Validate Auto Review and third-party agents for dogfood MVP #53

Current Status

Updated 2026-06-14 by Code.

State: Active umbrella. The native Auto Review dogfood spine is complete enough for daily workflow validation, the first third-party/external agent dogfood path is complete and hardened enough to rely on, and the Codex Lab home/repo-dotdir boundary has now landed.

Completed dogfood spine:

• Implement native Auto Review artifact storage #55 / PR Add native auto review artifact storage #60 added native Auto Review artifact storage in the small codex-auto-review crate.
• Wire native detached Auto Review runtime #56, Design safe background Auto Review trigger and debounce #61, and related follow-ups landed the native detached runtime, safe background trigger/debounce, lifecycle state, cancellation/control, and foreground race/performance fixes.
• Inject current Auto Review awareness into agent context #62 / PR Add bounded auto-review finding detail reads #79 landed bounded current Auto Review awareness and stable finding detail retrieval.
• Expose native Auto Review findings in app-server and TUI #57 is complete via PR Expose Auto Review summaries in app server and TUI #80 and PR Add TUI Auto Review summary surface #81: app-server summary/detail/status surfaces are available, and the TUI now renders background Auto Review status plus clean/findings/stale/failed/error summaries with stale-result suppression.
• Implement native third-party agent dogfood path #58 / PR Add external command agent dogfood path #82 delivered the first native external-command dogfood backend path with role-scoped config, lifecycle/status/list/cancel/close behavior, output surfacing, fallback behavior, subprocess runner coverage, and tests.
• Add Codex Lab dotdir ignore guardrails #84 added the first Plan Codex Lab home and repo dotdir boundary before Code Bridge #59 guardrails: repo-level .code/ ignore backstop and Codex Lab-owned .codex/.gitignore runtime/secret guardrails.
• Harden external-agent spawn graph and cancellation handling #85 / PR Harden external agent lifecycle #86 hardened the external-command lifecycle: persisted spawn graph edges before task launch, live external-child traversal, cancellation parent notifications, pre-launch cancellation guards, and close-vs-shutdown persisted edge semantics.
• Plan Codex Lab home and repo dotdir boundary before Code Bridge #59 / PR Use Codex Lab home boundary #87 landed the Codex Lab home boundary: CODEX_LAB_HOME is the only Codex Lab home selector, the default home is ~/.codex-lab, legacy CODE_HOME / CODEX_HOME selectors are ignored for Codex Lab runtime home selection, diagnostics were renamed, app-server daemon installer env bridging was handled deliberately, and exec-harness cache placement no longer falls back to legacy homes.

Current sequence:

1. Scope minimal Code Bridge vertical slice #36/Define Code Bridge protocol, trust, and payload contract #45-Gate thin app-server integration for Code Bridge #50 - Resume Code Bridge vertical-slice work now that Codex Lab home and repo-dotdir boundaries are settled.
2. Provider-agnostic worker decision spike #2 - Keep as the broader provider-agnostic worker strategy spike. Implement native third-party agent dogfood path #58 proves the first concrete dogfood backend and Harden external-agent spawn graph and cancellation handling #85 hardens its lifecycle; neither settles streaming, resume/full-history fork support, app-server runtime API, richer provider/auth strategy, usage/quota accounting, or broader UX polish.

Product boundary remains unchanged:

• Use native Codex review/runtime/context primitives wherever possible.
• Do not copy Every Code plugin systems.
• Removed auto-review@just-every and ultracode@just-every plugins remain reference-only.
• Code Bridge stays after the dogfood/home-boundary foundation, and is now the next MVP workstream.

Finish Line

Codex Lab is dogfood-ready by the user's definition: Auto Review works in the real daily workflow, and third-party agent providers are validated enough for daily use with clear limits and fallback behavior.

Acceptance Criteria

• Define the concrete dogfood workflow for Auto Review: when it runs, how findings are surfaced, how stale/detached findings are suppressed, and how details are recovered.
• Validate Auto Review outside the narrow exec-harness proof loop, or record the smallest missing implementation issue if runtime wiring is absent.
• Define the third-party agent provider dogfood bar: supported providers/models, auth/config source, lifecycle commands, cancellation/wait/output collection, usage/quota accounting, and failure/degraded behavior.
• Validate at least one non-OpenAI/third-party agent provider path end to end, or record the smallest missing implementation issue if runtime support is absent.
• Decide whether Provider-agnostic worker decision spike #2 should be promoted from waiting to active implementation, split into child issues, or closed by a narrower dogfood contract.
• Update Codex Lab MVP dogfood plan #28 and Run Codex Lab dogfood readiness sweep #52 language so GO means covered-spine-only unless this issue is also complete.

Relationships

Child of #28. Blocks treating Code Bridge (#45-#50) as the only remaining MVP work. Related to #2 and #35.