Hivemind Graph Agent Skill

Hivemind Code Graph

A deterministic, AST-derived map of the current repository — every function,

class, method, interface, type, enum, const, and module, plus the edges between

them (calls, imports, extends, implements, method_of). It is queried as

synthesized files under the Deeplake mount; there are no real files on disk and

no network call in the read path.

The graph builds and refreshes automatically (on Stop / SessionEnd, gated by

a rate limit + git diff). You never run a build command — just read it.

Use it as a fast INDEX to locate the few files/symbols that matter, then open

them with Read to answer. It is not a substitute for the source.

When to use this skill

Activate when the user asks a *structural / relational* question about the code:

• "What calls pushSnapshot?" / "Who uses this function?"
• "What does deeplake-pull.ts import?" / "What depends on X?"
• "Where is GraphSnapshot defined?" / "Find the function that handles Y."
• "What are the main subsystems / the architecture here?"
• "If I change this signature, what's affected?" → use impact/<symbol> (transitive blast radius)

When NOT to use this skill

• Reading the body of a symbol you already located → use Read on the real

source file. The graph gives location + relationships, not full source.

• Code that isn't committed/built yet — the graph can lag uncommitted edits.

If a file's mtime is newer than the build timestamp, read the live source.

• Languages outside TypeScript, JavaScript, and Python (Go, Rust, …) — the

extractor covers those three, with cross-file calls/imports resolved for

named imports. For anything else, fall back to grep/read.

Path cheat sheet

cat ~/.deeplake/memory/graph/index.md
#   Overview: node/edge counts, kind breakdown, top files by node count.

cat ~/.deeplake/memory/graph/query/<pattern>   # START HERE (the 2-in-1)
#   Search + expand the top matches with their 1-hop neighbors (callers,
#   callees, imports, heritage). Multi-token AND: query/<a>+<b>.

cat ~/.deeplake/memory/graph/find/<pattern>
#   Case-insensitive substring search on node id + label (max 50 hits).
#   Prints numbered handles [1] [2] ... saved for this worktree.

cat ~/.deeplake/memory/graph/show/<handle-or-pattern>
#   <handle>: a digit from a prior find/ (e.g. 3).
#   <pattern>: a substring → unique node detail, or a candidate list.
#   Output: the node + its 1-hop neighbors grouped by edge relation.

cat ~/.deeplake/memory/graph/neighborhood/<file>
#   Every symbol in a file + its cross-file neighbors (callers/callees/imports).

cat ~/.deeplake/memory/graph/impact/<pattern>
#   Transitive dependents — the blast radius of changing a symbol.

cat ~/.deeplake/memory/graph/path/<from>/<to>
#   Shortest dependency path between two symbol patterns (trace a flow across files).

cat ~/.deeplake/memory/graph/layers      # architectural layers / subsystems
cat ~/.deeplake/memory/graph/tour        # deterministic guided walkthrough

Workflow

• Broad? Start at index.md to see subsystems and the biggest files.
• Looking for a symbol? find/<name> (or query/<name>) → pick the handle.
• Want relationships? show/<handle> / neighborhood/<file> → callers/callees, imports.
• Tracing a flow? path/<from>/<to>. Change impact? impact/<symbol>.
• Need the actual code? Take the source_file:line and Read it — don't answer from the graph alone.

Anti-patterns (read these)

• "Incoming (0)" does NOT mean dead code. Cross-file calls are resolved for

*named imports* (TS/JS/Python), but instance-method dispatch (obj.method()),

dynamic calls, and nested/inner functions are NOT — a zero-incoming symbol may

still be reached via one of those. Confirm in the source before calling it unused.

• The graph can be stale. It rebuilds at most once per rate-limit window. The

SessionStart inject prints the build age; if it's old or you've just edited a

file, prefer the live source for that file.

• Don't try to build it. There is no user-facing build step in normal use;

the hooks handle it. Just read the mount.

• find/ is lexical, not semantic. It matches substrings, not meaning —

find/auth won't surface login/credentials unless those strings appear in

the id/label. Try multiple keywords if the first misses.