Skip to main content

Releases

0.8.0 changes and migration​

Download v0.8.0, published from 4b5b297 on 2026-11-13 for all four declared native platforms.

Version 0.8.0 combines guided first-run profile/plugin setup, immediate controls, recoverable conversation views, independent agent conversation tabs, saved agent activity, interactive agent/workflow tables and process CPU/memory charts. Automatic permission review can use the active provider and the actual task context. Coding completion now checks current verification evidence; edits that leave file text unchanged fail, and finite queries support bounded repetition recovery through optional ToolSpec.loopGuard: "query" metadata.

/btw <question> now opens or restores a persistent side-agent conversation. Use /btw --once <question> for the former isolated notification response. The bundled btw plugin is version 2.0.0; update the host before adopting the new plugin major. Existing scoped plugin settings remain separate from installation and activation. Conversation history persists; unsent side-agent drafts are preserved during navigation but do not survive a process restart.

OpenTUI migration and embedded CPU inference for onboarding remain planned. The setup guide itself requires no model call; completing agent tasks requires a configured provider. Local CPU diagnostics and passing deterministic checks do not establish general model quality or stable 1.0 readiness. See the acceptance gates and current evidence.

Distribution contract​

The workflows pin GitHub Actions to immutable commits for checkout 7.0.1, setup-node 7.1.0, upload-artifact 7.0.2, download-artifact 8.0.2, setup-bun 2.2.0 and git-cliff-action 4.9.1. The JavaScript actions use their own Node 24 runtime and require Actions Runner >=2.327.1; this is separate from the Node version used to build and test the application. Hosted runners satisfy that requirement. Artifact digest mismatches fail the release download step. Bun 1.4.4, native Callix 1.7.0 and git-cliff 2.14.2 remain explicit tool pins.

The Linux test job uses ubuntu-24.04, matching the native release baseline, with Node 26 and 22.19.0. It does not inherit distribution upgrades through the rolling ubuntu-latest label. Linux validation jobs install ripgrep 15.1.0 from its official GitHub release through node tools/release/ci-ripgrep.mjs, verify a pinned SHA-256 and the binary version, then add the isolated installation to GITHUB_PATH. This avoids Ubuntu package mirrors and sudo; installation failure still stops tests rather than skipping search coverage. The installer retains the upstream MIT license. Native archive jobs retain their declared platform-specific prerequisites.

Plugin release verification installs the Chromium revision selected by the locked Playwright package, using the system libraries already supplied by the declared hosted runners. It does not rerun playwright install --with-deps or depend on Ubuntu package mirrors for optional font packages. A separate browser preparation step must launch Chromium and load a page before mandatory terminal scenarios; missing browser binaries or libraries still fail the job. Browser preparation is limited to 10 minutes, source/platform certification jobs to 30 minutes each, and publication to 15 minutes. These are CI job limits, not application tool limits.

Release archives support Linux glibc 2.38+ (x64/arm64) and macOS (x64/arm64). Windows and Linux musl are not currently supported by the installer. The shipped Monty native module needs glibc 2.38+; the complete Linux bundle is tested on Ubuntu 24.04. Ubuntu 22.04 is not a supported full-bundle target. Each archive contains compositor/compositor, external standard plugins, mascot APNG assets, public host API modules, Monty's platform-specific worker, and a pinned native Callix executable with its upstream license/provenance. The Bun runtime is compiled into the executable. Keep the archive's directories together; copying only the executable drops the supplied plugin distribution.

The optional OpenTUI frontend ships as an isolated JavaScript bundle under node_modules/@agent-compositor/renderer-opentui/, together with its matching native library, parser worker, WebAssembly grammars and highlight queries. The native build bundles only its declared platform's library; Windows and musl packages remain outside the supported archive targets. The lazy renderer loader and native-library bootstrap resolve paths from the installed files rather than the current working directory. An explicit upstream OPENTUI_LIB_PATH override is preserved. Native FFmpeg license/source information and bundled dependencies' author and patent notices are retained under licenses/, including dependencies installed beside the renderer workspace at different versions.

