Skip to main content

Harness support

SuperDomestique's primary supported harness is Claude Code. Codex has driven a complete interactive prep-to-merge run, but that run still needed manual handoffs and exposed gaps in installation, containment, review, and run attribution. Selected read-only producer calls through Codex are experimental. Codex is not currently a supported unattended L3 or L4 path.

pi.dev is planned as a future harness target. It is not supported, and no implementation or verification work is currently scheduled.

This status was checked on 2026-08-09. The faff plugin, CLI, and configuration names are unchanged whichever harness starts the work.

What the labels mean

  • Supported means there is a documented, repeatable path with no known missing essential integration for the stated level.
  • Demonstrated with limitations means a recorded successful path exists, but important setup, lifecycle, safety, or automation gaps remain.
  • Experimental means a narrow implementation or observation exists. It is not an end-user support promise.
  • Planned means the path has not shipped.
  • Not supported means there is no current path to rely on.

A portable SKILL.md, a working CLI transport, and a supported delivery loop are different claims. SuperDomestique treats them separately.

Capability matrix

The short answer is to use Claude Code for the documented product today. Use Codex only when the limitations below fit the job and you are prepared to drive the handoffs yourself.

CapabilityClaude CodeCodexpi.dev
SetupSupportedDemonstrated with limitationsNot supported; no work scheduled
Skill discoverySupportedDemonstrated with limitationsNot supported; no work scheduled
Tracker accessSupportedDemonstrated with limitationsNot supported; no work scheduled
Interactive prepSupportedDemonstrated with limitationsNot supported; no work scheduled
Interactive build and mergeSupportedDemonstrated with limitationsNot supported; no work scheduled
Producer dispatchSupportedExperimental, read-onlyNot supported; no work scheduled
Independent reviewSupportedDemonstrated with limitationsNot supported; no work scheduled
Parallel review fan-outSupportedExperimentalNot supported; no work scheduled
L3 unattended entrySupportedPlannedNot supported; no work scheduled
L4 isolation and holdoutSupported as previewNot supportedNot supported; no work scheduled
Usage telemetrySupportedExperimentalNot supported; no work scheduled
Live event streamSupportedExperimental CLI streamNot supported; no work scheduled

L4 is preview-labelled on every harness. The mechanisms exist, but the public repository does not yet have enough external evidence for an unqualified L4-complete claim. The levels page describes the current boundary.

What differs under Codex

Installation and discovery

Codex can discover globally linked skills, and the installer and doctor now understand the shared agent skill locations. Repo-local and marketplace plugin installs remain unavailable to Codex. The standard marketplace installation in the README is therefore not a complete Codex setup path.

Tracker discovery has been corrected, but the tracker connector is still registered by the harness. A connector available in Claude Code is not automatically available in Codex.

Session lifecycle

The recorded Codex run reached prep, implementation, pull request, and merge. It did not prove unattended operation. The operator had to initiate each handoff, and that lifecycle gap remains open.

faff engine call can spawn codex exec for selected producer lanes. That child is deliberately read-only, ephemeral, and outside the working repository. It cannot act as the write-capable build lane or the whole-session unattended entry point. Build and evaluation routing remain planned work.

Safety and evidence

The demonstrated run was uncaged, and work escaped its intended worktree. A checked environment floor and non-Claude operator guidance have not shipped.

Codex events expose useful usage and tool information, but current run records do not consistently identify the harness and model. That attribution gap remains open.

Review and fan-out

Concurrent non-Claude review-lens dispatch shipped in PR #579. This improves the review path, but it does not establish full fan-out or isolation parity. Whether subagents actually ran remained opaque in the recorded workflow, and provider policy blocked some adversarial work.

Evidence

  • The Codex CLI observation records flags, event shapes, auth, state, and skill discovery observed with Codex CLI 0.145.0 on 2026-07-28.
  • PR #510 is the durable repository outcome from the interactive Codex run. Its description and committed report record both the successful delivery path and the failed L4 audit result.
  • The Codex engine implementation defines the experimental read-only producer boundary.
  • The harness coupling inventory classifies the architectural seams. It is a design inventory, not a support promise.

When this status can change

A completed ticket is not enough by itself. A support label moves only when the path is documented, repeatable, and backed by a committed observation, run result, enforcing mechanism, or merged implementation.

Codex L3 support needs a whole-session entry and run-lifecycle contract, plus a documented environment that meets the level's floor. Codex L4 support also needs the external B10 portability run and evidence that isolation, holdout, attribution, and review still work together. Until then, a successful interactive session does not imply unattended support. The harness coupling inventory records the current implementation owners.