io.github.aahlijia/dekko

Anchor a durable note to a symbol. Notes are committed to .dekko/notes.json and shown on the symbol's card and in its context pack. Use after a non-obvious change so the rationale survives.

Repo-wide resolver-trust summary: ambiguous call-site count and rate, top colliding names/files. Run before an impact decision built on get_callers/get_callees/workset in a repo with common method names: a low rate means trust the graph; a high rate concentrated in a few files means spot-check those call sites by hand.

Every parameter or return annotation that uses a type — 'what breaks if I change this struct/class's shape', which the call graph can't answer since code can use a type without calling it. Covers named functions/methods plus TS callbacks and interface method signatures. Not covered: fields, generic arguments, JSX.

The symbols that call into an external (out-of-repo) name — a function, an imported binding, or a whole module — with call sites and a summary line. Calls only: type positions, JSX and property reads are not recorded; the summary's importing-file count shows what a low count leaves out.

Every in-repo symbol a symbol calls — what this code depends on, without reading its body. Resolved call edges, not a grep of the body for call-shaped names.

Every symbol (and module-level site) that calls a symbol — exact call edges, unlike grep, which can't tell a call from a same-named string. Use for impact analysis before a change.

Compact signature neighborhood (callers/callees within N hops) for editing a symbol or file. Token-budgetable. Example: target="awardXp", task="who calls this".

What extends, implements, or is impl'd for a class/interface/struct/trait — the blast radius of changing a type declaration. Does not include each implementor's own callers; pair with get_callers for that.

What a class/interface/struct/trait extends, implements, or is impl'd for — its own declared heritage (Python/JS/TS/Java/Rust/C++). Go is not covered: struct embedding is not extracted and structural interface satisfaction has no declaring syntax to see.

Test files a runner should exercise after a change: reverse call-graph reachability from changed symbols plus an import-edge fallback. Leads, not verdicts (static analysis misses fixtures and dynamic dispatch), but it finds the indirect callers a grep for the changed name misses.

List notes anchored to a symbol, or every note in the repo when no symbol is given.

Report whether map.json is fresh or stale.

A file's (or directory's) structural outline — signatures + doc lines, no bodies — at roughly a tenth the cost of reading it. Prefer this before reading a file to learn what it contains.

Signature, kind, location, doc, fan-in/out, and notes for one symbol — the fast way to learn what a symbol is without reading its file.

Regenerate the map; set full=true to ignore the cache and re-parse every file.

Rank symbols by free-text relevance (BM25 over names, signatures and doc lines, not substring matching) — for when you know what the code should do but not its name. Zero hits is not an error: try broader terms. Once you have an exact name, use query_symbol/get_callers instead.

Compact repo digest (~40 lines): counts, language mix, per-directory rollup with coupling and purpose, load-bearing/orchestrating symbols, entry points, parse errors. Read this before exploring an unfamiliar repo.

Task work-set: for a change (git rev) or a symbol, bundle the touched files' outlines plus call-graph packs for the most central touched symbols under one token budget. One call replaces impacted_tests + N outlines + N packs when starting on a change.