Each native job extracts the archive outside the checkout and runs the actual OpenTUI app, managers, navigation, cancellation, request-window, selected private-controller and editor scenarios through the shared PTY/browser bench. Each job also verifies explicit recovery of a queued input after SIGKILL and a retry interrupted during a controlled journal write. Reports identify the extracted executable by SHA-256 and remain available on failure. These cases use a loopback provider and do not prove recovery at every external effect boundary. The selected controller gate uses isolated explicit approval/question/form providers with a bounded loopback model; it does not select a custom editor, history or placement implementation. The matching source command is npm run visual:opentui-controllers. npm run visual:opentui-child adds delegated and managed agent requests in native and selected-controller compositions. Source CI and all four native jobs run the same CLI scenario against their actual runtime and retain its reports, private synthetic receipts, terminal input/output and screenshots on failure. It verifies source separation, exact decisions, Main draft ownership, selected producer registries and managed input across Main resume and shutdown. npm run visual:opentui-btw checks the side conversation while Main is running, including separate drafts, window reopening and restart. Node 26 CI and all four native jobs run this scenario and retain its screenshots and report. npm run visual:opentui-resources checks live system charts and explicit sample sharing while Main is running. Node 26 CI and all four native jobs retain the report; the native report identifies the executable, renderer and bundled plugin. The additional npm run visual:opentui-editor gate selects a custom editor in two isolated compositions, covering F2/window ownership, real admission and recovery, Unicode checkpoints, canonical reload/resume and expired capabilities. It retains native and selected private requests, and checks the actual relocated executable and renderer identities. The emulator uses an explicit Unicode 11 width profile; this does not prove every native terminal's Unicode behavior. npm run visual:opentui-workspace-controllers adds selected history and placement to the same actual-app gate: browse and private forms during a Main question, canonical session mutations/resume, real Apply/Save rules, late-focus/target guards, hide/cancel/unload, and restart restoration. Source and every native job use the same bounded loopback fixture and retain input/output plus identities. npm run visual:opentui-conversation-copies creates persisted windows through the real Ink CLI, then launches OpenTUI against that same isolated storage. Its native and selected-controller compositions verify every saved window kind, closed identities, shared Main input/request/editor ownership, independent scroll, private forms and actual reload/resume/restart. No migration layout or journal is manufactured by the fixture. The source job and all four relocated native jobs require this gate and preserve its artifacts on failure. npm run visual:opentui-main-draft checks built-in Main draft recovery, explicit review and conflict selection, including keyboard and mouse, rapid submission, process termination and normal exit. Source CI and each relocated native job run all nine scenarios with a loopback provider and retain their reports. npm run visual:opentui-appearance adds native/selected compositions for actual bundled theme commands, all portable colors/chrome/presentation, contribution ordering/removal, exact title-control hits, live Main tool details, private requests, Unicode draft ownership and reload/restart. Its bounded loopback fixture uses normal plugin services and user input, not UI/Core injections. Source and all four relocated native jobs require the same appearance gate. Reports, terminal input/output, screenshots and executable/renderer identities are retained with the other installed-runtime evidence, including on failure.

npm run visual:opentui-render adds the public hook gate in native/selected compositions. It exercises the actual bundled theme status override, every public site and differently sized copies, default-first asynchronous rendering, normal hook-chain validation/effect guards, invalidation, rendered actions/table/tabs/ raster, private input isolation and stale results across lifecycle changes. Source and all four relocated native jobs require this gate and retain its actual input/output, screenshots, receipts and executable/renderer identities.

The configured application's OpenTUI checks use the native presentation profile of protocol 0.7 through the production asynchronous message path. A separate npm run visual:opentui-wire gate exercises the compatible public grid source/send client in a real PTY: schema-valid full/diff/flush output, focus-routed keyboard/paste/global mouse, and cleanup after EOF, rejection and a blocked source. Each native job points VISUAL_RENDERER_MODULE at the exact relocated archive's renderer entry, recording its hash with the input, ANSI, screenshots and process exits. This API-client fixture does not replace the configured application's lifecycle scenarios or physical-terminal media checks.

