PR #64568: fix(ollama): enable streaming usage for OpenAI-compat endpoint (closes #64326)

• Repository: openclaw/openclaw
• Author: xchunzhao
• State: closed | merged: False
• Link: https://github.com/openclaw/openclaw/pull/64568

Description (problem / solution / changelog)

Summary

• Problem: Ollama provider reports hardcoded 32768 input tokens for every message, triggering premature compaction and breaking multi-turn conversations.
• Why it matters: All Ollama local model users experience broken conversations — compaction fires after every response, wiping context.
• What changed: Added Ollama to the supportsUsageInStreaming list in resolveOpenAICompletionsCompatDefaults, so stream_options.include_usage=true is sent to Ollama's OpenAI-compat endpoint. This makes Ollama return real usage.prompt_tokens and usage.completion_tokens in the final streaming chunk.
• What did NOT change: Native Ollama API path (already correct), other providers' streaming usage behavior.

Change Type (select all)

• Bug fix
• Feature
• Refactor required for the fix
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Closes #64326
• This PR fixes a bug or regression

Root Cause (if applicable)

• Root cause: resolveOpenAICompletionsCompatDefaults in src/agents/openai-completions-compat.ts did not recognize Ollama as supporting streaming usage. Without stream_options.include_usage=true in the request, Ollama's OpenAI-compat endpoint omits usage data from streaming responses. The OpenAI transport stream then reports 0 input tokens (or falls back to incorrect values), while a separate code path uses contextWindow (32768) as the token count, triggering compaction.
• Missing detection / guardrail: No test for Ollama OpenAI-compat streaming usage support.
• Contributing context: Ollama's native API path correctly returns prompt_eval_count, but when accessed via /v1/chat/completions (OpenAI-compat), usage requires explicit opt-in via stream_options.

Regression Test Plan (if applicable)

• Coverage level that should have caught this:
  • Unit test
• Target test or file: src/agents/openai-completions-compat.test.ts
• Scenario the test should lock in: resolveOpenAICompletionsCompatDefaults({ provider: "ollama" }).supportsUsageInStreaming === true
• If no new test is added, why not: Keeping PR minimal — happy to add if requested

User-visible / Behavior Changes

• Ollama models via OpenAI-compat endpoint now report accurate token counts
• Multi-turn conversations no longer break due to premature compaction
• Compaction only triggers when context actually approaches the limit

Diagram (if applicable)

Before:
request to Ollama /v1/chat/completions (no stream_options.include_usage)
-> Ollama omits usage in stream -> reported as 0 or contextWindow -> compaction triggers

After:
request to Ollama /v1/chat/completions (stream_options.include_usage=true)
-> Ollama returns real usage in final chunk -> accurate token count -> no premature compaction

Security Impact (required)

• New permissions/capabilities? No
• Auth boundary changes? No
• Secrets/token exposure risk? No
• New external calls? No
• Sandbox/isolation changes? No

Evidence

• Code trace confirms Ollama was missing from streaming usage support list
• pnpm check passes
• AI-assisted: fix authored by Claude Code, reviewed and verified manually

Human Verification (required)

• Verified scenarios: Traced the full path from request construction through streaming usage parsing
• Edge cases checked: Native Ollama API (unaffected), non-Ollama providers (unaffected)
• What you did not verify: Live Ollama integration test

Review Conversations

• I replied to or resolved every bot review conversation I addressed in this PR.
• I left unresolved only the conversations that still need reviewer or maintainer judgment.

Compatibility / Migration

• Backward compatible? Yes
• Config/env changes? No
• Migration needed? No

Risks and Mitigations

• Risk: Older Ollama versions might not support stream_options.include_usage
  • Mitigation: Ollama silently ignores unknown request fields, so this is safe for older versions (they just won't return usage, same as before)

Changed files

• src/agents/openai-completions-compat.ts (modified, +4/-1)