Skip to content

Latest commit

 

History

History
61 lines (43 loc) · 3.37 KB

implementation_plan.md

File metadata and controls

61 lines (43 loc) · 3.37 KB

Implementation Plan - Heybro Architecture Improvements

Goal Description

Improve the Heybro Chrome extension architecture to handle long-running automations, optimize token usage, and ensure "Zero Point Failure" robustness.

User Review Required

Important

Breaking Change Potential: The "SmartLocator" changes might affect how elements are found. We will maintain backward compatibility with ID-based lookups but prioritize signatures when IDs fail.

Proposed Changes

Milestone 1: Token & Prompt Optimization

[MODIFY] planner.js

  • Optimize compressElements: Reduce the JSON footprint of element lists by using shorter keys and removing redundant attributes.
  • Refine buildSystemPrompt: Shorten the static instructions.
  • History Pruning: Update callGemini to accept a summarized history or implement the summarization logic before calling it.

[MODIFY] agent.js

  • History Management: Implement a sliding window for buildHistoryForPrompt. Keep the last 10 actions in full detail, and summarize or drop older ones.

Milestone 2: Robustness (Zero Point Failure)

[MODIFY] content.js

  • Enhance SmartLocator: Improve the scoring algorithm to better utilize the "Signature" (text, role, context) when the ID is invalid.
  • Add rebind capability: Allow the agent to re-scan and find the element if the initial lookup fails.

[MODIFY] agent.js

  • Retry Logic: If a tool execution fails with "Element not found", trigger a simplify (refresh) and retry the action using the signature before asking the Planner again.

Milestone 3: Long-Running Support

[MODIFY] background.js

  • State Persistence: Save taskContext and actionHistory to chrome.storage.local on every update. Load it on startup.

[MODIFY] lib/tools.js

  • Implement isIgnoredUrl to detect GTM and other service domains.
  • Update execute and probe to filter out ignored frames and prioritize the main frame (frameId 0).

[MODIFY] agent.js

  • Add a safety check in the main loop to detect if the agent is stuck on an ignored URL (like GTM) and provide a clear signal to the planner or auto-recover.

[MODIFY] content.js

  • (Optional) Add logic to getPageState to flag if the current context is a known service frame.

[MODIFY] agent.js

  • Step Limit: Change the hard 50-step limit to a soft limit that prompts the user or auto-continues based on settings.

Verification Plan

Automated Tests

  • We don't have a full test suite, so we will verify manually.

Manual Verification

  1. Token Usage: Check the logs/console to see the size of the prompt being sent to Gemini.
  2. Robustness:
    • Go to a dynamic site (e.g., YouTube, Twitter).
    • Start a task.
    • Manually modify the DOM (delete an ID) while the agent is "thinking".
    • Verify the agent still finds the element using the signature.
  3. Long Run: Run a task that requires > 50 steps (e.g., "Scroll down 60 times").