Files

T

Henry 0563bf7d39 fix(office): escalate reviewer and approver agent failures to the CEO (#3275 )

Co-authored-by: nova28 <17953305+nova28@users.noreply.github.com>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Co-authored-by: Carlos Florencio <carlosmflorencio@gmail.com>

2026-09-02 16:55:39 +00:00

2.5 KiB

Raw Blame History

status, system, created, owners

status

system

created

owners

draft

office

2026-05-04

cfl

Office Agent Runtime — Error Handling Contract Requirements

Overview

When an agent run fails mid-turn (invalid model, auth failure, malformed response, transient upstream), the failed wakeup is terminal. The workflow engine can still evaluate a configured on_agent_error action and queue a separate coordinator run for escalation.

Requirements

REQ-OFFICE-RUNTIME-001: Office Agent Runtime — Error Handling Contract

Intent: When an agent run fails mid-turn (invalid model, auth failure, malformed response, transient upstream), the failed agent run is terminal. A workflow can use on_agent_error to queue a separate coordinator run for escalation.

Acceptance criteria

AC-OFFICE-RUNTIME-001.1: No automatic retry of the failed agent run. Every adapter error is terminal for the wakeup that produced it.
AC-OFFICE-RUNTIME-001.2: The wakeup row is stamped with status = failed and error_message. The failure path does not queue another wakeup for the same failed agent. A configured on_agent_error action can queue a separate coordinator run.
AC-OFFICE-RUNTIME-001.3: Re-runs of the failed agent happen only via explicit user action: Resume session in chat, Mark fixed on an inbox entry, or task reassignment to a different agent. A coordinator run from on_agent_error is escalation, not a retry.
AC-OFFICE-RUNTIME-001.4: agent_run_failed - one entry per failed (task, agent) wakeup while the agent is below threshold. Title: " failed on ". Action: Mark fixed -> dismiss + retry.
AC-OFFICE-RUNTIME-001.5: agent_paused_after_failures - one entry per auto-paused agent. Title: " auto-paused after failures (tasks A, B, C)". Action: Mark fixed -> unpause + retry the affected tasks.
AC-OFFICE-RUNTIME-001.6: Changing assignee_agent_instance_id on a task fires the existing reactivity pipeline, which queues a fresh task_assigned wakeup for the new agent.
AC-OFFICE-RUNTIME-001.7: The existing staleness check (recovery-reliability spec) cancels the prior wakeup for the (task, old agent) since the assignee has changed.
AC-OFFICE-RUNTIME-001.8: Any per-task agent_run_failed inbox entry tied to the old (task, agent) auto-dismisses - the failure is no longer actionable on this task.

System design

The migrated technical source is split into part 1, part 2.

2.5 KiB Raw Blame History