Implement native third-party agent dogfood path #58

Current Status

Updated 2026-06-14 by Code.

The first external-command dogfood slice is merged to main via PR #82 (904228cfc50b99e4327284171820a0fdd412452b). This gives MultiAgent V2 a native role-scoped external_command backend proof path: spawn creates a synthetic registry-backed agent instead of an internal Codex thread, the backend receives a bounded one-shot JSON request over stdin, returns a bounded JSON response over stdout, and completion wakes the parent through the V2 mailbox/list/status path.

Delivered:

• [agents.<role>].backend config/schema support for type = "external_command", including command/timeout validation and generated schema updates.
• Synthetic external-agent registry/status/cancel/list/close handling and V2 follow-up rejection.
• One-shot subprocess runner with timeout/cancel coverage across stdin/write/wait, bounded stdout/stderr reads, and stale-completion suppression after close.
• Focused config tests for load/default/invalid backend values and a V2 external-command round trip through the parent mailbox.

Validation before merge:

• cargo fmt --all -- --check passed, with only the existing stable-rustfmt warning about imports_granularity.
• cargo test -p codex-core agent_role_external_command_backend_ --lib passed.
• cargo test -p codex-core multi_agent_v2_external_command_role_completes_through_parent_mailbox --lib passed.
• cargo check -p codex-core --tests passed.
• just write-config-schema was run after schema source changes.
• PR Add external command agent dogfood path #82 checks passed: codespell, cargo-deny, build-test, app build, and exec harness.
• Targeted read-only lifecycle/config review agents found issues that were patched before PR creation. Background auto-review had no active/current target findings at merge time.

Next action:

• Continue Implement native third-party agent dogfood path #58 with follow-up dogfood hardening only if needed after real use. Remaining intentionally out-of-scope areas are streaming, resume/full-history fork support, app-server runtime API, richer provider strategy, and broader UX polish; those should become separate plan/sub-issues rather than expanding the merged first slice.

Finish Line

Codex Lab has a native dogfood path for at least one third-party/external agent backend, with clear config, lifecycle, cancellation, output surfacing, and fallback behavior. This must not depend on the removed Ultracode plugin being installed.

Acceptance Criteria

• Decide whether Provider-agnostic worker decision spike #2 is superseded by this issue, remains the umbrella, or is split into implementation children.
• Wire a native external-agent dispatch point into the V2 multi-agent path without regressing internal agents.
• Define the request/response/lifecycle contract for external agent spawning, including role, initial context, cancellation, completion, failure, and output events.
• Add config for opting an agent role into the external path with self-describing fields rather than ambiguous booleans where possible.
• Implement or stub one concrete dogfood backend, preferably subprocess or app-server protocol based, and prove a full round-trip.
• Preserve sandbox/worktree safety and prevent orphaned external processes or stuck session registry state.
• Validate that existing internal multi-agent tests still pass and add integration coverage for external success, crash, cancellation, and config fallback.

Relationships

Child of #53. Related to #2 and #54. Blocks full dogfood MVP alongside native Auto Review.

• related: Provider-agnostic worker decision spike #2 - Provider-agnostic worker decision spike #2