dongsheng123132/dshx

dshx은(는) 무엇을 하나요?

Use dshx as the machine boundary around the upstream dsh --profile headless runtime. Keep the task bounded and make the workspace explicit.

Readiness

Run:

dshx doctor --json

Require node_supported, dsh_available, and api_key_present to be true before delegating paid/model work. Never print credential values.

If dshx is missing, explain the pinned install command and run it only with user authorization:

npm install -g @hfshfg/[email protected]

Delegate one task

Prefer stdin so task text does not depend on shell quoting:

dshx run --stdin --cwd ./workspace --timeout 1800000 --json

Pipe or redirect the task into stdin. For instructions longer than 24,000 characters, write the source material inside the authorized workspace and send a short task that references that file.

Choose a timeout proportional to the task. Keep the default 30 minutes for normal coding work; use a shorter bound for diagnosis. Do not set an unbounded timeout unless the user explicitly asks for persistent execution.

Interpret results

• ok: true, reason: completed: return or summarize text and report duration_ms.
• reason: failed: report the upstream exit code and concise diagnostics.
• reason: timeout: explain that the process tree was terminated at the declared bound; do not claim the task completed.
• reason: output_limit: narrow the task or ask DSH to write a result file in the workspace.
• Exit code 2 means local parameter/input error; fix the invocation before retrying.

Treat stdout as the result channel and stderr as diagnostics. Do not parse ANSI or spinner output; dshx removes it from machine results.

Boundaries

• Delegate only work inside the user-authorized workspace.
• Do not include API keys or other secrets in task text, argv, or output summaries.
• Do not use dsh web for agent-to-agent delegation; it is the human-facing browser UI.
• Review DSH changes and test evidence before presenting them as complete.
• Remember that DeepSeek Harness is Developer Preview and dshx 0.2.1 pins @deepseek-ai/dsh 0.1.0-rc.6.