v0.4.0

0.4.0 makes every write command preview by default: spec new and spec set now write only with --write, like the rest, so add --write where scripts and CI create or change documents. It also adds coverage audits, decision-maker import, [check] link_ignore and safer --write undo for files with uncommitted edits, and fixes about 100 issues across checks, hints and refusals.

Behavior changes

What an existing command does differs from 0.3.0; check scripts and CI that call these.

spec new and spec set now only preview without --write: add --write to every script, CI job and agent instruction that creates or changes documents with them. Like every other write command, a run without --write (or with --dry-run, which does the same) writes nothing, shows the change with its diff, and exits 0 with next: naming the same command with --write; --json reports written: false. Their hints, help, docs and agent skills now give the --write form. The 0.3 stderr warning is gone.

spec check --write-baseline with --rule or --skip is refused with exit 2: a baseline is recorded from a full run only (C181).

Two or more [[guards]] with the same baseline file are refused when the config loads (exit 2, naming both guards): each guard's --write-baseline used to overwrite the other's entries (C184).

A --write (fix, edit, set and the rest) over a tracked file that already has uncommitted edits first copies it to .specfold/trash/<timestamp>/, and its undo: copies that back instead of git stash, which also took the user's own edits; the preview names such files. Files without uncommitted edits keep the git stash undo (C183).

-C <dir> now resolves relative path arguments against <dir>, and next commands include -C <dir>.

Write previews show an update's changed lines as unified-diff hunks (@@ -old,n +new,m @@): changes with an unchanged line between them, such as the refreshed updated: line and an edited section, are separate hunks instead of one block. In --json, update items carry hunks (each with line, old_line, before, after) in place of the top-level before and after; line is still the first changed line.
spec sync --prune previews name the trash folder .specfold/trash/<timestamp>/ instead of the current second, so a preview and the refusal of the same write agree; the write itself uses the second it runs in and prints the real path. The pruned-files note names .specfold/trash/.
spec check --rule/--skip refuse (rc 2) a rule id that does not exist, as --fail-on does; a lower-case id is pointed at its upper-case form (C132).
spec fix --pick refuses (rc 2) an id that is not one of the finding's choices, listing them (C146).
spec log --since must be a real YYYY-MM-DD date, and spec log --type, spec find --type and spec find --status must name a known type or status; others are refused with rc 2 (C148, C153).
--limit 0 is refused (rc 2) by every command that takes --limit; it showed nothing and its --offset hint pointed at itself (C152).
spec new --governs and spec context refuse (rc 2) an empty path or one outside the repository, instead of recording it or suggesting spec new --governs for it (C130).
A guard whose range is not staged, push, all or a git range (a..b) is refused by every command that loads spec.toml, naming the guard; spec doctor reports it as a config finding (C141).

