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.

4.0 KB
Raw Blame

Harness lifecycle facts

Status: implemented for GitHub issue #1036. Diagnosis began against origin/master at da10a846783a0aaa0932050db04a812d9667868d.

Goal

Make Task and Watch lifecycle state follow persisted transitions. A definition must not become paused or finished because a clock passed, an unrelated edit moved updated_at, or a later cycle overwrote the transition owner.

Current-head diagnosis

  • Pause time is deliberately absent from every projection. The UI renders the paused state without a timestamp, and existing tests reject updated_at as a pause time. No consumer needs paused_at, so this change preserves that smaller invariant instead of adding unused state.
  • run_at already has one meaning at the two boundaries that interpret it: the scheduler and next-fire projection share resolve_run_at, and naive values use the definition's IANA timezone. Lifecycle rows and counts no longer infer an outcome from that clock, so they do not duplicate conversion logic.
  • Before this change, one-shot Tasks became finished from a clock comparison. A manual early run can therefore be mistaken for the scheduled outcome, and a missed fire can look successful.
  • A late Watch cycle preserves retirement timestamps but can still overwrite the exit code and error owned by the cycle that retired the Watch.
  • Definition writers already publish definitions.updated after successful commits. New lifecycle writes must use the same notification.

Model

retired_at remains the single definition terminal marker for both Tasks and Watches. A nullable retirement_reason only distinguishes Task terminal facts that cannot be reconstructed from run history:

  • schedule_consumed: the scheduler atomically consumed an at definition and queued its Run;
  • schedule_missed: APScheduler emitted a real misfire for the current stored one-shot schedule.

The reason does not form a second state machine. Lifecycle remains running | waiting | paused | finished, declared once by the storage query. An in-flight Run outranks retirement, then retired_at determines finished.

Manual vibe task run never writes either terminal fact. The scheduler enqueue transaction writes schedule_consumed together with the Run, so a crash cannot leave one without the other. Startup recovery relies on APScheduler's misfire event, not wall-clock inference; schedule_missed is therefore an observed scheduler outcome and never a fabricated success.

Run completion does not own Task retirement. This matters when the user replaces and resumes a consumed schedule while its Run is still active: that late result may update run history, but cannot disable or retire the replacement schedule.

A retired one-shot cannot be resumed in place: its old schedule has no future fire to restore. Both CLI and Workbench toggles reject that transition. Updating run_at or the schedule type first clears the terminal marker and creates the new lifecycle; unrelated edits preserve it.

For Watches, once a retirement commits, last_finished_at, last_exit_code, and last_error are that terminal outcome. A late cycle cannot modify them. The cycle's individual Run remains the place where that cycle's own result is retained.

Migration

Add nullable run_definitions.retirement_reason with no backfill. Existing retired_at, run history, updated_at, and wall-clock deadlines are not used to invent a reason. Legacy rows therefore remain unknown until a new lifecycle transition is observed and committed by its current owner. The one-time JSON import preserves terminal fields that are explicitly present, but never derives missing ones.

Evidence

  • Lifecycle contract: pause/resume/edit preservation, marker-driven Task rows and counts, early manual run, naive run_at timezone agreement.
  • Scheduler scenario: atomic scheduled consumption and explicit missed-fire recovery.
  • Watch concurrency contract: retirement outcome survives a late cycle result.
  • Migration: column addition and null legacy backfill.