# dsh-ui-tool-graph

A **tool-call value graph** tab for the [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) (dsh) Web UI.

Agent trajectories bury their own economics: which tool ate the wall time, where calls failed, how many tokens each round burned, which calls repeated themselves verbatim. This plugin adds a **Tool Graph (工具图谱)** tab to the conversation view ring that folds the session log into one weighted relationship graph:

- **Nodes are tools**, sized by the metric you pick (total duration, call count, errors) with an outer ring proportional to the error rate;
- **Edges are transitions**, one per consecutive dispatch pair within a turn, weighted by transition count;
- **Insights** call out the dominant bottleneck, error-prone tools, identical repeat calls, self-loops worth batching, the longest inter-call wait, and the step that produced the most output tokens;
- **Next-turn guidance** renders those metrics into a self-contained analysis prompt you can copy to the clipboard or queue as your literal next message.

Everything is folded client-side from the durable session log — deterministic, replay-safe, no extra model calls, no persistence, zero token cost.

![Tool graph in a real session](assets/tool-graph.png)

## Features

- **Weighted radial graph** — switch node weighting between total duration / call count / errors at any time; edge width ∝ transition count.
- **Optimization insights** — rule-derived findings with severity dots: bottleneck share, error-prone tools, verbatim repeats (cache/dedup candidates), consecutive same-tool chains, model-generation waits, token-heavy steps.
- **Per-tool ranking table** — total / average duration, calls, errors; click a row or node to drill into every individual call with its duration and argument excerpt.
- **Session totals strip** — calls, running, failures, tool wall time, longest average gap, output tokens.
- **One-click next turn** — copy the analysis prompt, or send it as your next queued message after an explicit second click.

## Installation

Tarball from a release (no build-script permission needed):

```sh
dsh plugin --profile web add https://github.com/rocklau/dsh-ui-tool-graph/releases/download/v0.1.2/dsh-ui-tool-graph-0.1.2.tgz
```

Or from source:

```sh
dsh plugin --profile web add github:rocklau/dsh-ui-tool-graph
```

A git install fetches sources, so pnpm runs this package's `prepare` script on first `add`; follow the CLI's prompt to allow the build in the profile's `pnpm-workspace.yaml`, then re-run the command. Restart `dsh web` afterwards — the **工具图谱** tab appears in every session header.

Uninstall: `dsh plugin --profile web remove dsh-ui-tool-graph`.

## How it works

| Metric | Definition |
| --- | --- |
| Call duration | `tool/result` time − paired `tool/call` time (wall clock) |
| Tool wall time | Sum over completed calls in the loaded window |
| Error rate | Error results ÷ completed calls per tool |
| Transition edge | Consecutive call pair within one turn, in log order |
| Average gap | Per-edge mean of result→next-start latency (dominated by model generation wait) |
| Tokens | Provider usage carried by each finalized assistant message, deduplicated per step |

The view target folds incrementally (`replace`/`apply` over the event window), so totals cover exactly what the history window holds; page older turns in the Trajectory tab to extend coverage.

## Known limitations

- Nested sub-calls (delegations) fold as their parent result only.
- The LLM-vs-tool wall-time split is approximated by inter-call gaps; per-request timing attribution is not yet surfaced.
- Single-tool sessions render one node without edges — by definition.

## Development

```sh
pnpm install
pnpm typecheck   # tsc --noEmit (strict)
pnpm test        # vitest: folding parity, layout purity, prompts, component behavior
pnpm build       # lib/index.js (node half) + lib/client.js (browser bundle)
pnpm pack        # installable tarball
```

The package is a dual-identity bundle: `dsh.bundle` ships the patch layer, `dsh.client` declares the web plugin. Runtime imports are limited to the host-shared module table (react/cordis/client-runtime); everything else is bundled.

## License

MIT