• An id held by two files (I1, e.g. a .spec/ and a .spec.local/ copy) is refused (exit 2) by every command that takes a document, listing both paths; name one by its path. spec doctor's fix names the path (C139).
• spec edit --section/--new-section refuses (exit 2) text holding a heading at the section's level or above, which would start or duplicate a section; deeper headings and fenced code are fine (C125).
• spec import --apply --write refuses (exit 4) while rows marked ! still need a decision, lists them and suggests --force; the new --force imports anyway. Previews are unchanged (C155).
• spec attach --into refuses (exit 2) an absolute folder or one with ..; it used to write outside the bundle (C142).
• spec new skill|rule|agent refuses (exit 2) a title with no ASCII words unless --slug is given, and a --slug that is not kebab-case (C147).
• spec init refuses (exit 2) an --owner-kind with an empty pattern or template, a {n} beyond the pattern's *s, or a suffix other than :manifest, and a --root that is not an owner id (C140).
• spec set <id> title= refuses (exit 2) emptying the required title (C128).
• spec export … --dry-run previews and writes nothing (it exported before): it lists files and, over an earlier export, replaces with exit 0; a plugin exporter is not called. spec check -q now shows config warnings too (C114, C115, C116).
• spec set <plan> status=archived --write refuses with exit 4 while the plan has open tasks and lists them; --force archives anyway, and --dry-run reports them with exit 0.
• Exit code 4 (CLI-17): a write flag was given but a safeguard refused it, so nothing was
  written (--write over [write] max_files, sync --prune without --yes, a hand-edited
  copy without --force, export over an earlier export or attach to the store without
  --yes). These exited 0 before.
• spec set <id> key=[a, b] stores a list, not the string "[a, b]".
• spec sync --prune --write --yes no longer deletes: each pruned file moves to .specfold/trash/<stamp>/ (ignored by git), and the undo line copies it back (cp -Rp .specfold/trash/<stamp>/. .), so a file never committed can be restored. The undo of a sync or copy lists only the files written, not their sources; a successful sync ends with next: git status --short (C100, C101, C106).
• spec check --write-baseline without [check] baseline says what to add to spec.toml before previewing, with next: spec explain config (C102). --dry-run is accepted on every command guarded by --yes or --write and writes nothing: check and guard ignore it, import previews even with --apply (C120), and export previews instead (C103, C114). An export whose --out is outside the repository suggests ls <out> instead of git status (C104).
• An invalid spec.toml points to spec explain config (C106).
• Overwrites need ownership or --force (CLI-4): spec sync assets --write overwrites a copy file only when its content is a version its source has had (git history or now); a hand-edited copy needs --force. spec fetch leaves a different existing file alone and fetches nothing unless --force. init upgrades and export are unchanged.
• Bulk writes are bounded (CLI-7): [write] max_files = 50 (0 = no bound). When spec mv, fix, sync or import would change more files, the preview says how many and --write alone writes nothing; narrow the scope or add --yes (new on mv, fix and import).
• One meaning per shared flag (CLI-11): spec guard --tree (was --all), spec find --body (was --text), spec init --owner-kind (was --owner). The old names keep working as hidden aliases with a stderr warning; help JSON lists them under each flag's deprecated_aliases.
• spec --help and spec help list the 12 core commands (context, find, show, graph, new, set, edit, check, guard, fix, explain, init); spec --help --all / spec help --all list every command. spec help --json, --markdown, --llms and MCP still describe every command; JSON adds core.
• spec guard --write-baseline without --yes previews the hit counts it would record.
• CLI conventions CLI-1–CLI-16 (AGENTS.md, tests/cli_contract.rs). Actions that delete,
  overwrite or send files out now need --yes besides their write flag; without it they write
  nothing and next shows the command with --yes: spec sync assets --prune --write,
  spec attach … --write when a file goes to the store, spec export markdown replacing an
  earlier export in --out, and spec check/spec guard --write-baseline.

Added

• spec audit --kind coverage: with [audit] coverage_types = ["sot"], lists each {#ANCHOR} heading of a current document of those types that no other document cites as [[<id>#ANCHOR]]. Off by default.

• Import keeps who decided: MADR decision-makers/deciders, consulted, informed and MADR 2/Log4brains * Deciders: body lines become decision_makers, consulted and informed lists, now declared adr fields (no S1 warning).

• [check] link_ignore: globs of link targets rule L8 skips.

• A config warning when require_when names a status the type does not have.

• A config warning when [audit] coverage_types names a type that does not exist.

• Rules L8 and L9 (warn). L8 checks plain Markdown links and images offline: a relative path
  must exist and a #fragment must match a heading's GitHub slug (repeats get -1, -2) or a
  {#anchor} in the target file; URLs and ids are skipped. L9 flags an image without alt text.

• New rule S8 (warning): a field that require_when requires for the current status must hold a usable value — not a placeholder ([types.<t>] placeholders, default TBD, Pending, TODO, ?), of the kind declared in fields = { approvers = "people", resolution = "url", review_by = "date", decision = "link" }, and not shared within a distinct = [["authors", "approvers"]] group. fields as a list works as before.

• spec help --json marks each command's deprecated_aliases (name, use, since). A deprecated name keeps running the new command and warns on stderr (none yet).

• spec set ID key= with an empty value deletes the key (an emptied rel is removed too).

• D1 also counts staged deletes and the old paths of renames (p (removed)). Its fix is a runnable spec new change ... --depends-on ID --write; spec explain D1 points to spec check --staged --rule D1. A staged/changed run without documents says so in its summary.

• spec new --depends-on ID (repeatable) writes rel.depends_on.

• spec export reports attachments it did not scan for secrets (not_scanned); binary files no longer get the scan_max_bytes hint. Success next is git status --short -- <out>.

