Windows fixes, agent-friendly MCP replies, simpler README

.gitattributes

@@ -0,0 +1,4 @@
+# Launchers: bash needs LF, cmd.exe needs CRLF (labels misparse in LF-only batch files).
+*.sh text eol=lf
+*.bat text eol=crlf
+*.ps1 text eol=crlf

AGENTS.md

@@ -1,6 +1,6 @@
 # AGENTS.md
 
-Entry point for any coding agent working in this repository, or using Wikifier on another tree. The day-to-day contract is `skills/run.md` (Agent Protocol v0.7, package 5.0.0). Read that file before substantive work. This file says what is live, what is history, and which rules are easy to miss.
+Entry point for any coding agent working in this repository, or using Wikifier on another tree. The day-to-day contract is `skills/run.md` (Agent Protocol v0.7, package 5.1.0). Read that file before substantive work. This file says what is live, what is history, and which rules are easy to miss.
 
 ## What this program is
 

AGENTS.md.wiki.md

@@ -2,4 +2,4 @@
 
 **Role:** Entry point for any coding agent in this repository.
 
-It points at `skills/run.md` (protocol v0.7, package 5.0.0) and names the live Rust binary, the daily loop, and the Python-era records kept under `Findings/archive/`. `pip install wikifier` is the archived Python 4.7 package. Map roots stay `src/` and `tests/`. The import scan stays comment-masked regex plus named re-exports. Rust also resolves `crate::`, `self::`, `super::`, root crate names, and `[workspace] members`, so `src/main.rs` and `tests/*.rs` are dependents of the library modules they import. `mark-green` checks Rust `pub` names, impl methods, and trait-impl items.
+It points at `skills/run.md` (protocol v0.7, package 5.1.0) and names the live Rust binary, the daily loop, and the Python-era records kept under `Findings/archive/`. `pip install wikifier` is the archived Python 4.7 package. Map roots stay `src/` and `tests/`. The import scan stays comment-masked regex plus named re-exports. Rust also resolves `crate::`, `self::`, `super::`, root crate names, and `[workspace] members`, so `src/main.rs` and `tests/*.rs` are dependents of the library modules they import. `mark-green` checks Rust `pub` names, impl methods, and trait-impl items.

CHANGELOG.md

@@ -9,6 +9,38 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [5.1.0] - 2026-11-05
+
+Agent-facing fixes found by driving the MCP server like an agent on real repositories, Windows fixes, and a speed pass. Protocol stays v0.7.
+
+**MCP clients, note:** `session_bootstrap` and `get_project_status` now list at most 100 files unless the call passes `limit` (`files_total` gives the full count, `limit: 0` lists every file), and `update_maps` leaves out the edge list unless `edges: true`. Orient with `map_index` and `list_paths`. The CLI defaults for both commands are unchanged.
+
+### Fixed
+- Windows: `wikifier.bat` passes the binary's exit code through. It read `%ERRORLEVEL%` inside a parenthesized block, so every failure exited 0.
+- Windows: health baselines hash CRLF as LF (`content_sha256_prefixed`). A `core.autocrlf` checkout no longer flips every Green file to dirty (51 files in this repo).
+- A file argument written with `\` separators or as an absolute path under the root names the same row as `src/a.rs`. Before, an absolute path returned empty lookups and a backslash path created a second health row.
+- `prepare_edit`, `get_dependencies`, `get_dependents`, and `get_file_wiki` return `FileNotFound` with a `list_paths` hint for a path that is neither on disk nor in the map, instead of `success: true` with empty fields.
+- MCP `tools/call` with a missing required argument (for example `file_path` instead of `file`) returns `InvalidArguments` naming the accepted and received arguments. Before, the tool ran with an empty path and reported success.
+- `health --limit N` (and MCP `limit`) caps the entry list and adds `entries_total`. The argument was advertised and ignored.
+- `pending_updates.md` no longer says "(no pending items …)" above a list of items. The note is dropped when the first item arrives and comes back when the last one is removed. This was carried over from the Python package.
+- A `record-change` pending item reads `- path: <reason> — refresh the wiki, then mark-green` instead of the Python wording `LLM/agent edit — <reason>`.
+
+### Performance
+Same outputs: on a 587-file Cargo workspace the cache files, `library.md`, folder cards, health hashes, and the `update-maps`, `cycles`, `dependencies --full`, `dependents`, `check-changes`, and `session-bootstrap` output were byte-identical before and after (CLI, timestamps aside). Warm MCP calls there, median of 5 on Windows 11: `check_changes` 353 → 53 ms, `update_maps` 708 → 50 ms, `get_cycles` 317 → 39 ms, `session_bootstrap` 33 → 16 ms and 21.7 → 4.7 KB. Cold `update-maps` (CLI, runs interleaved on a loaded machine, median of 3) went from 18.7 s to 6.4 s.
+- Derived cache files (`by_src`, `by_dst`, `payload`, `notes`, barrel records) are written on several threads through a temp file and a rename, so a reader never sees a torn file, and are no longer synced. `wk_cache/sizes.tsv` records each file's length; `update-maps` compares it with one directory listing per kind and treats a missing, resized, or unreadable file as a miss: the source is parsed again or the `by_dst` record is rewritten from the index. Without `sizes.tsv` every file is read and checked once. Lookups that meet a bad record answer from `index.tsv`. On this Windows machine the sync change made no measurable difference to a cold map (median 11.9 s with batched syncs, 12.5 s without, same copy, interleaved, machine at about 80% load): the remaining cost is creating about 2,000 small files.
+- `file_health.json`, `file_health.md`, the journal, and `pending_updates.md` are now replaced through a synced temp file and a rename. They were plain overwrites with no sync.
+- A full map scans each file once; the second JS/TS/Rust resolution pass reuses those scans. Import line numbers come from one substring search per distinct spec instead of a line-by-line scan per import. JS/TS barrel follow-through builds its sorted scan table once instead of once per imported name.
+- `cycles` and `dependencies --full` read edges from `index.tsv` instead of one `by_src` file per mapped file, and `cycles` loads edge flags only for files inside a cycle. `library.md` reads `by_src` records without the notes files.
+- `update-maps` keeps `library.md` and the folder cards when the cache commit changed nothing.
+- `check-changes` and health seeding take a mapped file's content hash from its cache row when size and mtime match (the rule `update-maps` already used), and read the file otherwise or when that hash is not the baseline. Health rows are found by path through a map, not a scan per file.
+- The file walk takes directory and file types from the directory entry instead of two `stat` calls per entry.
+
+### Changed
+- `init` seeds `map_paths.txt` and `monitored_paths.txt` with `src/` when it exists, otherwise with each top-level directory that holds source (a Cargo workspace's `crates/`, a Python package directory, `tests/`). Before, a project without `src/` hit a scope error on the first `update-maps`.
+- MCP `update_maps` omits the edge list unless `edges: true`; `edges_persisted` still gives the count. On a 587-file Cargo workspace the list was about 280 KB. The schema now lists `full`, `max_files`, and `edges`; `health` lists `limit` and `summary`.
+- `.gitattributes` keeps `*.sh` LF and `*.bat` / `*.ps1` CRLF on every checkout.
+- `session-bootstrap --limit N` (MCP `limit`) caps the `files` list and adds `files_total`. Over MCP the default cap is 100; `limit: 0` lists every file. `map_index` and `list_paths` are the way to orient on a large tree.
+
 ## [5.0.0] - 2026-11-03
 
