Deepseek Harness

DeepSeek Harness

Work against the exact Harness version the user owns. The bundled references were reviewed against a pre-stable release whose commands, package contracts, and persisted formats can change incompatibly; runtime help and a matching checkout or release are more authoritative than remembered examples.

Select the lane

• Operate or configure: install or launch Harness, choose a profile, inspect patches, configure providers, manage plugins, or troubleshoot runtime state. Load operation and configuration.
• Automate: drive headless, SDK, or ACP surfaces from another process. Load operation and configuration, then add the implementation skill for the caller language when code changes.
• Extend: author an out-of-tree Cordis plugin, tool, model adapter, capability, settings surface, or installable Bundle. Load extension and development.
• Maintain a source checkout: change Harness packages, applications, tests, or documentation only when the user explicitly targets that checkout. Load extension and development and follow the checkout's complete AGENTS.md hierarchy and exposed project skills.

Generic DeepSeek API calls do not select this skill. Use prompt-engineering as primary when prompt behavior is the deliverable, security when the user requests a security assessment, project-platform-diagnose while a platform-dependent failure remains unexplained, and project-release-maintainer when publishing or release machinery is the main job.

Ground the target

1. Identify whether the target is an installed CLI, source checkout, out-of-tree plugin project, or SDK/ACP caller.
2. Resolve the exact installed or checkout version before choosing commands, config fields, imports, or package examples. Use profile-free version help when available; for source work, inspect package.json, the applicable instructions, and version-matched documentation.
3. Prefer target-local source, runtime help, generated catalogs, and package READMEs over the snapshot bundled with this skill. If the executable or matching sources are unavailable, report the gap instead of installing or upgrading them without authority.
4. Name the Harness home, profile, and workspace explicitly before any action that can create state, load code, call a model, or change files. Distinguish persistent Harness state from the workspace the agent may inspect or edit.

Execute the smallest safe workflow

1. State the intended outcome and side effects. Classify package downloads, profile writes, patch changes, credential access, model/API use, network listeners, workspace mutation, and host-code execution before proceeding.
2. Choose one application surface. Use Web for interactive work, headless for one process-owned task, an SDK for application-controlled JSON-RPC, and ACP only for a trusted controller. Do not silently substitute one lifecycle for another.
3. Inspect before changing. Read existing profile and patch files directly. Use a configuration dump only after accounting for its ability to initialize missing profile state. Preserve the documented layer order and remember that a later patch replaces a row's complete config, not individual keys.
4. Prefer composition over core edits. Add behavior at a documented Cordis service, event, tool, or other extension point. Default durable third-party work to an out-of-tree plugin Bundle; do not treat the upstream internal package cookbook as an external scaffold.
5. Make only authorized mutations. Let the Harness CLI maintain Profile manifests. Review and pin third-party code before an install that may fetch packages or execute build scripts. Preserve unrelated workspace and Harness home state.
6. Verify through the real boundary. Inspect the resulting files and effective composition, boot the real Loader or application when permitted, and run the target repository's smallest relevant checks. A successful process exit, protocol completion, or agent completed state does not prove that the requested work is correct.
7. Report evidence precisely. Separate commands actually run, observed results, static source conclusions, skipped credentialed checks, and residual version or environment risk.

Protect trust boundaries

• Never print or commit credentials, tokenized Web startup URLs, authenticated state, or secret-bearing config. Treat logs containing a startup token as credential material.
• Treat external plugins, package lifecycle scripts, dynamic Cordis code, and MCP server commands as trusted code executing on the host, outside the agent's ordinary workspace sandbox.
• Do not describe the base write policy as read or network confinement. Use a disposable checkout or container when untrusted work needs stronger isolation.
• The minimal SDK composition has no ordinary approval or managed-credential services and grants unrestricted process access. Use it only with an explicitly accepted, disposable workspace.
• Confirm telemetry behavior before sensitive work. Feedback-triggered export can include messages, tool arguments and results, and workspace paths.
• Profile management and even configuration inspection can create persistent files. Record partial state after failure and either repair it deliberately or give the exact cleanup path; do not delete an existing Harness home speculatively.

Complete the task

Report:

• resolved Harness version and evidence source;
• selected lane, application surface, Harness home, profile, and workspace;
• persistent changes, installed code, credentials or network dependencies, and other side effects;
• exact validation performed and what it proves;
• unavailable runtime, model, platform, or credential checks;
• remaining cleanup, restart, compatibility, or security risk.

For maintainer provenance, prompt-routing fixtures, and evidence terminology, load coverage and validation.