Look up a single symbol by name in the local corbel index and return where it is defined (file, line, signature), everything that calls it (callers), and everything it calls (callees). Every caller and callee comes with a resolution field (e.g. same-file, scoped, global-unique) naming which index-wide lookup found the definition, rather than a guess from text matching — this is what makes the result trustworthy for navigation and refactoring, unlike a grep/text search which can't tell you if a match is actually the same symbol. scoped and global-unique both mean a single matching definition was found index-wide; the difference is only whether the caller's file imports that name (see docs/mcp-tools.md for details). Use this tool when you need to jump to a function's or type's definition, inspect its signature, or see who calls it and what it calls, before editing it. If name (optionally narrowed by file) still matches more than one symbol — e.g. overloaded declarations in the same file — pass line as well; the find tool's results already carry the exact name/file/line triple needed to pin down one match. The response can be truncated to fit within a token budget; when it is, truncated is set to true and truncated_count reports how many additional callers and callees together were left out. The budget is divided evenly across every matched symbol first (so if name is ambiguous and returns several results, no single match can consume the whole budget and starve the others), then each match's own share is split evenly between its callers and its callees (so a hot function's huge caller list can't crowd out its callees, or vice versa). The index is built ahead of time by corbel index and only reflects the state of the repository as of the last index run.

Trace the blast radius of changing a symbol: starting from the given symbol, walk the reverse call graph — direct callers, their callers, and so on across multiple hops — and return every symbol that could be affected by a change to it. Each affected symbol comes with a depth field (how many hops away it is) and a resolution field naming which index-wide lookup resolved that call edge (e.g. same-file, scoped, global-unique — scoped/global-unique both mean a single matching definition was found index-wide, see docs/mcp-tools.md), so results are grounded in real, resolved call relationships rather than a text search for the symbol's name (which cannot follow more than one hop and cannot tell a real call from a coincidental name match). Use this tool before refactoring — e.g. changing a function's signature or behavior — to find every place in the codebase that may need to change as a result, including indirect callers that a single-hop "find references" would miss. The response can be truncated to fit within a token budget and/or a hop-count depth; when either limit is hit, the corresponding flag is set (truncated/truncated_count for the budget, depth_truncated for depth) so a caller can tell which one cut the walk short. Results come from the local corbel index built by corbel index and only reflect the repository as of the last index run.

Search the local corbel index for symbols whose name matches a query, for when you don't know the exact symbol name to pass to get_symbol. Matching is substring-based and case-insensitive, ranked so exact name matches come first, then names starting with the query, then names merely containing it; within each of those tiers results are ordered by name, then file, then line for a stable, repeatable order. Case-insensitivity only folds ASCII letters (SQLite's default LIKE behavior) — it will not match differently-cased Unicode identifiers (e.g. a Python or JavaScript identifier using non-ASCII letters), so queries against such names must match the identifier's actual case. Each match reports name, file, and line, which together are the exact triple get_symbol needs to pin down that one symbol even when several symbols elsewhere share its name. This tool does not resolve or return call relationships — use get_symbol or impact on a specific match for that. The response can be truncated to fit within limit and a token budget; when it is, truncated is set to true and truncated_count reports how many further matches were left out. Results come from the local corbel index built by corbel index and only reflect the repository as of the last index run.