 First release of the std-only Rust rewrite. `[dependencies]` is empty: no crates.io crates and no Python runtime. The binary implements the Python 4.7 agent contract. Protocol v0.7 stays. `pip install wikifier` still installs that Python 4.7 package.

CLAUDE.md

@@ -6,7 +6,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 Wikifier is a **zero-dependency, agent-to-agent codebase wiki**: it gives AI agents a token-efficient, queryable map of a codebase (health matrix, dependency graph, per-file wiki summaries) so they can look things up instead of re-reading full sources. The human layer (`index.html` dashboard) is strictly secondary — a read-only viewer of the same artifacts. It is deliberately NOT a general human documentation tool.
 
-**This repo dogfoods itself.** Wikifier runs on its own source, and you are expected to follow its agent protocol when working here (see below). `AGENTS.md` is the entry point for any coding agent. The authoritative contract is `skills/run.md` (Agent Protocol v0.7, package 5.0.0). Read it before substantive work.
+**This repo dogfoods itself.** Wikifier runs on its own source, and you are expected to follow its agent protocol when working here (see below). `AGENTS.md` is the entry point for any coding agent. The authoritative contract is `skills/run.md` (Agent Protocol v0.7, package 5.1.0). Read it before substantive work.
 
 **USE WIKIFIER FIRST (non-negotiable in this repo, preferred on any managed target):** At session start call `session_bootstrap` (MCP or `wikifier session-bootstrap`). Prefer Core daily tools (`list_core_tools`): bootstrap, `check_changes`, `prepare_edit`, `suggest_next_actions` (json `actions[]`), `record_change`, `mark_green`. Look up via map/wiki/deps before re-reading full sources. Selective work only (🔴 / actionable 🟡). Never skip `record_change` after edits.
 
@@ -56,12 +56,12 @@ wikifier mcp                           # MCP server on stdio
 
 ## Architecture
 
-Pipeline: **scan → parse → resolve → cache → health → artifacts**, implemented in Rust under `src/` (version in `Cargo.toml`, currently **5.0.0**). The Python package is no longer in this repository. `pip install wikifier` still installs Python 4.7. CLI (`src/main.rs`) and MCP (`src/mcp.rs`) both call `src/api.rs`. Every function returns a JSON object with `success`. Operational failures return `{"success": false, "error", "error_type"}`.
+Pipeline: **scan → parse → resolve → cache → health → artifacts**, implemented in Rust under `src/` (version in `Cargo.toml`, currently **5.1.0**). The Python package is no longer in this repository. `pip install wikifier` still installs Python 4.7. CLI (`src/main.rs`) and MCP (`src/mcp.rs`) both call `src/api.rs`. Every function returns a JSON object with `success`. Operational failures return `{"success": false, "error", "error_type"}`.
 
