dsh-pilot

Drive DeepSeek Harness as a job: hand it a brief, let it work, and get an exit code that distinguishes the agent finished from the wait gave up — because every wrong call made against these harnesses came from an instrument that could not tell those apart.

Used alongside pi-pilot. Both harnesses load the workspace's AGENTS.md, so a brief describes the task and the repository carries the standard.

What is here

SKILL.md	the operating knowledge — what costs time, what a negative result does not prove
pilot.py	the driver: one task, start to finish, over ACP

Requirements

• Node ≥ 22. DSH uses createZstdDecompress and Promise.withResolvers. On Node 20 it dies during plugin load, and a driver that only watches for a reply will sit out its whole timeout waiting for a process that is already gone. The driver resolves a suitable binary itself and refuses early when it cannot.
• A DSH checkout (--repo, default /home/team/workspaces/deepseek-harness).
• Credentials for the provider the profile resolves — for AtlasCloud, ATLASCLOUD_API_KEY.

Use

python3 pilot.py --cwd <worktree> --brief <file> [--out <log>] \
  [--model <id>] [--provider <id>] [--permission-mode danger-full-access] \
  [--precheck 'npm test'] [--node /path/to/node22+] [--timeout 3600]

Exit codes are the point:

code	meaning
0	the turn settled and the work landed as a commit
1	it ended without settling, or nothing landed
2	the invocation itself was wrong
3	refused before starting — no Node ≥ 22, precheck failed, or the harness died at startup
124	no settlement inside the timeout

Two things that will cost you a run if you do not know them

git commit needs --permission-mode danger-full-access. Under the default workspace-write preset the escalation a commit requires is cancelled without ever reaching the client — no permission request arrives, so there is nothing to approve. The turn then ends normally and the worktree is quietly dirty. Measured on one brief and one model, changing only the preset: workspace-write → no commit; danger-full-access → commit landed.

It is not the default here because it turns approval prompts off on a host with no isolation.

session/prompt does not reliably return a stop reason. It returned end_turn on a read-only turn and nothing at all on a turn that used tools. The driver also reads the session/update stream and reports absent rather than inventing one.

Why the driver checks that a commit exists

"It finished" and "it produced something" are different questions, and a driver that answers only the first is the blind instrument this whole skill exists to avoid. --no-require-commit opts out when a run is genuinely read-only.