v0.6.0

Compatibility with opencode2 v2.0.27's built-in worktree tools (PR
anomalyco/opencode#53981), which moved worktree_create, worktree_list,
and worktree_remove into the host.

• Native creates produce cow clones. The built-in create carries no
  strategy field and uses the location's default strategy, which this plugin
  registers. Verified on 2.7.50: strategy-less creates report cow in the
  inventory and materialize Deep clones, including when the create is the
  server's first worktree call — no lazy-activation gap. The new native-path
  matrix (scripts/e2e/native-matrix.ts, 18 checks) pins the contract.
• remove_worktree now requires edit approval (permission: "edit",
  the host's removal convention since v2.0.27). With edit: deny the tool
  is hidden, same as the built-in remover.
• The strategy's force-less refusal reason carries through the native remove
  path verbatim, and force: true discards ignored files on both surfaces.
  Cow clones hold env files and local state that exist nowhere else, so
  treat force on a cow lane as destructive.
• Suite: 348 tests; the e2e harness (112 checks) and the native-path matrix
  (18 checks) pass against opencode2 2.7.50.

v0.5.0

The plugin is renamed: opencode2-cow-worktree is now opencode-cow-worktree.

• The package name, plugin id, and TUI id carry the new name. Install with
  opencode2 plugin add opencode-cow-worktree, or
  "package": "opencode-cow-worktree@latest" in opencode.json. The GitHub
  repository moved to rbelem/opencode-cow-worktree; GitHub redirects old
  links.
• CoW capability probes write scratch directories prefixed
  .opencode-cow-capability- (was .opencode2-cow-capability-). Leftover
  scratch from older versions is inert: cleanup is in-process and nothing
  scans for the old prefix.
• The dogfood install check now matches the 20260917 nightly's API: readiness
  polls GET /api/plugin (/api/health lost its handler on the 20260915
  nightly), activation is detected by polling the plugin list, and worktree
  endpoints carry projectID.
• Suite: 348 tests; the dogfood install check passes 12/12 against the
  20260917 nightly.

v0.4.1

0.4.1

Fixes a false refusal class in remove_worktree: a lane whose work already
landed on the remote is no longer refused.

• The landed-ness guard refreshes stale refs before refusing (#16). A cow
  lane is a separate clone whose origin/* refs freeze at clone time, so a
  commit that landed through a merged PR read as unlanded, and the guard
  refused — training operators to reach for force. On the refusal path the
  guard now runs one git fetch origin, re-judges on the fresh refs, and
  reports that verdict. A pass on stale refs is already sound for
  fast-forward remotes, so the happy path never fetches. A failed fetch
  keeps the refusal and names the verdict as stale-limited, preserving the
  guard's fail-closed posture.
• Suite: 348 tests. The live-git repro of the issue and three new
  decision-table rows pin the refresh, the failed refresh, and the
  unchanged verdict.

0.4.0

v0.4.0

0.4.0

The cow lifecycle is complete: worktrees can now be removed by the agent, not
only created and listed. And a class of project roots that could never clone
at all — those carrying a live socket — now clone cleanly.

• remove_worktree: agent-facing removal wrapping the strategy remove
  path. The call names the worktree by name or directory (exactly one of
  the two, resolved against the inventory), refuses anything the inventory
  does not know as a cow worktree, and refuses rows whose directory is
  already gone — until upstream candidate 8 lands, a dangling row cannot be
  de-registered through the API. The landed-ness guard fails closed: HEAD
  must be an ancestor of the landing ref (origin/HEAD, else main, else
  master), or the tree must be clean with no unique commits; force: true
  bypasses the guard and the result says so.
• list_worktrees gains the missing knob (ADR 0003's dangling-row
  story). An ENOENT on a listed directory still fails the listing loudly by
  default; missing: "report" returns the row flagged missing: true
  without a derived createdAt, and missing: "prune" de-registers the row
  through the inventory's own remove (forced) and drops it from the listing.
  Any non-ENOENT stat failure remains a hard failure in every mode. Prune is
  gated on upstream candidate 8: Worktree.remove resolves the real path
  before reading the record, so a gone directory answers 400 even at force.
• Clone skips sockets, FIFOs, and device nodes (#17). A live daemon
  socket at a project root (e.g. .codegraph/daemon.sock) classified as a
  plain file, failed FICLONE with EOPNOTSUPP, and rolled the whole clone
  back — every cow clone on such a root aborted. The entries are now
  classified "special" and omitted: runtime state the clone's own daemons
  recreate, never tree content. Pinned by a real node:net unix-socket
  integration test that runs on any filesystem.