 - `src/api.rs` — the library surface: `session_bootstrap`, `check_changes`, `prepare_edit`, `suggest_next_actions`, `record_change`, `record_changes`, `record_deletion`, `mark_green`, `verify_wikis`, `update_maps` / `update_maps_run`, `dependency_report`, `load_dependents`, `cycles`, `init_project`, `serve_loop`, `monitor_loop`, `daemon_ctl`. `check_changes` watches mapped files, Green files, and `monitored_paths.txt`. `record_changes` reads `git status` (git is optional).
 - `src/scope.rs` — project root, `map_paths.txt`, `monitored_paths.txt`, excludes, and the file walk. A missing configured map directory is a scope error. Three or more child projects make the root a container and the walk stops.
 - `src/parse.rs` — comment-masked regex scans for Python, JS/TS, Rust, Go, C/C++, C#, and Java. Named JS/TS re-exports are followed to depth 8 (`via_barrel`). An indented Python import is `conditional`. `import()` is `dynamic`. JS/TS resolution reads `tsconfig.json` `paths`, `package.json` `exports`, and npm `workspaces`. Rust resolves `mod`, `crate::`, `self::`, `super::`, and package or lib names from the root `Cargo.toml` and `[workspace] members` (`-` to `_`, globs included). `crate::` stays in the file's crate; a member name points at that member's library. This is not a CPython `ast` parser. The accuracy gate is `accuracy_fixtures_match_frozen_edges`.
-- `src/cache.rs` + `src/cache_commit.rs` — text index at `.wikifier_staging/wk_cache` (per-file payload, `by_src`, `by_dst`, notes, barrels). One commit per `update-maps`. Size and mtime skip unchanged files. `--full` rehashes. `src/legacy_cache.rs` imports an empty cache from `import_cache.sqlite` (the `sqlite3` CLI) or `import_cache.json`.
+- `src/cache.rs` + `src/cache_commit.rs` — text index at `.wikifier_staging/wk_cache` (per-file payload, `by_src`, `by_dst`, notes, barrels). One commit per `update-maps`. Size and mtime skip unchanged files. `--full` rehashes. Cache files are derived: they are renamed into place but not synced, `wk_cache/sizes.tsv` records their lengths, and a missing, resized, or malformed file is a miss that the next `update-maps` re-parses. Health, journal, and pending writes are synced. `src/legacy_cache.rs` imports an empty cache from `import_cache.sqlite` (the `sqlite3` CLI) or `import_cache.json`.
 - `src/artifacts.rs` — `file_health.json` (schema version 2) plus the `file_health.md` view, pending lines, and the journal. Identical journal retries within 120 seconds are skipped. `mark_green` stores `public_symbols`, `wiki_hash`, and `source_mtime_ns`, and refuses a Python or JS/TS wiki that misses the current public names, still mentions a removed name in backticks, or is unchanged after the public surface changed.
 - `src/maps_out.rs` — `library.md` (a file tree, a Mermaid block capped at 80 edges, and the dependency table) and sharded folder cards at `.wikifier_staging/maps/_root.json` and `maps/{prefix}/_index.json`.
 - `src/project_lock.rs` — advisory lock `.wikifier_staging/.lock`, taken before the cache lock, re-entrant per thread, 30 second wait, stolen when the recorded pid is dead.

CLAUDE.md.wiki.md

@@ -2,4 +2,4 @@
 
 **Role:** Repo guidance for an agent working in this tree.
 
-Hard constraints: empty Cargo dependencies, no hand-edits of `library.md` or the health and pending files, and the human dashboard stays secondary. The Python-era plan logs live under `Findings/archive/`. It points agents at `AGENTS.md` and `skills/run.md` (protocol v0.7, package 5.0.0). The architecture section names the Rust modules under `src/`, the library file tree, and this repo's map roots `src/` and `tests/`. The parse note includes Rust `crate::`, `self::`, `super::`, Cargo.toml crate names, `[workspace] members`, and `rust_public_symbols` for wiki checks. JSON failures use `success: false`.
+Hard constraints: empty Cargo dependencies, no hand-edits of `library.md` or the health and pending files, and the human dashboard stays secondary. The Python-era plan logs live under `Findings/archive/`. It points agents at `AGENTS.md` and `skills/run.md` (protocol v0.7, package 5.1.0). The architecture section names the Rust modules under `src/`, the library file tree, and this repo's map roots `src/` and `tests/`. The parse note includes Rust `crate::`, `self::`, `super::`, Cargo.toml crate names, `[workspace] members`, and `rust_public_symbols` for wiki checks. JSON failures use `success: false`.

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "wikifier"
-version = "5.0.0"
+version = "5.1.0"
 edition = "2021"
 description = "Wikifier — zero-dependency agent-to-agent codebase wiki."
 license = "MIT"