fix(tools): coerce string-typed integer tool arguments (Android tap tool error)

[philipph-askui]

Problem

Three Android test runs from 2026-09-22 failed with

Tool raised an unexpected error: '<' not supported between instances of 'str' and 'int'

every time the agent called tap_tool. In all eight occurrences the model (Claude via Bedrock) had emitted the integer arguments as JSON strings:

{ "x": "726", "y": "122", "repeat": "1", "repeat_delay_in_ms": "50" }

AndroidTapTool.__call__ compares repeat_delay_in_ms < 0 before doing anything else, and "50" < 0 raises the TypeError. ToolCollection._run_regular_tool passed the model's input straight through as tool(**tool_input), so the "type": "integer" in the tool's input_schema was only a hint to the model and never enforced. In one run the agent gave up on tapping after four failures, fell back to drag_and_drop_tool, and marked the test BROKEN.

Change

Generic coercion in the tool dispatcher. New askui.models.shared.tool_input_coercion.coerce_tool_input() walks the tool's JSON schema (refs resolved, anyOf/oneOf/type lists, nested objects and arrays) and performs only lossless conversions: numeric strings and integral floats to integer, numeric strings to number, "true"/"false"/0/1 to boolean, numbers to string. Anything ambiguous is left untouched so the tool's own validation still applies. It is wired into ToolCollection._run_regular_tool, which is the single place model-driven tool calls (including cached trajectory replay) go through, so every regular tool is protected, not just the tap tool.

The tap tool itself is unchanged. A local coercion there would be redundant: MCP exposure via to_mcp_tool is already validated by FastMCP/Pydantic in lax mode, and direct Python calls are covered by type hints.

Scope note: only regular tools dispatched through _run_regular_tool are coerced. Tools served by external MCP servers (e.g. Playwright MCP, which validates strictly) still receive the model input unchanged. Extending coercion to _call_mcp_tool is a possible follow-up.

Tests

• tests/unit/models/shared/test_tool_input_coercion.py: coercion rules, unions, nested objects/arrays, $ref resolution, no-op cases, and an end-to-end ToolCollection.run() case.
• tests/unit/tools/android/test_tap_tool.py: basic AndroidTapTool behaviour plus a regression test running the real tap tool through ToolCollection with the exact string payload from the failing reports.

Verification

• pdm run qa:fix clean (mypy, ruff format, ruff lint)
• pdm run test:unit: 811 passed

🤖 Generated with Claude Code