openclaw - 💡(How to fix) Fix [Bug/Performance]: Agent prep latency regression — 25-30s overhead per turn on v2026.5.7 (Ollama takes only 2.3s) [1 comments, 2 participants]

Bug type

Performance regression

Summary

OpenClaw v2026.5.7 adds 25-30 seconds of overhead per agent turn that does not exist when calling Ollama directly with the same prompt size. This makes isolated agents unusable in production.

Environment

• OpenClaw: 2026.5.7 (eeef486)
• Ollama: remote at 10.100.75.63:11434
• Model: gpt-oss:120b-cloud (also tested gemma4:31b-cloud, deepseek-v4-pro:cloud)
• OS: Linux 6.12.74+deb13+1-amd64, Node v22.22.2
• Agent: isolated agent "cobuart" with dedicated workspace (~10 workspace files, ~35KB total)

Measured Data

Net OpenClaw overhead: ~25-28 seconds per turn.

Steps to Reproduce

1. Create an isolated agent via openclaw agents add
2. Send a trivial message (e.g., "Hola") via the chat completions endpoint
3. Compare with direct Ollama call using the same prompt size

# Direct Ollama (fast, 2.3s)
curl http://10.100.75.63:11434/api/generate \
  -d "{\"model\":\"gpt-oss:120b-cloud\",\"prompt\":\"<long context>...\nHola\",\"stream\":false}"

# Through OpenClaw (slow, 27.0s)
curl -H "Authorization: Bearer <password>" \
  -d "{\"model\":\"openclaw/cobuart\",\"messages\":[{\"role\":\"user\",\"content\":\"Hola\"}],\"max_tokens\":30,\"stream\":false}" \
  http://10.100.75.9:18789/v1/chat/completions

What We Tried (no improvement)

• contextInjection: "continuation-skip" — same latency (overhead is not from re-injection)
• Reduced bootstrapMaxChars from 20000 to 8000
• Reduced bootstrapTotalMaxChars from 150000 to 40000
• Disabled bootstrap-extra-files hook
• Multiple models: gemma4:31b-cloud, gpt-oss:120b-cloud, deepseek-v4-pro:cloud
• thinking: off, reasoning: false

Analysis

The issue appears identical to #75329 and #55762. The function resolvePromptBuildSystemPrompt() seems to be spending 25+ seconds building the system prompt from ~35KB of workspace files + tool definitions, even on continuation turns where contextInjection is set to continuation-skip.

The 2.3s Ollama result proves the model is NOT the bottleneck. The entire latency is inside OpenClaw's prompt assembly pipeline.

Context Size Breakdown

• Workspace files (AGENTS.md, SOUL.md, IDENTITY.md, USER.md, TOOLS.md, MEMORY.md, HEARTBEAT.md): ~10KB
• Tool definitions (exec, read, write, edit, memory_search, etc.): ~8KB
• Runtime prompt (OpenClaw system instructions): ~8KB
• Skills list: ~3KB
• Total: ~30KB → assembled into ~29K tokens of system prompt

Impact

• Isolated agents are not viable for production with this latency
• A user uploading a single CV takes 40+ seconds just for the agent to acknowledge
• Multi-document workflows (typically 8-12 PDFs) would take 5+ minutes of wall-clock time
• Open WebUI users see "typing..." for 30 seconds before any response

Expected Behavior

Agent prompt assembly should take <500ms for 30-50KB of text, similar to how Ollama processes the same content in 2.3s total including inference.