Bug Report
Repro (v2026.7.2, PepsVida project, 2026-07-17 session): 'cleo add --type task --parent T555 ... --field /data/created/0', 'cleo verify T827 --gate implemented --evidence commit:...;files:...', and 'cleo complete T827' each hang past a 120s harness timeout and have to be backgrounded; some eventually succeed minutes later (the add DID create T827/T828), others produce no output at all. Every hung run prints: 'AI SDK Warning (ollama.chat / qwen2.5-coder:3b): The feature responseFormat is not supported. JSON response format schema is only supported with structuredOutputs' - suggesting the BRAIN memory/embedding hook on mutations is calling a local Ollama model synchronously and not timing out or failing open. Read-only commands (show, find, briefing) return in seconds. Expected: mutations return their LAFS envelope promptly and memory/embedding work happens async or with a hard timeout; a wedged local LLM sidecar should never block task CRUD. Impact: agents cannot complete the verify->complete gate ritual interactively; every session loses minutes per mutation.
Severity: Major
Area: cli
Are you using an AI agent?
Yes - AI agent filed this issue
Environment
| Component |
Version |
| CLEO |
2026.7.2 |
| Node.js |
v24.16.0 |
| OS |
linux 7.0.12-201.fc44.x86_64 x64 (x64) |
| Shell |
/bin/bash |
| gh CLI |
gh version 2.94.0 (2026-06-11) |
| Install |
/home/keatonhoskins/.npm-global/bin/cleo |
Bug Report
Repro (v2026.7.2, PepsVida project, 2026-07-17 session): 'cleo add --type task --parent T555 ... --field /data/created/0', 'cleo verify T827 --gate implemented --evidence commit:...;files:...', and 'cleo complete T827' each hang past a 120s harness timeout and have to be backgrounded; some eventually succeed minutes later (the add DID create T827/T828), others produce no output at all. Every hung run prints: 'AI SDK Warning (ollama.chat / qwen2.5-coder:3b): The feature responseFormat is not supported. JSON response format schema is only supported with structuredOutputs' - suggesting the BRAIN memory/embedding hook on mutations is calling a local Ollama model synchronously and not timing out or failing open. Read-only commands (show, find, briefing) return in seconds. Expected: mutations return their LAFS envelope promptly and memory/embedding work happens async or with a hard timeout; a wedged local LLM sidecar should never block task CRUD. Impact: agents cannot complete the verify->complete gate ritual interactively; every session loses minutes per mutation.
Severity: Major
Area: cli
Are you using an AI agent?
Yes - AI agent filed this issue
Environment