ClawRouter continues consuming credits despite primary model set to Codex
Problem Description
After configuring openai-codex/gpt-5.3-codex as the primary model in the configuration file, ClawRouter continues to intercept and consume credits for every request, completely ignoring the model setting.
Documented Credit Loss
Time | Remaining Credit | Loss
-- | -- | --
11:42 | $5.45 | -
11:56 | $4.02 | -$1.43
12:01 | $3.78 | -$0.24
12:07 | $3.19 | -$0.59
Total | | -$2.26+
Current Configuration (agents.defaults.model)
{
"primary": "codex",
"fallbacks": ["auto", "openrouter/auto", "gemini-2.5-flash"]
}
Expected Behavior
- Requests should use OpenAI Codex (OAuth authentication, included in subscription)
- No ClawRouter credit consumption for normal chat requests
Actual Behavior
- [ClawRouter] Received model: "auto" in logs despite
primary: "codex" - Usage of paid models (
xai/grok-4-fast-reasoning, moonshot/kimi-k2.5) - Cost per call: ~$0.0033
- Current session shows
model: "auto" instead of "codex"
Relevant Logs
[ClawRouter] Received model: "auto" -> normalized: "auto", isAuto: true
[ClawRouter] Tools detected (23), forcing agentic mode
[ClawRouter] Trying model 1/2: moonshot/kimi-k2.5
[ClawRouter] Success with model: moonshot/kimi-k2.5
[plugins] [REASONING] moonshot/kimi-k2.5 $0.0033 (saved 99%)
Steps to Reproduce
- Configure
primary: "codex" in agents.defaults.model - Restart OpenClaw (SIGUSR1)
- Verify
session_status shows model: "openai-codex/gpt-5.3-codex" - Execute any request
- Observe logs still show
[ClawRouter] Received model: "auto"
Attempted Solutions
- Changed
primary from blockrun/auto to codex - Removed
auto from fallbacks - Set
session_status with model: "codex" for current session - Restarted OpenClaw multiple times
- Completely disabled ClawRouter plugin - not attempted for fear of losing functionality
Environment
- OpenClaw Version: 2026.2.9 (33c75cb)
- OS: Linux 6.12.63+deb13-amd64 (x64)
- Node: v22.22.0
Additional Context
The issue appears to be that even when the session status shows the correct model (openai-codex/gpt-5.3-codex), the underlying tool calls or agent processing still routes through ClawRouter with model: "auto". This suggests the model override is not being properly propagated to all subsystems, particularly when tools are involved ("Tools detected (23), forcing agentic mode").
Possible Root Causes
- Tool calls force
agentic mode which overrides user model preference - Session model override doesn't propagate to ClawRouter plugin
auto alias in fallbacks causes ClawRouter to activate despite primary being set- Configuration patch applies to defaults but active sessions retain cached model
Request
Please investigate why ClawRouter continues to intercept requests when a specific non-ClawRouter primary model is configured. The expected behavior is that explicit model selection should bypass ClawRouter entirely unless explicitly requested via /model auto.
ClawRouter continues consuming credits despite primary model set to Codex
Problem Description
After configuring
openai-codex/gpt-5.3-codexas the primary model in the configuration file, ClawRouter continues to intercept and consume credits for every request, completely ignoring the model setting.Documented Credit Loss
Time | Remaining Credit | Loss -- | -- | -- 11:42 | $5.45 | - 11:56 | $4.02 | -$1.43 12:01 | $3.78 | -$0.24 12:07 | $3.19 | -$0.59 Total | | -$2.26+Current Configuration (
agents.defaults.model)Expected Behavior
Actual Behavior
primary: "codex"xai/grok-4-fast-reasoning,moonshot/kimi-k2.5)model: "auto"instead of"codex"Relevant Logs
Steps to Reproduce
primary: "codex"inagents.defaults.modelsession_statusshowsmodel: "openai-codex/gpt-5.3-codex"[ClawRouter] Received model: "auto"Attempted Solutions
primaryfromblockrun/autotocodexautofrom fallbackssession_statuswithmodel: "codex"for current sessionEnvironment
Additional Context
The issue appears to be that even when the session status shows the correct model (
openai-codex/gpt-5.3-codex), the underlying tool calls or agent processing still routes through ClawRouter withmodel: "auto". This suggests the model override is not being properly propagated to all subsystems, particularly when tools are involved ("Tools detected (23), forcing agentic mode").Possible Root Causes
agentic modewhich overrides user model preferenceautoalias in fallbacks causes ClawRouter to activate despite primary being setRequest
Please investigate why ClawRouter continues to intercept requests when a specific non-ClawRouter primary model is configured. The expected behavior is that explicit model selection should bypass ClawRouter entirely unless explicitly requested via
/model auto.