codebase-snapshot 0.0.0-stage → 0.9.0

package/CHANGELOG.md

@@ -0,0 +1,1242 @@
+# Changelog
+
+All notable changes to `codebase-snapshot`.
+
+---
+
+## 0.9.0
+
+### The package is now `codebase-snapshot`, unscoped
+
+**Breaking:** published as `codebase-snapshot` rather than `@ankurhalder/codebase-snapshot`.
+
+The old name is **deprecated, not unpublished**, so existing instructions keep working:
+
+```bash
+npx @ankurhalder/codebase-snapshot generate   # still resolves, now prints a deprecation notice
+npx codebase-snapshot generate                # the supported name from 0.9.0
+```
+
+Why: npm's search index ranks on text match rather than downloads — a package with a few thousand
+weekly downloads outranks one with a hundred million when the query matches its name better — so a
+name that *is* the phrase people search is the strongest free signal available. Doing this at
+version 0.9.0 with no dependents costs one migration; doing it after documentation, articles and
+tutorials all point at the scoped name costs far more.
+
+The rename is not cosmetic. The package name is written into every ownership marker and every
+manifest, and the output-directory guard compares it exactly, so renaming it naively would have made
+**every snapshot written by 1.0–4.4.x unrecognisable to its own tool** — the tool would have refused
+to update its own output. Ownership now accepts the legacy spelling and rewrites it to the current
+name on the next run, so an existing snapshot keeps working and is upgraded in place.
+
+### Metadata and discoverability
+
+- `keywords` went from 9 to 20, chosen by measured position rather than intuition — phrases like
+  `codebase-to-prompt` and `repo-to-llm` that describe how people actually ask for this.
+- The `description` now contains the word "codebase" and names the clients people search for.
+- The README's 16 relative links became absolute. On npmjs.com — the most-linked page the package
+  has — every documentation link was resolving to a 404.
+- The README states build provenance, which the release pipeline has been producing since 4.0.0
+  without ever mentioning.
+
+### Documentation site
+
+The docs now have a crawlable home at <https://codebase-snapshot.ankurhalder.com>, generated from
+`README.md` and `docs/` in this repository, with a sitemap, its own `robots.txt`, an `llms.txt` at
+the site root and a Markdown twin per page. GitHub's `robots.txt` blocks directory-tree and raw-file
+paths and publishes no repository sitemap, so `docs/` had no crawlable entry point before this.
+
+The GitHub wiki is closed. The same prose was being maintained in the wiki, `docs/` and the README
+with no canonical copy; `Home.md` is now a single pointer and the other pages are one-line stubs so
+inbound links keep resolving.
+
+### New guides
+
+- [Feeding a codebase to ChatGPT, Claude or Cursor](docs/how-to-feed-a-repo-to-claude.md)
+- [Budgeting LLM context](docs/how-to-context-budget.md)
+- Comparisons with [repomix](docs/vs-repomix.md), [gitingest](docs/vs-gitingest.md) and
+  [code2prompt](docs/vs-code2prompt.md), each stating where the alternative is the better choice
+
+### Tooling
+
+- `npm run seo-rank` measures this package's position in npm and GitHub search for a fixed query
+  set, appends to `seo-log.jsonl`, and `--check` fails when a query we hold slips. It runs weekly in
+  CI.
+- `npm run site` renders the documentation site; `scripts/check-site-links.mjs` fails the deploy on
+  an unresolvable internal link.
+
+---
+
+## 4.4.1
+
+### `setup` ran the install with npm's own environment, which could refuse it
+
+Running `setup` through `npx` — the documented way — put the parent npm's resolved configuration
+into `npm_config_*` variables, and the nested `npm install` inherited them. With an npm that
+supports an `allowScripts` switch the child aborted with
+`EALLOWSCRIPTS: --allow-scripts is not allowed in project-scoped installs`, so the install was
+skipped and only a warning said so. The nested run now gets the caller's environment minus
+`npm_config_*` (and npm's lifecycle variables), so it reads the project's own `.npmrc` and flags —
+which is the point of using the project's package manager at all. A shim-based test pins it: the
+package manager is invoked, and none of npm's config reaches it.
+## 4.4.0
+
+### One command that does the whole first run
+
+`npx @ankurhalder/codebase-snapshot setup` — detect the project, write the commented config (only
+when the repository has none; an existing one is kept byte for byte and the output says so), add
+the npm scripts, the devDependency and the `.gitignore` entries, install dependencies with the
+package manager the repository's lockfile names (`bun`, `pnpm`, `yarn` or npm), then generate
+through the normal `generate` command so every guard, warning and JSON field is the same one a
+plain run produces.
+
+- `--no-install` skips the install: CI, an offline machine, or a lockfile the project's own tooling
+  owns.
+- A failed install is a warning, not a dead end — the snapshot is still generated, because the
+  scripts resolve the package through `npx` until it is installed.
+- `--dry-run` prints the plan and the config it would write, and writes nothing.
+- `--format json` emits one document: `{ "setup": { … }, "manifest": { … } }`.
+- The ownership guard is untouched: an `outDir` holding files this tool did not write is still
+  refused (exit 1), and the refusal names the entries.
+
+`init` keeps its old behaviour — it refuses when a config exists and needs `--force`, because
+someone who typed `init` should be told. `setup` is the merge-friendly path.
+
+### An outDir that holds only empty directories is replaced
+
+Reported from a real repository: `generate` refused with "it holds no .codebase-snapshot marker
+(found parts, reports) — remove the directory yourself" for a directory that held **no files**, just
+`parts/` and `reports/` from an interrupted run. Replacing a tree with no files destroys nothing, so
+the guard now treats it as empty. The walk is iterative, a symlink counts as content, and a
+directory holding even one file is still refused without a marker.
+## 4.3.0
+
+### Node 18 and 20 actually run the suite now — and one of them was broken
+
+`engines` has said `>=18`, and until this release no run had ever executed on 18 or 20: the publish
+workflow ran `npm run verify` on one Node version only. The gate now runs on 18, 20 and 22 for every
+release, and the first honest run found a real defect — see the next section. Verified locally on
+18.20.4 and 20.19.4: 498 tests, 493 pass, 0 fail, 5 version-dependent skips.
+
+### `validate` called an unreadable file "missing from disk" on Node 18
+
+A Windows `Deny-Read` ACE makes Node 18's `statSync` throw `EPERM` and `existsSync` answer `false`,
+so a file the user could not read looked *deleted*: `validate` reported "no longer exists on disk"
+for a file sitting right there. The validator now asks why a path cannot be stat'd (`ENOENT` means
+gone, `EPERM`/`EACCES` means unreadable) instead of trusting a boolean, and both drift classes are
+deduplicated — an unreadable file is seen by the scan loop and the manifest loop.
+
+### A base64-wrapped private key is detected
+
+`LS0tLS1CRUdJTi…` is the base64 of `-----BEGIN `, so a wrapped key is found by its prefix without
+decoding anything. The new `base64-private-key` rule takes the built-in set to 30; archives and
+other encodings remain out of scope, and `SECURITY.md` now says which is which.
+
+### A dependency manifest is data, not documentation
+
+`requirements.txt` and its variants were classified kind `doc` because they end in `.txt`. They are
+machine input, like `package.json`, so they are `data` now.
+
+### The monolith explains the one field it cannot carry
+
+`full-codebase.json` embeds a copy of the manifest, and a file cannot record its own sha256. The
+embedded copy keeps `artifacts.fullJson: null` and carries a `selfReference` sentence pointing at
+`manifest.json`, which holds the real byteSize and digest, instead of leaving the reader to wonder
+why the two manifests disagree.
+
+### Empty and trailing-blank line accounting is documented
+
+`includedLines` counts content lines after CRLF normalisation and without a trailing empty element:
+an empty file records `0` while the fence shows one empty line. That is the intended meaning of the
+field, and `docs/output.md` now states it rather than leaving it to be rediscovered.
+
+### The audit release
+
+Every other change here came out of a full audit of the tool against eight real repositories and an
+adversarial review of every module. The three defects that mattered most were all in the two
+promises this package leads with — "never leak secrets" and "machine-readable output".
