Workers, watches, later checks, peer communication and browser reports around the native Codex CLI/TUI. Keep your continuing conversation and its existing daemon; choose explicit worker workflows or ask for useful follow-up in ordinary language.
Start with the setup guide to choose workers, monitoring or reporting. With the relevant components installed, open an ordinary Codex session and say, for example:
- “$luna-task Investigate the importer failure and return a reproduction.”
- “$luna-goal-loop Complete the approved importer migration and verify it.”
- “Watch the session doing the packaging work and tell me when it is ready.”
- “Tell that session the handoff should include the risks.”
- “Watch this run and tell me if it stalls or fails.”
- “Check this endpoint again later.”
- “Stop watching it.”
- “Make a browser report of what we learned.”
The model selects the source, filtering, cadence and continuation. It discovers independent peers from context and clarifies ambiguity; no UUID, exact title, JSON prompt or polling recipe is required from you.
Read the daily-use guide, open the browser walkthrough, or see how the package works.
Workers, monitoring and reporting share one Codex Plus version and release. Choose the component archives you need in the setup guide; an existing managed setup can use these requests directly. The public site includes the walkthrough and eight HTML guides. Their Markdown sources and technical evidence are linked below.
The preview is ready for daily-use trials. Choose workers, follow useful changes or publish a browser handoff around your existing Codex conversation.
v0.1.0-preview.10 includes bounded reminder history, rediscoverable outgoing messages with quoted purpose, clearer follow-up outcomes and reported MCP-version diagnosis. The guides explain how to return to unfinished work; see released and source features. Updating a guide or checkout does not upgrade an installed runtime.
| Capability | What is available |
|---|---|
| Watches | Quiet command/script observation, authenticated endpoint access, passive peer reads and configured MCP inspection. |
| Later checks | One-shot reassessment and model-selected continuation, with status and cancellation. |
| Returning to work | “What am I still waiting on?” combines this conversation's stored watches, reminder outcomes and outgoing receipts, with stale or missing evidence explained. |
| Peer coordination | Contextual independent-session discovery, attributed information/requests and handling receipts within one daemon. |
| Browser reports | Immutable static HTML, project catalog, four report MCP tools, a report skill and independent HTTP hosting. |
| Workers | Existing two agents and four worker workflows, with unchanged identities and explicit invocation. |
Command polling can stay quiet without a model turn per read. Configured MCP checks use scheduled native model turns under normal authentication/approval; each check has model cost. Explicit stop cancels selected observation without stopping the observed work. Esc interrupts the foreground turn and leaves independent registrations within their bounds.
Bounded synthetic native qualification is separate from real-source reliability,
credential/provider recovery, long unattended operation and personal UX acceptance.
Cross-host native messaging,
a recurring calendar and native /exit confirmation interception are outside
this preview. Use the soak checklist to collect actual-use
evidence without manufacturing workloads.
Managed deployment details: Snap, Starship and Carbon
Use codex for Snap OSS and codex-work for work, including their agents
subcommands. The old codex-oss compatibility entry point is retired. These are
separate homes and daemons; native --profile alone does not separate state.
See multiple environments for portable defaults and the
existing explicit Snap paths.
Snap, Starship and Carbon select v0.1.0-preview.10 for workers, monitoring and
reporting from source 078c630 through chezmoi. The scoped rollout on 2026-10-22
verified all component manifests and preserved native daemons, active MCP
processes, configuration and report hosting. Existing processes adopt new code on normal
reconnect. The implementation handoff records
source, release and installed state.
Desktop notification delivery and terminal customization are outside Codex Plus. Native Codex alerts remain selected. Notification enrichment and lifecycle inspection are parked; native terminal formatting is managed with Kitty/chezmoi.
| HTML guide | For | Markdown source |
|---|---|---|
| Get started | Choose components and follow the right setup path. | Source |
| Daily use | Natural requests, useful follow-up, stopping and reports. | Source |
| Luna & workers | Tasks, goal loops, adaptive delegation and worker installation. | Source |
| Multiple environments | Home selection, separate stores and same-daemon boundaries. | Source |
| How it works | Components, permissions, costs, receipts and Claude-workflow coverage. | Source |
| Monitoring installation | Immutable releases, MCP loading, managed setup and rollback. | Source |
| Browser reports | Publication, catalog, storage and independent hosting. | Source |
| Daily-use trial | Record real outcomes and friction during ordinary work. | Source |
For development and qualification, see the implementation handoff, managed activation results, source-access results and technical watch reference. The presentation notes and Pages deployment guide explain authoring and publication.
This repository was previously Codex Worker Kit. Compatible agent/skill paths remain in place. Current GitHub documentation tracks main; documentation changes do not repin qualified installed artifacts or update managed source archives. The implementation plan labels completed contracts and current work; Claude context retains the historical usage study. Active sessions supplied passive evidence; experimental messaging/delivery used owned synthetic sessions. Agent Observer is not a runtime dependency.
Reusable workers and execution workflows for a Sol XHigh primary with selective Luna Max delegation. The primary owns decisions, integration, review, and acceptance. Adaptive workflows keep work in the primary when the benefit of a handoff is uncertain. Workflows preserve the selected primary model.
| Role | Model | Reasoning | Responsibility |
|---|---|---|---|
| Recommended primary | gpt-6.1-sol |
xhigh |
Decisions, difficult portions, integration, and acceptance |
Cheaper worker luna-max |
gpt-6-luna |
max |
Clear, bounded investigation and implementation with practical verification |
Optional worker sol-xhigh |
gpt-6.1-sol |
xhigh |
Separable parallel work, context-isolated investigation, or independent review |
Delegate to Luna when the objective, responsibility boundary, and acceptance check are clear and expected savings justify packaging, review, and possible corrections. Confidence concerns the assignment and verification; the primary need not solve every implementation detail first. A substantial specified migration or focused investigation can qualify. Evolving diagnosis, shared decisions, and small remainders often fit primary execution better.
