Current preview guide · Source on GitHub

Using Codex Plus day to day

Codex Plus adds explicit worker workflows, watches, later checks, peer messages
and browser reports to the ordinary Codex CLI/TUI. If this is your first visit,
use the setup guide to choose the components you need.
For the managed Snap and Starship preview, open a fresh ordinary session in
the intended environment; monitoring and report MCP tools load from its
configuration. Existing opt-outs remain effective.

Describe what you need in normal language. The model chooses the source,
filtering, cadence and continuation. You do not need a thread UUID, exact session
title, JSON prompt or polling recipe. If the source or destination is unclear,
the model should ask a short clarification.

Choose the environment first

On Snap, codex / codex agents use OSS, while codex-work /
codex-work agents use work. codex-oss is retired. A native --profile
selects settings inside one home; it does not provide this separation.

Peers are discovered only in the selected daemon. A work conversation cannot
message an OSS conversation through the peer tools, even though both are on
Snap. Reports also use their own selected store and endpoint. Existing sessions
may keep older MCP code until reconnect; a source upgrade alone does not reload
an active process. See environment selection.

Native completion/approval alerts remain available. Codex Plus adds no terminal
relay, custom hook or desktop-notification publisher.

Start with a real task

For delegated implementation or investigation, choose an explicit
Luna or adaptive worker workflow. Use $luna-task for one bounded
assignment or $luna-goal-loop for a substantial goal with Luna. Your primary
keeps decisions and review. $worker-task and $worker-goal-loop let the primary
choose whether delegation helps. A generic “goal loop” does not select a worker.

Supply an endpoint, run or project reference when it is not already established
in the conversation. Keep credentials in the existing configured connection or
an authorized private reference rather than pasting them into the request.

Watch another session

Start two independent sessions within the same daemon on one host. One can do
the packaging work while the other observes it. A subagent spawned by a parent
is a different relationship; it is excluded from ordinary peer discovery.

The observing model uses project and recent task context to find the right
session. If two plausible sessions remain, it asks which one. Once selected,
“that session” keeps the same identity even if its title changes. A lost reference
does not silently switch to a different session.

Setup should give a short confirmation of the selected work and useful condition.
Follow-up should report changes relevant to that condition. Routine checks,
ordinary phase changes and successful renewals can remain quiet.
Later updates should explain a meaningful finding, such as readiness or a request
for input. Observation does not resume, interrupt or stop the other session.
Idle status alone does not prove that its task succeeded.

“Tell that session to include the risks” sends an attributed request. The receiver
evaluates it against its existing task and direct user instructions. A receipt
can show received, accepted or declined; accepted does not prove the requested
work finished. Ask “Did it accept that request?” for the cached handling receipt.

Peer communication is currently within one daemon. Installing Codex Plus on two
hosts does not add Snap-to-Starship native messaging.

Watch an endpoint or inspect through MCP

For an accessible endpoint, the model can select repeated short commands or a
continuing script. Unchanged reads can stay quiet without a model turn per poll.
Useful emitted observations return to your conversation for interpretation.

For a configured MCP source, the model schedules native inspection turns and
calls the selected tool through its normal approved connection. Each check has
model cost. “Check it less often” can guide the next cadence; “Stop watching it”
ends the selected watch. Missing access is explained rather than bypassed.

An observer has a finite lease and event allowance. Expiry means observation
ended, not that the task completed. The model can explicitly continue useful
observation under your existing request or report that it ended. Explicit stop
forbids renewal or replacement.

Check again later

“Check this endpoint again in a little while” registers one future reassessment
and returns now. The model should not hold the foreground turn open with sleep
or repeatedly poll while waiting. A due check follows native conversation
control and can wait behind active work; it is not an exact-time alarm.

The default expiry leaves up to two minutes for delivery after the requested
delay, within the configured lifetime limit. The receipt states that window;
the model can choose an explicit longer expiry when the request needs it.
If active work outlasts expiry, the reminder records that outcome.

A reminder is separate from a watch. Ask to list or cancel the relevant reminder
when you want to stop it. The preview does not provide a recurring calendar or
cron service.

Inspect follow-up health

Ask “What are you watching?” for bounded active and unresolved watch status.
Source failures, observation gaps, control holds and uncertain delivery are
reported with cached receipt freshness. An offline coordinator makes a previous
running receipt unknown. Ask for watch history to page through settled entries;
history and handling receipts remain retained.

Return to unfinished work

The following inspection improvements are available in preview.10 and later; see
released and source features.

For example, a recap could say:

I’m still watching packaging for a blocker. The CI reminder was received, but
its follow-up turn hasn’t finished. Packaging accepted the release-note request;
completion is unconfirmed. These are cached receipts, not a fresh check of CI
or packaging.

This is an authored example. Inspection reads the current receipts before
summarizing; it leaves existing follow-ups unchanged and does not poll a source,
resend a message or restart an observer. Saved references survive context
compaction in this conversation. The recap covers its own recorded follow-ups;
it does not inventory other conversations, environments or hosts.

A reminder's acknowledgement confirms receipt before follow-up work finishes.
Pending or unknown native turn outcomes remain visible. A failed or interrupted
follow-up needs attention until you explicitly cancel that reminder; its receipts
stay in history. Cancellation preserves an already consumed native turn, whose
unknown outcome remains visible until a terminal receipt is recorded. A missing
receipt stays unknown even if the coordinator has exited; it is not inferred as
success or automatically retried. Unresolved delivery/withdrawal also stays visible.

Detailed inspection shows up to three event receipts with unresolved delivery
first, plus bounded renewal details. A receipt records observation, admission or
handling; it does not establish that the watched work succeeded.

Stop, interrupt or leave

Cancellation cannot undo an already handled message or an external action.
The preview does not intercept native /exit to add a background-task
confirmation dialog. Treat viewer detachment and daemon shutdown separately.

Make a browser report

Ask for a report when a result benefits from a shareable view. The report skill
and four report MCP tools publish immutable static HTML, list/get reports and
delete a specifically selected report. Reports are grouped by project.

Publishing can start or reuse the separately managed report server; read-only
listing does not start it. The returned link remains useful independently of
TUI attachment while its host and serving endpoint are online. From an SSH Kitty
terminal, Ctrl+Shift-click opens it in the local desktop browser.

See the report guide for hosting and storage details.

What daily use is testing

Useful watches, messages, endpoint access, configured MCP inspections and explicit
stop passed bounded native acceptance on both hosts. Those checks used synthetic
sources and owned sessions. Real credentials/provider recovery, long unattended
operation and your judgment of the interaction still need daily-use evidence.

Use the soak checklist during normal work. See
how the package works for component boundaries, Claude
workflow coverage and evidence, or the browser walkthrough
for illustrative examples. Low-level CLI diagnostics remain in the
technical watch guide.