recovery

Recovery Skill

Structured recovery for browser automation failures. Covers execution errors, semantic failures, state divergence, tool limitations, and environmental blockers.

When to Use

Load this skill when any of these occur:

• A tool returns an error (is_error in tool result).
• A post-action screenshot shows no change or an unexpected change.
• A verification step (snapshot / verify) contradicts the expected state.
• An action succeeded but the effect is wrong (filled wrong field, clicked wrong element, upload didn't trigger, text truncated).
• An unexpected element blocks progress (modal, banner, CAPTCHA, cookie consent).

Recovery Loop

Follow three phases in order. Do not skip OBSERVE.

Step 1 — OBSERVE

Re-perceive the actual page state before making any recovery decision.

Batch these read-only tools in a single turn (the engine runs them concurrently):

• take_snapshot — current DOM / accessibility tree.
• take_screenshot — current visual state.
• list_console_messages — JS errors that may explain the failure.
• list_network_requests — failed API calls or unexpected redirects.

Before / after comparison: Compare the current state against what the page looked like before the failed action. Ask:

1. What changed? (anything at all — URL, DOM, visual layout)
2. What should have changed but didn't?
3. Are there new elements that weren't there before (modals, errors)?

Step 2 — DIAGNOSE

Classify the failure and assess progress.

Error classification — load the error-taxonomy reference for the full list: load_skill(skill_name="recovery", reference="error-taxonomy")

Key questions:

• Is this an execution error (tool reported failure) or a semantic error (tool succeeded but wrong effect)?
• Is the root cause a tool limitation, a wrong selector, a page state change, or an environmental blocker?

Progress assessment — did the action make any progress toward the goal?

• Partial progress (e.g. 3 of 5 fields filled): preserve what worked, recover only the failed part.
• Zero progress (nothing changed): the approach itself may be wrong; escalate sooner.
• Negative progress (broke something): undo if possible (evaluate_script({ code: "history.back()" }) or navigate to the previous URL), then re-observe.