Telegram routed agent sessions do not honor agents.list[].model per-agent override; runtime still uses fallback/default model

Summary

When I create a dedicated agent with its own Telegram bot binding and set agents.list[].model.primary to an Ollama model, the routed Telegram session does not actually run on that configured per-agent model.

Instead:

• the session appears to target the configured Ollama model
• session_status reports the configured Ollama model
• but the actual session log and runtime replies are generated by openai-codex/gpt-5.4

This makes the active runtime model differ from the model reported by status/tooling.

Environment

• OpenClaw version: 2026.3.11
• Host: macOS arm64
• Channel: Telegram
• Providers involved:
  • openai-codex/gpt-5.4
  • ollama/gpt-oss:120b
• Ollama endpoint:
  • remote Ollama at http://172.28.28.158:11434
• Routing style:
  • top-level bindings[] mapping Telegram accountId to a dedicated agentId

Expected behavior

A Telegram message routed to an agent with:

{
  "id": "ollama-agent",
  "model": {
    "primary": "ollama/gpt-oss:120b",
    "fallbacks": [
      "openai-codex/gpt-5.4",
      "github-copilot/gpt-4o"
    ]
  }
}

should actually execute on ollama/gpt-oss:120b, unless that model truly fails.

If fallback occurs, the failure reason should be surfaced clearly, and status output should distinguish configured vs active runtime model.

Actual behavior

For a Telegram-routed dedicated agent:

• the bot replies successfully
• session_status reports:
  • Model: ollama/gpt-oss:120b
  • Fallback: openai-codex/gpt-5.4
• the agent then answers ollama/gpt-oss:120b when asked which model it is using
• however, the actual session log shows:
  • provider: openai-codex
  • model: gpt-5.4
• openclaw status also shows the active session model as gpt-5.4

So the configured model / reported model and the actual generation model diverge.

Reproduction steps

1. Configure a custom Ollama provider reachable from OpenClaw.
2. Add a dedicated agent in agents.list[], for example ollama-agent.
3. Set:

"model": {
  "primary": "ollama/gpt-oss:120b",
  "fallbacks": [
    "openai-codex/gpt-5.4",
    "github-copilot/gpt-4o"
  ]
}

1. Bind a Telegram bot/account to that agent:

{
  "type": "route",
  "agentId": "ollama-agent",
  "match": {
    "channel": "telegram",
    "accountId": "ollama-agent"
  }
}

1. Send a message to that Telegram bot.
2. Observe:
   • the reply works
   • openclaw status shows the session model as gpt-5.4
   • the session JSONL entries show provider=openai-codex, model=gpt-5.4
   • session_status inside the same session reports ollama/gpt-oss:120b

Relevant config shape

Simplified config:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "openai-codex/gpt-5.4",
        "fallbacks": [
          "ollama/gpt-oss:120b",
          "github-copilot/gpt-4o"
        ]
      },
      "models": {
        "openai-codex/gpt-5.4": {},
        "github-copilot/gpt-4o": {},
        "ollama/gpt-oss:120b": {}
      }
    },
    "list": [
      {
        "id": "ollama-agent",
        "workspace": "...",
        "agentDir": "...",
        "model": {
          "primary": "ollama/gpt-oss:120b",
          "fallbacks": [
            "openai-codex/gpt-5.4",
            "github-copilot/gpt-4o"
          ]
        }
      }
    ]
  },
  "bindings": [
    {
      "type": "route",
      "agentId": "ollama-agent",
      "match": {
        "channel": "telegram",
        "accountId": "ollama-agent"
      }
    }
  ],
  "channels": {
    "telegram": {
      "accounts": {
        "ollama-agent": {
          "botToken": "configured"
        }
      }
    }
  }
}

Evidence

openclaw status

Shows a routed session like:

• agent:ollama-agent:telegram:direct:...
• model shown as gpt-5.4

Session metadata

The session metadata includes:

• fallbackNoticeSelectedModel = ollama/gpt-oss:120b
• fallbackNoticeActiveModel = openai-codex/gpt-5.4

Session JSONL

The actual assistant messages are recorded with:

• provider: openai-codex
• model: gpt-5.4

session_status inconsistency

Inside the same session, session_status reports:

• Model: ollama/gpt-oss:120b
• Fallback: openai-codex/gpt-5.4

But the actual generation entries are still gpt-5.4.

Why this seems buggy

This appears to be one of:

1. agents.list[].model is not actually applied for Telegram-routed sessions
2. runtime falls back immediately but does not expose the actual failure reason clearly
3. session_status reports configured/selected model rather than active generation model
4. model selection and model reporting are inconsistent for routed agent sessions

Suggested fixes

1. Ensure Telegram routed sessions actually honor agents.list[].model
2. If fallback occurs before first generation, surface the real reason clearly
3. Make session_status distinguish:
   • configured target model
   • selected model
   • active generation model
4. Align openclaw status, session_status, and JSONL runtime metadata

Additional note

I also observed similar behavior when trying to apply per-agent model overrides to main, where new Telegram sessions still used the global default model rather than the agent-specific primary model. That may indicate this is not limited to a single custom agent, but a broader issue with routed session model resolution.