openclaw - 💡(How to fix) Fix [Bug]: Codex app-server runs ignore agent thinkingDefault unless thinking is explicitly set [1 comments, 2 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

On the Codex app-server run path, agent thinkingDefault is ignored. /status shows Think: medium even when agentCfg.thinkingDefault = "xhigh" and agents.defaults.models[codex/gpt-5.5].params.thinking = "xhigh" are both set, because the default-aware thinking level is computed but not passed into the Codex app-server attempt params.

Related to #67199 (same symptom class — thinkingDefault resolved but not passed — on the HTTP /v1/chat/completions → embedded-runner path). This report is the Codex app-server path.

Steps to reproduce

1. Configure a main agent bound to codex/gpt-5.5 with Codex harness and thinkingDefault: "xhigh":

   {
     "agents": {
       "list": [
         {
           "id": "main",
           "default": true,
           "thinkingDefault": "xhigh",
           "model": { "primary": "codex/gpt-5.5" },
           "embeddedHarness": { "runtime": "codex", "fallback": "none" }
         }
       ],
       "defaults": {
         "models": {
           "codex/gpt-5.5": { "params": { "thinking": "xhigh" } },
           "openai-codex/gpt-5.5": { "params": { "thinking": "xhigh" } }
         }
       }
     }
   }

2. Also set ~/.codex/config.toml:

   model = "gpt-5.5"
   model_reasoning_effort = "xhigh"

3. Open a fresh DM session (Feishu in the original report, any channel routed to this agent will do). Do not use /think and do not touch sessionEntry.thinkingLevel.

4. Run /status.

Expected behavior

The Codex app-server turn/start request should receive effort: "xhigh" because agentCfg.thinkingDefault = "xhigh" is configured. /status should show:

Model: codex/gpt-5.5
Think: xhigh

Actual behavior

A fresh DM session shows:

Model: codex/gpt-5.5
Think: medium

openclaw config validate passes, all three levels of thinking config (agent.thinkingDefault, agents.defaults.models.codex/gpt-5.5.params.thinking, agents.defaults.models.openai-codex/gpt-5.5.params.thinking) are set to "xhigh", and ~/.codex/config.toml has model_reasoning_effort = "xhigh". Only an explicit /think xhigh or a persisted sessionEntry.thinkingLevel override brings the session back to xhigh.

Root cause (source-grounded)

In dist/get-reply-DIRI0vf6.js, OpenClaw computes two values:

const resolvedThinkLevel = directives.thinkLevel ?? targetSessionEntry?.thinkingLevel;
const resolvedThinkLevelWithDefault =
    resolvedThinkLevel ??
    await modelState.resolveDefaultThinkingLevel() ??
    agentCfg?.thinkingDefault;

but the Codex app-server attempt is built with the non-default-aware value:

thinkLevel: resolvedThinkLevel,  // ← ignores thinkingDefault

Downstream in dist/extensions/codex/run-attempt-sbLvYJcF.js:

function buildTurnStartParams(params, options) {
  return {
    threadId: options.threadId,
    input: buildUserInput(params, options.promptText),
    cwd: options.cwd,
    approvalPolicy: options.appServer.approvalPolicy,
    approvalsReviewer: options.appServer.approvalsReviewer,
    model: params.modelId,
    effort: resolveReasoningEffort(params.thinkLevel)
  };
}

When params.thinkLevel is undefined, effort becomes null, and the Codex app-server falls back to its own default (observed as medium).

Minimal fix

Pass the default-aware value into the Codex app-server attempt params:

- thinkLevel: resolvedThinkLevel,
+ thinkLevel: resolvedThinkLevelWithDefault,

Local hotfix verification

After patching dist/get-reply-DIRI0vf6.js with the diff above and running node --check to validate syntax, plus patching the active session with:

openclaw gateway call sessions.patch \
  --params '{"key":"agent:main:feishu:default:direct:<peerId>","thinkingLevel":"xhigh"}' \
  --json

/status then reports Think: xhigh for Codex app-server runs without any per-message /think directive.

Suggested regression test

Codex app-server run attempt where:

1. agentCfg.thinkingDefault = "xhigh"
2. No /think directive
3. No sessionEntry.thinkingLevel
4. Model is codex/gpt-5.5

Expected:

attemptParams.thinkLevel === "xhigh"
// and
turnStartParams.effort === "xhigh"

Impact

Users can correctly configure thinkingDefault = "xhigh" and still silently receive medium reasoning on Codex app-server sessions. /status disagrees with the declared configuration, reasoning effort is silently downgraded, and the bug is especially confusing because /think xhigh and explicit session patches both work — only the declarative thinkingDefault is ignored on this path.

OpenClaw version

2026.4.22 (00bd2cf)

Operating system

macOS (tested on an M-series Mac Mini with fnm-managed Node v22.22.0)

Install method

npm

Model

codex/gpt-5.5 (Codex app-server path)