PR #68742: fix(agents): restore streaming usage tracking for non-native openai-completions providers

• Repository: openclaw/openclaw
• Author: armorbreak001
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/68742

Description (problem / solution / changelog)

Summary

Fixes #68707

When using custom openai-completions endpoints (llama-cpp, LM Studio, DashScope, etc.), stream_options.include_usage: true was never injected into the request payload, causing /status to always show 0% context usage.

Root Cause

In resolveOpenAICompletionsCompatDefaults(), supportsUsageInStreaming was set to false when usesConfiguredNonOpenAIEndpoint was true (i.e. the user pointed to a custom endpoint) and supportsNativeStreamingUsageCompat was not explicitly set. This caused the transport layer to skip injecting stream_options: { include_usage: true }.

Fix

• openai-completions-compat.ts: Changed the default from opt-in to opt-out. supportsNativeStreamingUsageCompat now defaults to true via ?? true, since most OpenAI-compatible APIs handle stream_options.include_usage correctly (as established by #46142).

• provider-model-compat.ts: Updated the fallback in the compat override layer from ?? false to ?? true to match the new default.

Impact

• Restores context token tracking for llama-cpp, LM Studio, and other custom openai-completions providers
• No behavioral change for native OpenAI endpoints (already had true)
• No behavioral change for known non-standard providers (already forced false)
• Aligns with prior fix (#46142) which established that most compat APIs support streaming usage

Testing

• Updated unit tests to reflect the new default
• Existing ollama tests remain passing (explicit isOllamaCompatProvider branch unchanged)

Changed files

• src/agents/model-compat.test.ts (modified, +7/-3)
• src/agents/openai-completions-compat.test.ts (modified, +2/-2)
• src/agents/openai-completions-compat.ts (modified, +1/-1)
• src/plugins/provider-model-compat.ts (modified, +2/-2)

---

PR #68746: fix: always send stream_options.include_usage when streaming (openai-completions)

• Repository: openclaw/openclaw
• Author: kagura-agent
• State: closed | merged: True
• Link: https://github.com/openclaw/openclaw/pull/68746

Description (problem / solution / changelog)

Fixes #68707

Problem

buildOpenAICompletionsParams() only included stream_options: { include_usage: true } when compat.supportsUsageInStreaming was true. For non-standard/custom endpoints (llama-cpp, LM Studio, etc.), this flag resolved to false, so the gateway's resolveIncludeUsageForStreaming() never saw the field and context token tracking was always 0%.

Fix

Always include stream_options: { include_usage: true } in streaming request payloads, matching the OpenAI SDK's default behavior. Backends that don't support the field simply ignore it.

Changes

• src/agents/openai-transport-stream.ts — unconditionally include stream_options instead of gating on compat.supportsUsageInStreaming
• src/agents/openai-transport-stream.test.ts — added test for non-standard backends, updated existing assertion

Changed files

• CHANGELOG.md (modified, +2/-0)
• src/agents/openai-transport-stream.test.ts (modified, +28/-1)
• src/agents/openai-transport-stream.ts (modified, +1/-3)

---

PR #68749: fix: enable stream_options.include_usage for local openai-completions endpoints

• Repository: openclaw/openclaw
• Author: Kailigithub
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/68749

Description (problem / solution / changelog)

Problem

Closes #68707

When using local OpenAI-compatible backends (llama.cpp, LM Studio, Ollama), stream_options: { include_usage: true } is never sent in streaming requests. This causes context token usage to always show 0% in /status and the web UI.

Root Cause

In detectOpenAICompletionsCompat(), supportsUsageInStreaming evaluates to false for local endpoints because:

• usesConfiguredNonOpenAIEndpoint = true (endpointClass "local" ≠ "default")
• supportsNativeStreamingUsageCompat = false (only moonshot/modelstudio qualify)
• Result: true && (!true || false) = false

However, llama.cpp and LM Studio both implement the OpenAI-compatible streaming API and support stream_options.include_usage.

Fix

Add endpointClass === "local" to the supportsUsageInStreaming condition. This allows local OpenAI-compatible endpoints to receive usage data in streaming responses.

Note: LM Studio extension already works around this in extensions/lmstudio/src/stream.ts:259-264 by overriding supportsUsageInStreaming. This core fix covers llama.cpp, Ollama, and other local backends that do not have their own workaround.

Changed files

• src/agents/openai-completions-compat.ts (modified, +1/-1)