Port ComputerUseMixin workflow learning and replay from gaia6

[kovtcharov]

Summary

Port ComputerUseMixin from gaia6 (agents/base/computer_use.py, ~1,176 lines). Enables agents to learn browser workflows by recording actions, then replay them with parameter substitution.

Tools Provided

• learn_workflow(name, url) — Record browser actions as a replayable skill
• replay_workflow(name, params) — Execute a learned skill with parameter substitution
• list_workflows(filter) — List all learned skills
• test_workflow(name) — Replay in visible mode for verification

Architecture

• Uses PlaywrightBridge for browser automation (abstracted for testability)
• Skills stored in KnowledgeDB as category="skill" with metadata.type="replay"
• Screenshots stored in ~/.gaia/skills/{insight_id}/step_N.png
• Depends on MemoryMixin (KnowledgeDB for skill storage)

Reconciliation with Existing CUA Issues

Existing issues (#224, #458-#461) approach CUA from an MCP perspective. This gaia6 approach is mixin-based with workflow learning. Need to reconcile:

• PlaywrightBridge (gaia6) vs BrowserToolsMixin with Playwright (Implement BrowserToolsMixin with Playwright for web browsing agent #458)
• Workflow recording/replay is additive to both approaches
• Security policies (Add browser use security policy and URL allowlisting #459, Add desktop control security policy and opt-in model #461) apply to both

Source

Port from gaia6/src/gaia/agents/base/computer_use.py

Acceptance Criteria

• Workflow learning records browser actions
• Replay executes with parameter substitution
• Skills persist across sessions via KnowledgeDB
• Screenshots captured per step
• Unit tests (port from gaia6: test_computer_use.py)
• Integration tests (port from gaia6: test_computer_use_e2e.py)