+
+### Two ways a secret kept its tail
+
+A quoted credential containing an escaped quote was masked only up to the backslash:
+`JSON.parse('{"password": "p\"ssw0rd"})` was emitted as
+`"[REDACTED: potential secret-like value omitted from snapshot]"ssw0rd"` — most of the value in
+clear, and the line left syntactically broken. An unquoted value stopped at the first character
+outside `[A-Za-z0-9_+=/-]`, so a percent-encoded token kept its tail:
+`?access_token=abc%2Edef&scope=read` lost `abc` and shipped `%2Edef`. Both value classes now
+describe the value they are masking — escaped pairs are consumed by the quoted rule, and `%`,
+`:` and `~` are token characters — and the mask replaces the whole value.
+
+### A compressed fixture was emitted as mojibake, at twice its own size
+
+`.br` was missing from `BINARY_EXTENSIONS`, so ripgrep's 186-byte `sherlock.br` failed UTF-8
+decoding instead of being recognised as binary: it was emitted `full` as 341 bytes containing 78
+U+FFFD — the one case where the artifact is larger than its source and says nothing. `.br` (and
+`zst`, `lz4`, `lzma`) are now in the list, and a second guard catches the next unknown format:
+at least 32 undecodable bytes **and** at least a third of the file means binary. The bar is high
+on purpose — a Latin-1 document is legitimately full of invalid UTF-8 bytes and must stay
+`utf-8-lossy`.
+
+### A config that exported nothing ran on the built-in defaults
+
+`export default null;` was erased by a nullish fallback (`imported.default ?? imported`), so the
+module namespace stood in for the config, every setting was ignored, and the run generated with
+the defaults at exit 0. The loader now distinguishes "no default export" from "the default
+export is null" and refuses with the same message every other wrong export gets.
+
+### An unknown setting was dropped in silence
+
+`{ maxPartByte: 4096 }` generated with the default budget at exit 0 — the typo was invisible, and
+it is the silent fallback this project forbids. A config file (module, JSON or `.snapshotrc`)
+now rejects any key that is not a setting, naming it and pointing at the generated config as the
+list of what exists.
+
+### `--format json` produced nothing when the arguments were wrong
+
+Every usage error — unknown flag, unknown command, bad value, missing `--cwd`, wrong arity —
+exited 2 with **0 bytes** on stdout, while the docs promised one parseable document per JSON-mode
+invocation. The format is now read from raw argv before validation, so the error document is
+emitted for usage errors too; the human line stays on stderr.
+
+### An async hook could append to the JSON payload
+
+The stdout redirect that keeps `console.log` inside a hook out of the payload wrapped only the
+synchronous call. A hook that logged after an `await` appended its text after the JSON, so
+`--format json` output was no longer parseable while the command still exited 0. The redirect now
+lives until the returned promise settles, and the payload is written to the real stdout exactly
+once.
+
+### Truncation was invisible in the audit
+
+`reports/audit.md` printed "✅ No files exceeded the large file budget" while a third of a
+document was missing from the snapshot, because the cut came from `maxPartBytes` rather than
+`largeFileBytes`. It is structural for the `docs` preset, whose 100 KB part budget sits *below*
+the 200 KB large-file threshold. Section 3 now reports truncation from either cause, with bytes,
+lines and a `Cut by` column, and `generate` warns when `maxPartBytes < largeFileBytes`.
+
+### A part-budget cut carried no marker
+
+When the part budget was the binding constraint the block simply stopped — no
+`[N lines omitted]` anywhere, and the manifest said `truncated: true` while the artifact read as
+complete. Every budget-forced cut now carries a marker naming the budget
+(`… [19940 lines omitted: part budget 1887 B] …`), recomputed correctly when a second shrink pass
+runs.
+
+### The preview tail window lost lines on newline-terminated files
+
+`content.split("\n")` yields a trailing empty element, which consumed a tail slot: a 10-line file
+at `previewHeadLines: 2`, `previewTailLines: 2` emitted `line 1`, `line 2`, an omission marker and
+**no tail at all**, while `includedLines` claimed one more line than the block carried. The tail
+window and the omission count now describe what is emitted.
+
+### A filename could inject a link into `reports/navigation.md`
+
+The link label escaped `|` and control characters but not brackets, so a file named `b](b).ts`
+produced `[src/b](b).ts](…)` — a live Markdown link to `b` — and `x](/evil).ts` linked to `/evil`.
+The label is escaped; the link target and anchor are unchanged.
+
+### A deep directory tree crashed the run
+
+The project-tree walker recursed per directory, so a ~6,000-level chain overflowed the stack after
+76 s and wrote **no snapshot**. The walk is iterative, the output order is unchanged, and the tree
+is capped at 512 levels with a marker and a warning.
+
+### One oversized text file aborted the whole run
+
+A file over ~512 MiB of *text* threw `Cannot create a string longer than 0x1fffffe8 characters`
+from inside the run, with no `skipped[]` record; the equivalent *binary* file was skipped
+correctly. A decode failure is now a per-file `read-error` skip and the run continues.
+
+### `validate`, `report` and `diff` now agree on what a snapshot is
+
+`validate` called a manifest "valid and up to date" with `fingerprint` deleted, `report` printed
+`undefined`, and `diff` rejected the same file. One shared guard now requires `generatedAt`,
+`schemaVersion`, a 64-character lowercase `fingerprint`, `counts`, `totals`, `layers` and the
+`files`/`skipped`/`parts` arrays, and all three commands use it.
+
+### `diff` notices a schema or a parts change
+
+A copy with `schemaVersion: "9.9.9"`, a deleted `parts` array, or a changed part digest was
+reported as `+0 -0 ~0 =N unchanged`. `diff` now reports a schema mismatch as a difference and
+compares part metadata by filename.
+
+### `--discovery walk` now honours `.gitignore` escapes
+
+`\#hashfile.ts`, `\!bangfile.ts` and `spaced\ .ts` were all *included* in walk mode while git
+ignored them — the parser trimmed and sniffed first characters, and the matcher normalised `\` to
+`/`. The parser follows gitignore(5)'s line grammar and an escape-aware matcher handles patterns
+containing a backslash; plain `!` negation is unchanged.
+
+### `init` writes the repository's real branch
+
+The generated config always said `branch: "main"`, so on any repository whose default branch is
+`master` or `vNext` the first `validate` printed a branch-mismatch warning the user never caused.
+It now records the branch `getGitInfo` reports and falls back to `"main"` only when git cannot
+report one.
+
+### The contract publishes `check` and `--mode`, and stops over-promising `--option=value`
+
+`check` is documented as an alias of `validate`, the hidden `--mode` selector is in
+`describe.options` with its values and contradictions, and the help sentence now says
+"value-taking options" — booleans never accepted `=true`.
+
+### Documentation
+
+The complete list of corrections is in the wiki and `docs/`: option counts, the removed `agents`
+skill target, `--tmp-dir` (a config key, not a flag), the `--watch` message, the shipped README's
+404 links (`CHANGELOG.md` and `SECURITY.md` now ship; the rest are absolute URLs), and
+measurement-backed statements for determinism, text-mode output streams, monolith byte counts and
+ignore granularity.
+
+### The generated config was a lint warning in every Next.js project
+
+`init` wrote `export default { … }`. `import/no-anonymous-default-export`, which
+`eslint-config-next` enables by default, reports that inline form — so in the framework this tool
+detects most often, the one file we tell people to keep and edit arrived flagged the moment it was
+created. The object is now assigned to `config` and exported by name:
+
+```js
+const config = { /* every setting */ };
+
+export default config;
+```
+
+Both forms load identically; only one is quiet. The reason is in the file itself, so deleting the
+comment does not invite the warning back.
+
+### A clean repository reported `Warnings: 156`
+
+Every masked line was appended to `manifest.warnings` as `<rule> at <path>:<line>`, so a
+repository with nothing wrong with it reported 156 "warnings" — the first thing a new user sees
+after `init` + `generate`, and the reason a 583-file project looked broken. Worse, the summary
+printed ten of them under that count with nothing to say the other 146 existed.
+
+Redaction is not a warning: a masked value is the tool doing its job, while a warning says the
+snapshot may surprise you. The two are now separate — the summary prints
+`Redacted: N line(s) — values masked; see reports/audit.md` and keeps `Warnings:` for the things
+worth acting on, and a warnings list longer than the ten-row preview ends with
+`… and N more in manifest.warnings[]`. When `reports/audit.md` is disabled the manifest says the
+per-line detail was not written instead of losing it silently. `manifest.totals.redactedLines`
+carries the count for machines.
