A hard dollar cap per Qwen Code session with ${session_id} in customHeaders

[domondi1]

The ${session_id} placeholder in customHeaders (with outboundCorrelation.allowDynamicHeaderValues) is enough to give every Qwen Code session its own hard dollar ceiling, config only. Point a provider at a local gateway, send the session id as the gateway's run id and a static budget, and the gateway refuses a call with HTTP 402 before it reaches the provider once that session's budget is used. Useful for long headless runs and loops where one session can spend far more than you meant. Disclosure: I maintain the gateway (Inferrail, open source).

pip install inferrail
export OPENAI_API_KEY=sk-...
inferrail serve --quickstart --app-mode      # http://127.0.0.1:8000

~/.qwen/settings.json:

{
  "modelProviders": {
    "openai": [
      {
        "id": "gpt-4o-mini",
        "name": "gpt-4o-mini (capped)",
        "envKey": "INFERRAIL_PLACEHOLDER_KEY",
        "baseUrl": "http://127.0.0.1:8000/v1",
        "generationConfig": {
          "customHeaders": {
            "X-Inferrail-Attribute-Work-Id": "qwen-${session_id}",
            "X-Inferrail-Budget-Usd": "2.00"
          },
          "samplingParams": { "max_tokens": 4000 }
        }
      }
    ]
  },
  "outboundCorrelation": { "allowDynamicHeaderValues": true },
  "security": { "auth": { "selectedType": "openai" } },
  "model": { "name": "gpt-4o-mini" }
}

export INFERRAIL_PLACEHOLDER_KEY=unused      # the real key stays in the gateway
qwen -p "..."
inferrail report --by work_id                # one row per session: calls, tokens, dollars

What I checked (Qwen Code 0.25.0 from npm, inferrail 0.4.15, qwen -p against a stand-in OpenAI endpoint with fixed token counts):

• Two headless sessions: each got its own id (qwen-<session id>), 3 calls each, recorded separately with tokens and cost.
• A budget too small for the next call: [API Error: 402 budget 'work_id:qwen-...:per_work' ... would be exceeded], exit 1, about 2 s, 1 request, nothing reached the upstream, no retry loop.

Caveats: each call reserves its worst case up front from max_tokens, so keep samplingParams.max_tokens realistic. A dollar budget needs a price for the model on the gateway side (most OpenAI and Anthropic models are built in, inferrail models; for DashScope or other providers you add the endpoint and a price in inferrail.yaml). As the docs note, enabling dynamic values sends the session id to that provider entry's host, here only the local gateway, which stores usage and cost per call, never prompts or code. I didn't test /new and /resume; per the docs they rotate the id, which would start a fresh budget.