whetstone_DSL/AGENT_NOTES_2026-02-22.md

Agent Notes 2026-02-22

Use this file as the primary handoff for the next agent.

Commits made

• 33103b5 Add migration planning tools and step899-908 test suite
• f4f9605 Ignore local logs, training data, and build/runtime artifacts

Main focus this session

• Improved MCP/Ollama taskitem bridge logging and dataset tooling.
• Added native Ollama tool mode prototype (OLLAMA_INTERFACE_MODE=native_tools).
• Added anti-loop / escalation logic and improved execution-stage capture.
• Added batch graceful stop + preflight + executor timeout controls.

Files added/changed (core)

• tools/mcp/ollama_mcp_taskitem_executor.sh
• tools/mcp/run_slm_training_batch.sh
• tools/mcp/validate_slm_runtime.sh
• tools/mcp/export_mcp_call_quality_for_lora.sh
• tools/mcp/export_malformed_mcp_calls_for_lora.sh
• tools/mcp/build_multimodel_lora_dataset.sh
• tools/mcp/reclassify_tool_execution_quality.sh
• docs/mcp_logging_data_strategy.md
• docs/multimodel_lora_data_playbook.md
• training_data/lora/README.md

Run status at handoff

• Last batch stopped gracefully by stop-file:
  • logs/taskitem_runs/batches/slm_batch_20260222_191303.jsonl
• Observed mixed behavior:
  • Some rounds produced valid tokens + execution events.
  • Some rounds timed out (exit_code=124, empty run_dir).

Known issue to fix first

• Batch mode still intermittently stalls/produces nested executor hangs.
• Priority: stabilize orchestration before long unattended runs.

Environment note (critical)

Ollama daemon must be started from host shell with local model store:

export OLLAMA_MODELS=/home/bill/.ollama/models
export OLLAMA_HOST=127.0.0.1:11434
nohup ollama serve > /tmp/ollama_serve.log 2>&1 &

Validate:

curl -s http://127.0.0.1:11434/api/tags | jq '.models[].name'

Recommended next command sequence

cd /home/bill/Documents/CLionProjects/whetstone_DSL
./tools/mcp/validate_slm_runtime.sh
RUN_PREFLIGHT=0 REPEATS=2 MAX_TURNS=3 EXECUTOR_TIMEOUT_SECONDS=180 \
  CURL_TIMEOUT_SECONDS=25 MCP_TIMEOUT_SECONDS=15 JOB_SPECS='126:0' \
  ./tools/mcp/run_slm_training_batch.sh

Then inspect if each round gets non-empty run_dir and nonzero tokens before scaling repeats.