Log per-action tool calls in CuaAgent templates

[dprevoznik]

Summary

The anthropic-computer-use, openai-computer-use, and gemini-computer-use templates build on CuaAgent but never subscribed to its events, so they only printed the final answer — no visibility into the steps the agent took.

This adds a shared logAgentEvent helper per template (logging.ts / lib/logging.ts, matching each template's file layout) wired up with agent.subscribe(logAgentEvent). It logs:

• the model's narration between steps (agent> ...) from message_end
• each concrete browser action (→ click (85, 267), → drag (780,385) → (1100,385), → goto https://...) from tool_execution_start
• failed actions (✗ click failed) so retries are visible

Anthropic batches actions under computer_batch and OpenAI wraps navigation in computer_use_extra; both are unwrapped so every provider's logs read the same. Byte-count/base64 noise from the raw tool results is not logged.

This brings the CuaAgent templates in line with the readable per-step logs the hand-rolled yutori/tzafon templates and the Python openai-computer-use template already produce.

Test plan

Deployed and invoked all three end-to-end against live Kernel browsers (multi-step task: navigate, filter, drag cards). Sample output:

agent> I need to filter by Alice Johnson. Let me click the Filters button.
  → click (638)
  ✗ click failed
  → click (638, 125)
agent> The filter dropdown appeared. Clicking Alice Johnson.
  → click (706, 271)
agent> Now dragging her in-progress tasks to Done.
  → drag (780, 385) → (1100, 385)

• bun install && bunx tsc --noEmit passes in all three template dirs
• anthropic (computer_batch), openai (computer_use_extra), and gemini (top-level tools) all verified live

---

Note

Low Risk
Template-only console logging with no changes to agent execution, auth, or data handling.

Overview
The anthropic, gemini, and openai TypeScript computer-use templates now subscribe to CuaAgent events so runs print step-by-step output instead of only the final answer.

Each template adds a logAgentEvent helper (under logging.ts or openai-computer-use/lib/logging.ts) and registers it with agent.subscribe(logAgentEvent) before prompt. Logs include assistant narration (agent> … on message_end), each browser action at start (→ … on tool_execution_start), and failures (✗ … when tool_execution_end is an error). describe unwraps computer_batch and computer_use_extra so Anthropic/OpenAI nested actions read like flat click/drag/goto lines, with long argument values truncated.

^{Reviewed by Cursor Bugbot for commit a489ac7. Bugbot is set up for automated code reviews on this repo. Configure here.}