• --dry-run on every write command (edit, mv, fix, init, attach, sync, audit;
  import --plan takes it as an alias; new and set keep theirs).

• spec help --json gives each command a risk: read, write, destructive or external. MCP tools
  carry the matching annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint).

• Writes report an undo command (git stash push --include-untracked -- <paths>).

• Rule D1 (warning, spec check --staged and --changed only): code changed under an active
  document's rel.governs while neither the document nor a change note naming it changed.

• Rule U2 runs the built-in secrets check (the one spec export always applies) over every
  document file, so the commit hook (spec check --staged) refuses what export would refuse.
  spec guard secrets runs it over any range, and spec guard with no secrets guard configured
  includes it. A configured secrets = true guard keeps reporting through U1 instead.
  Hits in local (git-ignored) documents are warnings; in committed documents, errors.
  U2 also reads each document's text attachments kept in the repository (up to the new
  [files] scan_max_bytes, 1 MiB by default; binary files are skipped), so a secret attached
  with a warning keeps being reported by spec check and spec guard secrets. spec export
  reads the same attachments of the documents it exports and refuses on a hit.
  Attachments it skips are named in one U2 warning (and a not_scanned line in
  spec guard secrets) that points at raising [files] scan_max_bytes.

• spec attach runs the built-in secrets check on every file, with no [[guards]] needed: a hit
  refuses a store upload (--allow-hit secrets lets it through) and warns for a repository file.

• A slug id (plan:root/payment-rollout for a timestamp-keyed plan) resolves in [[…]]
  references, Markdown links, export and spec find; L1 suggests the canonical id.

• spec help <command> shows the command's notes (limits, settings, what a refusal means)
  after its arguments and examples; --markdown and --json (notes) carry them too.

• The attachments guide, the config reference and spec help attach show a [files.store] example.

• Built-in secrets (secrets = true in a guard) now find connection strings with a password
  (scheme://user:password@host, any scheme: postgres, mysql, mongodb, redis, amqp …) and
  three-part JWTs. Not flagged: example and loopback hosts (example.com|org|net, *.example,
  *.test, *.invalid, localhost, 127.*, ::1, 0.0.0.0); placeholder passwords
  (${VAR}, $VAR, %VAR%, {{var}}, <var>, one repeated character such as xxxx, words
  such as password or changeme, a password equal to the user name such as
  postgres:postgres, fewer than 4 characters); and the jwt.io sample token.

• spec attach runs every guard that reads content (secrets, denied terms or patterns, denied
  files) over files bound for the store before uploading; an error-level hit refuses, a warn
  hit is shown, and --allow-hit <guard> lets one guard's hits through. Matched values are
  never printed.

• spec export runs the built-in secrets check, and the guards marked public = true, over
  what leaves (the built-in tree, or the source files of the documents a plugin receives) and
  writes nothing, nor calls the plugin, on an error-level hit. The secrets check runs even when
  no guard is configured; --no-guard skips the check and the guards.

• Rule L1 warns about a Markdown link whose target is a document id ([text](plan:root/…))
  and offers [[id|text]] (spec fix --rule L1 --write); the built-in export rewrites the
  ones that resolve into the target's link.

• A timestamp-keyed document resolves by its id with the slug in place of the timestamp:
  plan:root/payment-rollout names plan:root/20261030020644.

• spec export markdown ends with a next command, and its missing --target/--out and
  non-empty --out errors offer a runnable one. The export guide documents how same-named
  files get -2, -3.

• The Markdown source of the export guide is published at /guides/export/index.md.

• spec mcp replaces built-in secret matches with [redacted: <pattern>] in tool results and
  reports the count as redacted; [mcp] redact = false turns it off. spec edit and
  spec set refuse text holding a redaction marker the document does not already hold.

• spec export markdown --target obsidian|mkdocs|docusaurus|github-wiki --out <dir>: a
  built-in exporter that writes plain Markdown with each target's links, flattened
  frontmatter, a .specfold-export.json manifest, and a report of references written as text
  (--fail-on unresolved to fail instead). --local adds .spec.local documents; --banner
  names each page's source. The name markdown is reserved for it.

• Exporter plugins can declare wants: ["body", "links"] to receive each document's body,
  source_rev hash and resolved references with byte spans; past --max-bytes, bodies come as
  body_path.

• spec graph consumers prints a header, a count and scope footer, and searched in JSON.

