delegate-to-deepseek-harness

简体中文

Codex keeps the judgment. DeepSeek Harness does the bounded work.

Codex              = orchestrator + reviewer
DeepSeek Harness   = worker

User → Codex ── bounded work ──→ DeepSeek Harness worker
                                          │
                              changes + tests
                                          ↓
                                  Codex reviews ✓

delegate-to-deepseek-harness is a small Agent Skill for handing well-scoped coding work from Codex to DeepSeek Harness, then bringing the result back for review and verification.

It is deliberately not a framework. No task queues, no dashboards, no extra config format — just three questions, answered honestly: should this be delegated, what exactly should the worker do, and how do you check the result before you believe it.

The worker itself can run through a DeepSeek Harness MCP bridge or the official headless CLI. If one path breaks tomorrow, the workflow doesn't.

Try it

git clone https://github.com/LomoMao/delegate-to-deepseek-harness.git
cd delegate-to-deepseek-harness
./scripts/install_skill.sh   # installs to $HOME/.agents/skills

Then in Codex:

Use $delegate-to-deepseek-harness to fix the failing parser tests.
Keep the public API unchanged, then review the diff and rerun the focused tests yourself.

Why

Sometimes the expensive part of a coding task is not the final decision — it is reading a large repo, making repetitive edits, repairing tests, or implementing a clearly specified slice of work.

This skill keeps that split simple:

You → Codex → DeepSeek Harness → workspace changes
              ↓
         Codex reviews
         diff + tests

Codex stays responsible for scope, risky judgment, integration, and the final answer.

Methodology first, backend-agnostic. The reusable part of this skill is the delegation contract and the review gate, not any particular worker path. If the MCP bridge changes or disappears, the headless CLI fallback (or any future Harness interface) keeps the same workflow intact.

Good fits

  • bounded implementation work
  • mechanical refactors and migrations
  • test repair and test generation
  • repository exploration with a concrete question
  • independent second attempts
  • parallel work in separate worktrees

Not a good fit for architecture decisions, secrets, deployments, destructive operations, or work with no meaningful verification path.

Other ways to install

Codex loads user skills from $HOME/.agents/skills (repo-local skills live in .agents/skills).

Manual copy: copy SKILL.md, agents/, references/, and scripts/ into $HOME/.agents/skills/delegate-to-deepseek-harness/.

Codex's built-in skill installer:

$skill-installer install https://github.com/LomoMao/delegate-to-deepseek-harness

Worker backends

Preferred: a local DeepSeek Harness MCP server exposing agent_run, task_inbox, and task_result.

Fallback: the official one-shot CLI:

dsh --profile headless "run the tests"

See setup for the current MCP example and the safer defaults.

One rule that matters

A worker saying “done” is not evidence that the task is done.

After delegation, Codex verifies the worker's result against a machine-checked verification contract (changed files in scope, no new dependencies, tests pass) via scripts/verify_workspace.sh, then reviews by risk level — and stops when the contract passes.

Delegate the work. Verify the evidence. Don't redo the work.

Why that last sentence exists, from one real delegation run:

planning         8%
worker          12%
manager review  80%   ← the worker had already passed 16/16 tests

The expensive part wasn't delegation. It was verification without a stopping rule.

Status

Small on purpose, and staying that way. The core workflow — contract, delegation, verification — is stable. The DeepSeek Harness ecosystem around it moves fast, so MCP tool names are pinned to @chushixixin/dsh-harness-mcp-server v0.1.x in setup and may need a quick re-check after upgrades.

Issues and small PRs are welcome.

Security

A Harness worker may be able to edit files and run shell commands. Treat third-party Harness plugins as trusted local code, keep MCP endpoints on loopback, and scope writable workspaces as narrowly as practical. See SECURITY.md.

Acknowledgements

Inspired by the broader agent-delegation pattern used by projects such as delegate-skills, delegate-to-pi, and other reviewer/worker workflows.

This project is independent and is not affiliated with OpenAI or DeepSeek.

License

MIT