User and project plugin directories take precedence over the distribution's supplied packages. Since v0.4.0, plugins remain inactive until explicitly enabled in user or project scope; installation alone never activates them. The repository marketplace ships offline with the distribution. Existing explicit activation settings are preserved. Use /extensions enable NAME --project (or --user) and /reload, including for the optional mascot and plugin-author packages. Plugin versions/releases and catalog versions are independent of the host. See plugin standards and mandatory testing. Plugin version/changelog preparation is automated by the release PR workflow; application version preparation below remains separate. Plugin publication extracts only the selected version's notes and retains all certification gates. Monty's bridge needs Node.js 22+ or AGENT_NODE_BIN. Callix 1.7.0 is bundled under plugins/callix/bin/; the installer needs no separate Callix download. CALLIX_BIN and plugin settings still override the bundled executable. The already-published v0.1.1 archive predates this addition and needs explicit setup. Native third-party plugins may declare additional runtime dependencies.

Independent plugin archives retain license notices for dependencies selected by the bundler and separately copied native runtime packages. Unrelated workspace dependencies, including the documentation site, are excluded so they do not inflate plugin installation inventories. Native Callix provenance and its upstream license remain alongside its executable.

The POSIX installer uses curl, tar, awk and sha256sum (or macOS shasum). --prefix defaults to $HOME/.local; --version selects a release tag or latest. COMPOSITOR_INSTALL_PREFIX supplies the default prefix; COMPOSITOR_RELEASE_BASE_URL can point to a release mirror with the same URL layout. A checksum mismatch stops installation before extraction. Downloads allow 15 seconds to connect and stop after 60 seconds below 1 KiB/s. Each attempt has a 30-minute limit, with up to three retries admitted within 30 minutes; an attempt already started can finish after that retry window. Network failures leave the active installation and application state untouched. Checksums validate download integrity, not publisher identity or signatures.

Installations live in unique directories under PREFIX/share/agent-compositor/, with PREFIX/bin/compositor pointing to the active binary. Updates switch that link and retain older versions. Application state is never copied or removed. To uninstall, remove that symlink and the PREFIX/share/agent-compositor distribution directories; user state remains in its configured home. The installer does not edit shell profiles or use sudo.

Build and smoke test​

dist/release/discovery-smoke-PLATFORM.json records the installed discovery/read probe: real empty text search, broad glob, rejected zero-based read pagination, valid ranged read and its guarded edit anchor, and rejection of an edit that leaves the actual file unchanged. It checks the actual model-facing schema and response guidance and preserves project/user files. Each native job runs it both normally and with the explicitly selected project inside a parent repository's ignored directory. The latter adds a semantic query that must find a visible symbol and exclude an .agentignore symbol, using seven bounded loopback completions instead of six. It also checks path-specific ancestor, repository and global exclusions in both file and semantic results. Its report is dist/release/discovery-ignored-smoke-PLATFORM.json. Both reports identify the executable by SHA-256. The probes establish packaged tool behavior, not live model coding quality.

dist/release/edit-inspect-smoke-PLATFORM.json records a separate installed loopback-provider flow: an edit duplicates adjacent code, its returned inspect arguments read the current file, and a fresh guarded edit restores the original. The probe checks a separate draft survives. It verifies packaged tool cooperation, not autonomous model recovery.

dist/release/query-smoke-PLATFORM.json records installed query-loop recovery. Actual built-in reads and Callix searches must each receive a recovery reminder and stop after one novel plus six repeated query steps, with seven successful tool results verified in the session journal and exit code 1. A separate case uses a new read after the reminder and must finish normally. The three cases use at most 20 loopback completions, preserve source/user files, and are mandatory on every native platform. These controlled responses do not establish autonomous model recovery quality.

The installed-runtime suite also runs a coding-preset completion probe with no saved task plan. An unchecked write must retain progress and fail after bounded feedback; a second case must execute a real project verification command before successful completion. Two further cases rerun a real check: a pass followed by a failure must reject completion; a third successful run must resolve that failure. All four preserve the user draft and checker source and record the real run count. The probe uses bounded local HTTP responses, records exit/error/result evidence in dist/release/coding-smoke-PLATFORM.json, and is required on every declared native release platform. This deterministic policy check does not establish live model coding quality.

