tritonkit-autonomous-cruise

Define the cruise goal, budget, forbidden actions, and stop conditions up front, then let the agent work in small verifiable slices. Each slice needs acceptance criteria, focused validation, and a checkpoint commit so progress stays reversible and auditable.

An autonomous cruise is a structured long-running session where the agent converts ordinary uncertainty into explicit assumptions and keeps working without constant user input. It only stops for credentials, destructive operations, unauthorized pushes or releases, and unresolvable repeated failures.

The agent stops when it needs credentials, private data, or production access, when the next step is destructive or irreversible, or when unauthorized branch, tag, or release operations are required. It also stops when the same validation failure persists despite reasonable investigation.

The default limit is three parallel subagents per wave unless the user or environment sets a lower threshold. Subagents must have non-overlapping file scopes, and the controlling agent verifies their output against diffs and tests before accepting it.

The agent stops taking new feature slices, writes a wrap-up report to the cruise space, and runs the appropriate verification gates such as the local verify script or docs checks. It then records memory, makes a final wrap-up commit, and reports whether any push, tag, or release occurred.

No. Pushing, tagging, releasing, and updating Homebrew are forbidden unless the user explicitly authorizes them. The default final report states that none of these actions were performed.