[!NOTE] 📦 This plugin lives in the dsh-plugins monorepo — file issues & pull requests there. npm: @hy-sde-org/dsh-session-intelligence

dsh-session-intelligence — session health intelligence for DeepSeek Harness

English · 中文（子包 README）

Session health intelligence for DeepSeek Harness — port of agentsview's (MIT, Kenn Software) internal/signals engine: outcome classification (completed/abandoned/errored/unknown), penalty-based A–F health grades with basis + penalties, tool-health signals, prompt/workflow heuristics, and context-pressure signals. Powers the session_health tool and a standalone CLI. See packages/session-intelligence/README.md for the engine API, tool argument tables, and the CLI reference.

Identity	Value
Package	@hy-sde-org/dsh-session-intelligence
Plugin row id	hy-sde-tool-session-intelligence (host row)
Tools	session_health (action: "recent" / "session"), session_recent_edits, session_insights
CLI	dsh-session-intelligence (bin) over the same engine

Based on agentsview (MIT, Kenn Software) — the signals engine (internal/signals/* → src/engine/*) and the DSH event mapping (internal/sync/signal_compute.go → src/adapter/dsh.ts) are ports of agentsview's Go implementation; session_insights was merged unchanged from the superseded @hy-sde-org/dsh-tool-session-insights; the CLI's projectKey path encoding ports @deepseek-ai/dsh-session-persistence-jsonl. See THIRD-PARTY-NOTICES.md.

Why

How did past sessions in this workspace actually go? Without signals, the answer is a guess. session_health classifies each session's outcome (completed / abandoned / errored / unknown with confidence) and grades it A–F with the basis categories and per-signal penalties behind the grade — tool failures/retries/edit churn, prompt-quality heuristics (short prompts, unstructured starts, missing success criteria, duplicate prompts, runaway tool loops), and context pressure (compactions, mid-task compactions, peak tokens vs the model's window). Check it before repeating a failed approach, after a crash to see how far the session got, or to audit which tools burn errors (session_insights) and which files past sessions touched (session_recent_edits).

Prerequisites

• Node.js 22.19 or newer with npm and pnpm on PATH;
• a DeepSeek Harness release carrying the 0.2.0-rc.2 peer range — @deepseek-ai/cordis ~4.0.4 and @deepseek-ai/dsh-llm, @deepseek-ai/dsh-session-query, @deepseek-ai/dsh-system-prompt, @deepseek-ai/dsh-timeout, @deepseek-ai/dsh-tools ^0.2.0-rc.2 — including the standard dsh CLI;
• the host services tools, systemPrompt, and sessionQuery mounted — the stock base bundle does (the row consumes them; it needs no realm, no filesystem access, and no persistence of its own);
• no API keys — nothing else.

Install the Harness CLI and pnpm before continuing:

npm install --global @deepseek-ai/dsh pnpm
dsh --version

Quick start

Route A — published npm package (recommended)

dsh plugin --profile web add @hy-sde-org/dsh-session-intelligence

Route B — from source (validate this checkout or hack on the plugin)

git clone git@github.com:hy-sde/dsh-plugins.git
cd dsh-plugins
pnpm install
PACKAGE_TARBALL="$(cd dsh-session-intelligence/packages/session-intelligence && pnpm pack --pack-destination /tmp | tail -n 1)"
dsh plugin --profile web add "$PACKAGE_TARBALL"

pnpm pack runs the normal prepack build and produces a tarball containing dist/. A direct github:<this-repo> dependency does not contain built output and is not a supported install path — always install the built tarball (or the published package).

Verify the composed configuration

dsh web --dump-config

The composed tree must show the hy-sde-tool-session-intelligence row loading @hy-sde-org/dsh-session-intelligence.