+
+### One line of code was counted twice
+
+`scanAndRedactSecrets` pushed a finding per (line, rule) pair, so an unquoted
+`JWT_SECRET=<value>` — which matches both `credential-assignment-unquoted` and
+`env-secret-assignment` — appeared twice in `reports/audit.md` on the same file and line and
+inflated the total, making a 78-site repository report 156 findings. A finding is now one *line*
+with `rules[]` naming every rule that matched it; the audit table has one row per line and a
+`Rules` column that lists them.
+
+### Artifact counts followed the generating machine's locale
+
+`reports/audit.md` and the snapshot `README.md` called `toLocaleString()` with no locale, so the
+same repository reported `1,24,976` lines on an `en-IN` machine and `124,976` on an `en-US` one —
+different bytes from the same commit, which defeats diffing two snapshots or caching one. Both
+generators now share `formatNumber()` with the CLI, which already pinned `"en-US"`.
+
+## 4.1.1
+
+### `init` left the package undeclared, so nothing reached `node_modules`
+
+`init` wrote `npx @ankurhalder/codebase-snapshot generate` into `package.json` and no dependency
+at all. `npm install` therefore installed nothing, every `npm run snapshot` resolved `@latest`
+from the registry instead of a pinned copy, and the `npm install && npm run snapshot` it printed
+as the next step could not have installed it. `init` now records
+`devDependencies["@ankurhalder/codebase-snapshot"]` in the same single write as the scripts, and
+leaves a range that is already declared — including a deliberate pin — untouched.
+
+### `--minimal` wrote a config that documented nothing
+
+The flag wrote only the settings that differed from the defaults, so a Next.js project received
+
+```js
+export default { preset: "nextjs" };
+```
+
+— 179 bytes listing none of the twenty-six settings, in the one file a user opens to change
+them. The flag is gone: `init` always writes the complete commented config, and `describe` no
+longer advertises an option the parser rejects.
+
+### `skill --install` wrote a tool's documentation into the project's `AGENTS.md`
+
+`AGENTS.md` is a repository's own agent contract — build commands, conventions, house rules —
+not a registry for third-party skills. The install created it when missing, which meant a file
+whose name claims the repository documents itself to agents, holding only this tool's text; and
+it merged a marked section into it when present, editing a tracked file on the user's behalf.
+It now writes only the two registries keyed by skill name —
+`.claude/skills/codebase-snapshot/SKILL.md` and `.cursor/rules/codebase-snapshot.mdc` — and
+appends exactly what it *creates* to `.gitignore`, so an install no longer turns a one-line
+command into a decision about which untracked files to commit. A file that already existed is
+never ignored, and a section left behind by 4.1.0 is reported rather than rewritten. To put the
+skill in `AGENTS.md`, print it and paste it: `codebase-snapshot skill >> AGENTS.md`.
+
+`init` also stops printing the tip that led there: an install's whole footprint is now the
+config, `package.json` and `.gitignore`.
+
+### A file's recorded bytes and lines described the file, not the block
+
+`includedBytes` counted the final newline that the closing fence swallows — 6 290 of 6 344 files in
+a real repository were off by exactly one byte — and a CRLF checkout, which is written as LF, was
+over-counted by one byte per line. A bounded preview reported `head + tail + 1` lines, one more than
+its block carried. Both numbers are now computed once, from the body after normalisation and after
+the closing newline is dropped, and an invariant test compares every `### FILE:` block in every part
+against the manifest entry.
+
+### A large file was parsed for a summary the snapshot could not carry
+
+A file above `largeFileBytes` is emitted as a head/tail excerpt, yet it was still parsed into an AST
+and had every identifier masked so a whole-file summary could be attached — a summary describing
+lines no artifact contains, which the partitioner then discarded. On a 1.9 MB source file that was
+70% of the wall clock: 1.8 s / 283 MB peak RSS became 0.5 s / 105 MB. The manifest names every
+skipped preview in one grouped warning; `summaryOmitted` is deliberately not reused for them,
+because it means "dropped for budget" and the part prints that reason.
+
+### `--format json` wrote nothing at all when a command failed
+
+Progress goes to stderr and payloads to stdout, which meant a failed JSON-mode run produced an empty
+stdout: an agent had to notice the empty string, then parse stderr prose and guess. Every failure now
+writes one parseable document — `{ "error": { "message": …, "exitCode": … } }` — while the human line
+stays on stderr and the exit code is unchanged. A successful run never carries an `error` key.
+
+### Smaller corrections
+
+- Masking no longer walks left through a URL path: `https://example.com/v1/<key>` was emitted as
+  `https://example.[REDACTED]`, contradicting "only the matched span is masked".
+- `explain` answers a file inside an ignored directory with the record its directory carries,
+  naming the ancestor, instead of the weaker `not-discovered`.
+- `package.json#snapshot` as an array is reported like any other wrong type instead of being
+  silently ignored, and an unreadable config file names itself in the error.
+- `init` no longer fails with a raw `EISDIR` when `.gitignore` is a directory — after the config had
+  already been written. The `.gitignore` line is a convenience, so it is a note, and `init` finishes.
+- `--format`'s scope is the same machine-readable eight-command list as every other option, and a
+  test now fails if any option's scope regresses to prose or names a command that does not exist.
+- The hook documentation says what "not awaited" does not mean: Node keeps the process alive while
+  the hook's own timers and sockets are outstanding, so `afterGenerate` should stay short.
+- The generators agree with each other: `full-codebase.md` escapes control and bidi characters like
+  every other artifact, the snapshot `README.md` links only to reports that were actually written,
+  and a layer or kind name containing `|` or a backtick can no longer break the table or the inline
+  code span it is listed in.
+- The wizard no longer treats `package.json` `"snapshot": []` as an existing configuration.- The README and `llms.txt` describe what 4.1.0 actually guarantees — the eleven skip reasons, the
+  `encoding` field, discovery modes, the emitted-byte accounting — instead of the feature list from
+  before the release.
+
+---
+
+## 4.1.0
+
+### A UTF-16 source file was invisible, and a Latin-1 one was quietly mangled
+
+Two encoding decisions were silent, and one was wrong in both directions. `generate` decided
+"binary" from a NUL in the first 8 MB and then decoded with `buffer.toString("utf8")`, so a UTF-16
+source file — which is half NUL bytes — was skipped as `binary` and vanished from the snapshot,
+while a Latin-1 file was emitted with U+FFFD in place of every non-ASCII byte, and `validate` then
+reported the snapshot as current because the hash covers the raw bytes.
+
+Decoding is now explicit. A BOM, or the NUL padding of a BOM-less UTF-16 file, is recognised and
+the file is transcoded; a file that is not valid UTF-8 is kept, decoded lossily, and labelled
+`encoding: "utf-8-lossy"` with the count of undecodable bytes in `warnings`; a NUL anywhere in the
+file — including past the 8 MB head scan, which used to be the end of the check — marks it binary
+and `skipped[]` records the offset. `files[].encoding` is present only when a file was not plain
+UTF-8, so its absence keeps meaning "clean UTF-8". `explain` answers the same question the same
+way. Verified: 12 new tests covering both endiannesses, BOM-less detection, five classes of
+invalid sequence, a NUL at byte 8388608, and hash stability across two runs.
+
+### A file the tool did not write could be deleted without a word
+
+Ownership checking had a hole with a nice name. `.gitkeep` and `.gitignore` were treated as
+placeholders and subtracted from the set of entries belonging to somebody else, so a snapshot
+directory this generator owned and that a user then added a `.gitignore` to was replaced — atomically,
+exit 0, `manifest.warnings` empty — and the file was gone. A directory holding nothing but
+`.gitkeep`, or nothing but a `.gitignore` (`*`, `!.gitignore`), was replaced without any ownership
+evidence at all, because the placeholder shortcut fired before the marker was ever read.
+
+Both are files a user creates; the generator never writes either into an output directory. They are
+foreign entries now, so the run refuses and names them, exactly as it already did for `notes.md`.
+`.DS_Store` remains the only tolerated entry, because Finder writes it and it carries no intent.
+Verified: 6 new tests — owned directory plus `.gitignore`, bare `.gitkeep`, bare `.gitignore`, the
+refusal naming the entry, `.DS_Store` still discarded, and a genuine snapshot still replaced.
+
+### A directory named `snapshot` disappeared, and the manifest said nothing
+
+The output directory was excluded from the scan by name, at every depth: `snapshot`,
+`snapshot.tmp` and `snapshot.backup-*` were dropped before the branch that records a `skipped[]`
+entry, so a directory the user committed under one of those names was missing from the snapshot
+with no record anywhere — the exact omission `skipped[]` exists to prevent, and it contradicted
+the documented promise that every discovered path is accounted for. `generate` and `explain` also
+disagreed: one dropped the path, the other called it `not-discovered`.
