[Bug]: Codex subagent rows omit model and reasoning-effort labels

[tlmader]

Before submitting

• I searched existing issues and did not find an open issue for this exact Codex behavior.
• I included enough detail to investigate the problem.

Area

apps/desktop

Steps to reproduce

1. Open a thread with the Codex provider.
2. Ask the parent agent to spawn several subagents with specified models and reasoning effort levels.
3. Open the Agents panel.
4. Check the labels on the subagent rows during activity and after an agent becomes idle.

These steps describe the reported workflow. This report uses an observed screenshot; an independent reproduction has not been run.

Expected behavior

Each Codex subagent row shows its actual model and reasoning effort, when those values are available.

The labels remain visible while the agent runs and after it becomes idle or completes.

Actual behavior

The Codex subagent rows appear under DIRECT SPAWNS, but they have no model or reasoning-effort labels.

The screenshot shows five rows with agent names, activity or status, token counts, and elapsed time. Some rows also show a run count. One row shows Idle · resumable, several show assistant message, and another shows collab agent tool call. All five rows omit both labels.

The missing labels prevent users from checking which model and effort each agent uses, including whether requested overrides took effect.

Impact

Minor bug or occasional failure.

Users cannot verify subagent model and effort settings from the Agents panel.

Version or commit

T3 Code (Alpha) 0.0.38 is installed on the reporting machine. The screenshot does not include a version string.

Environment

macOS desktop app, Codex provider. Exact Codex CLI version and agent model settings were not captured with the screenshot.

Screenshots, recordings, or supporting files

The reporter supplied a screenshot of the Agents panel. Its visible state is described above; the image is not attached to this issue.

Workaround

No verified workaround.

Related reports

• [Feature]: Always show each subagent's model in the Agents panel #6120 requested consistent model labels and was converted to a discussion on August 15, 2026. It describes a similar missing-model symptom.
• Agents panel shows the parent session's model/effort for every subagent instead of the subagent's own #7281 is open, but describes Claude subagents showing the parent's model and effort values.
• [Bug]: Subagents not showing in Agents Panel with Codex #8499 is open, but describes entire Codex subagent rows missing from the panel.

This report concerns Codex rows that appear correctly but omit both model and reasoning-effort labels.

[juliusmarminge]

Triage

Valid Codex bug. The Agents panel is rendering these rows correctly; the model and reasoning-effort chips are omitted because those fields never arrive on the child task events.

This is not a duplicate of #7281 (Claude rows show the parent model/effort) or #8499 (Codex rows missing entirely). #6120 described the same missing-label symptom and was converted to a discussion before the Codex fix landed.

What the code does today

#8502 (fix(codex): show sub-agent models) already shipped in desktop 0.0.38. It deliberately leaves unknown Codex child values out instead of inheriting the parent session.

The panel only prints a label when agent.model is set (formatSubagentModelLabel in packages/client-runtime/src/state/subagentRuntime.ts; AgentRow in apps/web/src/components/AgentsPanel.tsx). Desktop wraps that panel.

Codex spawn signals (thread/started, subAgentActivity) do not include model or effort. CodexSessionRuntime therefore does a one-shot background thread/resume (excludeTurns: true) after the child registers, and also listens for thread/settings/updated / model/rerouted. Lookup failures are swallowed, and the lookup is not retried after idle. Official thread/resume params do not define excludeTurns.

That matches the report: rows under Direct spawns with names, activity, tokens, and elapsed time, and no model/effort chips while running or after Idle · resumable.

Suggested fix

Keep the “blank until known” rule. Do not seed the parent model.

Next: reproduce with current Codex CLI and confirm whether thread/resume on a live child returns model / reasoningEffort or errors. If it fails, fix the metadata source (correct resume/read params, retry after the child settles, or read spawn/settings) so task.updated carries the child’s values through CodexAdapter.mapCollabAgentEvent.

Codex CLI version from the reporting machine would help, but it is not required to start.

Surfaces: web + desktop Agents panel. Mobile does not render this panel. Provider-specific to Codex.

[juliusmarminge]

I checked this against current main f8b4c46. The Codex paths are unchanged from d92dca7, where all 113 focused runtime, adapter and client-state tests passed.

Those controls confirm successful child model/effort lookup, newer settings winning over delayed lookup results, and metadata surviving idle/completed states. They use controlled protocol inputs, so they do not reproduce the affected desktop/CLI session. The existing recorded Codex fan-out does not include a child lookup response.

One caution on the earlier diagnosis: excludeTurns being absent from generated types does not prove the CLI rejects it. #8629 documents it as an experimental field. I have not removed it or added speculative retries.

Keeping this open. The next useful evidence is the affected Codex CLI version and a sanitized child thread/resume result or error while active and after idle. Preserve whether model and reasoningEffort were present, but omit prompts, credentials, private paths and private model identifiers. An already-captured trace is enough; please do not post a full session log.

Audited by GPT 6 Astra via Codex in T3 Code.