AgentInterface.

Overview

A local workflow connecting code discovery, read-only analysis, approved changes, and test feedback through conversation.

• Public source code, configured and run locally.
• Not a hosted demo. Real agent access depends on the user’s environment.

The problem

Users do not always know which file holds the problem. And an AI suggestion should not automatically become a file change.

• Find relevant source code from a natural-language question.
• Distinguish a suggestion from approval and an execution result.

My role / What I built

I connected the product interface and backend capabilities around actual user journeys.

• The conversation workspace, persistence, search, archiving, and presentation.
• The handoff from project search and trusted reading to a patch approval card.
• Test feedback and post-action conversation context.

Key capabilities

Capabilities are organized around concrete tasks, rather than unrestricted model access to tools.

• Bounded project search, file reading, and code navigation.
• Follow-up suggestions resolve within the same conversation, with file and diff previews.
• Approved changes are executed through a controlled path, with recorded test feedback.

Architecture

The product layer owns business state. Reading, authorization, and execution have separate responsibilities.

1. Web UI
2. Product API / Facade
3. ConversationCoordinator
4. Trusted Reader / Approved execution
5. Results & conversation state

• The Web UI reaches the coordinator through the Product API / Facade.
• Read-only analysis and approved execution use distinct paths.

Safety design

Checks surround the operation instead of relying on prompts alone.

• No evidence, no definitive claim.
• Changes require approval and a file-freshness check. Rejection does not execute.
• No generic shell, arbitrary test commands, or Git writes.
• Completed operations are not replayed after restart.

Engineering challenges

The difficult part is not only the answer. It is keeping product state consistent.

• Preventing file evidence from crossing conversation boundaries.
• Persisting test outcomes as trusted action records for later conversation.
• Showing useful results without exposing internal technical metadata.

Validation

Core flows have been checked through automated acceptance and real user journeys. Neither replaces the other.

• Automated acceptance uses isolated conversations and deterministic synthesis, with no real agent calls by default.
• It checks search, reading, approved fixture changes, controlled tests, and result recovery.
• Fixtures are restored afterwards. Visual quality still requires human review.

Tech stack & boundaries

Python, the standard-library HTTP server, a JavaScript Web UI, pytest, and external OpenClaw / agent integration.

• This is not a FastAPI service.
• A local single-user tool, without claims of multi-tenancy or production security guarantees.