AI tool-use history

AI4 keeps a record of the tools it calls during a conversation. If a call fails, an agent repeats an unproductive action, or a conversation runs out of tool rounds, you can inspect that record to see what happened. Field can also read recent logs and ask a model to propose a source change based on them.

What is logged

Each AI4 send normally creates a JSONL file under ~/.afield/logs/ai4_tool_io/, whether you use OpenAI, OpenRouter or Anthropic. It records the prompt, response shape, tool calls and results, including exceptions and tool-round limits. Large text fields are clipped.

Logging is enabled by default. Set AFIELD_AI4_TOOL_IO_LOG=0 to disable it, AFIELD_AI4_TOOL_IO_LOG_PATH to choose another directory, or AFIELD_AI4_TOOL_IO_LOG_MAX to change the per-string clipping limit.

These files can contain code, prompts, tool arguments, and tool results. Treat the directory as sensitive local diagnostic data, not as an anonymized event log.

Review and proposal runs

The Agent tool use card on /welcome offers:

• Review logs: runs improveAiToolUseDryRun against at most the 12 most recent logs. Source edits occur only in an in-memory overlay. The task writes build/ai-tool-use-improvement-report.md and, when edits are proposed, build/ai-tool-use-improvement-dry-run.patch.
• Generate proposal: creates one or more generated git worktrees and runs improveAiToolUse separately in each. The task's edits are restricted to src/; successful proposal changes are committed in their generated worktree, leaving the current checkout untouched.
• Mark reviewed: moves the 12 most recent JSONL files into a timestamped archived-reviewed directory. It does not delete them.

Open the detail page at /ai-tool-use-improvement.html to see the logs, run status, report and patch paths, and controls for generated worktrees.

The review CLI summarizes tool-call counts, failed results, exceptions, and max-round events before asking the configured model for a small source-level improvement grounded in those logs. Use AFIELD_TOOL_USE_IMPROVEMENT_MODEL to override its model.

Code-use telemetry is separate

The Code-use telemetry card opens /code-use. This is a local SQLite record of Field execution activity: snippets, compile/evaluation outcomes, launch lifecycle events, symbol evidence, and related source. It is not part of the AI tool-I/O JSONL log and has its own storage and retention behavior.

See Semantic retrieval for how local workspace and source indexes differ from both histories, and AI documentation polish for the separate generated-completion-doc workflow.

To read whole saved chats rather than single tool calls, and to have AI4 summarize them, see Reading Saved AI4 Chats.