The "stale symbol" detection is a nice touch - one of the failure modes I've noticed with AI coding agents is them operating on outdated mental models when files have changed mid-session. Color-coding what's potentially stale vs fresh gives you a sense of when to nudge the agent to re-read.
The granularity levels (unseen → name-only → overview → signature → full body) also map nicely to how humans skim code. Wonder if this could eventually feed back into the agent itself - like a "coverage indicator" that helps it decide what to read next when context is limited.
Currently Rust and Python via tree-sitter - Serena MCP integration should help with other languages. Would be interesting to see TypeScript support given how much Claude Code is used in JS/TS projects.
Marker – visualize Claude's symbol understanding · HackerTrans
The granularity levels (unseen → name-only → overview → signature → full body) also map nicely to how humans skim code. Wonder if this could eventually feed back into the agent itself - like a "coverage indicator" that helps it decide what to read next when context is limited.
Currently Rust and Python via tree-sitter - Serena MCP integration should help with other languages. Would be interesting to see TypeScript support given how much Claude Code is used in JS/TS projects.