• spec help lsp takes --dir and points to the editor setup guide.

• spec new warns when the slug leaves out words in another script ("결제 비교 Gateway" gives
  gateway), not only when most of the title is lost.

Fixed

• spec import --apply --write of a CRLF source writes the new document CRLF throughout, its generated frontmatter included, instead of LF frontmatter over a CRLF body; LF sources stay LF (C243).

• A usage error missing some other required option (spec edit <id> --cell S=x --text x without --row) no longer suggests spec edit <id> <id> ...; it points to spec help <command>. --format json next to an option left without its value is no longer rewritten into a rerun that fails the same way (C241).

• spec attach refuses (exit 2, nothing written) the document's own file, or a file that would land on its bundle index name, which used to record a hash that could never match (C239).

• A heading's optional closing # run (## Context ##) is dropped from its text, so the outline shows "Context" and [x](#context) links pass L8 (C240).

• A line indented four or more spaces, or starting with seven or more #, is no longer read as a heading: outline, show/edit --section, edit's heading guard, required sections (S7) and anchors share one CommonMark heading parser (C238).

• spec.toml regexes and globs ([verify] serial, full_run_on, [forbid], [cites], guard and asset paths) are checked at load: a bad one stops with its line (exit 2) instead of being dropped. An empty or missing [verify] shell is reported by spec doctor, and spec verify names the shell and points to spec doctor (C237).

• Language tags compare case-insensitively in spec census --translations and R1's one-translation-per-language check, as spec new --translation does; --lang that is not a language tag exits 2 (C236).

• spec doctor on a spec.toml that does not load says to fix it at the line TOML names; spec init --upgrade on such a file fails with exit 2 instead of calling it current (C235).

• spec mv of a bundle document: the undo stashes the old and new folder and runs cleanly, and relative links to the bundle's attachments follow the move (C233, C234).

• Release notes come from the CHANGELOG section for the tag, falling back to the commit subjects since the previous tag.

• spec help new examples and the spec agents lines that start documents carry --write (C232).

• spec guard --range A..B reads files at B (HEAD when empty) instead of the checkout, and also reads every version a commit in the range wrote, so a value added and later removed inside the range still fails (C230, C231).

• Suggested next commands write option values starting with - as --opt=value, put such words after --, and a usage error's next keeps -C <dir> (C229).

• C227: every spec set that a fix, next, rule explanation or help example suggests carries --write, including both steps of the T1 fix for a draft, so it still writes once set previews by default

• C226: "section '…' matches N headings (line …)" from spec edit and spec show --section names file lines, not body lines

• C223: spec set <doc> updated= takes only a date, an RFC 3339 time or today (rc 2 otherwise, preview included), refuses updated= as it does title=, and its next runs S5

• C224: spec mv counts a referring document once when both its links and its rel lines change, in the preview, the write and --json

• C225: the undo of spec attach --write --yes over other content also puts the document file back, so one cp -Rp restores the files list and the replaced files

• C220: spec guard --range refuses an empty value and ..X with their own message (rc 2, next <base>..HEAD); the <rev>..HEAD suggestion is built only for a single revision

• C222: spec export markdown in a repository with no documents creates the folder with the manifest alone (rc 0, next spec new) instead of a raw os error

• C221: spec export markdown --out resolves the nearest existing folder through symbolic links, so a link into .spec, .spec.local or .git (or one containing the repository) is refused like the plain path

• C216: spec guard <unknown> lists secrets (built-in) among the guards it has, with next spec guard secrets

• C214: spec fix resolves its targets like spec check (a missing file, a directory, a non-document or :abc is rc 2; ids are accepted)

• C204: T1 on a draft names the two-step fix (status=active --write && status=superseded), and T1 and spec set status=superseded ignore a cross-type rel.supersedes that L3 rejects

• C213: spec graph chain walks a supersede cycle once, reports current: null ("a supersede cycle, nothing is current") and points at spec check --rule T1

• C212: spec census --translations rows are paged items (--jsonl, --limit, --offset), and spec graph --jsonl streams one row per node, edge and governed path

• C211: spec import walks sources in path order, so records sharing a date get ids in file-name order on every filesystem

• C210: spec init --owner-kind and spec.toml refuse an owner pattern segment that mixes glob characters (**, svc-*) with rc 2, and {N} counts whole * segments

