Finish Line

The external_command dogfood backend is lifecycle-safe enough to rely on during Codex Lab dogfood work: external children are tracked in the same spawn graph paths as internal children, cancellation cannot race into launching a subprocess after cancellation was already requested, and parent wait/mailbox flows receive a terminal notification when an external child is cancelled or closed.

Acceptance Criteria

• Persist external children in the spawn graph before the launch path can return control or enter a state where cancellation/parent lookup may race.
• External children are discoverable by the existing descendant/list/close paths that rely on spawn graph state.
• Cancelling or closing an external child sends the parent a terminal notification so wait_for_agent or equivalent waiting flows do not hang.
• Check cancellation immediately before launching the external subprocess, and do not start the subprocess when cancellation was already requested.
• Cover the regression cases with focused tests: external child appears in the spawn graph, close/cancel notifies the parent, and pre-launch cancellation prevents subprocess spawn.
• Keep scope on the PR Add external command agent dogfood path #82 external_command backend and existing MultiAgent V2 lifecycle paths; do not broaden into provider-agnostic worker redesign, streaming, resume/fork semantics, or Code Bridge.
• Validate with focused codex-core tests and a broader MultiAgent V2 regression command before merge.

Current Status

Created 2026-06-14 by Code after background Auto Review reported three target-applicable external-agent lifecycle findings on the PR #82 / #58 dogfood path:

• P1: external children are not persisted in the spawn graph before returning.
• P1: cancelling or closing an external child does not notify the parent path that waiting flows depend on.
• P2: cancellation can still race into launching the subprocess.

Implementation Notes

• The issue should fix the existing external_command lifecycle behavior, not reopen Implement native third-party agent dogfood path #58 or expand Provider-agnostic worker decision spike #2's provider strategy scope.
• The spawn graph fix should preserve the internal-agent descendant/list/close semantics for external children and avoid orphaned external registry state.
• Parent notification should avoid double terminal messages while ensuring cancelled/closed external children still unblock waiting flows.
• Cancellation-before-launch tests should avoid process-leak flakiness by asserting deterministic terminal behavior rather than relying only on eventual process cleanup.

Relationships

• Child/follow-up under Validate Auto Review and third-party agents for dogfood MVP #53.
• Follow-up to Implement native third-party agent dogfood path #58 / PR Add external command agent dogfood path #82; Implement native third-party agent dogfood path #58 remains complete.
• Related to Provider-agnostic worker decision spike #2 only as later provider-worker evidence; this issue is not blocked by Provider-agnostic worker decision spike #2.

Relationships

Child/follow-up under #53.

Follow-up to #58 / PR #82; #58 remains complete.

Related to #2 only as later provider-worker evidence; this issue is not blocked by #2.

• related: Implement native third-party agent dogfood path #58 - Implement native third-party agent dogfood path #58

Finish Line

The external_command dogfood backend is lifecycle-safe enough to rely on during Codex Lab dogfood work: external children are tracked in the same spawn graph paths as internal children, cancellation cannot race into launching a subprocess after cancellation was already requested, and parent wait/mailbox flows receive a terminal notification when an external child is cancelled or closed.

Acceptance Criteria

• Persist external children in the spawn graph before the launch path can return control or enter a state where cancellation/parent lookup may race.
• External children are discoverable by the existing descendant/list/close paths that rely on spawn graph state.
• Cancelling or closing an external child sends the parent a terminal notification so wait_for_agent or equivalent waiting flows do not hang.
• Check cancellation immediately before launching the external subprocess, and do not start the subprocess when cancellation was already requested.
• Cover the regression cases with focused tests: external child appears in the spawn graph, close/cancel notifies the parent, and pre-launch cancellation prevents subprocess spawn.
• Keep scope on the PR Add external command agent dogfood path #82 external_command backend and existing MultiAgent V2 lifecycle paths; do not broaden into provider-agnostic worker redesign, streaming, resume/fork semantics, or Code Bridge.
• Validate with focused codex-core tests and a broader MultiAgent V2 regression command before merge.

Current Status

Created 2026-06-14 by Code after background Auto Review reported three target-applicable external-agent lifecycle findings on the PR #82 / #58 dogfood path:

• P1: external children are not persisted in the spawn graph before returning.
• P1: cancelling or closing an external child does not notify the parent path that waiting flows depend on.
• P2: cancellation can still race into launching the subprocess.

Agent planning review recommends keeping this issue narrow: fix the existing external_command lifecycle behavior, add focused regression tests, and do not reopen #58 or expand #2's provider strategy scope.

Implementation notes:

• The spawn graph fix should preserve the internal-agent descendant/list/close semantics for external children and avoid orphaned external registry state.
• Parent notification should avoid double terminal messages while ensuring cancelled/closed external children still unblock waiting flows.
• Cancellation-before-launch tests should avoid process-leak flakiness by asserting deterministic terminal behavior rather than relying only on eventual process cleanup.