Run these commands on the target platform with Node.js >=22.19.0 (Node 26 in release CI), Bun 1.4.4 and Python 3 for the release verifier's real POSIX PTY:

npm ci
npm run build:release
npm run release:smoke

Building downloads the platform-specific official Callix 1.7.0 archive and checks its pinned SHA-256 before extracting only the binary and upstream MIT license. plugins/callix/bin/release.json records the source URL, platform, version, archive digest and binary/license digests. No developer-local binary is copied into releases. For an offline build, set CALLIX_RELEASE_ARCHIVE to the matching official archive; the same pinned checksum is mandatory. Update the pins in plugins/callix/distribution.mjs when upgrading Callix.

The archive is written to dist/release/compositor-PLATFORM.tar.gz. Native approval probes check effective canonical file paths. Discovery and cross-project probes consume the current metadata/verbatim read frame; guarded edits copy its complete editRef into the required model target parameter. The smoke test serves a local release mirror and model endpoint, rejects a corrupt checksum, installs from a versioned URL and latest, checks default plugin discovery outside the repository, executes a headless model request and the packaged Monty bridge, and checks state preservation across reinstall. It also drives the installed binary through a write, completion feedback and automatic compaction using a local model fixture. The original task must remain verbatim in the next main request and every summary reference, the write must not repeat, and an unrelated user draft must survive. This deterministic check does not measure live-model development quality. Its report is written to dist/release/context-smoke-PLATFORM.json and printed in the smoke log. It also starts a real owned background shell, delivers its completion while keeping the original task for permission review and automatic compaction, then resumes the same session and delivers another plugin-owned follow-up. The original task, journal provenance, request previews and unrelated draft must survive. A small public SDK fixture records permission context; it does not replace core, context or session services. All model traffic goes to a bounded local HTTP fixture. The report is dist/release/background-smoke-PLATFORM.json. Another installed-CLI case journals a controlled tool call, kills the process after its file effect but before its result, and resumes the session. The next local model request must retain the original task and report the outcome as unknown; the effect must not repeat and an unrelated draft must survive. The report is dist/release/effect-crash-smoke-PLATFORM.json. A second variant first requests an unknown tool and verifies that its original arguments and rejection survive the same process death; its report is dist/release/rejected-call-smoke-PLATFORM.json. This fixture does not prove recovery for arbitrary external effects. These controlled checks do not establish long-session or live-model quality. Plain-session startup is checked with a real SDK plugin too: immediate EOF initializes the plugin once without a model request, and startup finishes before the first model request in a second process. Its report is dist/release/plain-startup-smoke-PLATFORM.json. The installed headless signal gate runs six actual CLI cases: SIGINT and SIGTERM against an immediate SDK command, an ordinary foreground exec, and a held model with native and plugin-owned background work. It observes owned worker readiness before signaling the CLI, then requires their termination before fixture cleanup, an unrelated live control process, exactly one unsuccessful final result and exit 130 or 143. All model traffic is loopback. Reports are written to dist/release/headless-signals-smoke-PLATFORM.json; the matching directory retains per-case input, stdout/stderr, journals, HTTP and PID evidence with hashes. CI preserves both the report and raw directory even when a gate fails. This is graceful signal-cleanup evidence, not a guarantee for SIGKILL or OS containment. The installed plain interface also receives a pasted main prompt plus an immediate control that must release a held local request, and a separate EOF case that must wait for a pending command after the main response completes. The bundled /btw --once must return its notification and exact provider-reported usage in headless JSON; unavailable agent support must produce a failed result without a model request. Finally the existing Python PTY bridge drives a real raw-mode approval: an explicit y applies one controlled write and never becomes another model prompt. The CLI must then actually exit, rather than only print bye. The raw response remains streaming until the first Ctrl+C, so the verifier waits for the observable interrupted state before the second Ctrl+C exits. Text output alone is not evidence that turn cleanup has completed. Each native job preserves its smoke JSON reports as installed-runtime-PLATFORM, including on failure. The PTY bridge reaps a finished child before group cleanup and propagates its actual exit status. The native gate requires a successful raw-process exit; printing bye or a bridge error cannot count as success. The report is dist/release/plain-controls-smoke-PLATFORM.json. This uses bounded loopback responses, not live-model quality or graphical TUI evidence. Python is a verification dependency; the installed application does not require it for plain input. The installed bundled auto reviewer is exercised through /auto on and three controlled write proposals: task-context medium-risk approval executes the write, structured refusal stops it without a manual question, and an explicit host deny prevents both review and execution. Two additional cases propose a real local npm test command: a structured approval runs the fixture check, while a structured refusal leaves its execution marker absent. Each case preserves an unrelated draft; review requests must contain the exact original task. This uses the actual packaged plugin and active-provider fallback with local structured responses, not a substitute permission-review plugin. The report is dist/release/approval-smoke-PLATFORM.json. Classification quality, uncertainty dialogs and cancellation require their separate checks. The smoke test also checks Callix provenance, version and license, then invokes code_query through the installed compositor against a two-file TypeScript project. The check requires a resolved cross-file call with CALLIX_BIN cleared and PATH empty, proving use of the bundled binary. It requires no provider credentials or external model service.

