Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

Auto-compaction in OpenClaw 2026.6.1 does not resolve a configured compaction model alias before dispatch, causing the alias gpt54mini to be sent as openai/gpt54mini and rejected as an unknown model.

Steps to reproduce

1. Start OpenClaw 2026.6.1 with a model alias and compaction override shaped like this sanitized config:

   {
     "models": {
       "providers": {
         "openai": {
           "api": "openai-chatgpt-responses",
           "agentRuntime": { "id": "pi" }
         }
       }
     },
     "agents": {
       "defaults": {
         "model": {
           "primary": "gpt55",
           "fallbacks": ["gpt54"]
         },
         "models": {
           "openai/gpt-5.5": {
             "alias": "gpt55",
             "agentRuntime": { "id": "pi" }
           },
           "openai/gpt-5.4": {
             "alias": "gpt54",
             "agentRuntime": { "id": "pi" }
           },
           "openai/gpt-5.4-mini": {
             "alias": "gpt54mini",
             "params": { "thinking": "high" },
             "agentRuntime": { "id": "pi" }
           },
           "openai/*": {
             "agentRuntime": { "id": "pi" }
           }
         },
         "compaction": {
           "reserveTokens": 30000,
           "keepRecentTokens": 20000,
           "maxHistoryShare": 0.1,
           "model": "gpt54mini",
           "notifyUser": true
         }
       }
     }
   }

2. Run a long openai/gpt-5.5 session until embedded auto-compaction is triggered.

3. Observe the gateway/agent log line shown below.

Expected behavior

The auto-compaction path should resolve agents.defaults.compaction.model = "gpt54mini" through the same model alias resolver used by normal agent model selection, then dispatch compaction to canonical openai/gpt-5.4-mini.

Actual behavior

Auto-compaction dispatches or reports the alias as openai/gpt54mini, which is not a configured model id and is rejected as unknown.

OpenClaw version

2026.6.1

Operating system

macOS 15.x / Darwin 24.x

Install method

npm global, Node.js 24

Model

Primary session model: openai/gpt-5.5

Compaction override: alias gpt54mini for canonical openai/gpt-5.4-mini

Provider / routing chain

OpenClaw -> OpenAI plugin -> ChatGPT Responses runtime (agentRuntime.id = "pi")

Additional provider/model setup details

The alias is defined under agents.defaults.models as:

"openai/gpt-5.4-mini": {
  "alias": "gpt54mini",
  "params": { "thinking": "high" },
  "agentRuntime": { "id": "pi" }
}

Changing only agents.defaults.compaction.model from gpt54mini to canonical openai/gpt-5.4-mini avoids the failure.

Logs, screenshots, and evidence

15:17:24+02:00 warn agent/embedded {"subsystem":"agent/embedded"} auto-compaction failed for openai/gpt-5.5: Unknown model: openai/gpt54mini

Sanitized verification after workaround:

agents.defaults.compaction.model = "openai/gpt-5.4-mini"
config valid: true
issues: []
warnings: []

Impact and severity

• Affected users/systems/channels: OpenClaw sessions using agents.defaults.compaction.model set to a configured alias instead of a fully-qualified provider/model id.
• Severity: Medium. Normal chat continues, but auto-compaction fails, so long sessions can lose the intended context-management behavior.
• Frequency: Observed when auto-compaction runs with the alias override shown above.
• Consequence: Context compaction does not complete until the config is changed to a canonical model id.

Additional information

This appears to be either:

1. the embedded auto-compaction path not calling the normal model alias resolver, or
2. the call site passing the configured alias into provider/model normalization before alias resolution.

The workaround is to avoid aliases in agents.defaults.compaction.model and use openai/gpt-5.4-mini directly.

No credentials, user paths, account names, chat IDs, or local file paths are included in this report.