• C207: context and graph governs on a folder also list documents whose governs path or glob lies inside it, and no longer suggest spec new sot then

• C206: concurrent spec new runs reserve their timestamp id with create-new and take the next second on a clash; I1's explain gives a repair path for timestamp ids

• C201: inline code spans close only at a backtick run of the same length (CommonMark), so a link after a ` b is checked and counted

• C200: S1 reports a title that is not a non-empty string and S4 a status that is not a string; spec set title=[x, y] writes quoted text, not a list

• C199: set, edit and mv keep a document's leading UTF-8 BOM; the preview no longer shows line 1 changing

• C198: a section name picks an exact heading or {#anchor} before a heading that only starts with it; several prefix matches or duplicate headings are rc 2 listing them

• C215: an argument that is not UTF-8 is rc 2 with the argument named, instead of a panic

• C218: spec help --markdown/--llms put --flag words in prose in code spans, so the docs site no longer renders --write as an en dash

• spec guard --range <rev> without .. is refused rc 2 (next: --range <rev>..HEAD), as the config already did; it used to scan the history of <rev> and pass the new commits.

• spec edit --text/--file refuses text that opens a code fence and never closes it (rc 2), which used to turn every later heading into code and hide sections and open tasks.

• Undo lines work for git-ignored paths (the local tier: rm -f instead of a stash git cannot reach) and for documents never committed (set/edit copy the old file to the trash instead of a stash that removes the whole document).

• spec attach no longer silently overwrites a bundle file with other content (repo or store): the preview marks it replace, the write needs --yes (and --force for a file the document does not list, as spec fetch), and the old file is copied to the trash with an undo: line.

• spec mcp refuses a tool argument whose type does not match the schema (a string for a boolean, a non-integer limit) and a positional value starting with -, so a value can no longer become a flag such as --write or --yes.

• spec audit record and spec audit --record accept --slug (validated like spec new), so a title with no ASCII words can be recorded; --write now prints an undo: line.

• A guard baseline that passes through a symbolic link into a folder that does not exist yet and back out (lnk/sub/../.guard-base.json) is compared where the write would land: the deepest existing part of the path is resolved before .. is applied. Such baselines are refused as shared instead of overwriting another guard's baseline and creating sub/ (C195).

• Two guards whose baselines reach one file through a symbolic link folder (lnk/.guard-base.json and ./.guard-base.json with lnk -> .) are refused as sharing it: an existing folder is resolved before the paths are compared (C194).

• spec graph --format dot names nodes d0/p0 with the text as label, like mermaid, so a governs path spelled like a document id no longer redeclares that document as a folder or lands its edge on it. Scripts that matched quoted id node names must match labels now (C190).

• spec init --upgrade reads a spec.toml that starts with a UTF-8 BOM (and keeps the BOM when it writes), so the upgrade check asks for no longer reports "is current" and loops (C191).

• spec edit --new-section refuses (exit 2) a name that is empty or blank, holds a line break, or starts with #, as new --title does, instead of writing an empty, doubled or ## ## heading (C188).

• -C <file> is refused (exit 3) instead of running against no repository, and an empty -C value points at the help of the command typed rather than an unrelated one (C193).

• -C <subfolder> now acts like cd: relative path arguments are read from it and rebased to the repository root (context, new --governs, …), and undo/fix shell commands run from the root (git -C <root>), so the printed undo works (C192).

• spec export --local --no-local is refused (exit 2) in either order instead of silently dropping the local documents the preview counted (C189).

• spec context and spec check hint --governs with glob characters escaped (src/[[]a]/z.rs), and spec new --governs escapes an existing literal path holding [ * ? {, so the document governs that file and not the ones a glob reading would match (C185).

• The shared-baseline check resolves . and .. and paths under the repository root before comparing, so x/../b.json and ./b.json count as one file (C186).

• Two writes in the same second no longer share a .specfold/trash/<timestamp>/ folder: each write claims a fresh one (<timestamp>-1, -2, … when taken), so an earlier trash copy is never overwritten. This applies to edit, fix, set, mv, import and sync --prune (C187).

• spec edit --row with a key that several rows share now says so, lists their line numbers and points at editing the whole section; it asks for --section only when the rows sit in different sections (C182).

• spec mv rewrites reference definitions, angle-bracket and percent-encoded relative links too, and no longer touches links inside code (C180).

