Skip to content

Releases

v1.1.0: approval subjects + approval-mode opt-out

Choose a tag to compare

@pgquiles pgquiles released this 16 Nov 19:42

v1.1.0

Contributor release — thank you to Bearmancer for #53.

Approval prompts name their subject (#52, via #53)

Conversational approval prompts now show what is being approved — the network destination (protocol://host:port), shell command, file path, or process target — plus each choice's rulePreview and the reviewer's reason. Distinct approvals are distinguishable at a glance instead of looking like the same prompt repeated.

Approval-mode opt-out (#52, via #53)

New OPENCODE_MUSE_APPROVAL_MODE knob (also as the approvalMode plugin option if core ever forwards options). Default stays the vetted forced onRequest; host-default omits approvalMode so the user's own Muse permission profile applies instead of being silently overridden, and the other host-defined modes (allowAll, promptUnmatched, denyUnmatched) can be selected explicitly. Anything else fails setup fast. The opencode tool backend ignores the knob. Design record amended accordingly.

Muse CLI 1.4.20 classified (additive, SDK stays 1.4.14)

Schema ledger entry, compatibility matrix, and README pin extended to 1.4.20 under the recorded fingerprint warning; exact-pin equality with 1.4.14 remains. Catalog, lifecycle + explicit-cut fork, and approval-mode probes re-verified on Linux as well as Windows.

Opencode backend gated on unverified CLIs (#58)

1.4.20 rejects session MCP servers in run.toolset (configError: unknown tool), and without the pin MCP tools never load — so the backend fails setup fast with a clear MuseInvalidConfigError on any served fingerprint without a passing backend live-turn probe (seed: 1.4.14), instead of dying obscurely per turn. Upstream regression reported as meta-models/muse-code-sdk#153. The default muse backend is unaffected on all verified CLIs.

Full test suite

431 tests green on both backends plus live turns per gate; see docs/compatibility.md for the probe record.

v1.0.0: opencode tool backend

Choose a tag to compare

@pgarciaq pgarciaq released this 01 Nov 23:59

Highlights

  • Opencode tool backend, opt-in via OPENCODE_MUSE_TOOL_BACKEND (#37): private per-session config, loopback MCP bridge, parked-call attach. Default muse backend unchanged. Proven with a live turn per backend plus fail-closed setup.
  • Rewind/fork on the pinned Muse 1.4.14 runtime (#20): anchor ledger, explicit-cut fork, fail-closed unmappable rewinds.
  • Registry lifecycle (#49) and loopback hardening (#50) found by adversarial review, fixed before release.

Gate

tsc clean, full suite green on both backends, probe:host-override ok on 1.4.14. Requires Muse Code 1.4.14 and @muse-code/sdk 1.4.14.

Known limitations

  • #36: no sampling support in the pinned runtime.
  • #35: no responseFormat support in the pinned runtime.
  • #22: provider approval UX waits on upstream OpenCode core support.
  • #6: open UI issue, waiting on upstream.

See docs/compatibility.md for pins, probe results, and details.

v0.9.0

Choose a tag to compare

@pgarciaq pgarciaq released this 01 Nov 18:29

What is new since 0.8.0

  • Rewind/fork re-enabled on the pinned Muse 1.4.14 runtime (tracks #20): rewinding the OpenCode conversation forks the mapped Muse session at the retained turn boundary via an explicit cut point; unmappable boundaries keep the fail-closed fresh-session error. OpenCode forks with a uniquely matching transcript adopt via fork instead of starting blank.
  • Forks do not preserve source turn ids (offline probe check 4b): re-pointed entries drop anchors and post-fork turns re-anchor; the fork's own turns stay forkable.
  • Rewinding never rolls back files Muse changed, and every rewind orphans the pre-rewind Muse session in the user's Muse home (MSP has no delete).

Strict build, 368/368 tests, and offline probe (override, rawLog, resume, fork, fork-of-fork, steer) pass. Install snippet pinned to 0.9.0 in the same commit (lockstep rule).

v0.8.0

Choose a tag to compare

@pgarciaq pgarciaq released this 01 Nov 12:43

What is new since 0.7.1

  • Muse CLI and @muse-code/sdk are both pinned to 1.4.14: the SDK expected fingerprint equals the served stable fingerprint, so exact-pin equality is restored and no startup warning is recorded (tracks #47).
  • Schema delta 1.4.13 -> 1.4.14 classified additive-only; snapshot ledger verified against a fresh export.
  • Steering (#44), per-model thinking levels (#45), and lease drainer counting (#48) now run against the matching SDK.
  • Probe helper accepts -32030 commandRejected as rename-unavailable alongside the old -32603.

