轨迹视图

Render the AI execution flow of a DeepSeek Harness Web GUI session as a replayable, horizontal timeline — every step's actions and every raw request sent to the model, laid out in front of you. Built for learning and tuning prompts: wondering why the AI answered that way? Open Trace View to see what it reasoned about, which tools it called, what results it got, and the actual message payload it saw at the time.

---

1. The problem it solves

Reconstructing "how did the AI actually execute this session?" with the built-in trajectory tab means flipping through a bare event list; and "what exactly was sent to the model for that request?" (system prompt, tool catalog, rebuilt messages) is invisible — the data lives in the session's internal state with no picture to look at.

This plugin adds a "Trace View" tab right next to the "轨迹" tab, rendering the same session state as an interactive, searchable, exportable horizontal timeline:

session trajectory (views.get("trajectory") snapshot — reuses official data as-is)
  │
  ├─ top minimap: time-ordered station blocks, click to jump
  ├─ horizontal timeline: real wall-clock labels, ◀ ▶ step-by-step replay
  ├─ bottom [Step Detail]: reasoning / tool args / full results / sub-calls / timings
  ├─ bottom [Raw Request]: the full context sent to the model for that step (prompt study)
  └─ header KPIs: turns / steps / tools / errors / duration / tokens + 📋 one-click Markdown

Read-only: it only re-renders the TrajectorySnapshot the session view has already produced; no file is modified, no network is touched, and the host half is a no-op stub.

2. Features

Flow overview

• ✅ Horizontal timeline + top minimap: one color block per station (user=blue / step=green / tool=orange / compaction=purple / error=red / retry=yellow / command=cyan); click a block to jump instantly; the current station is highlighted during replay.
• ✅ Full event stream: user messages, assistant steps, tool calls, context compaction, turn failures, model retries, max-output cutoffs and commands all land on the timeline in time order; compaction / command events are pinned to their owning turn by seq.
• ✅ KPI bar: turns / steps / tool count / errors / total tool time / request & generation tokens, a live "current step" indicator, and the active model name.
• ✅ Stations are labeled with real wall-clock time (not event indexes).

Finding what you want

• ✅ 🔍 Search: filter stations by tool name / args / body text / error name; hits get a gold outline plus a minimap highlight; Enter / Shift+Enter or ↑ ↓ step through hits; / or f focuses the search box, Esc clears.
• ✅ ⇤ / ⇥ turn jumps (PageUp / PageDown): hop between turn start points.
• ✅ ⇣ follow mode: while the AI is working, new stations auto-scroll into view; any manual navigation pauses following, ⏭ "show all" resumes it.
• ✅ Inspect deep-link contract: implements the session view's inspect protocol — if a tool-inspection request arrives while this tab is active, it locates the station containing that call.
• ✅ ⬆ load older history: one-click paging for events outside the long-session window (reuses the session's hasMore / loadOlder).

What each step did (bottom "Step Detail")

• ✅ Assistant output text and 💭 reasoning content (collapsible).
• ✅ Every tool call: args (formatted JSON), full tool result, a complete sub-call card (each sub-call's own args / result / duration / error), a duration bar, and the error code.
• ✅ Step-level metadata: ⏱ total step duration, time-to-first-token (TTFT), ⚡ decode throughput tok/s, token usage, and a "stopped" interruption flag.
• ✅ The detail panel can be ⤢ expanded to full height / ⤡ collapsed.

Raw request (bottom "Raw Request" tab, shortcut r)

• ✅ Model config: provider / model / temperature / reasoningEffort / thinking / maxTokens / stop / status / duration / retry info.
• ✅ ⚡ Prompt diff: what changed versus the previous request — system-prompt character delta and + added / - removed tool names (initial = the first system prompt).
• ✅ Full system prompt: character count + one-click copy (includes the tool-catalog JSON).
• ✅ Tool catalog: every tool schema carried by that request, expanded one by one; copyable.
• ✅ Message list: the model-visible messages approximately rebuilt from the session log (user / assistant / tool results, with compaction markers), each expandable in full; the whole request payload JSON is copyable. It is rebuilt only while this tab is open (lazy — long sessions stay snappy).
• ✅ Token usage: inbound (incl. cache reads / writes) and outbound (incl. thinking) breakdown.
• ✅ Context-compaction stations have their own "Raw Request" too: the compaction request's model / duration / usage plus the full compaction output.

Export study notes

• ✅ Header 📋 Markdown: one click copies the whole trace as Markdown — each step's reasoning / text / tool args / results and a per-step request-config summary, plus an appendix with the first request's system prompt and tool catalog.

