akb v0.22.0

Changed

• BREAKING: akb init inside a git repository refuses until the KB's versioning mode is
  chosen. The invocation exits 2 naming the repository it found and what each flag does:
  --embed puts the KB into that repository's history — akb commits only kb/, raw/ and
  .agent-kb/, never the host's other worktree changes — and --no-git leaves the KB
  unversioned and out of the host's git status through .git/info/exclude, a file local
  to the clone. Previously the same invocation created a repository nested inside the
  host's, and akb init wrote user.name into it. A KB that is not inside a repository is
  unchanged: akb init <name> with no flag still creates a repository of its own.

• The versioning mode is recorded in .agent-kb/akb.yaml as versioning: git or
  versioning: none, written explicitly at init in both modes. A config that predates
  the key means git — the mode those bases were created in — so an existing KB keeps
  working unchanged; any other value is rejected rather than read as git, so a misspelled
  mode never commits under the wrong assumption. A none base invokes no git at all:
  mutations write their files and stop, the base takes its lock on .agent-kb/akb.lock
  instead of on the host repository, and akb status reports Versioning: none in place of
  the git line. --no-commit is a no-op there, as it is for any base that never commits.

• Commit identity resolves from the environment, then the base, then git, and is exported
  to the commit instead of passed as -c user.name=…. AKB_AUTHOR_NAME/AKB_AUTHOR_EMAIL
  win, then git-author/git-email in the base's akb.yaml, then git's own resolution
  (GIT_AUTHOR_* in the environment, then the merged repository and global configuration).
  An identity akb supplies is exported as GIT_AUTHOR_* and GIT_COMMITTER_*; when it
  supplies none, git resolves the identity itself, which leaves a user-set GIT_AUTHOR_*
  working as a documented escape hatch. No source naming an identity is a usage refusal
  (exit 2) that fires before anything is written, replacing the raw
  Committer identity unknown fatal an agent could only hit after its page was staged.

• A mutation blocked by an in-progress merge is refused up front instead of failing at
  the commit. Detecting a clean in-progress merge (MERGE_HEAD), not only unmerged paths,
  closes the gap where a git status --porcelain scan passed and
  fatal: cannot do a partial commit during a merge surfaced after the page had been
  written and staged. The refusal names the host repository and the path it changed, and
  states that the merge belongs to the host project rather than to the KB.

• Every mutating command goes through one mode-aware storage helper. storage.OpenStore
  selects the provider from the recorded mode and owns the lock, the identity and merge
  preflight, and the commit, replacing the hardcoded NewGitProvider call sites and the
  direct LockRepo/CommitFiles/NothingToCommit calls across the command packages. The
  base lock is now held across the kb/log.md appends too.

Added

• akb init gained --embed, --no-git, --force, --author-name/--author-email,
  and akb init .. --force bypasses the refusal to write .agent-kb/, kb/ and
  raw/ directly into a repository root; it selects no versioning mode, so a repository-root
  invocation still names one (akb init . --embed --force). --author-name/
  --author-email supply the init commit's identity for that commit only and are never
  written to akb.yaml: the akb default (agent-kb <agent@agent-kb>) is recorded there
  only when neither the environment nor git config names an identity, because an identity
  baked into a file that travels with the base would be mis-attributed after a clone.
  akb init . initializes the directory it runs in and adopts that directory's basename as
  the KB name, reported on the exit-0 line. akb init honors --no-commit: it scaffolds and
  stages the base and leaves the commit to the caller, so an orchestrator can bundle init
  with templates and an overview into one commit.

• KNOWN-LIMITATIONS.md records that a KB cannot be a git submodule of the host
  project, with the reason (a base's repository is local to the machine that initialized
  it and akb has no remote or push machinery, while git refuses the local-path clone form
  since CVE-2022-39253: fatal: transport 'file' not allowed) and the supported answers
  (--embed, --no-git, or a human-run git submodule add over a hosted repository).

Fixed

• akb init no longer clobbers an existing .gitignore. Both files it writes append
  the lines they need and keep every rule already there, so re-running init over a
  directory that carries an ignore file is safe instead of destructive.

• Repository-root .git/info/exclude entries are anchored with a leading slash
  (/.agent-kb/, /kb/, /raw/). Unanchored, they also matched a same-named directory
  anywhere below the root, hiding unrelated host files from git status.

• The integration suite no longer takes its git identity from the machine it runs on.
  Once akb init stopped writing repo-local user.name (identity moved to akb.yaml),
  the testscripts that shell out to raw git commit had no git-visible identity of their
  own: they passed on hosts whose system-level git config supplies one (NixOS does, even
  with HOME moved) and failed on CI runners, which have none. The testscript Setup
  now injects an ambient GIT_CONFIG_* identity every scenario inherits; scenarios that
  exercise identity resolution itself override or empty those same slots. The release
  preconditions gained a test gate that runs the suite under a CI-simulated git
  environment (no system or global config, no identity environment), so a
  machine-dependent scenario blocks the tag instead of the published release workflow.

• A --no-git base at a repository root no longer writes a root .gitignore for
  exclusions the host's exclude file already carries.

• An invocation with nothing to record no longer fails on a commit identity it would
  never use. akb raw sync, akb index remove and akb template write settle the
  no-change case under the base lock before the preflight; a sync that finds nothing also
  stops rewriting the manifest.

• Post-write failure guidance is mode-accurate: an unversioned base is no longer told to
  resolve a merge or finish a commit by hand.

akb v0.20.0

Changed

• BREAKING: the knowledge-base state directory is renamed .akb/ → .agent-kb/,
  and its config file loses the redundant leading dot: .akb.yaml → akb.yaml.
  The directory now carries the project's name; the tool-facing files inside it
  keep the tool's. akb init creates the new layout, and every command —
  discovery, path resolution, read/write, lint, templates, search — recognizes
  only it. There is no compatibility shim: a base in the old layout is reported
  as not a knowledge base, with the expected marker path named. Migrate an
  existing base by renaming the directory and config file (mv .akb .agent-kb && mv .agent-kb/.akb.yaml .agent-kb/akb.yaml) and updating the .akb/search.db*
  line in the base's .gitignore. Tool-named surfaces are unchanged: the
  AKB_KB environment variable, the *.akb.bak backup extension, and the
  page.akb CEL map key.
• The on-disk layout now has one definition site. internal/path exports
  StateDirName, ConfigFileName, and the StateDir/ConfigPath/
  TemplatesDir/SearchDBPath helpers, replacing the ".akb" string literals
  scattered across the command and internal packages.

akb v0.19.1

Fixed

• akb init stages and commits the new base through the index-lock retry. A concurrent
  process holding the git index lock no longer aborts the initial staging of base files;
  the retry now covers both stages.
• The compute-budget error is identified by sentinel, not by text. internal/cel
  exports ErrComputeBudget, and a rule's evaluation error embeds the key names the rule
  reads — so text matching handed the budget remedy to rules whose field names collide with
  the phrase. Write-path dispatch now matches with errors.Is.
• akb template get --example warns about every broken mockup rule, not just the first.
  Revalidation previously stopped at the first unevaluable rule, hiding later failures and
  dropping the variant's staleness headline. Every failed rule is collected, then named.
• The template-load failure is the error reported, not a stale page-read error. When
  akb index add falls back to templates and the template set fails to load, the wrapped
  load error is surfaced — it names the cause where the old message named only the page;
  the per-file guard at the append, index-add, lint, and type-directory call sites keeps
  its own error, its classification, and its exit code.