Strict build, 358/358 tests, offline probe all true. Live catalog, text turn, and explicit-cut fork PASS on subscription and PAYG (user-reported keyed-shell runs). Install snippet pinned to 0.8.0 in the same commit (lockstep rule).

0.7.1

Choose a tag to compare

@pgarciaq pgarciaq released this 27 Oct 19:43

Patch release: approval-bridge reliability fixes, all verified live against OpenCode 2.1.3.

  • Fix conversational deny-loop trap: after a second unrecognized answer the denial stands, the parked turn is interrupted, and the message runs as a fresh turn, so a re-asking model can never trap the conversation.
  • Recovery choice (#38) defers to live bindings: no more false recovery prompts on slow turn/start; the turn-lifecycle rule decides in resolve.
  • Verdicts are consumed before session resolve and matched against the latest user message only, fixing WorkspaceBusyError deadlocks and stale-text joins.
  • Park wire-object approval requirement ids under the canonical key, fixing the round where every approval was unroutable.
  • Retry worktree-bearing session/start once on explicit rejection (non-git/SVN/plain checkouts).
  • Funnel all usage through nested V3 containers, fixing the flat-usage crash.
  • Brief the model against prose permission asks (#39), ending the double approval prompt.
  • Derive Allow/Deny labels when the CLI offers no presentable options.

v0.7.0: per-session worktrees

Choose a tag to compare

@pgquiles pgquiles released this 25 Oct 17:37

What is new since 0.6.0

  • Each OpenCode session starts its Muse session on a server-managed worktree instead of sharing the checkout, so concurrent sessions stop serializing at the workspace lock (tracks #34, which stays open for the remaining lease-protocol and orphans work).
  • The human merges through normal source control; the bridge never auto-merges, and purge-on-close removes only provably-empty worktrees.
  • Graceful fallback: untrusted workspaces and CLIs without the experimental surface keep the single-writer lock with a conversational notice.
  • First-turn labels tell the agent which worktree it is on; PDF staging follows the session onto its worktree.

Strict build, 310/310 tests, and package-contents checks pass. Install snippet pinned to 0.7.0 in the same commit (lockstep rule).

v0.6.0: audio transcription and PDF staging

Choose a tag to compare

@pgquiles pgquiles released this 25 Oct 16:35

What is new since 0.5.2

  • Audio attachments are transcribed in-plugin through muse voice transcribe (voice mode is subscription-included) and enter the turn as declared labeled text; untranscribable clips fail loudly, never silently dropped (#32, closed).
  • PDF attachments are staged into <workspace>/.opencode-muse-uploads/<uuid>.pdf for the agent to read with its own tools; staged names are always generated, retention is purge-on-close (#33, closed).
  • Install snippet pinned to 0.6.0 in the same commit (lockstep rule).

No changes to approvals, billing lane, or session management. Strict build, 301/301 tests, and package-contents checks pass.

v0.5.2: install snippet matches release

Choose a tag to compare

@pgquiles pgquiles released this 25 Oct 15:37

Docs-only follow-up: the README install snippet now names opencode-muse@0.5.2 in the same commit as the version bump, so the npm page never points at an older plugin. No code changes; strict build and 284/284 tests pass.

v0.5.1: README refresh

Choose a tag to compare

@pgquiles pgquiles released this 25 Oct 15:24

Docs-only release so the npm package page reflects the 0.5.0 functionality: conversational approvals (#30) and option-question answers (#31) are now claimed as Features, Contributor-by-default is stated up front, and Current limitations keeps only genuine residuals. No code changes since v0.5.0; strict build and 284/284 tests pass.

v0.5.0: conversational option-question answers

Choose a tag to compare

@pgquiles pgquiles released this 25 Oct 13:58

What is new since 0.4.0

  • Conversational answers for structured option-questions (#31, closed): userInput/request parks in a per-session exchange instead of failing with method-not-found; the oldest unanswered question renders as a lettered list and the next message answers it (exact-one-match, re-ask once then cancel with a reason). Multi-select and option-less questions cancel with a named reason. userInputDialogs stays false so honoring servers see no behavior change.

Known limitations (unchanged)

  • Rewind/fork is unavailable on the pinned Linux runtime (tracked in #20, upstream bug filed).
  • Provider approvals depend on core support (tracked in #22, upstream feature request filed); the conversational workaround in #30 stands in meanwhile.
  • Structured input beyond single-select option-questions cancels with a reason (tracked in #6, upstream support awaited).
  • Live event-gap recovery is unverified; fake-host coverage stands in (tracked in #23, closed as unobservable).

Strict build, 284/284 tests, and package-contents checks pass on the tag via the publish workflow.