3. Preview

Trace View overview (timeline + minimap + bottom detail / raw-request panels)

4. Directory layout

dsh-trajectory-traceview/   # repo root = npm package root
├── package.json            # dsh.bundle.patch + dsh.client (browser declaration) + exports["./client"]
├── cordis.patch.yml        # composition row: a single plugin record (no route, no config)
├── LICENSE                 # MIT
├── screenshot/             # screenshots for the README "Preview" section
├── test/
│   └── smoke.test.cjs      # zero-dependency helper unit tests (plain node)
└── lib/
    ├── index.js            # host half: no-op stub (browser-only plugin)
    └── client.js           # browser bundle: the trace view (timeline / minimap / detail / raw request)

5. Quick start

One-line install (GitHub):

dsh plugin --profile web add github:AFAP/dsh-trajectory-traceview

Then restart dsh web to take effect.

After installation the plugin lives at $DSH_HOME\profiles\web\node_modules\dsh-trajectory-traceview (cloned by pnpm from GitHub), independent of the source checkout location.

Upgrade:

dsh plugin --profile web update dsh-trajectory-traceview

Uninstall:

dsh plugin --profile web remove dsh-trajectory-traceview

Manual install from a source directory (for equivalent verification)

dsh plugin --profile web add "G:\path\to\dsh-trajectory-traceview"

Verify it loaded

Open any session with an AI execution trace → a "Trace View" tab appears next to the built-in trajectory tab → clicking it shows the horizontal timeline and minimap.

6. Usage

1. Open any session that has an AI execution trace.
2. Click the "Trace View" tab: the minimap at the top thumbnails every station; the timeline below is ordered by real wall-clock time.
3. Browse / replay: ◀ ▶ or the arrow keys step through stations; ▶ auto-replays (×1 / ×2 / ×4 speeds); Space toggles play / pause.
4. Search: press / to focus the search box, type a tool name / arg / body text / error name, and press Enter to hop between hits.
5. Select any assistant step: the bottom "Step Detail" shows its reasoning, text and every tool call's args / results / sub-calls / timing; switch to the "Raw Request" tab (r) to study the full context that request sent to the model.
6. Export notes: click the 📋 header button to copy the whole trace as Markdown.
7. Long sessions: click ⬆ to load older history; switch on ⇣ follow mode while the AI is working.

7. Shortcuts

Key	Action
← / →	previous / next station
Home / End	jump to the start / end
Space	play / pause replay
PageUp / PageDown	previous / next turn start
/ or f	focus the search box
↑ / ↓ (searching)	previous / next search hit
Enter / Shift+Enter (searching)	next / previous hit
Esc	clear search / dismiss
r	toggle the "Raw Request" panel

8. Implementation notes (data sources)

Concern	Approach
Data source	Reuses ONLY the views.get("trajectory") snapshot (TrajectorySnapshot) of the DSH Web session state: eventNodes (with timing / usage / provenance / interrupted), requests (assistant + compaction requests), callSchemas (tool schemas at call time), partial / runningCalls (live streaming state), plus the session-level hasMore / loadOlder paging.
Raw request	The "message list" is an approximately rebuilt payload from the session log (the original messages array is not snapshotted); the panel marks it "approximate" — precise enough for prompt study.
Laziness	The message list is rebuilt only when the "Raw Request" tab is opened — long sessions stay responsive.
Mounting	Registers the new tab beside the session view's built-in trajectory tab via dsh.client.inject (dsh-client-runtime / dsh-client-locale / dsh-client-ui-conversation).
Deep link	Implements the session view's inspect protocol: when a tool-inspection request arrives, it locates the station containing that call.

9. Troubleshooting

Symptom	Where to look
"Trace View" tab missing	Confirm dsh web was restarted; search F12 Console for dsh-trajectory-traceview.
Timeline empty / no stations	The session needs an AI execution trace (user message → assistant steps); brand-new sessions show the empty-state hint.
Search finds nothing	The search term must match a tool name / arg / body text / error name.
Message list differs from the real request	The list is an approximate rebuild (the original messages array is not snapshotted) — the panel marks this; treat the system prompt and tool catalog as authoritative.
Old sessions feel slow	Only opening the "Raw Request" tab rebuilds the message list; plain timeline browsing is unaffected.

10. Security & compliance

• Read-only: only renders the TrajectorySnapshot the session view already produced; no business-DOM mutation, no file modification or deletion, no host route, no network access.
• Local data: everything is processed inside the browser; the raw request is rebuilt only when you open its panel, and nothing is written to storage.
• No persistence: nothing written to localStorage / cookies; uninstalling leaves no trace.

11. Development & build

Plain JS, no build step. lib/index.js (host half) is a no-op stub; lib/client.js is a classic script (window.__ModuleLoader__.load) served directly by the client module system.

node test/smoke.test.cjs

The tests run the helper region against mock snapshots covering: tree building / request attachment / compaction-request pairing / sub-call detail / prompt diff / step timing / search matching / Markdown export / token stats.

12. License

MIT © AFAP

---

Disclaimer: This plugin is a local learning / debugging tool. The authors are not responsible for any decision made based on trace data (including approximately rebuilt message payloads); please don't share session traces containing sensitive content on shared / multi-user screens.