Alicia Klein

Something that surprised me while building this week: "OpenAI-compatible" is a statement about the API shape, not about how code edits are expressed inside it.

Aider does not use tool calls. It sends edits as structured text blocks SEARCH/REPLACE delimiters inside the assistant message. The filename is on the line before the fence, not inside it. Cline and Continue use proper tool calls. A response from some backends carries both.

The adapter we shipped this week (`openai_chat.py` in the LineageLens proxy) handles all three patterns: tool-call edits, text-content edits (Aider SEARCH/REPLACE, unified diffs, fenced blocks), and mixed responses where both appear together. It covers Aider, Cline, Continue, Copilot CLI, Goose, Windsurf, and any OpenAI-compatible backend Azure, groq, fireworks, mistral, together.ai.

The hardest design decision was not the parsing. It was fail-open vs. fail-safe. If capture logic throws an error, do you surface it (fail-safe) or swallow it and pass through (fail-open)? We chose fail-open. A proxy that occasionally interrupts a developer's coding session gets uninstalled. Governance with a 60% retention rate is strictly worse than governance with 95% retention.

Last quarter, our team spent three weeks building a feature that does... nothing.

When a user writes something deeply painful in their journal grief, heartbreak, feeling lost Murror's AI used to respond immediately. Empathetic words, gentle reflections, suggested reframes. Users loved it in surveys. Our response quality scores were excellent.

Since our May 31 launch (v1.6), we shipped two major releases:

v1.7 added OAuth 2.0 on the MCP endpoint and importance-score-aware smart merging, so high-value memories resist being collapsed by near-duplicates.