+
+Exclusion is now evidence-based. A directory is this tool's output when it carries the
+`.codebase-snapshot` marker, and only the exact `outDir`/`tmpDir` paths a run is about to write are
+excluded by name — so a first run still does not record its own output (two identical runs keep
+producing identical manifests) while a user's directory with the same name is scanned like any
+other. `snapshot` and `snapshot.tmp` also left the built-in ignore list, where they had been
+second-layer protection for the same mistake. Verified: 5 new tests, including the same fixture
+through the git and the walking discovery paths, a full generate that now lists
+`snapshot/keep.md`, and the determinism guard that the second run does not record the first.
+
+### A hook could kill a run that had already succeeded
+
+`hooks.afterGenerate` ran after the snapshot was swapped into place. A hook that returned a
+rejected promise was never awaited and never caught, so Node turned it into an unhandled rejection:
+the process exited 1 *after* printing "Snapshot generated … Integrity: passed" and writing its
+`--format json` payload, and nothing about the failure reached `manifest.json`. A hook's own files
+were also missing from the ownership inventory, so the next run refused a directory the tool had
+written itself, and a `console.log` inside a hook landed on stdout in the middle of the JSON
+payload.
+
+A returned promise is now tracked rather than awaited — this is a synchronous API — and its
+rejection is recorded as a warning on the manifest and in the marker, which are re-shipped when it
+lands. The marker is written after the hook for the same reason it now includes the hook's files.
+stdout is redirected to stderr for the duration of the hook and restored afterwards, so the payload
+channel stays pure. Verified: 6 end-to-end tests through the built CLI, including a rejected async
+hook that must exit 0 with a parseable payload and a recorded warning.
+
+### `validate` would read any file the manifest named
+
+A manifest is an input, and `validate` treated it as trustworthy. It joined `parts[].filename`
+onto `snapshot/parts` with `path.join`, which accepts `../../../some/file` and absolute paths, so a
+manifest naming a file outside the snapshot made `validate` read that file and then report its byte
+size — a read primitive, with a size oracle for the file it named, reachable by anyone who can
+write a manifest. Part names are bare names by construction, so a name that does not stay inside
+`snapshot/parts` is now a drift error and is never read. `explain`'s text form also re-spells
+control and bidi characters in paths, layer names, kind names and part names, because a committed
+`snapshot.config.*` can put a U+202E in a layer name and the terminal renders it.
+
+Verified: 5 new tests — a traversal name, an absolute name, the byte-size oracle specifically
+proven closed, a legitimate snapshot still current, and control characters escaped while accented
+text passes through untouched.
+
+### Also in this release
+
+- **`--watch` reloads a module config.** Node's ESM loader caches by URL, so `snapshot.config.ts`
+  (exactly what `init` writes) was never re-read: an edited config was ignored and a broken one
+  was invisible while the run printed `passed`. The loader now evicts the CJS cache and imports
+  with a per-load token. 6 tests.
+- **`--watch` never drops an edit.** A change during a run, a change inside the post-run quiet
+  window, and a source directory whose name matched the outDir basename were all lost silently
+  while the manifest kept reporting `Integrity: passed`. 6 tests plus live repros.
+- **JWTs are redacted.** Only three-segment JWTs with long signatures were masked; two-segment,
+  padded, empty-signature and short-signature forms reached the parts verbatim. 8 tests.
+- **Display hygiene.** Control, C1 and bidi characters in file and layer names are escaped wherever
+  a path is displayed (parts, tree, navigation, audit, explain), raw bytes are kept in
+  `manifest.json` and file bodies, and the snapshot README and part header say that contents are
+  untrusted repository data. 12 tests.
+- **The hint after `init` is runnable.** It printed `npm run snapshot` with no `package.json`, and
+  a bare package name on re-runs. 4 tests. The scaffolder now records the tool in
+  `devDependencies` at its own version instead of resolving it from the registry on first run.
+  7 tests.
+- **The linter ignores prose.** Content rules matched comments, so documenting why `console` is
+  banned failed the lint that bans it.
+
+### `validate` and `report` could be talked into saying the state was fine
+
+Four ways the state checks answered confidently and wrongly.
+
+A run killed mid-write left the output directory holding a marker and no manifest, and `validate`
+answered "run generate first" for ever — an instruction that could never resolve, because
+generation had already run and failed. The marker records the state, the run id and the start time,
+so the message now names them and says whether to wait or regenerate.
+
+A manifest with no `config.digest` fell back to comparing a handful of fields, so a changed
+`maxMonolithBytes`, `layers`, `kinds` or custom secret rule passed as "valid and up to date". A
+configuration that cannot be verified is not a match: the absence of the digest is itself reported
+as drift.
+
+A file that existed but could not be read was reported as "no longer exist on disk", sending the
+reader to look for a file that was right there. Unreadable and deleted are separate findings now,
+and the summary says which.
+
+`report` printed whatever numbers the manifest contained — `-5 files included`,
+`Total lines: 1.5`, `Total size: Infinity MB` — and both `report` and `validate` accepted
+`fingerprint: "12345"`. Counts must be non-negative integers and a fingerprint must be a sha256
+digest, or the file is refused as not being a snapshot manifest.
+
+Verified: 17 new tests across `tests/snapshot-incomplete.test.js`,
+`tests/validate-drift-classes.test.js` and the guards suite, including a real Windows ACL deny for
+the unreadable case and a control that an untouched snapshot still reports no drift at all.
+
+### `validate` called a binary file a new file
+
+`generate` decides "binary" twice: from the extension and the 8 MB head scan, and again on the
+bytes it reads, where a control-byte payload or a NUL past the head scan only reveals itself.
+`validate` re-evaluated skipped paths with the cheap pre-read question alone, so a file `generate`
+had just skipped came back as `new file` and the first `validate` after a `generate` failed on an
+untouched tree. Both commands decide the same way now.
+
+### `--discovery` chooses how files are found
+
+A `.git` directory that exists but cannot be listed made discovery refuse, and there was no
+override — the only way forward was to repair the repository or copy the tree out of it. The
+refusal also justified itself with a gap that no longer exists (a walk does honour nested
+`.gitignore` files now), so it names the differences that remain: `info/exclude`, the global
+excludes file, tracked files that a `.gitignore` now matches, and a submodule as one entry.
+
+`--discovery auto|git|walk` (config `discovery`) makes that choice explicit. `auto` keeps the old
+behaviour, `git` requires git, `walk` scans the tree and reads `.gitignore` directly. The mode is
+validated, scoped to the commands that scan, published in the contract, written to the init
+template, and recorded in `manifest.config`, so changing it is config drift. 10 tests.
+
+### A workspace is classified by what its packages are
+
+Detection read only the repository root, so a Next.js app under `apps/web` in a pnpm workspace
+detected as `monorepo` and a docs repository whose markdown lived under `docs/` detected as
+`generic` — both contradicting the published preset table. Detection now probes a bounded number
+of workspace packages and markdown directories, so a workspace is classified by the frameworks its
+packages use. The bounds keep it honest: a bare `apps/` without a workspace declaration does not
+trigger it, and a framework two levels down still detects as `monorepo`. 8 tests.
+
+### `skill --install` could write outside the repository
+
+The installer built its three targets from `--cwd` and wrote them without checking where they
+landed, so a pre-existing junction (`.claude` itself, or `.claude/skills/codebase-snapshot`) or a
+hard-linked `AGENTS.md` made it write outside the repository. Targets are resolved and refused
+when they leave it. The section merge also repairs malformed marker layouts — duplicated blocks,
+an inverted pair, a missing end marker — instead of appending another one, while a well-formed
+file stays byte-identical on a rerun. 11 tests.
+
+### `diff` saw no change when the payload changed
+
+`diff` compared `relativePath` plus the source hash, so two snapshots of identical sources with
+different `--max-part-bytes` or `--large-file-bytes` — a 4 KB preview against a 36 KB full body,
+275 KB of parts against 307 KB — reported "0 added / 0 removed / 0 changed". The per-file answer
+stays the same, and a payload block now says the configuration differs and why (the config digest,
+the generator version, part and layer totals), which is what "what changed in the context we hand
+to the model?" was asking. 5 tests.
