v2.3.0 — token budget, duplicate detector, monorepo hints

Added

--budget=<tokens> — fit any context window

Trims the least-imported exports until the snapshot fits a token ceiling. Usage ranking only counts imports that actually resolve to the scanned files (with /index fallback) or name a workspace package — import { format } from 'date-fns' can never inflate a local export's rank. --budget also lifts the scan cap from 300 to 5,000 files, since large repos are exactly who needs it. Config key: "budget".

--dupes — find duplication before your AI adds more

Flags likely duplicate exports: identical names in different files (parseISO ≈ parse_iso) and same-words-different-order names (formatDate ≈ dateFormat), each with file:line locations. Exits 1 when any are found — slots into CI next to --check.

Monorepo awareness

npm/yarn/pnpm workspaces are auto-detected (package.json workspaces, pnpm-workspace.yaml). Import hints for files in another workspace package now say @acme/ui instead of ../../packages/ui/src/button — in markdown, JSON, and MCP output.

Fixed

• MCP root parameter on all tools: Claude Desktop starts MCP servers in / or the home directory, so cwd-based scans silently found nothing there. Pass your project path once and the tools work from any host.
• Version-stamped snapshots: the header now carries contextsnap@<version> (JSON gets a "version" field). --check detects a generator/checker version mismatch and tells you to pin the version in CI, instead of reporting a misleading "stale".
• Prompt-injection hardening: JSDoc is capped at 600 chars and the snapshot preamble instructs the AI to treat comment text as reference material, never as instructions.
• README no longer claims --budget "guarantees" a ceiling — token counts are ~4 chars/token estimates.

🤖 Generated with Claude Code

v2.2.0 — MCP query tools, CLAUDE.md inject, CI drift check

Added

MCP query tools

• search_exports — find an export by name, signature, or JSDoc substring. Returns file:line locations so your agent jumps straight to a definition instead of pulling the whole snapshot.
• get_file_context — compact contract block (signatures, line numbers, JSDoc) for a single file. Much cheaper than get_context when the agent already knows the file.

--inject[=<file>]

Writes the snapshot between <!-- contextsnap:start/end --> markers in CLAUDE.md / AGENTS.md (auto-detected) or any file. Your content outside the markers is never touched; re-runs are idempotent. Every Claude Code / agent session picks the context up automatically — no clipboard, no MCP setup. Composes with --watch. Config key: "inject": true | "<file>".

--check

CI drift guard: regenerates the snapshot in memory and compares it (generation dates ignored) against the committed .ai-context.md — or the injected block when combined with --inject. Exits 1 on drift, like a lint step:

- run: npx contextsnap src --check

Fixed

• MCP scans no longer crash with ENOENT when an auto-detect directory (e.g. src/lib) doesn't exist.
• MCP server version now reads from package.json instead of a stale hardcoded string.

🤖 Generated with Claude Code

v2.1.0 — per-export line numbers

Added

Source line numbers on every export — requested by a user so AI agents can jump straight to an implementation instead of reading the whole file.

• Markdown output: each signature ends with // :N (its 1-based line in the source file), and the preamble tells the agent what the suffix means:

  export function formatLocal(isoString: string): string; // :12

• JSON output (--format=json) and the MCP get_context tool: each export gains a "line" field.
• Zero config — always on, ~2 tokens per export.

🤖 Generated with Claude Code

v2.0.1 — parser validation layer

Fixed

Silent-failure gaps in export extraction. The fast regex parser could silently drop or corrupt exports — giving the LLM false confidence about what exists in your codebase. This release closes that class of bug:

• New export forms extracted: export default <expression> (identifiers, object literals, arrow functions), export enum, export abstract class, export let/var, generator functions, export * [as ns] from, export type { ... }, export declare
• Regex-literal safety: source containing /"/ or /{/ no longer derails string/brace scanning
• No phantom exports: export syntax inside string values (e.g. templates that build export statements) is no longer matched
• Validation layer: every file is cross-checked against es-module-lexer. Any ESM export the fast parser misses is still emitted as a name-only stub (export host; // signature unresolved) with a CLI warning — the context never silently omits an export.

Notes

• New dependency: es-module-lexer (tiny, zero-dep, wasm-based)
• parseFile() results now include a warnings array
• Known limitations documented in README (TS generic object constraints, decorators, TS lexer coverage)

🤖 Generated with Claude Code