Capture tool errors with Fix-it

Fix-it captures tool failures and recovery evidence for review as training examples. Capture is a separate opt-in for Codex CLI, OpenCode, native OpenPond and Claude Code 2.1.292. It uploads bounded, sanitized tool arguments, failure excerpts and a bounded window of follow-up actions under the selected private Model. Full conversations, credentials and native permission state are excluded. Ordinary successful controls are sampled; after a failure, at most 32 following eligible observations within ten minutes are retained as recovery candidates. A successful follow-up is not automatically a verified recovery.

Codex CLI and OpenCode connections are capture-only on macOS and Linux. Codex uses an explicitly launched captured CLI task; it does not observe an existing Desktop chat. OpenCode uses a project plugin and needs a restart after installation. Captured examples reach the private Fix-it Dataset; training requires separate review and enrollment. These adapters do not install model consultations or approve actions.

Connect your coding agent

Open Fix-it connections, choose your coding agent and copy its setup prompt into the workspace you want to connect. You can also use Set up tool learning in the Model Overview to copy a prompt containing the selected workspace and Model. Choose the current client, reuse the existing account/home, and inspect openpond tool-learning --help. The setup command requires explicit upload consent and the actual application/version context. Update to a release whose help includes tool-learning setup before configuring capture.

Prepare a small JSON context file, using the real application's identity and only the state needed to understand its tools:

The CLI computes a fingerprint when omitted. Never include credentials, environment dumps, private file contents, or a guessed version. Choose exact allowed tool names from the current client; for Claude's shell tool this is Bash. Then run, substituting the selected Model, workspace, file, account origin, and home:

For native OpenPond use --client openpond and the exact native tool names being enabled. Setup preserves unrelated Claude hooks/MCP entries and the existing permission system. Restart the agent, accept its normal workspace/MCP trust prompts, perform the intended task, and run tool-learning doctor <installation-id>. Doctor checks actual lifecycle observations, queue status, and the local MCP handshake; configuration files alone do not prove capture works. A changed Claude version pauses capture until that version is qualified and registered.

Check capture and manage access

Use tool-learning status, episodes, and episode to inspect private owner records, and flush to deliver queued evidence. disable stops new admission; revoke permanently revokes the installation credential; purge removes hosted raw events; uninstall revokes and removes only this installation's owned configuration. Reviewed Dataset material and existing model weights have separate retention.

Review recoveries before training

An owner explicitly selects a failed observation and a later successful observation with tool-learning project --input-file .... The host verifies retained event/receipt lineage and saves a pending private target. Review that target through the existing learning commands before sealing/publishing a supervised batch. Capture and projection never grant training authority or start a GPU. Internal SFT consumes the reviewed batch through the existing training flow; no external Python SDK is required.

Continue with Set up continual learning to configure compatible training, review, schedules and budgets.

Enable consultations separately

Consultations stay inactive until the owner uses tool-learning configure to select a model, completed evaluation evidence, narrowly declared application/version/tool capabilities, and call/session/day limits. An integration selection identifies an unchanged model; a learned selection additionally requires the real completed SFT run, held-out comparison, and matching served adapter. Agents receive only the current capability brief and call openpond_consult after an observed failure. The model's advice does not execute or approve actions. Recover a lost response with openpond_consult_status and the original operation ID; do not start a replacement paid request. Setup, event capture, human review, training, qualification, and serving are separate visible states.