Changelog and publishing​

git-cliff reads Conventional Commits and renders cliff.toml. Its npm dependency is pinned so local and CI output use the same version. feat, fix, docs, perf, refactor, test and maintenance commits are grouped; breaking changes are highlighted. Release bookkeeping commits are excluded. It generates notes; it does not build binaries or publish assets. Only tags beginning with v followed by a digit delimit host releases; independent plugin/NAME/vVERSION tags do not consume pending host changes.

  1. Update the root and packages/app/package.json versions together, then run npm install --package-lock-only.

  2. Run npm run changelog:preview to inspect pending changes. Generate the next entry with npm run changelog -- --tag vX.Y.Z.

  3. Run type checking and tests, commit the release changes, and push main.

  4. Create and push an annotated tag matching the package versions:

    git tag -a vX.Y.Z -m "Release vX.Y.Z"
    git push origin vX.Y.Z

The release workflow runs the quality suite before building on four native runners. Every archive passes the installer/model/Monty smoke test on its own platform. Only then does git-cliff generate the tagged release notes. CI creates a draft, uploads all four archives, install.sh and SHA256SUMS, and publishes it once uploads finish. A manually dispatched workflow with a blank release_tag builds and verifies without publishing. To rebuild an existing unpublished tag through a repaired pipeline, dispatch the branch with release_tag=vX.Y.Z. Publication requires the package version to match that tag and every source file to match its commit, except this release workflow and tools/release/smoke.mjs. The publish job checks out the immutable tag for release notes and installer content. All four native build and smoke gates still apply; downloaded CI artifact digests are verified before publication.

If publication fails after creating a draft, inspect the failed run and draft before retrying. The workflow intentionally refuses to overwrite an existing release. Never move a published version tag; fix the issue in a new version.

Unpublished v0.6.0 candidate​

The v0.6.0 tag was not published: native smoke checks exposed a journal payload containing a live AbortSignal in a custom repository. Bun rejects cloning that capability. The corrected core serializes detached JSON records before the journal service chain, retaining fail-closed writes and live execution signals. The corrected release is v0.6.1, which passed all four native build and installer smoke gates. v0.6.0 stays immutable and is ignored by git-cliff so its unpublished feature/fix commits appear in the v0.6.1 release notes.

v0.1.1 dependency audit​

The original v0.1.1 Zod 3 dependency was flagged by npm audit for GHSA-mrhx-c7xc-m3w5, a memory-exhaustion issue during validation of oversized collections. The published fix requires Zod 4.6.6 or later. The subsequent dependency refresh migrates the validators to Zod 4.6.7 while preserving frozen event contracts; the refreshed dependency tree reports zero advisories. See the dependency review for versions and validation.