Coding Owl is archived. It did not fulfill its core loop, so rather than pivoting or refactoring it, I am building something new from scratch with the lessons learned. Why Coding Owl is archived

ADR-0017

Owl appends a standing unattended contract to every Run

Context

Interactive Claude Code asks when it is unsure. Unattended it cannot: Owl runs Agents with --permission-prompts none (ADR-0012), so anything that would prompt is denied silently, and there is nobody awake to answer a clarifying question anyway.

An Agent that meets ambiguity at 2am therefore either guesses, or stops having done almost nothing. The user finds out at 9am either way, and only one of those outcomes is recoverable.

Decision

Every Run carries Owl’s standing unattended contract, injected by the Driver in whatever way its tool supports - --append-system-prompt for claude-code (ADR-0018):

  • you are running unattended; nobody can answer a question
  • make reasonable assumptions rather than stalling, and write them down
  • commit incrementally, with messages that explain the reasoning
  • keep .coding-owl/HANDOFF.md current as you go, not at the end - the next Run starts with no memory of this one (ADR-0026)
  • when genuinely blocked, stop and state the blocker plainly
  • never guess at anything destructive or irreversible

A Project may append its own clauses through .coding-owl.yaml. owl jobs show <id> prints the effective system prompt, so nothing Owl injects is hidden.

Consequences

  • Every Job inherits sane unattended behaviour without the user restating it, and the night they forget to is no longer the night that is wasted.
  • “Commit as you go” is load-bearing elsewhere: it is what makes ADR-0011’s SIGTERM escalation cheap, and what lets ADR-0015’s garbage collection tell finished work from stray uncommitted changes.
  • “Record your assumptions” gives the morning review something specific to check, which is most of the difference between a diff you can trust and one you have to re-derive.
  • Owl now injects text the user did not write, making it a hidden variable in every Job’s behaviour. That is mitigated by printing it on demand and by keeping it short and stable - it is a contract, not a style guide.
  • Changing the contract changes behaviour across every Project at once, so it is versioned with the config apiVersion (ADR-0014) and changed deliberately.
  • A per-Project append can weaken the contract. Because Project config is read from the base branch (ADR-0014), an Agent cannot weaken the contract it is itself running under.

Alternatives considered

Inject nothing; the Job prompt is the whole contract. The most transparent arrangement possible - an Agent that misbehaves does so because of text the user can read. Rejected because it pushes a fixed checklist into every prompt by hand, and the failure mode of forgetting is an entire wasted night.

Per-Project preambles only, with no global default. Rules live next to the code they govern and are version-controlled with it, which fits ADR-0014 neatly. Rejected because a freshly registered Project would have no guardrails at all, which is precisely when they matter most.