+
+---
+
+## 4.0.9
+
+### The tool executed a config it had been told to ignore
+
+`generate --watch --no-config` printed "config file ignored" and then reloaded the repository
+config on every regeneration, running its side effects and applying its include filter (4 tracked
+files became 1). The watch loop now honours the flag.
+
+`diff` also loaded the repository config, so comparing two manifests executed repository code —
+and `--no-config` was rejected for that command, leaving no way to opt out. `diff` reads two
+manifests and nothing else now; it never loads a config, and the flag is accepted there for
+consistency. Verified: a config that writes a file on import no longer does so during `diff`, and
+still does during `generate`.
+
+### Accepted risk: the ownership marker is evidence, not authentication
+
+A hand-written marker plus a stub manifest with a matching digest can still pass the ownership
+check, because there is no secret involved — anyone who can write those two files can also delete
+the directory directly. The check exists to stop the tool from deleting a directory *it* did not
+create, not to resist an attacker with write access. `SECURITY.md` now says so.
+
+---
+
+## 4.0.8
+
+### A dry run could write and delete
+`--dry-run` printed "nothing was written" and then ran `hooks.afterGenerate`. A hook can write
+files anywhere and delete parts of the live snapshot — one did exactly that during a dry run. The
+hook no longer runs when nothing is written.
+
+### Two runs to one output directory both claimed success
+Runs were serialised on `tmpDir` only, so two invocations with different `tmpDir`s but the same
+`outDir` each took their own lock, both exited 0, and one snapshot was silently lost. Both keys are
+locked now, in a fixed order so two runs cannot deadlock: one succeeds, the other refuses.
+
+### A git failure no longer changes what gets scanned
+When `git ls-files` failed (corrupt index, git missing) the scan fell back to a directory walk,
+which honours only the root `.gitignore` and ignores nested ones — a gitignored secrets file
+entered `files[]` with no warning, and the snapshot looked clean. A git checkout whose files git
+cannot list is now an error: fix the repository rather than silently scanning something different.
+---
+
+## 4.0.7
+
+Wave-2 findings: nine independent critics, one dimension each.
+
+### Credentials
+- **A branch name is user input.** `git checkout -b "release2/<ghp_…>"` copied the token raw into
+  `manifest.git.branch`, the README and the audit report — a channel that never passed through
+  redaction. It is masked now.
+- **Azure had no rules.** DevOps PATs, AD client secrets (`…8Q~…`) and SAS `sig=` values reached
+  parts and both monoliths verbatim. Added.
+- **Missing prefixes** in otherwise exhaustive alternations: `ABIA` for AWS bearer keys, `glrt-`
+  (GitLab runner), `hvb.` (Vault batch), `1//0…` (GCP refresh), `sk_test_` (Stripe test).
+- **`exclude` in a config** is the flag's spelling of `excludePatterns` and used to be ignored
+  silently, so a directory the user excluded was still emitted. It is accepted as an alias now.
+
+### The documented programmatic flow scanned the wrong project
+`resolveConfig(await loadConfigFile(dir))` ignored `dir` and generated into `process.cwd()`, exit 0.
+The loaded config now carries the directory it came from (non-enumerable), so the documented flow
+scans the project the config describes.
+
+### Small lies in the CLI
+- `--mode` was discarded when a command word was present: `generate --mode validate` wrote a
+  snapshot. A contradiction is now exit 2, and an unknown mode is an error either way.
+- `--only-artifacts monolith` wrote nothing but the marker, because the artifact was off. Naming
+  the monolith now enables it.
+- `init` wrote `"snapshot": "codebase-snapshot generate"` into `package.json` without installing the
+  package, so the very next line it printed (`npm run snapshot`) failed with "not recognized".
+  The scripts are `npx @ankurhalder/codebase-snapshot …` now, which works either way.
+---
+
+## 4.0.6
+
+### A NUL byte past the first 8 KB
+Binary detection sampled 8 KB, so a file with a NUL at byte 9000 was emitted as text into a part.
+The whole file is scanned now, up to 8 MB, in 64 KB chunks. Verified: a 20 KB file with one NUL at
+offset 9000 is skipped as `binary`.
+
+### Layer and kind names are masked wherever they are derived
+Masking `config.layers` was not enough: the *derived* values were raw — `manifest.layers` keys,
+`files[].layer`, part headers, the README parts table, the audit layer table, navigation, report and
+`config.budgets.byLayer` keys. Eleven sinks, one root. The name is now masked where it is assigned,
+so every derived field inherits the mask. Verified with a `ghp_…` layer name: zero occurrences in
+any artifact.
+---
+
+## 4.0.5
+
+### `validate` now notices every configuration change
+
+It compared a handful of settings field by field and silently ignored the rest: `layers`, `kinds`,
+`secrets.customPatterns`, `disableDefaultPatterns`, `maxMonolithBytes`. With a config that is not
+itself in the snapshot (`include: ["src/**"]`), changing classification or redaction rules still
+produced "✅ valid and up to date" on a stale snapshot — the "config changes are drift" promise
+was false for exactly the rules that decide what goes in.
+
+The manifest now records a digest of the effective settings, computed from the raw values, and
+`validate` compares it. That covers every setting, including the ones no field comparison lists,
+and it works where masked values cannot: two different credential-shaped exclude patterns both
+mask to the same placeholder, so a masked comparison would call them equal.
+
+```
+$ validate        # same config
+✅ Snapshot is valid and up to date with disk.
+
+$ validate        # after adding a layer rule
+❌ config: configuration: the settings differ from the ones this snapshot was generated with (run generate)
+```
+
+### Still open, recorded rather than hidden
+100k files take about 67 s (I/O bound); parallel runs with different `outDir`s can still hit the
+staging lock and exit 1 instead of waiting; `--watch`'s quiet window can defer an edit made during
+a run; layer and kind names supplied in the config are still emitted raw in derived fields; binary
+detection samples the first 8 KB; the monolith is built before `--max-monolith-bytes` rejects it.
+---
+
+## 4.0.4
+
+Five defects found by driving the published package from another agent's environment.
+
+### A stale shim exited 0 without doing anything
+`dist/cli.js` exported `main()` and never called it. `dist/bin.js` is the shipped entry point, but a
+`node_modules/.bin` shim left by an earlier install can point at `cli.js`, which then imported, did
+nothing and exited 0 — indistinguishable from a successful run with empty output. `cli.js` now runs
+itself when it is the entry file, and only then.
+
+### Redaction was breaking the parser
+Structural summaries were built from the **redacted** text, so masking an identifier such as
+`const secret = previousSecrets[i]` produced source the TypeScript parser could not read: 969
+"parse errors" on a codebase that has none, and incomplete summaries everywhere. The parser now
+reads the original source, and every string that leaves a summary — exports, variables, functions,
+classes, headings, JSON keys — is masked before it is written. Structure stays correct; nothing raw
+reaches an artifact.
+
+### Audit diagnostics
+`reports/audit.md` gained a **Parse & AST Diagnostics** section listing every file whose summary
+was incomplete, with its parse-error count. The CLI warned about these; the artifact an auditor
+reads did not.
+
+### Formatting
+- Summary lines no longer mix separators (`snapshot\parts` instead of `snapshot\parts/`).
+- Line counts use `en-US` grouping rather than the host locale, so a CI runner and a developer
+  machine print the same number.
+
+### Known gaps, not fixed here
+`validate` does not compare `layers`, `kinds` or `secrets.customPatterns`, so changing those does
+not register as drift; a changed credential-shaped value compares equal because both sides mask;
+layer and kind **names** from your config are still emitted raw in derived fields; the ownership
+marker can record a hook-written file whose name contains a token; and binary detection samples the
+first 8 KB only.
+---
+
+## 4.0.3
+
+### A snapshot could fail its own `validate`
+
+4.0.2 masks config values in the manifest, which is right — an `excludePatterns` entry or an
+`include` glob can contain a credential-shaped string. But `validate` compared that masked value
+against the **raw** current config, so a repository that generated cleanly failed validation one
+second later, and the mismatch line printed the very token the mask was hiding.
+
+Both sides are now masked before comparing, and the mismatch text uses the masked form. Verified:
+`generate` then `validate` with `excludePatterns: ["ghp_…"]` exits 0, and the token appears in no
+file under `snapshot/`.
+
+Known and unchanged: a layer or kind **name** supplied in the config is still emitted raw in the
+derived fields (`manifest.layers` keys, `files[].layer`, part headers). It is your own string
+rather than a discovered secret, but it is the same class and will be fixed with the next batch.
