dsh-autopilot

Most agent workflows end with a confident "Done." A few files changed, a command run, and the model declares the task finished — yet behind that sentence may hide a broken build, a wrong status code, a missed requirement, or a leftover TODO. dsh-autopilot intercepts exactly that moment: it does not trust the model self-assessment and only issues a VERIFIED COMPLETE certificate after every acceptance criterion has independently passed.

The core mechanism is a single gate of criteria → independent verification → release. You declare mandatory checks in a project-level .autopilot.yml. Autopilot hooks into the official agent/turn-stopping lifecycle and uses deterministic evaluators (command exit codes, file content, regex, git workspace state, HTTP responses, manual confirmation, and an independent agent judge) to validate each item. Fixable failures produce structured feedback and the current turn continues via agent.steer. Regressions, repeated no-progress, and oscillation are flagged automatically and trigger a one-time policy reset before escalating to BLOCKED. On the security front, the config is treated as untrusted input, HTTP defaults to localhost-only, cloud metadata addresses are blocked, and evaluators carry no workspace write permission.

It is built for individuals and teams running agent workflows on DeepSeek Harness. If your pain point is the agent insisting all tests passed while the exit code says 1, or you need completion to shift from probabilistic to auditable and deterministic, Autopilot is that gate.