Also in this release:

• Docs: upstream candidates 9-11 filed from the 2026-09-20 field run.
• The coverage gate is green again at 100% lines on all 22 gated files:
  remove_worktree's plugin wiring (the live ctx.worktree remove calls,
  force pass-through, result text, and the prune seam) and the live git
  binding are now pinned by tests. Suite: 344 tests.

v0.3.0

0.3.0

Adapted to the projectID-era opencode2 worktree API (20260915 nightlies and
later; verified against 0.0.0-next-20260917). The API now requires the
project id on every worktree call, dropped the per-request strategy field,
and dropped the location query. spawn_workspace and list_worktrees
derive the id from the plugin's own location context, pass the source as
from, and still send strategy for 2.0.2-era binaries, where it is what
selects the git fallback's mechanism.

Behavior notes:

• On projectID-era binaries the git fallback cannot be requested through
  the create API: a non-CoW source fails with the cow refusal, and the
  refusal is final. The option still works on 2.0.2-era builds. Documented
  in the README.
• Create no longer assembles <parent>/<name> verbatim when the path is
  taken; the server suffixes name-2 through name-10. Named
  spawn_workspace calls are unaffected: the tool refuses anything it
  cannot attach to before create runs.

Also in this release:

• Issue #14 (Desktop hardcodes strategy: "git") resolved upstream by the
  same refactor and closed.
• The e2e harness speaks the projectID-era API end to end: 112 checks, 0
  failed against a real 0.0.0-next-20260917 server
  (docs/e2e/run-2026-09-17.md). Unit suite: 307 tests, coverage gate at
  100% lines on all 21 shipped files.

v0.2.0

0.2.0

spawn_workspace refuses to attach a new session onto a cow worktree a live
session is still using (ADR 0004). The tool records occupancy itself: on both
the create and attach flows it writes .cow-session.json at the worktree root
(kept out of git status through .git/info/exclude), and at attach time it
probes the recorded session through the session API. Only a positive proof
that the session is gone lets attach proceed: a 404, or silence past a
60-minute window. A live session, a malformed marker, or a probe failure that
proves nothing refuses the attach and names the occupying session plus the
three recoveries (pick another name, delete the session out of band, remove a
known-stale marker). Legacy worktrees without a marker keep today's fail-open
behavior and gain a marker; the git fallback is untouched. A failed marker
write never fails the call; the result carries an "unguarded" warning.

Verified against a real opencode2 serve (0.0.0-next-20260912.3): the
attach scenario refuses on a live session, proceeds after the occupying
session is deleted, and backfills markers onto legacy worktrees
(docs/e2e/run-2026-09-16.md). Unit suite: 307 tests, coverage gate at 100%
lines on all 21 shipped files.

Also in this release:

• The e2e harness stays runnable across the 20260915 nightly split; its
  readiness probe polls /api/plugin because the nightly answers 404 on
  /api/health. The pinned 0.0.0-next-20260912.3 is unaffected.
• Docs: ADR 0004, the announcement draft, the ecosystem listing draft, and
  upstream filing candidates 6-8 (the nightly's /api/health 404, its
  projectID requirement that breaks plugin worktree creates, and the
  realpath-first Worktree.remove that strands dangling rows).

v0.1.1

0.1.1

Documentation-only release; no code changes. It carries the new install
story to the npm package page: the plugin installs as a single config entry
(opencode2 downloads it from npm at startup) or one
opencode2 plugin add command, with options in the same entry and the
checkout recipe kept for development. Also new: user-side verification with
opencode2 plugin list, the plugin-cache reset for stuck updates, the
recorded release procedure, and the ecosystem listing draft.

v0.1.0

0.1.0

The first release: tagged v0.1.0 and published to npm.

• cow Strategy: Deep clones via forced reflink (Linux
  COPYFILE_FICLONE_FORCE, macOS copyfile(3) through bun:ffi), a fail-loud
  capability probe, and a same-device pre-flight. Refuses an occupied target
  before the first write, and a failed create removes only what it cloned.
• spawn_workspace: create-plus-session with attach-to-existing semantics
  (attached: true), mechanism reporting, an opt-in git fallback, and a
  loud refusal of anything at the target path the inventory does not know as a
  cow worktree.
• list_worktrees: the location's cow worktrees, derived from opencode2's
  inventory alone.
• Post-create hooks (hooks.postCreate): sequential sh -c in the new
  worktree, COW_WORKTREE_PATH/COW_SOURCE_DIRECTORY in the environment, a
  five-minute per-command timeout, stdin detached, captured output capped in
  errors, and a first failure that rolls the create back.
• Removal: an uncommitted-work guard that fails closed on unknown state, an
  identity-captured quarantine rename, and two-phase deletion with an async
  tail for node_modules.
• Every plugin option validated once at setup; misconfiguration fails the
  plugin load rather than surfacing mid-session.
• macOS backend verified on real APFS by CI (arm64 and x86_64, shared extents
  measured). A 31-scenario parallel-agent swarm and a four-scenario e2e
  harness are recorded under docs/e2e/.
• Development: scripts/lane.ts spawns and absorbs parallel CoW lane clones
  (origin-stripped, --no-ff absorb); conventions in docs/lane-workflow.md.