+---
+
+## 4.0.2
+
+### The last place a path secret could reach an artifact
+
+4.0.1 masked the path field of every skip record. It did not mask the **detail** beside it, and
+that text names the link target or the exclude pattern: a symlink to `real-ghp_…` and an
+`excludePatterns: ["ghp_…"]` both printed the token in `manifest.skipped[].detail`, the audit
+table and `explain`. Config values recorded for drift detection (`include`, `excludePatterns`,
+layer and kind names) had the same problem. All of them now go through the same rules.
+
+### `--watch` idles again
+
+After 4.0.1's staging changes the watcher fed itself: 15 regenerations in 8 seconds with no edits,
+because Windows reports its own writes with paths the segment filter did not recognise. The
+watcher now also stays quiet for a short window after each run. Measured: one regeneration in nine
+idle seconds, and an edit still triggers exactly one more.
+
+### Honest answers where there used to be silence
+
+- `loadConfigFile()` returned `null` for a config that existed but could not be parsed, so
+  `resolveConfig(await loadConfigFile(cwd))` — the documented example — silently generated on
+  defaults. It now throws, and `loadConfigFileDetailed` already reported the reason.
+- Custom secret rules are no longer stored in a module-level variable: a `collectFiles` call for
+  one project applied the previous project's rules. Every caller passes its own.
+- `describe` lists `.snapshotrc.json` before `.snapshotrc`, matching the real precedence.
+
+### Measured this round
+
+100k files: 56.6 s (was 336.4 s in 4.0.0), 20k: 11.8 s, 100k `validate`: 50.3 s, 100k `report`:
+1.28 s, 50 MB single TypeScript file: 1.33 s. About 60% of the remaining time is Windows file I/O
+— read, stat and open — not parsing or redaction.
+---
+
+## 4.0.1
+
+### A credential in a file name no longer reaches the snapshot
+
+A path that contained a token was masked only when the file was skipped *for that reason*. If it
+was skipped earlier — as a credential-bearing file (`certs/AKIA….key`), by `.gitignore`, by a user
+exclude or by the built-in ignore list — the raw path was written into `manifest.skipped[].path`
+and the audit table, and a binary file whose name carried a token still reached the manifest as
+`binary`. Custom `secrets.customPatterns` were not applied to paths at all, so a file named
+`src/CORP-<32>.ts` was *included* and the token landed in the parts, the navigation index and the
+project tree — the payload written to be pasted into a model.
+
+Path masking now runs over every skip record regardless of reason, and the included-file check
+uses the default **and** custom rules, so a custom-rule hit becomes a `sensitive-path` skip with a
+masked path. A tested fixture with all four shapes produces zero occurrences of either token in
+any file under `snapshot/`.
+
+### `report` no longer re-scans the repository
+
+`report` printed an integrity verdict by calling the validator, which re-reads every file: 4.5 s
+on a 20 000-file repository to print a summary already on disk. It now reads `manifest.json` and
+nothing else — **0.54 s** on the same repository — and says plainly that drift is `validate`'s job.
+---
+
+## 4.0.0
+
+### `snap` is back
+
+3.0.0 removed the `snap` alias after I misdiagnosed why `npx @ankurhalder/codebase-snapshot`
+printed nothing. **The removal was unnecessary and this release reverses it** — both
+`codebase-snapshot` and `snap` are published again.
+
+The real cause, confirmed by reading npm's own resolver
+(`libnpmexec/lib/get-bin-from-manifest.js`): npm resolves a package from its `bin` entries, and
+when every entry points at the same file it treats them as aliases and resolves one. Two
+aliases were always fine. What actually broke the bare command was a **stale global install on
+the test machine** — a leftover `npm link` back to the development checkout, which made `npx`
+run that instead of the published package. A clean profile runs the same command correctly, with
+either one or two aliases.
+
+If you ever hit a silent exit 0:
+
+```bash
+npm ls -g --depth=0 | grep codebase-snapshot
+npm rm -g @ankurhalder/codebase-snapshot
+```
+
+### Security: the redaction layer was rebuilt
+
+Independent adversarial review found credentials reaching `parts/*.md` while `audit.md` reported
+them as redacted. That is the worst failure this tool can have — its output is written to be
+pasted into a model — so the layer was rewritten rather than patched.
+
+Leaks that were real and are now closed, each with a regression test:
+
+- **CRLF checkouts.** Line-oriented rules ended in `$`, and a `\r` before the newline stopped the
+  match, so `export TOKEN=…` in a Windows checkout printed in the clear while the byte-identical
+  LF file was redacted. Scanning now splits on `\r\n`, `\n` and `\r` and preserves line numbers.
+- **Trailing comments and trailing spaces.** `export TOKEN=abc # rotate quarterly` leaked because
+  the value had to be the last thing on the line. The rule now matches the value, not the tail.
+- **camelCase keys and short values.** `awsSecret`, `dbPassword: "hunter2"` and similar were never
+  matched because only snake_case spellings were. Credential **names** now drive matching, and the
+  value is masked regardless of length.
+- **Unquoted YAML / docker-compose.** `POSTGRES_PASSWORD: hunter2` is now covered, while
+  expressions (`password: req.body.password`, `getToken()`, `process.env.SECRET`) are explicitly
+  not.
+- **Sensitive file names.** `.env~`, `.env.swp`, `.npmrc ` (trailing space), `.vault-token`,
+  `prod.env`, `secrets.toml` and friends are denied; deny-list matching normalises the name
+  instead of anchoring on the exact spelling.
+- **Secrets in a path.** A filename containing a token is skipped as `sensitive-path` with the
+  path span-masked, instead of being printed in the tree, the navigation index and the manifest.
+- **Authorization headers, and nine vendor prefixes** (`glpat-`, `hvs.`, `s.`, `xoxc/xoxd`,
+  `whsec_`, `dckr_pat_`, `hf_`, `ya29.`) are now detected.
+
+Masking is **span-only**: `dbPassword: "[REDACTED…]"` keeps the key and the rest of the line. That
+also fixed the opposite defect — over-redaction that corrupted ordinary code (`id-token: write`,
+`{ pass: 3 }`, and an unrelated string on a line that also contained a credential) is now
+explicitly negative-tested.
+
+The built-in rule count is now **29** (was 21).
+
+**Known limitations, stated rather than hidden** (see [SECURITY.md](SECURITY.md)): detection is
+pattern-based, so a secret deliberately split across two string literals is not caught, and
+unusual credential shapes can still be missed. Do not treat a clean audit as a guarantee.
+
+### The ownership marker is now evidence, not a filename
+
+`generate` deletes the previous snapshot, and it decides whether it is allowed to by checking
+whether it wrote that directory. The check was forgeable: an empty `.codebase-snapshot` file, or a
+three-key `manifest.json`, was enough to make the tool delete a directory full of someone else's
+files.
+
+`.codebase-snapshot` is now a JSON record — generator name, schema version, a random run id, a
+state, a digest of the manifest and an inventory of everything written. Replacement requires the
+marker to be complete and every file in the directory to be in that inventory; anything else
+blocks the run and is named in the message. The same evidence rule now applies to `tmpDir`, which
+used to be deleted wholesale. Concurrent runs take a per-`outDir` lock, so four simultaneous
+`generate` invocations leave one complete snapshot rather than a manifest describing parts that
+are gone.
+
+### Correctness
+
+- **JSONC comments in strings.** A valid `package.json` containing `//` anywhere in a string
+  (`"build": "tsc && echo done//ok"`) made every command fail with "the config could not be
+  read". The stripper is now a state machine that respects strings and escapes.
+- **Configs that are not objects.** `export default Promise.resolve({…})` was silently treated as
+  an empty config; a thenable, array or primitive is now a hard error naming what it received.
+- **Wrong-typed settings.** `excludePatterns: "src/**"` and `previewHeadLines: 0` were silently
+  ignored. Every setting is now validated, and the error names the setting, the expected shape and
+  what arrived.
+- **`layers` rejected the documented shape.** `{ name: "api", match: … }` — the shape the type
+  declares and the generated config recommends — crashed `generate` with a `TypeError`. Both
+  `layers` and `kinds` are validated up front, `name` is canonical for layers with `layer` as an
+  alias, and a bad rule is an error naming the index rather than a crash.
