[Bug] v3.8.1: Thought-level changes silently discarded after first selection (stale_runtime_refresh_ignored) + reasoning_effort never sent for custom OpenAI-compatible providers

提交前确认 · Pre-submission checklist

• 我已搜索过现有 issue，确认这不是重复 / I searched existing issues and confirmed this isn't a duplicate.
• 我已阅读 CONTRIBUTING.md / I've read CONTRIBUTING.md.

问题类别 · Category

模型设置 / 切换 · Model config

涉及的 Agent 框架 · Agent framework

ZCode Agent（自研）

严重程度 · Severity

阻塞使用 · Blocking (无法使用核心功能 / core function unusable)

复现频率 · Reproducibility

必现 · Always

问题描述 · Description

Custom OpenAI-compatible models (stealth/ox-alpha via OpenRouter, x-preview-f-free via opencode zen, qwen/qwen3.8-max-free via tokenrouter) configured with reasoning levels (provider.<id>.models.<model>.reasoning + modelCatalog.overrides) show a Thought-Level selector, but it never controls actual requests.

Bug A — level changes silently discarded after the first selection.
First selection in a session applies (previousThoughtLevel: null). Every later change logs session.reasoning_effort.updated but the runtime keeps the old value and the dropdown reverts. Host log shows the cause — 13 occurrences today, blocking BOTH directions:

warn zcode_protocol.session_thought.stale_runtime_refresh_ignored
"Ignored stale session runtime thought refresh"
{"currentThought":"high","incomingThought":"max"}

A retry loop repeats the same pair every few minutes without converging; later selections log a stale previousThoughtLevel pinned to an old value — two state layers permanently diverged.

Bug B — reasoning_effort never reaches the wire.
Even in fresh chats with Max explicitly selected (host logs confirm runtimeModel carries thoughtLevel:"max"), outgoing request bodies never contain reasoning_effort. Verified via model-I/O capture: 0 of 265 requests in one day contained the parameter, across all sessions/models/config paths tried (v2 config, ~/.zcode/cli/config.json, legacy provider options).

Bug C — model selection never persists. Every new session creation logs session.model_selection.persist_failed {"error":"FOREIGN KEY constraint failed"} (77× today) → last-used model/level is never remembered.

Related: #306, #223, #319, #202.

复现步骤 · Steps to reproduce

1. Add a custom OpenAI-compatible provider/model with reasoning variants + modelCatalog.overrides (see attached config-redacted.txt)
2. Create a new chat with that model, select Thought Level = Max → applies correctly (first selection)
3. Send one message
4. Switch Max→High→Low→Max repeatedly → dropdown keeps reverting to High, Max never sticks again
5. Inspect outgoing request bodies (model-I/O logs) → no reasoning_effort parameter at any point

期望表现 · Expected behavior

A selected Thought Level must always apply to the session runtime, and every subsequent chat-completions request must carry reasoning_effort: .

实际表现 · Actual behavior

Only the first selection per session takes effect. Afterwards all changes are discarded by zcode_protocol.session_thought.stale_runtime_refresh_ignored (13x/day), the UI reverts to the old level, and 0 of 265 captured outgoing requests contained reasoning_effort — even though the host log confirms thoughtLevel:"max" was applied. Additionally, session.model_selection.persist_failed ("FOREIGN KEY constraint failed") fires on every session creation (77x/day).

ZCode 版本 · ZCode version

v3.8.1 (3.8.1.5310)

设备 / 系统 / 浏览器 · Device / OS / Browser

Windows 10 x64 (build 26100) · ZCode Desktop

截图 / 录屏 / 日志 · Screenshots / Recordings / Logs

Attached:

• 🎥 Screen recording: fresh-chat Max selection OK → after first message, repeated changes revert
• 📄 zcode-2026-08-21-bug-evidence.txt — filtered log, 564 events, PII redacted:
  195× reasoning_effort.updated · 78× persist_failed(FOREIGN KEY) ·
  13× stale_runtime_refresh_ignored · 3× thought_unavailable-on-resume
• 📄 config-redacted.txt — exact repro config, all credentials redacted

Key lines from the log:

warn zcode_protocol.session_thought.stale_runtime_refresh_ignored
"Ignored stale session runtime thought refresh"
{"currentThought":"high","incomingThought":"max"}   ← same pair retried ~every 5 min, never converges

warn session.model_selection.persist_failed
{"error":"FOREIGN KEY constraint failed"}

log.zip
https://github.com/user-attachments/assets/d39c9bf6-78a3-480c-b50e-13f254079490

Start with the attached config-redacted.txt and reproduce the selection sequence while reviewing zcode_protocol.session_thought.stale_runtime_refresh_ignored and session.model_selection.persist_failed in the supplied logs. Use model-I/O capture to verify the selected level reaches every request and determine whether model selection persists without the reported foreign-key error; done means all level changes stick, reasoning_effort is sent, and persistence succeeds.