Checkpoint Validation

CRITICAL RULES (NEVER VIOLATE)

1. "Artifact exists" ≠ "completed" — ONLY trust checkpoint.status == "completed"
2. Never run 2 agents of the same role simultaneously — confirm prior agent terminated before respawn

Validates agent checkpoint.json to detect idle/fake completion/error states.

Input

• $ARGUMENTS: path to checkpoint.json

Procedure

Step 1: Read checkpoint.json

Read $ARGUMENTS

Step 2: Status Determination

checkpoint.status	Verdict	Action
"completed"	Verify all expected_artifacts exist	All present → PASS / Missing → FAKE COMPLETION
"in_progress"	Check timestamp	Last update > 5min ago → FAKE IDLE / Recent → WORKING
"error"	Analyze error field	ERROR + output recovery suggestion
File missing	Agent never started	NOT STARTED — respawn immediately

Step 3: Artifact Verification (when completed)

# Check each expected_artifact exists and is non-empty
for f in <expected_artifacts>; do
  if [ ! -f "$f" ] || [ ! -s "$f" ]; then
    echo "FAKE: $f missing or empty"
  fi
done

• 0-byte file = fake artifact → FAKE COMPLETION

Step 4: Output

[CHECKPOINT] Path: <path>
[CHECKPOINT] Agent: <agent_name>
[CHECKPOINT] Status: <status>
[CHECKPOINT] Phase: <phase>/<phase_name>
[CHECKPOINT] Completed: <completed_list>
[CHECKPOINT] In Progress: <in_progress>
[CHECKPOINT] Artifacts: <produced> / <expected>
[CHECKPOINT] Result: PASS / FAKE IDLE / FAKE COMPLETION / ERROR / NOT STARTED
[CHECKPOINT] Recovery: <specific recovery suggestion>

Step 5: Recovery Actions

State	Recovery
FAKE IDLE	Send "read checkpoint.json and continue" message once → still idle → respawn
FAKE COMPLETION	Respawn with missing artifact list (checkpoint injection)
ERROR	Resolve error cause, then respawn (env issues = Orchestrator fixes)
NOT STARTED	Spawn new agent immediately

REMINDER: Never trust "artifact exists = done". Only checkpoint.status == "completed" is truth.

Gotchas

• A 0-byte checkpoint.json that exists is still a fake completion — always verify status field value
• "artifact file exists" does NOT mean completed — only trust status=="completed" in checkpoint
• Agents may write checkpoint with in_progress then crash — check timestamp staleness (>10min = likely dead)