reask: [0, 2, 4, 8]           # corrections per 10 user turns
    toolLoad: [0.1, 0.5, 8, 20]   # tool calls per turn: [lowC, lowB, highC, highD]

Any subset may be overridden; unlisted dimensions keep the documented defaults. Unknown keys, wrong arity, non-numeric, negative, or non-ascending values fail fast at plugin load with a readable error — sessions are never graded against bands nobody chose.

Typical use: grade the ten sessions before and after a plugin or prompt change; a persistent Reliability regression is an early warning that the change hurt real work, not just benchmarks.

Dimensions & thresholds

All grades derive from three deterministic metrics. n/a when the session lacks the relevant activity.

Dimension	Metric	A	B	C	D	E
Reliability	tool error rate	≤5%	≤10%	≤20%	≤35%	>35%
Re-ask	corrections per 10 turns¹	0	<1	<2	<4	≥4
Tool load	tool calls per turn	0.5–8	0.1–0.5 or just above 8	<0.1 or 8–20	>20	—

¹ A correction signal is each extra user message inside a consecutive-user run — the human re-asking before any assistant reply, a behavioral proxy for "the agent didn't get it right the first time."

Overall is the worst graded dimension — conservative by design. Thresholds are constants in src/grade.ts; making them configurable is on the roadmap (see issues).

Diff verdicts include a noise band (±2pp error rate, ±0.5 re-ask, ±1 tool load) so trivial jitter doesn't cry regression.

Determinism guarantee

Same session logs in → same grade card out, every time, forever. There is no model in the loop, no sampling, no clock dependence in the metrics (wall time is reported, never graded). This is what makes /eval-diff trustworthy as a regression signal.

Development

npm ci
npm run typecheck   # tsc --noEmit
npm test            # vitest run
npm run bundle      # tsdown → lib/

License

MIT

Community

• Contributing · Security policy · Code of Conduct