name

code-map

description

Read the exact source of KNOWN symbols (functions, methods, classes) cheaply via the code-map `read` MCP tool instead of grep + sed/cat. Use PROACTIVELY when: (1) You already know a symbol's name, id, or file:line and need to see its body. (2) You need to read SEVERAL known symbols — batch them in one `read` call. (3) You are about to grep for a name and then sed/cat the matching lines to read a body (code-map returns that exact slice in one call — skip the grep+sed dance). Use your normal shell search ONLY to discover candidates whose name/location you don't know yet; once you know the names, switch to code-map read for the bodies.

code-map: exact source slices for known symbols

code-map is an MCP server exposing one tool, read. It returns the RAW source slice of a symbol (its own bytes — the function/method/class body), not the whole file, so it is far cheaper than grep-then-read and never floods context with unrelated matches.

Routing rules (when to use what)

Known symbol body → read. If you know the name / id / path#name, call read. Do NOT read a known body with grep/sed/cat.
Several independent known symbols → ONE batch read. Pass refs: ["a", "b", "c"] in a single call rather than one call per symbol (fewer round-trips, lower cost). Split only when a later read depends on an earlier result, or the batch exceeds 64 refs.
Unknown target → grep to DISCOVER, then read. Use shell search only to find the name/line of something you can't yet name; once discovered, read its body with read.
Interpret the raw code, not index metadata — read gives coordinates, you do the judging.

How to call it

The tool is read on the code-map MCP server. Examples:

one symbol: read({ ref: "dispatch_hook" }) or read({ ref: "requests/hooks.py#dispatch_hook" })
several: read({ refs: ["get_connection", "dispatch_hook", "resolve_redirects"] })

If a name is ambiguous it returns candidates; pick the right path#name and read that. If the file drifted since indexing, read re-anchors on the signature and flags it.

code-map: exact source slices for known symbols

Routing rules (when to use what)

Known symbol body → read. If you know the name / id / path#name, call read. Do NOT read a known body with grep/sed/cat.

Several independent known symbols → ONE batch read. Pass refs: ["a", "b", "c"] in a single call rather than one call per symbol (fewer round-trips, lower cost). Split only when a later read depends on an earlier result, or the batch exceeds 64 refs.

Unknown target → grep to DISCOVER, then read. Use shell search only to find the name/line of something you can't yet name; once discovered, read its body with read.

Interpret the raw code, not index metadata — read gives coordinates, you do the judging.

How to call it

The tool is read on the code-map MCP server. Examples:

one symbol: read({ ref: "dispatch_hook" }) or read({ ref: "requests/hooks.py#dispatch_hook" })

several: read({ refs: ["get_connection", "dispatch_hook", "resolve_redirects"] })

If a name is ambiguous it returns candidates; pick the right path#name and read that. If the file drifted since indexing, read re-anchors on the signature and flags it.