Configuring dsh-crew
DeepSeek Harness creates workers from its configured agent roster when tier is flash or pro. The dispatch controls documented by the README are:
| Key |
Values or inputs |
Behaviour |
tier_policy |
a tier policy |
Clamps every dispatch at the tool layer |
escalate_on_failure |
enabled or disabled |
Retries a failed flash run once on pro |
effort |
off through max |
Sets the worker's reasoning effort |
cwd |
a working-directory path |
Groups an in-host session and applies workspace checks |
worker |
agy or grok |
Explicitly selects the corresponding external CLI |
The README does not specify defaults for these settings. DeepSeek Harness sessions use the Agent preset selected for each tier.
dsh-crew Commands
The MCP command dsh_run_worker accepts tier, effort, cwd and worker. A normal dispatch runs inside the DeepSeek Harness hub when that hub is reachable. The fallback is the standalone dsh-jsonrpc-agent runtime, using worker.cordis.yml. The agy and grok values make the selected external CLI run the task instead of a DeepSeek Harness worker.
How dsh-crew behaves
DeepSeek Harness exposes each in-host worker as a first-class session in the Web UI, grouped by working directory. The host integration reports dispatch count, current step, tool calls and token usage. The jobs board retains running and finished jobs, shows tier, effort, progress and tokens, and marks a job as an orphan ghost if it disappears during a hub restart.
Before spawning, dsh-crew rejects worker-to-worker chains beyond origin depth 3, cycles, and a second job targeting a workspace already held by another job. describe_image prefers deepseek-v4-flash-vision-exp when a key is available, then can use configured CLIs or an OpenAI-compatible provider. generate_image uses the same provider routes.
dsh-crew Requirements and limits
DeepSeek Harness was tested with 0.1.1-rc.1; the plugin release shown in the README is 0.1.0-rc.11. Vision and image generation depend on an available key, CLI or configured provider. DSH models are text-only, so multimodal work uses the bridge rather than the worker model directly. The plugin can also run configured local command templates, which is the relevant terminal permission surface.
Written from the project's own documentation and kept in sync with it. Where the two disagree, the source is authoritative — read the README on GitHub