A Sol worker needs a concrete benefit from another execution context. Difficulty beyond Luna normally stays with the primary; the same-model worker is not a stronger tier. Separate files may still depend on shared decisions. An independent review adds a perspective, and its findings still need evidence. Terra and Astra are outside automatic routing.
Escalation requires evidence: an unresolved decision, contradictory observations, a verification gap, or repeated failure without new evidence. Cross-module work, migrations, concurrency, security, file count, or unfamiliarity alone do not exclude Luna. A worker may reason about options; the primary retains authority over decisions outside the execution contract.
| Workflow | Scope | Delegation |
|---|---|---|
$luna-task |
One bounded task; no goal management | Luna only; primary handles difficult portions |
$luna-goal-loop |
A substantial goal across cohesive assignments | Luna only; primary handles difficult portions |
$worker-task |
One bounded task; no goal management | Primary or selective Luna; Sol for independent execution |
$worker-goal-loop |
A substantial goal, routed per assignment | Primary or selective Luna; Sol for independent execution |
The Luna workflows preserve the explicit worker choice without requiring a workflow switch when the primary needs to help. The adaptive workflows choose whether delegation helps and which worker to use. An unavailable optional worker returns work to the primary; an explicitly requested unavailable Luna worker is reported without silently substituting another model.
Activate adaptive workflows with $worker-task, $worker-goal-loop, or an
explicit request to use the named skill. A generic "goal loop" uses the ordinary
workflow. Discussing or editing a skill does not activate it. The adaptive skills
disable implicit invocation in their metadata. Skills are enabled by default
when installed; an existing local disable preference must be changed separately.
Use one cohesive assignment and one writer by default. Reuse workers for related corrections, keep context compact, and transfer write ownership before taking over. Ordinary defects usually call for corrections. Failed attempts without new evidence call for primary diagnosis rather than repeated worker cycling.
Choose components in the setup guide. The worker installation steps install two agents and four worker workflows; the HTML report skill is a separate optional addition. The runtime guide and report guide cover their independent artifacts. With a dotfile manager, update its source and published pins through that owner.
Bryan's chezmoi distribution uses .chezmoiexternals/codex-plus.toml with pinned
Codex Plus URLs and checksums. It retains ~/.local/share/codex-worker-kit as the
managed archive location for compatibility; the installed agent and skill paths
also stay the same. The repository rename does not activate a watcher or change
worker routing.
Select the primary separately in your client or configuration:
model = "gpt-6.1-sol"
model_reasoning_effort = "xhigh"Each custom agent pins both its model and reasoning effort. The client and account must support these models, custom agents, and skills. A task label alone does not select a model; workflows use the named custom agent with a compact execution contract and minimal inherited context. See the official custom agent documentation.
GPT-5.6 Terra compatibility files live under legacy/ and are deprecated. They
retain the terra-xhigh, $terra-task, and $terra-goal-loop identities and do
not alias Sol. Install them separately only to preserve an explicit Terra workflow:
cp codex-plus/legacy/agents/terra-xhigh.toml ~/.codex/agents/
cp -R codex-plus/legacy/skills/. ~/.agents/skills/Copying the current default bundle does not remove previously installed Terra files. When migrating, retire the old agent and both Terra skill directories through the configuration manager or installation method that owns them. Check both supported personal skill roots if an earlier installation used a different one. For archive consumers, publish the kit before changing external URLs and checksums; do not point a deployment at an unpublished checkout.
Run rtk proxy scripts/check after changing the kit. It validates the default and legacy
inventory, agent model/effort selections, skill metadata, and explicit-only
adaptive invocation policies. It does not prove
routing quality or runtime model availability.
Run rtk proxy python3 -B tests/test_check.py for metadata regression checks. The validator
accepts the kit's two-level interface and policy mappings, quoted strings,
and lowercase booleans. Unsupported YAML forms are rejected; extend the validator
deliberately before adding other metadata formats or fields.
Review representative behavior as well:
- A specified migration with clear boundaries and useful checks can suit Luna.
- An unclear failure stays in the primary until a useful investigation can be confidently assigned and verified.
- A deeply coupled design decision stays in the primary, which may delegate resulting implementation when the savings justify the handoff.
- Separable parallel work or a noisy investigation can use Sol when another execution context helps.
- An ordinary failed test goes back for correction; repeated failure without new evidence returns to the primary.
- An explicit Luna workflow never silently creates a Sol or Terra worker.
Compare total primary-plus-worker usage, corrections, elapsed time, and accepted results on representative tasks. Use runtime cost data when available; do not infer savings from the number of Luna assignments or from token rates alone.
HTML report design and the v1 implementation plan retain the original defaults, interfaces, publication recovery and staged acceptance. The implemented and installed behavior is described in the usage guide and report results.
Worker routing design records selective Luna delegation, the separate purpose of Sol workers, and the rollout plan. Worker routing research reviews the evidence, its limits, and a lightweight policy to assess in practice. The active workflow instructions above describe the current implementation.