How it works

Model and schema: What it is. This page is the machine.

Parse

Tree-sitter grammars compiled to WASI, hosted by wazero. CGO_ENABLED=0. One instance per worker. Embedded: Go, Python, JS, TS/TSX, Java, C, C++, C#, Rust, Zig, SQL. Anything else is a pack (<name>.wasm + <name>.json) in ~/ctxoptimize/grammars/ or .ctxoptimize/grammars/. languages add builds one; Zig is fetched once, sha256-checked.

Kubernetes gather: 58% of time is tree-sitter inside wasm. The remaining lever is parsing fewer files. GOMAXPROCS caps workers (container-aware). That bounds wasm, not the graph — linux (~2.85M nodes) still wants ~14 GB.

Markdown

goldmark AST. Headings are sections. Fences are not. A link is an edge only if the target resolved in the walk. The old line-regex minted 618 phantom sections (a # inside a bash fence). 10 ms → 24 ms on this repo’s 3.5 MB of docs.

Store

~/ctxoptimize/<repo>/graph/{nodes,edges}.ndjson   # sorted, atomic rename
                  manifest.json                   # sha256 per file

add uses producer-scoped Replace (re-parse code, keep your adapter). Shrink >50% refused without --force. Merge (--json) never steals another producer’s stamp.

card index

graph/index/{labels,ids,edges-by-source,edges-by-target}.idx — sorted text, binary search + ReadAt. Kernel card <20 ms. Header is size+mtime of the graph; any mismatch falls back to a full scan. query does not use this index. How query scores.

Confidence

EXTRACTED	parsed
INFERRED	name-matched
AMBIGUOUS	shortlist; out of affected / path / change-plan unless --include-ambiguous

Blast radius is a floor. ? rows are candidates, not callers.

Boundaries

Rules are JSON over AST shapes, evaluated on the same walk that builds calls. TypeScript repo: 12.4 s regex → 0.06 s. A variable argv is visible as AMBIGUOUS, not missing. Boundaries.

Flow

GET /api/scene derives the picture in Go. 7,663 nodes → 4.4 KB. See.

No LLM, no embeddings, no MCP, no credentials at rest. Network only when you ask (update, grammar, remote script).