Skip to main content

Checkpoint Handoff Proof

This page is the public evidence for BUG-23: AgentXchain can drive a writable multi-role continuous session without stopping for manual repo commits between authoritative role handoffs.

Date of proof: 2026-04-18

Latest checked-in evidence artifact: examples/live-governed-proof/evidence/checkpoint-handoff-proof.latest.json

Back to Live Governed Proof.

What It Is

  • Category: live proof example
  • Surface: agentxchain run --continuous --auto-checkpoint
  • Runtime shape: local_cli for all writable roles
  • Default runtime: _checkpoint-proof-agent.mjs wrapper over cli/test-support/mock-agent.mjs
  • Source harness: examples/live-governed-proof/run-checkpoint-handoff-proof.mjs

Workflow Or Proof Shape

The harness scaffolds a temp governed repo, writes the minimum planning artifacts required for a real run, initializes git, and executes the real continuous command path:

agentxchain run --continuous --auto-checkpoint --vision .planning/VISION.md --max-runs 1 --max-idle-cycles 1 --poll-seconds 0

The default proof is intentionally self-contained:

  • pm, dev, and qa all run through local_cli
  • the wrapper appends phase-specific files so the working tree actually becomes dirty between handoffs
  • mock-agent.mjs still produces the governed gate artifacts, so the proof exercises real acceptance + checkpoint behavior instead of faking the post-accept state

The proof passes only if the temp repo records all of the following:

  • at least one turn_checkpointed event
  • at least one checkpoint: <turn_id> (role=<role>, phase=<phase>) git commit
  • zero Authoritative/proposed turns require a clean baseline errors
  • zero manual checkpoint-turn --turn reminders

Observed Harness Invocation

node examples/live-governed-proof/run-checkpoint-handoff-proof.mjs --json --output examples/live-governed-proof/evidence/checkpoint-handoff-proof.latest.json

Add --keep-temp when you want the temp repo preserved for inspection, or --runtime-command "codex exec --dangerously-bypass-approvals-and-sandbox" when you want all writable roles to run through a real local runtime instead of the default mock wrapper.

Evidence Summary

MetricValue
Session IDcont-d49f1fc5
Started2026-04-18T17:27:50.312Z
Finished2026-04-18T17:27:57.242Z
Session statuscompleted
Runs completed1
CLI version2.134.1
Checkpoint commits3
Checkpoint events3
Phases checkpointedplanning, implementation, qa
Clean-baseline errors0
Evidence artifactexamples/live-governed-proof/evidence/checkpoint-handoff-proof.latest.json

Git Log Evidence

The proof repo recorded one checkpoint commit per accepted writable turn:

checkpoint: turn_1754945b4734ceff (role=qa, phase=qa)
checkpoint: turn_b1b4533bfc375f0f (role=dev, phase=implementation)
checkpoint: turn_76e36e3e78fcf1a9 (role=pm, phase=planning)

This matters because BUG-23 was not “checkpoint command exists.” The bug was that true multi-role continuous execution still stopped on a dirty worktree between accepted turns. The evidence above proves the writable handoff is now resolved inside the run loop itself.

Event Evidence

The same run emitted turn_checkpointed events with the repo commit SHA attached:

RoleTurn IDCommit SHA
pmturn_76e36e3e78fcf1a938132cb5…
devturn_b1b4533bfc375f0fe712b935…
qaturn_1754945b4734ceff30af1ca1…

That event trail is the externally-visible proof surface for auto-checkpoint, while the git log is the repo-native proof surface.

Run It

node examples/live-governed-proof/run-checkpoint-handoff-proof.mjs --json --output examples/live-governed-proof/evidence/checkpoint-handoff-proof.latest.json

No API keys are required for the default proof.

Key Takeaways

  • BUG-23 now has a checked-in public evidence artifact, not just a harness and a claim.
  • The proof exercises the real run --continuous --auto-checkpoint surface instead of direct state edits or staged JSON.
  • Checkpointing is visible in both git history and events.jsonl, which is the right proof pair for a writable-handoff feature.