Recovering Observal

Execution contract

1. Execute commands with a 60 second timeout unless the operation documents a longer wait.
2. Use machine output by default: add --output json whenever supported. Parse list results from items and pagination fields.
3. Run --help before acting when a path or flag is uncertain.
4. Use dry run before reconciliation when scope or session volume is uncertain.
5. Supply documented force flags so version operations never prompt.
6. Verify cursor, outbox, installed version, checksum, and rollback state after recovery operations.
7. Fail openly. Never hide an unavailable server behind automatic local writes.
8. Never retry reconciliation, version changes, or fallback writes without checking resulting state.

Read Recovery workflows completely before executing.

Decision rules

• Healthy telemetry does not need routine reconciliation.
• Reconcile repairs missed local session delivery. It does not replace hook or extension installation.
• Upgrade, downgrade, and rollback are distinct requests. Do not substitute one for another.
• Local fallback is allowed only after JSON reports error.category: unavailable with exit code 9, or error.category: authentication with operation Load authenticated CLI configuration and exit code 3. The user must still request local files.
• Local fallback creates harness-native Agent files. It does not publish Registry state and must be reported as local-only.
• Never invent telemetry environment variables or wrappers.

Completion

Report sessions discovered, queued, skipped, or failed for reconciliation; old and new versions for CLI changes; or exact local paths for fallback. Include unresolved warnings and the command needed once the server is reachable.