| name | auto-checkpoint |
| description | Portable "loop kit" — inject durable state, terminal-condition, auto-resume and auto-escalation into any session. Makes the loop reliable without a daemon. |
| providers | ["claude","copilot","codex"] |
auto-checkpoint — the portable loop kit
A kit injectable into any session to make it loop-reliable without a daemon.
Implements the contract in loop/loop-protocol.md:
durable state, terminal-condition, auto-resume, auto-escalation.
What it does
- Writes/reads durable state —
state.db at a known path + .agent-state/<task>.jsonl (append-only cursor).
- Defines the terminal-condition — an empirical check against ground truth, not an estimate.
- Enables auto-resume — on startup, re-reads the state, resumes from the last
done step.
- Enables auto-escalation — 2 failures on the same problem → opus + log the reason.
State (daemon-optional)
state store: ~/.convergio/v3/state.db (if present)
~/.roberdan-os/state.db (fallback)
cursor: .agent-state/<task>.jsonl (gitignored, 1 record/step + evidence)
timestamp: RFC3339
Convergio, if active, reads the same state — optional observer, never a dependency.
Loop (pseudo)
on start:
state = read(state.db, cursor) # idempotent resume
step = last_done(state) + 1
loop:
result = execute(step)
append(cursor, {step, result, evidence}) # checkpoint = 1 commit/phase, evidence-first
if terminal_condition(): break # empirical verification (thor / job-specific check)
if failed_twice(step): escalate(opus); log(reason)
if no_progress(2 passes): STOP; report_wedged(); break
step += 1
on each phase end:
post_task_sync() # vault + cvg + repo
Per-platform driver
- Claude Code:
/loop + ScheduleWakeup for external waits (CI/deploy/embed) —
submit → wakeup +Nmin → check terminal-condition → done | re-arm.
- Others:
launchd/cron re-read the cursor and relaunch until the terminal-condition.
Reporting
Every checkpoint = an evidence-first update: [phase N/M ✓] commit <sha> · <check> · next: ….
Never "still working."
