Skip to content

armgit / avibe

Public Imported from GitHub
0

Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.

3.0 KB
Raw Blame

Agent Run Terminal Lifecycle

Background

vibe agent run currently records an agent_run row and marks it succeeded as soon as the harness request submits the prompt to the backend path. For async backends such as Codex, submitting the prompt only starts the native turn; tool events and the terminal result arrive later through emit_agent_message.

This makes run.status=succeeded and completed_at appear before the run's final result_text is available.

Goal

For direct agent_run requests, the run lifecycle must model one concrete Agent turn:

  • queued before claim
  • running while the backend turn is active
  • terminal only after the backend emits its real result

The fix should use the existing terminal-result signal. It must not add sleeps, post-completion polling, or Codex-specific behavior.

Design

Keep the existing request queue and run storage. Change only the direct agent_run lifecycle:

  1. Build the same scheduled/harness MessageContext with task_execution_id.
  2. For private/non-avibe runs, run the turn through dispatch_turn(..., source=SOURCE_SCHEDULED, on_chunk=noop) so dispatch returns only after the terminal result.
  3. For avibe/workbench sessions, keep routing through the session turn gate so queueing, Stop, and browser lifecycle remain unchanged.
  4. When emit_agent_message(..., "result", ...) carries task_trigger_kind="agent_run", update the run record with result_text, message id, terminal status, and completed_at.
  5. Treat terminal delivery and run completion as separate concerns: even when every IM send/upload fails, the terminal result still releases the turn waiter and records the run result.
  6. When an avibe agent run is queued behind an active workbench turn, return the run row to queued and let the workbench queue hold it until flush time. The background drain skips those held rows so the same run is not claimed twice.

This makes the outbound terminal-result chokepoint the source of truth for direct Agent Run completion, while preserving the existing avibe gate for workbench sessions.

Scope

This change is intentionally limited to request_type == "agent_run". Stored tasks and watches may still mean "trigger/follow-up submitted"; changing that semantic together would be a broader product migration.

Validation

  • Unit test that a direct private agent_run stays running until a delayed terminal result is emitted.
  • Unit tests that suppressed and visible agent_run terminal output update the run status.
  • Unit test that avibe agent_run still routes through the session gate and does not complete at submit time.
  • Unit test that failed terminal results remain failed and are not overwritten by request-submission success.
  • Unit test that synchronous dispatch errors are recorded as failed runs.
  • Unit test that terminal delivery failure still releases the waiting turn.
  • Unit test that busy avibe agent runs remain queued until the workbench queue actually starts them.
  • Existing scheduled task and dispatcher tests remain green.