• spec import --apply --write now rewrites reference definitions, angle-bracket links and percent-encoded links (keeping the encoding), still skipping code; L8 also checks reference definitions (C180).

• spec show <id> --body cut by --max-bytes now continues with spec show <id> --section … instead of putting the body text where the id goes (C179).

• spec new with an option it does not have (--tier local, --bogus x) now points at spec help new instead of a next that repeats the same error (C178).

• spec export markdown --dry-run (and the replace refusal) now print a next that keeps --banner, --local and --fail-on, so following it writes what was previewed; --no-guard is still left for you to type (C177).

• Next steps and fix commands shell-quote every path and id through one function, so spec context and audit gap suggestions run as printed for paths with spaces, quotes, or non-ASCII names.

• Audit: --kind stale shell-quotes governed paths in its git diff next step.

• Audit: --kind stale suggests only spec set <id> verified_at=HEAD when verified_at is not a commit, instead of a git diff that cannot run.

• Help: show and edit examples name sections new documents have (rule, Steps).

• spec context: the next names --section rule only for a sot; other types get spec show <id>.

• spec check --format sarif: artifactLocation.uri percent-encodes each path segment (a space is %20).

• spec init --upgrade: a spec.toml from a newer version is refused with the same message and exit code as spec check, not reported as current.

• spec init --upgrade: a table header with a trailing comment ([spec] # main) is read as TOML reads it, so an old schema under it is raised as spec check asks.

• spec fix --write with nothing to fix says "no change" and reports written: false, with no empty undo line (as mv and set).

• spec fix --rule: an unknown or lower-case rule is refused (exit 2, "no such rule", upper-case hint) as in spec check --rule.

• spec export: a refused run's next keeps --dry-run, --local, --banner and --fail-on, so following it from a preview no longer writes.

• spec new: the draft-supersede refusal now names spec new adr --title '…' --from <id>, and the missing---title next keeps --from and the other options given.

• C164: spec mv to the document's own id and path reports "no change" with written: false and no undo, as set does, instead of 0 change(s) written.

• C163: with -C <dir>, shell commands in next: and undo: run from where you typed them: git ones get git -C <dir>, others cd <dir> &&; check fix: hints get spec -C <dir>. Output without -C is unchanged.

• C162: new --title and set title= also refuse U+0085, U+2028 and U+2029 (YAML 1.1 line breaks), as for \n and \r.

• C161: with -C <dir>, relative path arguments (attach files, edit --file, import --map, export --out, audit --from) are read from <dir> instead of the shell's directory, and next: commands keep -C <dir>.

• spec new --title and spec set title= refuse a title with a line break (exit 2, with a next
  command to give it on one line); it split the body heading and one-row-per-document lists.

• spec export markdown looks up relation targets by id instead of scanning every document for
  each one: 10,000 documents export in about 1 s instead of 4.8 s, with the same files.
  A write preview of a very large change no longer runs out of memory: hunks are computed only while the compared region stays under about four million line pairs, and a larger change is shown as one hunk; 30,000 replaced lines had aborted (C158).
  Preview hunk headers match git diff -U0: an empty side names the line before the change, a count of 1 is left out, and an insertion or deletion among repeated lines sits where git puts it; JSON line and old_line follow the same rule (C159).
  The pre-push snippet from spec hook print no longer fails on a branch without an upstream: it compares with the merge-base of the default branch (origin/HEAD), and when there is neither it prints a note and skips (C126).
  The spec edit preview diff on a CRLF document compares lines with their original endings, so it shows only the changed lines instead of the whole document (C157).
  A refused --slug (e.g. spec new skill --slug 'Bad Slug') is replaced in the suggested rerun instead of given twice (C156).
  A failing spec verify suggests spec show with the failing document's id instead of a literal <id> (C129).
  spec log suggests spec show with a real id, and no next when nothing was listed (C149).
  spec graph --format box writes a cycle note on one line with ids only (C131).
  spec check --write-baseline without [check] baseline now carries the reason in JSON error (C134).
  LSP diagnostic messages keep a real line break before fix: instead of \u{000A} (C144).
  The spec fix preview's next keeps --staged, --pick and the targets, so following it writes what was previewed (C145).
  The usage-error next of spec new no longer takes option values as type or title, and keeps --title/--governs (C133).
  A plugin that writes a lot to stderr no longer stalls until its timeout; stdout and stderr are read together and the error shows the last stderr line (C150).
  A plugin that leaves a background child holding its pipes now stops at its timeout_ms; spec plugins reports a bad action before starting any plugin (C151).