+- **`preset: "detected"`**, which `describe` published as the default, was rejected as an unknown
+  preset. `detected` and `auto` are now accepted.
+- **Budget precedence.** `--max-part-bytes` lost to `budgets.default`, so the banner said "Part
+  budget: 4 KB" while 151 KB parts were written. The order is now CLI > config `maxPartBytes` >
+  `budgets.default` > 300 KB, and `budgets.byLayer` is capped by the chosen budget.
+- **The output directory was scanned into the next snapshot.** `--out-dir .context` produced a
+  snapshot that included its own previous manifest, so the fingerprint changed on every run and
+  `validate` failed immediately after `generate`. The resolved `outDir`/`tmpDir` are excluded as
+  tool-owned paths, silently and identically on the first run.
+- **The monolith contained previews.** `full-codebase.md` was supposed to be the whole repository
+  in one file, but a 400 KB file contributed about 3 KB — its head/tail preview. The monolith now
+  carries full redacted bodies; `largeFileBytes` governs the parts only.
+- **A failing hook disappeared.** `hooks.afterGenerate` throwing printed a warning, but the
+  manifest had already been written, so the warning was not in it. The manifest is now rewritten
+  with the final warning list (the hook still runs once).
+- **`--out-dir` with a missing parent** failed with a raw `ENOENT` from the rename; parents are
+  now created, an ancestor that is a file is refused cleanly, and a failed run no longer leaves
+  `snapshot.tmp` behind.
+- **`validate` exempted skipped files forever.** A file skipped as binary, then rewritten as
+  source, was never reported as new. Skips are re-evaluated.
+- **`--dry-run` skipped the output guard**, claiming it would write into a directory the real run
+  refuses. A dry run now runs the same guard and exits the same way.
+- **`diff` reported `before undefined undefined file(s)` and exited 0** on a manifest that was not
+  a snapshot. Both sides are validated and the malformed one is named.
+- **`reactComponents` listed plain classes.** A component now requires JSX/`createElement` or a
+  React base class.
+- **A filename could escape a code fence.** A file named ``` ``` ``` closed the `## Files` fence
+  and let the next filename be rendered as live markdown — a prompt-injection vector in a file
+  built for LLMs. Fences are now chosen from their content.
+- **`--watch` never reloaded the config.** Editing the config while watching appeared to work
+  while the old settings stayed in force; a broken config regenerated silently instead of
+  reporting the error.
+- **`reports/audit.md` said "Redacted Content Bytes"** when it meant emitted bytes. Renamed
+  "Emitted Content Bytes".
+
+### CLI honesty
+
+- **Options that do not apply to a command are now rejected** (exit 2) instead of being parsed and
+  ignored: `--watch` is generate-only, `--dry-run` is generate/init, config and size options do
+  not apply to `diff`/`describe`/`skill`. The published `appliesTo` column is now generated from
+  the same table the parser enforces, so the contract cannot drift from the CLI again.
+- **`--no-config`** ignores `snapshot.config.*` and runs on the built-in defaults — useful in a
+  checkout you do not trust, because a `.ts`/`.js`/`.mjs`/`.cjs` config is imported and executed.
+- **`explain` always exits 0** when it can answer; the verdict is `included`/`reason`. Exit 1 now
+  only means the command could not run.
+- **`skill --format json`** emits JSON instead of markdown.
+- **`init` writes `includeMonolith`** (the file now lists every setting, as it claims), honours
+  `--out-dir`, refuses to silently override a `package.json#snapshot` config, and no longer says
+  `Next: npm run snapshot` in a directory with no `package.json`.
+
+### Upgrading from 3.0.0
+
+- **Old snapshots are refused, not replaced.** A snapshot written by 3.0.0 carries a marker in the
+  old format, so `generate` will stop and ask you to remove the directory or point `--out-dir`
+  somewhere else. This is deliberate: the old marker could not prove the directory was ours.
+- `explain` no longer exits 1 for a file that would be skipped.
+- Attention-only flags (`--watch` on `validate`, for example) are now usage errors.
+- With `--include-monolith`, the monolith is much larger, because it now actually contains the
+  code.
+
+### Documentation and packaging
+
+Every document was rewritten against the current behaviour rather than patched:
+
+- **New** [docs/cli.md](docs/cli.md) — every command, every option with its default and which
+  commands it applies to, exit codes, output streams, worked examples.
+- **New** [docs/output.md](docs/output.md) — the snapshot layout and the complete
+  `manifest.json` schema, field by field, including how to read `skipped[]`.
+- **New** [docs/troubleshooting.md](docs/troubleshooting.md) — every message the tool prints,
+  its cause and its fix.
+- **Rewritten** README, [docs/install.md](docs/install.md),
+  [docs/configuration.md](docs/configuration.md) and [docs/agent.md](docs/agent.md).
+- **New** [AGENTS.md](AGENTS.md) for agents working in the repository and
+  [llms.txt](llms.txt) for AI discovery.
+- **New** CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md, issue forms and a pull request
+  template.
+
+Corrections worth calling out, because the old text was wrong:
+
+- The generated config was documented as importing `defineConfig`; it has been a plain object
+  with no imports since 2.2.0, so that it loads when the package is only used through `npx`.
+- The "bin-less placeholder" explanation for the bare `npx` failure has been removed. The
+  placeholder version is unrelated and harmless.
+- `snap init` writes an `.mjs` config for a CommonJS project, not a `.ts` one — the
+  documentation now states the rule.
+
+Packaging: `prepare` was replaced with `prepack`, so installing the package no longer triggers an
+install-script warning, and the published tarball is still built from source. The test suite now
+includes an end-to-end agent simulation that follows the six workflow steps published by
+`describe` and asserts the resulting snapshot.
+
+---
+
+## 3.0.0
+
+### BREAKING: the `snap` executable is gone; use `codebase-snapshot`
+
+The package now ships **exactly one** executable, `codebase-snapshot`.
+
+This is what makes the bare form work. `npx @ankurhalder/codebase-snapshot generate`
+used to exit 0 printing nothing, on every npm version and every OS, because of how npm
+picks a binary: it matches the command you typed against the package's `bin` keys, and no
+key can contain an `@scope/`. With two executables (`snap` and `codebase-snapshot`)
+nothing matched, so nothing ran. With one, npm uses it regardless of its name.
+
+Verify the rule against packages you know: `@sentry/cli`, `@babel/cli` and
+`@changesets/cli` all ship one executable and run fine by bare name;
+`@swc/cli` ships three and fails with "could not determine executable to run";
+`npm-check-updates` ships two but is unscoped, so its name matches one of its bins.
+
+Migration:
+
+| Before | After |
+| :--- | :--- |
+| `snap generate` | `codebase-snapshot generate` |
+| `npx snap` | `npx codebase-snapshot` |
+| `npm run snapshot` (generated by `init`) | unchanged — it already used `codebase-snapshot` |
+
+Nothing else in the API changed: the library exports, the config format, the manifest
+schema and every command and option are identical.
+
+### Changed
+
+- The agent contract no longer claims the bare spec is unusable. It documents one
+  executable and explains why, so an agent does not add a pin it does not need.
+- `docs/install.md` explains the single-executable rule instead of the previous
+  (incorrect) "bin-less placeholder" theory.
+
+### Note on `0.0.0-stage`
+
+Earlier notes suggested unpublishing a `0.0.0-stage` placeholder to fix the bare form.
+That diagnosis was wrong — the placeholder is unrelated and can stay. The npm behaviour
+above is the real cause, and the single-executable change fixes it for everyone with no
+registry surgery.
+
+---
+
+Built for AI agents. Everything an agent needs is now available as data, and the tool can
+install its own skill where agents look for it.
+
+### Added
+
+- **snap describe** — a machine-readable contract of the whole tool: the invocation that
+  works, an ordered workflow array, every command, every option with its default, the
+  exit-code contract, all 25 presets, the artifact list, config file names, output layout,
+  safety rules and a troubleshooting table. --format json for machines, text for humans.
+  Generated from the same code the CLI runs; a test compares its documented defaults
+  against a real `generate --dry-run` manifest so it cannot drift into lying.
+- **snap skill** — prints the bundled agent skill (to stdout, so it can be redirected),
+  or with `--install` writes it to `AGENTS.md`, `.claude/skills/codebase-snapshot/SKILL.md`
+  and `.cursor/rules/codebase-snapshot.mdc`. `--target claude|agents|cursor|all`.
+  The install is idempotent and replaces only its own marked block, so an existing
+  `AGENTS.md` keeps its conventions.
