Skip to content

v1.2.0 — GOAL_CHECK every round, Resume button, responsive transport

Choose a tag to compare

View all tags
@pekinlcc pekinlcc released this 19 Apr 06:45
v1.2.0
6450494

Makes the "distance to goal" panel actually update per cycle, fixes the
pause/stop UX so clicks feel responsive, and adds a Resume button so a
paused session can be continued without losing state.

Changed

  • GOAL_CHECK now runs after every REVIEWING, regardless of verdict.
    Previously the state machine only entered GoalCheck phase when Codex
    returned Approved — which in practice could never happen for dozens
    of rounds (Codex kept finding one blocking item, loop stayed in
    REVIEWING ↔ REFINING). The Remaining-to-goal panel kept showing
    "waiting for first goal check" all the way to round 167+ in real user
    testing. Now every REVIEWING → (verdict) cycle launches GOAL_CHECK on
    both agents in parallel, so the panel refreshes every 2–10 minutes.
    DONE still gates on both agents independently reporting done=true.
    Trade-off: +2 CLI invocations per round (one Claude goal-check, one
    Codex goal-check), each short. Worth it for the continuous signal.
    Blocked verdict still skips straight to Refining (unchanged) — if
    the review machinery itself is malformed, asking "are we done?" at
    that point is pointless.

Added

  • Resume button. Once a session enters PAUSED, the big triangle
    Play button in the transport lights up again as Resume; clicking it
    calls startSession(goal, workdir) (idempotent on the AppState side)
    so the orchestrator picks up from the recorded state/phase. Before
    this, pause was effectively a one-way gate into Stop territory.
  • Optimistic transport feedback. Clicking Pause or Stop now
    immediately shows ⏸ PAUSING… / ■ STOPPING… in the title-bar badge
    and disables the button, instead of waiting 5–10 seconds for the
    orchestrator to actually unwind the current turn. The label clears
    itself once the reducer confirms the real state_changed event.

Fixed

  • Pause and Stop appeared to "do nothing" because the orchestrator
    needs up to 5s (SIGINT grace) + next step-loop iteration to actually
    transition state, and the UI gave no in-between signal. This is now
    fixed by the optimistic feedback above.

Known limitations

  • Rate-limit / quota-exhaustion auto-pause still not implemented. When
    a CLI's 5-hour quota runs out you still see a crash loop until the
    flapping limit triggers ERRORED. Planned for v1.3.
Assets 2
Loading