Background session cold start ignores saved Medium reasoning effort and runs gpt-6-astra at High

**提交人**: liangdong
**客户端版本**: 0.1.82

---

## Summary

Background/delegated Codex sessions reportedly start with High reasoning effort after cold start or lazy resume, although the persisted session setting and UI show Medium. This causes unexpected usage consumption and makes continued background development difficult to control.

These observations and the suspected cause come from the reporter’s investigation; they have not been independently reproduced in this submission.

## Reported affected environment

- Observed September 15, 2026, on macOS.
- Agent: Codex; model: `gpt-6-astra`.
- Installed source commit: `db19a86bf319b195763d9929dd436a0844d4587f`.
- Build timestamp: `2026-09-14T17:09:36.760+08:00`.
- Workflow: background tasks, delegated sessions, and lazy resume.

## Expected versus actual

A session persisted with `model=gpt-6-astra` and `effort=medium` should use Medium from the first runtime turn, including `collaboration_mode.settings.reasoning_effort`. This should remain consistent across creation, cold start, lazy resume, continuation, and pause/resume.

The reporter found multiple affected child sessions between approximately 05:13 and 10:38 China Standard Time (UTC+8) on September 15. Persisted session rows and teammate/default settings showed Medium, while actual Codex turn contexts showed High for both effort fields.

## Suspected cause (needs verification)

The reporter identified a lazy-resume/cold-handoff path in `apps/desktop/src/main/maker-ipc/register.ts` that builds session creation options without forwarding persisted effort. In `packages/maker-core/src/agents/codex/index.ts`, the reported initialization is `opts.effort ?? 'high'`.

If that path omits effort, the runtime may default to High. Later hydration into the UI/bridge/store may not update the native runtime before its first turn. The existing `reconcileCreateOptsAgainstDb` helper reportedly is not called before session creation on this path.

## Suggested investigation and regression coverage

- Reconcile persisted effort before native session creation, preferably through the canonical helper rather than duplicated field mapping.
- Audit background dispatch, delegated continuation, paused/archived resume, and remote/Orca handoff for similar omissions.
- Preserve intentional runtime overrides and define their precedence.
- Test cold-start/lazy-resume of a persisted Medium session against the first actual turn context, asserting both effort fields remain Medium.
- Verify model, provider, permissions, working directory, and resume ID remain unchanged.
- Test persisted High and intentional live effort changes so the fix does not force every session to Medium.

## Additional observation: pause semantics

The same investigation reported a background task marked waiting/paused/stopped with its queue held at approximately 11:00:30 China Standard Time on September 15, while child-tool evidence files continued updating until approximately 11:02. A later process check at approximately 11:17 found no remaining related Godot/build/debug processes.

This may reflect graceful safe-point pausing rather than immediate child-process termination; the interpretation needs verification. Please clarify whether pause completion guarantees process-group cleanup, or explicitly indicate that the model session is paused while child tools may still be finishing or cleaning up.

No raw logs, database, usage CSV, screenshots, or private project files are attached; observations are summarized to avoid exposing unrelated session information.
---
## 提交时的任务环境

仅代表提交时快照,不一定是故障环境。OS 来自提交客户端本机,不含 SSH 远端主机;Harness / 模型来自当前任务。与运行环境无关的反馈可忽略本段。
**版本区域**: CN
**OS**: darwin arm64 (25.5.0)
**Harness**: Codex
**Model ID**: ` gpt-6-astra `
**界面语言**: zh-CN

Start with apps/desktop/src/main/maker-ipc/register.ts and packages/maker-core/src/agents/codex/index.ts, then inspect whether reconcileCreateOptsAgainstDb is used before session creation. Trace background, delegated, paused/resume, and remote handoff entry points. Done means a persisted Medium session uses Medium in both effort fields on its first turn while High and intentional overrides remain unchanged; verify pause cleanup semantics separately.

Written by the indexing model from the issue text.