• Writes keep a file's CRLF line endings instead of mixing in LF (C154).

• spec mv rewrites relative Markdown links to the moved file (fragment kept), and its next includes L8 (C137).

• spec audit --record cites findings as [[id]] so the record passes L2, and audit records are no longer reported as orphans (C136).

• spec attach previews and refusals keep --into in their next (C143).

• Code fences close only with the same character and at least the same length, so a ```` fence holding ``` lines or a ~~~ fence is read alike by every command (C135).

• spec import counts the alias table toward [write] max_files (C138).

• spec fetch downloads to a temporary file and replaces the target only after the checksum matches; on a mismatch it removes the temporary file, leaves the target alone and exits 3 (C127).

• spec sync assets: a path that is a folder on one side and a file on the other is named in the preview and the write alike (write refused, exit 4); --prune moves the copy's entry to the trash first. Folder errors name the path (C124).

• spec import --apply --write's undo brings back source documents that were never committed: they are copied to .specfold/trash/<stamp>/ before the write and copied back by the undo, alongside the git stash for committed sources (C123).

• spec import --apply --dry-run previews (rc 0, nothing written) instead of failing with a clash against --plan; --dry-run is listed in spec help import, and still refuses --yes (C120).

• spec import --apply --write lists the files it created, moved and updated (alias table included) and prints undo:; JSON gains written, undo and changes. With no source documents it writes nothing, not even the alias table (C121).

• spec mv --write on a document not yet committed offers an undo that works: a copy of the uncommitted file is kept under .specfold/trash/ and the undo removes the new path and copies it back, instead of a git stash that failed on the old path and left a stray stash (C122).

• spec set with values the document already has writes nothing: it reports no change, exits 0, and neither bumps updated nor offers an undo.

• spec import plans note body lines (* Status:, * Deciders:, * Date: …) that repeat values moved into frontmatter, and an H1 that keeps a number the title dropped.

• spec audit suggests checking open tasks (or --force) before archiving an active plan that still has them; spec check's summary counts spec.toml warnings (+N in spec.toml, field config_warnings); a mistyped [audit] coverage_types entry gets a did-you-mean, and spec doctor points at the type name to fix.

• spec mv previews keep --owner, --slug, --tier and --cited-paths in the next command.

• After spec set changes a status or a required field, the next check includes S6 and S8.

• Nexts add --force (sync over a hand-edited copy) and --yes (fix over max_files); the S6
  fix names every missing field.

• Docs say --owner-kind where they still said --owner.

• An attachment that changed because a secret was added to it lists U2 before E1, and E1
  points at spec explain U2 instead of suggesting to attach the leaky file again. Secret
  findings now come first among findings of their level.

• spec help <command> gives the first example as next without its column padding and
  # ... comment (spec help guard ended with the padded example line).

• spec export secret refusals name the source document file and line (not the exported copy)
  and suggest spec guard secrets.

• spec find with no current match suggests the same search with --all, runnable as written.

• spec attach writes updated as an RFC 3339 timestamp, like spec new and spec set.

• Refusing text that holds the mcp redaction marker now suggests spec edit <id> --append.

Changed

• spec new and spec set accept --write (they always write).
• A missing-section error from spec edit and spec show no longer lists the H1 title.
• spec new --translation no longer suggests --title when the title comes from the source.
• An error that ends with a runnable spec … command offers it as next.
• Exporter plugins no longer receive .spec.local documents unless --local is given; edges to
  them are left out and references to them arrive unresolved (reason: "not exported").
• spec edit --text, --new-section, --row and spec new --title take values that start with
  -, such as --text "- [ ] wire gateway".
• spec edit --text - is refused and points to --file - instead of writing a lone -.
• A file that is not UTF-8 is reported by check (S1), fix and show with its path and line,
  instead of "no frontmatter block".
• After spec set <id> aliases=…, the suggested check includes I2 (duplicate aliases).
• spec edit --section on a missing section offers the --new-section command as next;
  spec graph --format json says JSON output is --json and offers that command.