DSH Plugin: Trajectory Debug Workbench (trajectory-debug)

English | 中文

Target host: DeepSeek Harness (@deepseek-ai/dsh v0.1.0-rc.x, developer preview)

Turn DSH's event-sourced sessions into debuggable assets: a waterfall trajectory view, deterministic single-step replay (zero token cost), breakpoints, sandboxed edit-and-rerun, fork comparison, and performance analytics — plus OTel GenAI trace export and trajectory_* model tools for agents to introspect their own runs.

Packages

packages/
├── trajectory-debug/              # Service Definition + wire types (pure types, no plugin entry)
├── trajectory-debug-host/         # Host provider: replay/perf/compare engines, breakpoints, projections, commands, RPC transport
├── trajectory-debug-remotes/      # dual-face typert skeleton (browser RPC actually rides the host transport)
├── client-ui-trajectory-debug/    # browser "Debug" tab: waterfall + perf dashboard + replay/breakpoint/rerun/compare console
└── trajectory-debug-bundle/       # installable bundle: dsh.bundle + cordis.patch.yml

Commands

corepack pnpm install       # requires corepack; Node >= 22.19
corepack pnpm -r build      # emit lib/ (lib-first manifests)
corepack pnpm -r typecheck
corepack pnpm test          # 56 vitest cases
corepack pnpm check         # build + typecheck + test
corepack pnpm check:publish # pre-publish manifest validation
node scripts/smoke.mjs      # real dsh process load smoke (rebuilds the td-smoke profile)

Feature Status

Capability	Notes
Waterfall engine	buildTrajectoryPage: turn/step/tool rows, status, timing, tokens, error codes, filters, paging
Deterministic replay	stepContextAt: model view + action per step; ReplayCursor step/seek — zero token, zero tool execution
Performance analytics	analyzePerf: success rates/percentiles, failure taxonomy, token distribution, TTFT/decode, turn stats (optional price table → cost)
Fork compare	compareTrajectories: step alignment, tool changes, result diffs, summary
Breakpoints	BreakpointManager: agent/pre-step waterfall short-circuit + timeout auto-resume
Edit-and-rerun	rerunTool through the full ctx.tools pipeline; record|sandbox|ask (ask fails closed without a live agent)
Fork + live resume	sessions.fork + agents.resume + followup; cascade: truncate|preserve
Projections	trajectoryDebug/trajectory + trajectoryDebug/perf on the session-projection registry — the browser consumes them with zero client folding
Commands	/trajectory [stepIndex], /perf
Model tools	trajectory_search / trajectory_step / trajectory_perf (opt-in via enableModelTools)
Trace export	export('trace'): OTel GenAI semantic-convention spans (Langfuse/LangSmith ready)
Sidecar persistence	FileSidecar atomic JSON writes (sidecar: 'memory'|'file')
Browser UI	"Debug" conversation view tab: waterfall + perf dashboard (projection-pushed, live) + replay/breakpoint/rerun/compare console
Browser RPC	host registers POST /api/trajectory-debug/rpc (webserver custom route, typert-independent); the browser calls it with fetch
Real load	scripts/smoke.mjs installs into a standalone profile and boots in a real dsh process

Browser RPC transport (typert-free)

DSH's typert Remote chain depends on build-time codegen. This plugin instead uses the sanctioned webserver custom-route extension point:

• Endpoint: POST /api/trajectory-debug/rpc — body { method, params }, response { ok, value | error };
• Methods: trajectory.list / step.context / perf / replay.start|step|seek / breakpoint.set|remove|list|resume / intervention.rerunTool / variant.fork|list|compare / export;
• Security: the server binds loopback by default; the route adds no extra trust.

Client bundle build

pnpm -r build && node scripts/bundle-client.mjs produces the browser bundle:

• entry src/client/index.ts → esbuild (CJS) → wrapped as window.__ModuleLoader__.load({ id, factory }) (the official DSH client format);
• runtime externals: only react (shell seed); all dsh references are type-only (erased);
• output packages/client-ui-trajectory-debug/lib/client.js, served via exports["./client"];
• after installing into a profile, sync the bundle into the profile copy (pnpm file: deps are copies): robocopy packages\client-ui-trajectory-debug\lib <profile>\node_modules\dsh-client-ui-trajectory-debug\lib /MIR

Install into DSH

Published on npm — install directly (no build required):

dsh plugin --profile web add dsh-trajectory-debug-bundle
dsh web --dump-config   # expect trajectory-debug-host / -remotes / ui-trajectory-debug rows

From a source checkout:

corepack pnpm check
node scripts\smoke.mjs
dsh plugin --profile web add ./packages/trajectory-debug-bundle

Restart dsh web: the Debug tab appears in the conversation view ring; /trajectory and /perf work in the input box.

Publish status

All five packages are published to npm as v0.1.0: dsh-trajectory-debug, dsh-trajectory-debug-host, dsh-trajectory-debug-remotes, dsh-client-ui-trajectory-debug, dsh-trajectory-debug-bundle.

Release pipeline:

corepack pnpm check            # gates
corepack pnpm check:publish    # manifest validation (no file: deps, valid versions, bundle ships patch+index.js)
corepack pnpm publish:all      # pnpm -r publish: workspace:* → version ranges; prepublishOnly builds first

• Internal deps use workspace:* (pnpm rewrites to version ranges on publish); file: deps are rejected by the validator;
• Add the dsh-plugin topic to the repository (auto-listed on deepseekdocs.com/ecosystem); curated listing via docs/awesome-submission.md.

Ecosystem Comparison

See COMPARISON.md: positioning vs dsh-message-edit / dsh-plugin-cost / dsh-deeplink / dsh-eval, and the improvements already shipped.

Design Notes

• Self-contained event model: engines fold a minimal DebugEvent model; adapt.ts is the only boundary touching DSH SessionEvent — immune to DSH's preview-period breaking changes;
• Projection state is plain JSON (records/arrays, no Maps/Sets), satisfying the projection-cache persistence contract;
• "Model-visible == recorded" invariant: branch execution only ever uses the existing agent channels; the source session log is never rewritten;
• Replay is zero-cost: engines never call the LLM or tools;
• Self-built client bundle: no dependency on DSH's tsdown chain.

Compatibility

• Deps: @deepseek-ai/cordis + dsh-session / dsh-agent / dsh-commands / dsh-session-projection (all ^0.1.0-rc.6, tracking rc releases);
• Unloading the plugin withdraws every registration (effect-owned); source sessions are untouched.