Troubleshooting, Benchmarks, and Versions

Tested June 2026 on a Mac mini M2 Pro 32G and a MacBook Air M4 16G.

Symptom: file writes fail or diffs corrupt indentation. Cause: context too small for the file tree plus logs. Fix: restart Ollama with a larger context (OLLAMA_CONTEXT_LENGTH=32768 ollama serve on 32GB, 16384 on 16GB) and retry. For critical files use the Write-and-Notify strategy: write a new tmp_* file and merge by hand.

Symptom: generation slows mid session. Cause: macOS is swapping. The model plus IDE plus browser no longer fit in RAM. Fix: check Activity Monitor or htop for swap, close browser tabs, stop idle containers (docker system prune), switch to a terminal editor, or drop to a smaller model (gemma4:12b-it-qat).

Symptom: the run hits the execution timeout with no output. Cause: a bare sbcl call opened a REPL or debugger waiting on input. Fix: rerun headless per the common-lisp.md skill, always ending with --eval "(quit)", and read stderr for the condition.

Symptom: build overhead or missing project.clj errors. Cause: the agent guessed the legacy runner. Fix: point it at the clojure.md skill rule: default to clojure -M:test unless project.clj exists on disk.

Symptom: stray .js files, or green runs with type errors. Cause: the agent compiled with tsc into the tree or skipped checks. Fix: run with npx tsx <file>.ts and gate done on npx tsc --noEmit.

Symptom: packages land in the global interpreter or runs use the wrong env. Cause: the agent mixed pip, conda, and uv. Fix: per the python.md skill, use uv run when pyproject.toml or uv.lock exists, else a local python3 -m venv. Never install globally.

Symptom: the tree fills with Write-and-Notify outputs. Cause: normal use. Each agent task mints new files. Fix: delete merged tmp_* files at session end.

Dear reader, my rule from the Preface: split work by need. Local models own small scoped tasks with fast tests. Frontier models own wide refactors and novel design. When the 12B model stalls twice on one task, I shrink the task (one file, one function, paste the failing test output into the prompt) or escalate that single task to Gemini, DeepSeek, or Claude and bring the result back for local tests. I will add a worked failure log with timings in a later update. If you hit a good one, file it per WrapUp so I can include it.

Run scripts/smoke.sh to see versions on your system.