Planning workflow
Interactive actor, gate, execution, cooperation, close and watcher paths are documented and updated for the delivered runtime.
CLI interface verification
Native WSL profiles are configured and actual read-only launches passed for Codex and Hermes. Claude is blocked by an expired Claude OAuth session; Gemini is blocked until an auth method/API credential is configured in WSL.
Up-front node allocation — mandatory start gate
The runtime requires every workflow node to have one approved allocation and rejects duplicate nodes plus cross-agent file/output overlap before activation.
Continuous workflow state — mandatory
State records workflow/node/task IDs, agent, model, status, percent, action, blocker, next step, evidence, heartbeat and centrally generated timestamps.
Live workflow rendering
The generated local page shows ownership, action, status, progress, blocker, next step and last update, and refreshes every 15 seconds.
Wrapper and configuration
The guarded run command activates only an approved owner, records outcome, and can use a user-verified local profile.
Assessment and safety gates
Preflight reads relevant board data without acknowledging it, labels it untrusted, and assessment records impact decisions through the board.
Cursor and failure semantics
The runtime never advances an Agent Message Board cursor; acknowledgement remains an explicit board operation after assessment.
Watcher refresh
The optional watch command checks stale heartbeats and regenerates the live view. Scheduling policy stays under user control.
Node and workflow retirement
not-required visibly records a reason for a superseded node or whole workflow, retains owner/history/evidence, and rejects retirement while a node remains active.
Four-agent readiness
All demonstration workflow nodes are complete after safe Hermes/Codex handoff and reassignment. Codex and Hermes completed real smoke launches. The separate Claude and Gemini profile smoke nodes remain blocked until their respective providers authenticate successfully.
Acceptance evidence
Automated tests cover allocation, ownership boundaries, handoff, stale detection and rendering. run_agent.py remains unchanged.