Context and Problem Statement

auto-dev keeps N issues in flight by dispatching a worker per issue, and today it dispatches them as one-shot claude -p processes: a worker that ends its turn is killed, cannot be messaged mid-flight, and a correction means re-dispatching a fresh session with a tail prompt. In-process sub-agents are addressable, resumable and observable from the supervisor. (Recorded from the owner’s rule of 2026-08-31, the 2.0 breaking change; the opposite is still written into commands/auto-dev-worker.md and skills/auto-dev/SKILL.md as of this record.)

Considered Options

• One-shot claude -p worker processes, as today.
• In-process background sub-agents the supervisor can message and resume.
• A mix: sub-agents for phase 1, claude -p for the merge phase.

Decision Outcome

Workers are in-process background sub-agents; auto-dev never shells out to claude -p, so a supervisor can message a worker mid-flight, a correction is a message rather than a re-dispatch, and a worker that stalls is a live agent to be asked rather than a dead process to be re-run.

Consequences

skills/auto-dev/SKILL.md and commands/auto-dev-worker.md were rewritten to the sub-agent shape by #314 (PR #328, the 2.0 breaking change): a worker is dispatched through the Agent tool with a model parameter, can be messaged with SendMessage, and its final message is its report. Neither file describes claude -p any more — tests/auto-dev-never-wait/test.sh pins the worker prompt — which is what moved this record from proposed to accepted (2026-09-02). The cost: the trigger bench evals/run_all.py, which spawns real claude -p processes to measure descriptions, can never run inside a fleet and stays an owner-run step.

A further consequence surfaces when a worker’s phase has ended and it has returned its final report: a retired worker cannot be messaged any more, so its in-process nature is complete — an agent that has already returned is neither addressable nor resumable. When a worker holds a checked-out PR-branch worktree and cannot be asked to release it, the supervisor must call skills/auto-dev/scripts/release-branch.sh directly against the retired worker’s tree instead, as #510 established. This is not a new decision, only a discovered corollary of the in-process design, and it determines when and how PR branches are cleaned up at the end of a two-phase worker run.