A field guide to any codebase, written by the code itself: what depends on what, what changes together, and who knows it.
Field Guide presents 6 well-named, read-only tools with clear, domain-specific descriptions (avg 156 chars). All tools have proper JSON Schema input definitions with type constraints and parameter descriptions. Naming follows verb_noun convention (overview, describe_file, impact_of, changes_with, hotspots, reindex). Schemas include min/max bounds on numeric parameters (limit 1-200, max_depth 1-50, min_confidence 0-1). However, output schemas are not documented, LLMs cannot predict response structure for planning. Error handling is minimal; no recovery guidance or actionable error messages visible. Tool composition is sound (each does one thing), but descriptions lack explicit WHEN-to-use guidance that would optimize LLM selection.
Files that are historically committed together with the given one, with how often. Derived from git history rather than from imports, so it surfaces couplings no static analysis can see — duplicated logic, a schema and its migration, a component and its test, two copies of the same helper in different packages. Use when a change feels like it should have a counterpart somewhere.
Everything the map knows about one file: what it imports, what imports it, how far a change to it reaches, how often it changes, who has written it, and which other files tend to change alongside it. Use before editing an unfamiliar file.
The files changing most, weighted so that recent work counts for more than old work. These are where the current effort is, and usually where the current bugs are.
Lists every file that imports the given one, directly or through a chain, nearest first. Answers "if I change this, what else might break". This is transitive, which a text search for the file name is not.
Describes the repository as a whole: where execution starts, which modules everything depends on, which files are changing most, which third-party packages are used, and where knowledge is concentrated in one person. Start here when you are unfamiliar with a codebase, before searching it — it answers in one call what grepping around answers in twenty.
Output schemas not documented. LLMs cannot predict response structure (fields, types, nesting) for downstream planning or data extraction.
Error handling and recovery guidance absent. No visible error classification (retryable vs fatal) or actionable error messages to guide LLM recovery.
Tool descriptions lack explicit WHEN-to-use guidance. Descriptions explain WHAT but not when to prefer one tool over another (e.g., when to call impact_of vs changes_with).
reindex description is vague ('Rarely needed'). Does not explain prerequisites, side effects, or when freshness checks fail and reindex becomes necessary.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Forces the map to be rebuilt. Rarely needed: freshness is checked before every other call, so edits and new commits are picked up on their own. Use it if you suspect the map is behind.
No pagination or result-limiting guidance in tool descriptions. Tools returning lists (overview, describe_file, impact_of, changes_with, hotspots) should document max result count and pagination strategy.