capto-dsh-plugin
DeepSeek Harness (dsh) 插件：通过其 CLI 控制平面驱动仅限本地使用的 Capto 屏幕录制器的一等 capto_* 工具
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add capto-dsh-plugin@0.1.0Capto Dsh Plugin 以 capto-dsh-plugin 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
First-class Capto screen-recording tools for DeepSeek Harness — Everything is a Plugin.
capto-dsh-plugin registers a set of typed capto_* tools (capto_status, capto_record_start, capto_shot, …) so a Harness agent can drive the local-only Capto Windows screen recorder from any conversation — no shelling out, no FFmpeg, no cloud.
Error: capto exited <code> (<errorCode>): <message>; desktopUnavailable (exit 2) tells the model exactly what to do next (capto_open, wait, retry)capto CLI and adds cli\ to PATHdsh web, dsh --profile headless, …)npm install --prefix ~/.dsh/profiles/web capto-dsh-plugin
Enable it in the profile's cordis.patch.yml:
- insert:
- id: capto
name: capto-dsh-plugin
config:
command: ['capto'] # or an absolute path to capto.exe
timeoutMs: 120000
noLaunch: false
autoOpen: false
Verify the composed tree, then restart the GUI:
dsh --profile web --dump-config
# restart `dsh web` — the capto_* tools appear in the agent's toolset
The package declares a dsh.bundle manifest, so the official pnpm-based flow wires everything automatically:
dsh plugin --profile web add capto-dsh-plugin
dsh --profile web --dump-config # a "# == capto-dsh-plugin" layer appears
The npm-based flow above is the equivalent for npm-managed profiles.
npm install --prefix ~/.dsh/profiles/web "file:D:/path/to/Capto/packages/capto-dsh-plugin"
npm ≥ 7 links file: dependencies, so source edits are picked up on the next GUI restart.
Uninstall: npm uninstall --prefix ~/.dsh/profiles/web capto-dsh-plugin (or dsh plugin --profile web remove capto-dsh-plugin), then drop the patch row.
| Key | Default | Meaning |
|---|---|---|
command | ['capto'] | CLI argv prefix: ['capto'] (installed PATH), ['D:\\…\\capto.exe'], or ['cargo','run','-p','capto-cli','--'] in a dev checkout |
timeoutMs | 120000 | Per-call timeout in ms. Must exceed the CLI's 45 s desktop auto-launch wait so a cold start can finish |
noLaunch | false | Always pass --no-launch — never auto-start the Capto desktop |
autoOpen | false | On exit 2 (desktopUnavailable): run capto open, wait ~3 s, retry once |
Invalid configuration (empty command, non-positive timeoutMs, non-boolean flags) fails at plugin load with an actionable error.
| Tool | Purpose |
|---|---|
capto_status | Session snapshot { state, elapsedMs, outputPath, … } — check before recording |
capto_doctor | Environment readiness { ffmpegOk, controlPlane, … } |
capto_open | Open the desktop window (recovery entry for exit 2) |
capto_list | Enumerate displays / windows / audio / encoders |
capto_shot | Screenshot → { path } (absolute PNG) |
capto_record_start | Start recording (source, display/window/region, format, fps, quality, encoder, mic, loopback, cursor) |
capto_record_stop / capto_record_pause / capto_record_resume | Recording control |
capto_config_get / capto_config_set / capto_config_path | Read / patch settings (camelCase keys) |
capto_outputs_recent / capto_outputs_open | Recent outputs / open files in Explorer |
Every tool declares a canonical JSON output and a pretty renderer; read-only tools are marked concurrency-safe. The tool:capto prompt section teaches the agent: check capto_status before starting, never start twice, recover exit 2 with capto_open, always end with capto_record_stop, and find files with capto_outputs_recent.
tool:capto guidance section: a fixed, small per-request token cost while the plugin is active.Error: … carrying the CLI exit code and error code. desktopUnavailable messages give the model its next step (capto_open, wait ~3–5 s, retry; ask the user if it still fails).AbortError and never leaves an orphan CLI child.cd packages/capto-dsh-plugin
npm install # test deps (@deepseek-ai/dsh-tools, schemastery)
node test/smoke.mjs # 11 checks: contract, config, arg mapping, failure
# normalization, timeout, autoOpen recovery, real CLI
test/fixtures/fake-capto.mjs is a fake capto CLI speaking the JSON envelope contract, so the suite runs without the desktop. When target/debug/capto.exe exists, one real-CLI check runs too (accepting both control-plane states). CI runs the suite against the freshly built CLI on Windows and fake-only on Ubuntu, plus npm pack --dry-run.
capto_shot returns a path; image rendering in tool results is not available in the harness yet.autoOpen is best-effort — it retries once after capto open; a wedged desktop (stale process) still needs a manual kill + relaunch per the Capto skill docs.