BYOK generic-chat-completion-api can crash during streaming with `null is not an object (evaluating 'H.usage')`

[chindris-mihai-alexandru]

Summary

Droid CLI crashes during BYOK streaming for a custom generic-chat-completion-api model after the request has already reached first token.

In my case, the model is claude-opus-4-6 routed through AgentRouter, but the failure appears to be in Droid's stream/response handling rather than an early auth/config issue.

The user-visible failure is:

TypeError: null is not an object (evaluating 'H.usage')
MetaError: Exec failed

Environment

• Droid CLI: 0.100.0
• OS: macOS darwin 25.4.0
• Terminal: Warp
• Mode: droid exec
• Model provider: BYOK custom model with provider: "generic-chat-completion-api"

Config shape

I used a temporary settings overlay rather than editing my real ~/.factory/settings.json.

{
  "customModels": [
    {
      "model": "claude-opus-4-6",
      "displayName": "AgentRouter-Claude-Opus-4-6",
      "baseUrl": "https://agentrouter.org/v1",
      "apiKey": "${AGENT_ROUTER_TOKEN}",
      "provider": "generic-chat-completion-api",
      "maxOutputTokens": 64000
    }
  ]
}

Invocation:

droid exec \
  --settings /path/to/temp-settings.json \
  --model 'custom:AgentRouter-Claude-Opus-4-6-0' \
  --output-format debug \
  'Reply with the single word OK.'

Reproduction

1. Configure a BYOK custom model using provider: "generic-chat-completion-api"
2. Point it at an OpenAI-compatible endpoint that accepts the request and starts streaming back content
3. Run droid exec against that custom model
4. Observe that Droid gets past request start and records time-to-first-token
5. Then Droid crashes with H.usage

Expected behavior

If the upstream endpoint is valid enough for Droid to start streaming tokens, Droid should either:

• complete successfully, or
• surface a normal provider error

It should not crash internally with:

null is not an object (evaluating 'H.usage')

Actual behavior

droid exec exits with:

{"type":"error","source":"agent_loop","message":"null is not an object (evaluating 'H.usage')"}
{"type":"error","source":"cli","message":"Exec failed"}

Evidence that the request gets far enough to stream

From ~/.factory/logs/droid-log-single.log:

[2026-04-20T22:18:49.079Z] INFO: [LLM] sendMessage | Context: {"messageThreadLength":2,"toolCount":9,"modelId":"claude-opus-4-6",...}
[2026-04-20T22:18:52.593Z] INFO: [metrics_log_chat_client_time_to_first_token] | Context: {"modelId":"claude-opus-4-6","modelProvider":"generic-chat-completion-api","apiProvider":"fireworks","value":3.513,...}
[2026-04-20T22:18:52.677Z] WARN: [useLLMStreaming] LLM error | Context: {"error":{"name":"TypeError","message":"null is not an object (evaluating 'H.usage')","stack":"TypeError: null is not an object (evaluating 'H.usage')'..."},"attempt":1,"modelId":"claude-opus-4-6",...}
[2026-04-20T22:20:31.219Z] ERROR: [Agent] runAgent error
TypeError: null is not an object (evaluating 'H.usage')
    at X$L (src/models/chunkProcessing.ts:903:7)
    at <anonymous> (src/hooks/createLLMStreamingCore.ts:1511:9)

And the debug output from droid exec --output-format debug ends with:

{"type":"system","subtype":"init",...,"model":"custom:AgentRouter-Claude-Opus-4-6-0","reasoning_effort":"none"}
{"type":"message","role":"user",...,"text":"Reply with the single word OK."}
{"type":"error","source":"agent_loop","message":"null is not an object (evaluating 'H.usage')"}
{"type":"error","source":"cli","message":"Exec failed"}

Notes

• This is not the earlier custom-model selection bug; the custom model is accepted and invoked correctly.
• This is also different from the h.type streaming parser issue in v0.99.0 regression: undefined is not an object (evaluating 'h.type') — streaming chunk parser crashes on tool-call output index gaps #937, although it feels related in that it looks like a null-safety problem in chunk processing.
• The failure is reproducible for this BYOK setup and happens after first-token timing is recorded, which suggests Droid is assuming usage exists on some streamed/final object where it may actually be null or absent.

If useful, I can provide a reduced reproduction against the same endpoint shape without the vendor-specific token.

Co-Authored-By: ForgeCode noreply@forgecode.dev