Implementation Logic

Possible Interruptions

• Network issues causing repeated LLM API call failures.

• Parsing errors during action execution.

• Manual interruptions (e.g., Ctrl-C).

Serialized Storage Structure

All serialized data is stored in an integrated JSON file to minimize future changes to the storage structure. Example:

./workspace
  storage
    team
      team.json  # Contains team, environment, roles, actions, etc.

Recovery Scenarios

1. Role A fails to select an action:

   • Upon re-execution, Role A observes the unprocessed message and retries the action.

   • Role B resumes processing based on the updated message and continues.

2. Role B fails during action execution:

   • Upon recovery, Role B resumes directly from the failed action.

Key Recovery Steps

• Re-execute from the Message before the interruption:

  • Messages act as communication bridges between roles. To resume execution, interrupted messages must be reloaded and processed.

• Re-execute from the Action before the interruption:

  • Actions are granular execution units. Recovery ensures actions resume at the exact point of failure.

---

Result: Breakpoint Recovery Workflow

Command to enable recovery:

strataai "xxx" --recover_path "./workspace/storage/team"

Serialized data is stored in ./workspace/storage/team by default.

Example Execution Flow:

• A test case simulates an interruption and recovery:

  Execution case: test_team_recover_multi_roles_save
  - RoleB fails at ActionRaise.
  - Upon recovery, RoleA and completed actions of RoleB are skipped.
  - RoleB resumes from ActionRaise.

Recovery Logs:

2023-12-19 10:26:02.476 | DEBUG | strataai.environment:publish_message - publish_message: {...}
2023-12-19 10:26:12.518 | INFO  | strataai.roles.role:_act - RoleB ready to ActionRaise
2023-12-19 10:26:12.519 | ERROR | strataai.utils.utils:wrapper - Exception occurs, serialize the project...
2023-12-19 10:26:12.517 | DEBUG | strataai.roles.role:_observe - RoleB resumes execution from ActionRaise

---