# Agent Operating Instructions

## Continuous milestone execution

When the user asks to continue, work autonomously through the active milestone
and its verification. Do **not** send progress-only, acknowledgement, empty, or
status final responses. Reply only when:

1. the user asks for status;
2. a real product, security, credentials, or destructive-action decision needs
   the user's input; or
3. the requested milestone is fully implemented and verified.

Use one foreground milestone batch where possible. If work must run in a
background subagent, avoid notifying the user manually; inspect and verify the
result before replying at the same completion boundary.

A chat turn still necessarily ends after a model response. This file prevents
unnecessary model-generated completion messages; it cannot suppress
harness-generated tool or subagent notifications.

## Verification

Before claiming a milestone boundary, run `make check` and record focused
acceptance evidence under `docs/release/`.