+- `snap init` now points at `skill --install`.
+- [docs/agent.md](docs/agent.md) explains the contract and the skill for humans.
+
+### Why
+
+`--help` is prose for a person. An agent told "use npx @ankurhalder/codebase-snapshot"
+had to guess option names and failure modes. `describe` removes the guessing, and the
+skill removes the discovery problem: once installed, an agent finds the tool without being
+told it exists.
+
+### Tests
+
+138 pass (130 before): the contract's shape and defaults, skill output and frontmatter,
+installation to all three targets, idempotency, and that an existing `AGENTS.md` is
+preserved.
+
+---
+
+## 2.1.1
+
+A correctness and contract pass driven by an independent CLI audit of 2.1.0. No new
+features; every item below was reproduced against the released build first.
+
+### Fixed
+
+- **`validate` never printed the bookkeeping mismatches it counted.** It reported
+  "1 manifest bookkeeping value(s) disagree" without saying which one, while
+  `--format json` carried the detail. The text output now names them
+  (`counts.totalFiles: manifest 999 vs actual 5`).
+- **`generate` claimed artifacts it had not written.** `--only-artifacts parts` still
+  printed a `Reports:` line and a `Manifest:` line; `--skip-artifacts manifest` still
+  printed `Manifest:`. The summary now reflects the selection, and an all-disabled run
+  says so instead of listing files that do not exist.
+- **`diff` leaked raw Node errors** — `ENOENT: no such file or directory, stat …` for a
+  missing path, and `Cannot read properties of undefined (reading 'totalFiles')` for a
+  parseable but foreign manifest. Both are actionable messages now.
+- **`validate` double-counted a tracked-but-deleted file**: it appeared in
+  `missingOnDisk` twice, so the text said "2 file(s) no longer exist" for one file.
+- **A bad `--cwd` exited 1** although the documented contract reserves 2 for usage
+  errors. A missing directory or a file passed as `--cwd` is now exit 2.
+- **`--max-part-bytes -5`** produced the parser's "option argument is ambiguous" text;
+  it now says the value looks like another flag and points at the `--option=value` form.
+- **Empty option values silently meant "use the default"**, so `--out-dir "$UNSET"`
+  wrote to `./snapshot` instead of failing. Every string option now rejects an empty
+  value with a usage error.
+- **A detached HEAD was recorded as branch `"HEAD"`**, then `validate` warned about it
+  forever. A detached HEAD cannot name a branch, so the configured fallback is used.
+- **A read-only output directory produced a bare `EPERM`**. It now says the directory is
+  not writable and suggests `--out-dir`; the previous snapshot still survives.
+- **A wrong-typed `"snapshot"` key in `package.json` was silently dropped.** It now
+  warns. It stays non-fatal on purpose — other tools use that key too.
+- The all-artifacts-disabled warning claimed the directory "will be empty", which
+  stopped being true once the ownership marker was introduced.
+
+### Tests
+
+130 pass (119 before), including a POSIX-only check that a read-only output directory
+fails with a clear message.
+
+---
+
+## 2.1.0
+
+Multi-project support and an interactive setup.
+
+### Added
+
+- **Interactive setup.** `npm create codebase-snapshot` (or `npx @ankurhalder/codebase-snapshot init`)
+  detects the project, walks every tunable parameter, and writes the config. Every
+  prompt has a flag, so it works non-interactively in CI.
+- **A self-documenting config.** The generated `snapshot.config.mjs` lists *every*
+  setting at its current value, commented and grouped — not just the ones you
+  changed. Deleting any line is safe. `--minimal` writes the terse variant.
+- **25 presets**, up from 4: nextjs, nuxt, sveltekit, remix, astro, angular,
+  solid, vue-vite, react-vite, nest, node-express, node-lib, python-django,
+  python-fastapi, python-lib, go, rust, java-spring, dotnet, php-laravel,
+  ruby-rails, monorepo, infra, docs, generic.
+- **Per-ecosystem ignores** contributed by each preset — `__pycache__`, `.venv`,
+  `.mypy_cache` (Python), `target` (Rust), `.gradle` (JVM), `bin`/`obj` (.NET),
+  `vendor`/`storage` (PHP), `.svelte-kit`/`.angular`/`.astro` (JS frameworks) —
+  merged with `excludePatterns` and still recorded per file.
+- **Structural summaries for non-TypeScript languages.** Python, Go, Rust,
+  Java/Kotlin/Scala, C#, PHP, Ruby and C/C++ now get imports, declarations and
+  exports. Vue/Svelte/Astro components have their `<script>` block handed to the
+  real TypeScript compiler.
+- **`summaryQuality`** (`ast` / `structural` / `heuristic`) recorded per file and
+  rendered in every part section, so a line-scan is never presented as a parse.
+- **`preset`** option to force a preset by name; an unknown name is an error with
+  the full list rather than a silent fallback.
+
+### Changed
+
+- Preset detection now reads the repository once per run and shares that view
+  with every detector. `peerDependencies` and `optionalDependencies` count.
+- A bare `package.json` ranks below any definitive language manifest, so a Go or
+  Rust repo with a `package.json` for tooling is not mistaken for a Node library.
+
+### Fixed
+
+- **`outDir`/`tmpDir` pointing at an existing directory could destroy it.** A
+  config with `outDir: "src"` or `tmpDir: "src"` deleted the user's source tree
+  and exited 0: the scratch directory was removed outright, and the atomic swap
+  renamed the directory aside, moved the snapshot in, then deleted the backup. A
+  regular file named `snapshot` was destroyed the same way. Both directories are
+  now refused unless they are empty, hold only placeholder dotfiles, or carry
+  proof that a previous run wrote them (a `.codebase-snapshot` marker or a
+  parseable `manifest.json`). Filenames alone are never treated as proof, so an
+  existing `docs/README.md` is safe and hook-written artifacts do not block
+  reruns.
+- **A config that could not be parsed was silently ignored** — a typo produced a
+  snapshot with none of the developer's settings and exit 0. Unparseable configs
+  are now a hard error (exit 1, nothing written). A `.ts` config on a Node without
+  type stripping is still only a warning.
+- **Absolute `--out-dir` was honoured by `generate` but not by the readers.**
+  `validate`, `report` and `explain` joined instead of resolving the path, so they
+  looked for `<cwd>\<absolute path>` and reported a missing manifest immediately
+  after a successful generate.
+- **`--preset`/`--include`/`--exclude`/`--yes`/`--minimal`/`--force` were accepted
+  and ignored outside `init`.** `snap generate --preset nextjs` silently used the
+  detected preset. These are now a usage error (exit 2) on any other command.
+- **`init --preset <typo>` silently fell back to detection**, writing a config for
+  a preset nobody asked for. It is now a usage error listing every valid preset.
+- **Astro components were summarized as markup.** The leading `---` fence — where
+  the TypeScript actually lives — was not parsed, so the common case degraded to
+  `heuristic` quality. It now goes through the compiler like Vue and Svelte.
+- Summaries for `.jsonc` strip comments instead of reporting `invalid-json`.
+- `export default function Foo` records both `Foo` and `default`.
+- Parser diagnostics are counted and surfaced as warnings.
+
+---
+
+## 2.0.0
+
+Configurable snapshots, `explain`/`diff` commands and correctness fixes.
+
+### Added
+
+- `artifacts`: switch each of the seven artifacts off individually, with
+  `--skip-artifacts` and `--only-artifacts`.
+- `include`: allow-list globs, applied before `excludePatterns`.
+- `budgets.byLayer`: per-layer part budgets.
+- `layersMode`: user layers merge with the detected preset by default.
+- `hooks.afterGenerate`.
+- `snap explain <path>` and `snap diff <a> <b>`.
+- `generate --watch`.
+- Byte budgets accept human sizes (`120k`, `1.5mb`) from CLI and config.
+- Globs support brace groups and character classes (`*.{ts,tsx}`, `[abc]`).
+
+### Fixed
+
+- The monolith is built from pre-partition content, so `full-codebase.*` really is
+  the whole codebase even when a part truncates a file.
+- The part byte bound is enforced after the split-pass cap instead of being
+  silently exceeded.
+- Default-ignored files are recorded with `reason: "default-ignore"` instead of
+  vanishing, including leftover `snapshot.backup-*` directories.
+- Watch mode no longer retriggers itself on its own writes.
+
+---
+
+## 1.0.0
+
+Initial release.