Run

dsh web

Then, in the session (the tools are scoped to the caller workspace):

session_health { action: "recent", limit: 20 }                   # health table: outcome, grade, signals
session_health { action: "session", sessionId: "session-<id>" }  # full report behind a grade
session_recent_edits { path: "src", limit: 30 }                  # recently edited files, newest first
session_insights { action: "tool-frequency", top: 20 }           # tool ranking with error counts

action: "recent" returns a health table across the most-recent sessions; action: "session" returns the full report (basis categories, penalties, tool-health and heuristic details) for one sessionId. Sessions that cannot be read are counted and reported, never fatal.

The same engine ships as a standalone CLI (bin: dsh-session-intelligence, or node dist/cli.js from the package directory) — it reads session.jsonl.zstd (multi-frame zstd, torn-tail tolerant) or session.jsonl from the sessions directory and honors DSH_HOME (default ~/.dsh):

dsh-session-intelligence --cwd /Users/you/Documents/my-project --recent --limit 20
dsh-session-intelligence --cwd /Users/you/Documents/my-project --session session-<id>
dsh-session-intelligence --cwd /Users/you/Documents/my-project --edits --path src --limit 30
dsh-session-intelligence --cwd /Users/you/Documents/my-project --frequency --top 20

Uninstall

dsh plugin --profile web remove @hy-sde-org/dsh-session-intelligence

What the bundle does

The package declares a DSH bundle (dsh.bundle.patch → cordis.patch.yml), so dsh plugin installs it and applies its patch: one plain host row (hy-sde-tool-session-intelligence), like the official harness's own tool-session-query row. It consumes the deployment's host services — tools (stock base), systemPrompt (stock base), and sessionQuery (stock base mounts @deepseek-ai/dsh-session-query with the sqlite index backend) — so it needs no realm, no filesystem access, and no persistence of its own: the tools are scoped to the caller workspace by sessionQuery.filterSessions with a cwd clause, and unreadable sessions are counted and reported.

Each session's event stream (user/message, assistant/message, tool/call, tool/result, compaction/*) is reduced into: outcome + confidence, health score + grade + basis + penalties, tool-health signals, heuristic signals, compaction/mid-task counts, peak context tokens, the model (when recorded), and the context-pressure ratio. Inherited seeded prefixes are skipped, so subagent children ignore their parent's events.

Configuration

All options are optional; the row's config: fills the defaults.

Option	Default	Purpose
maxSessions	200	maximum sessions scanned
recentLimit	20	default table size for session_health { action: "recent" }
timeoutMs	60000	tool-call budget

- id: hy-sde-tool-session-intelligence
  name: '@hy-sde-org/dsh-session-intelligence'
  config:
    maxSessions: 200
    recentLimit: 20
    timeoutMs: 60000

Compatibility

Component	Supported contract
Node.js	22.19 or newer (engines.node >=22.19.0)
DeepSeek Harness	0.2.0-rc.2 peer range (@deepseek-ai/cordis ~4.0.4; dsh-llm, dsh-session-query, dsh-system-prompt, dsh-timeout, dsh-tools ^0.2.0-rc.2)
Host services	tools, systemPrompt, sessionQuery (the stock base bundle mounts all three)
Session artifacts	session.jsonl.zstd (multi-frame zstd via @hy-sde-org/dsh-zstd-frame) or session.jsonl (CLI path)

Upstream seam-contract changes require a new package release and contract review.

Development

pnpm install
pnpm -r check        # tsc --noEmit
pnpm -r test         # vitest run
pnpm -r build        # tsc -p tsconfig.build.json

The engine is importable with the same numbers the tool reports: analyzeSession({ id, createdAt, events }) from @hy-sde-org/dsh-session-intelligence (see the sub-package README for the pure entry points and the full tool-argument tables).

License and attribution

MIT — see LICENSE. Ports of agentsview (MIT, Kenn Software) and dsh-session-persistence-jsonl (MIT) are attributed in THIRD-PARTY-NOTICES.md; zstd artifacts are read via @hy-sde-org/dsh-zstd-frame.

This plugin is a separate installable package; the harness remains the property of its own project.