Skip to content

0.22: ship a mobile-first, plugin-free core and measure size reduction #4591

Open

Description

@thymikee

The 0.22 goal is a lean core with no default-plugin implementations shipped from the core entrypoints. Lazy imports alone do not meet that goal if those implementations still enter the core tarball. Approximately 3 MB unpacked is aspirational, not a release gate. The maintainer expects a substantial 30–50% overall size reduction as non-mobile/default-plugin implementations leave core; measure the achieved result against a declared pre-extraction baseline. Compressed download size is tracked separately.

Current measured core: 6,098,557 B unpacked / 1,850,043 B compressed at 716dddd (same tree as current integration root 0adf3b48). CI Size artifact contains the package-entry inventory. The distance to the aspirational 3,000,000 B reference is about 3.10 MB; this is context, not a required byte budget. These figures exclude separately installed plugins.

Final 0.22 delivery and consumer-management contract

Only declarative capability/catalog metadata needed for help, documentation, schemas and installation/recovery guidance may describe an absent plugin in core. No plugin factory, provider/engine implementation, plugin-only worker or dedicated runtime SDK may ship in the agent-device tarball or its default runtime dependency graph. A monorepo package or development build is not permission to bundle it. optionalDependencies and dynamic imports of bundled code do not satisfy extraction.

Plugin code is downloaded and installed separately, on demand, by agent-device's plugin manager. Keep the current v2 factory and ./agent-device export. Do not restore old manifest adapters, builtin factories or hidden fallback implementations. A user who never uses Maestro must pay no Maestro engine/worker/dependency bytes in the core install. Shared contracts and independently required local-device implementation remain with their actual core owners.

T3 Code and e2e consumers are first-class clients of the same management owner. They must be able to request a reproducible compatible plugin set, inspect installed versus active versions in machine-readable form, and obtain typed missing/incompatible/broken/stale-selection recovery. They must not implement npm installation, infer private store paths, or carry their own copy of the plugin installer. CLI and any required SDK adapters must share that owner. Do not publish speculative SDK function names as implemented interfaces.

Existing CLI interface example (use a real compatible published version):

agent-device plugins add @agent-device/maestro@<exact-version> --json
agent-device plugins list --json
agent-device replay --maestro flow.yaml

Keep installation explicit: a user or embedding framework requests it. Help, documentation and an ordinary missing-plugin error must not silently download code. Install on the daemon/execution host; a local client installation does not imply remote availability. Preserve host-owned sessions and explicit restart guidance rather than replacing or killing another consumer's daemon.

Acceptance includes a clean core consumer with no plugin bytes but working help, a real manager-installed package, and representative T3-style client plus e2e/embedded-host workflows using the same installation/selection lifecycle. Prove removal from all packed core entrypoints, not only the eager startup closure. If the current manager cannot express a required framework or packed-artifact workflow, extend its existing owner rather than adding a bespoke framework/test installer. Do not claim unsupported tarball-path syntax already works.

Required behavior:

Completion:

  1. Record a source-backed inventory of remaining built-in plugins, packed entrypoints and plugin-only dependencies, with attribution to owning issues.
  2. Turn any additional legacy build/bundling findings into small concrete fixes with prerequisites; delete superseded paths instead of retaining a second owner.
  3. Produce genuine equivalently prepared before/after core tarballs at named SHAs, with JavaScript/declaration/native/other breakdowns and unpacked/download percentage changes against a declared pre-extraction baseline. Evaluate the maintainer’s expected 30–50% overall reduction and explain the measured residual owners. Approximately 3 MB unpacked is aspirational; it is not a hard acceptance ceiling or reason to cut required mobile behavior or weaken checks.
  4. Verify core-without-plugins and isolated installed-plugin behavior; run required exact-head checks without weakening gates. Keep hosted/live/package readiness separate.

Dependencies: the extraction issues above retain their sequencing and publication prerequisites. This is final core-bundle acceptance plus additional proven legacy cleanup; it does not block experimental internal nightlies. The initial read-only audit established no additional obsolete configuration outside existing extraction ownership; final byte attribution and this delivery contract remain open. The maintainer now explicitly includes removal of non-mobile implementation from the default core, with web tracked in #4597. Inventory other candidates before proposing bounded follow-ups; do not infer blanket platform removals, a generic bundler migration or unrelated web feature work.

Activity

  1. added this to the 0.22 - lean core milestone on Nov 16, 2026
  2. apex-sentinel-staging commented on Nov 16, 2026

    @apex-sentinel-staging

    Apex Sentinel · Issue review

    Suggested label: feature

    The issue requests a 0.22 enhancement to ship a mobile-first, plugin-free core, remove default-plugin implementations, and measure size reduction with acceptance criteria and dependencies.

  3. thymikee commented on Nov 16, 2026

    @thymikee
    MemberAuthor

    Decision: plugin distribution: final 0.22 extraction means separate agent-device-managed installation, not merely lazy loading. Core retains declarative help/docs/catalog metadata, with no shipped plugin implementation or dedicated dependency closure. T3 Code and e2e use the same management owner and observable install/selection/recovery contract. The issue body now states this explicitly; #4591 owns shared consumer acceptance and the approximately 3 MB unpacked-core target. Earlier seam-only/static-registration or v1 manifest wording is historical preparation, not the final requirement. Existing extraction owners, sequencing, host/session ownership and publication prerequisites remain in force.

  4. thymikee commented on Nov 16, 2026

    @thymikee
    MemberAuthor

    Current state: the maintainer’s final delivery contract is now explicit in #3309, #3310, #3311, #3377, #3871, #3759 and #4583, with the shared contract here. Core ships help/docs/catalog metadata only; executable plugin code is installed separately by agent-device, with T3 Code/e2e management acceptance. Earlier lazy/static seam preparation is not final extraction.

    The initial read-only audit at 0adf3b48 established no additional independently obsolete bundling configuration beyond cleanup already owned by the extractions. In particular, #3311 owns the Limrun SDK export, ambient-declaration workaround and built-in removal. Other examined settings have current consumers or need more evidence; no speculative wrapper/deletion or byte-savings claim is justified. This does not establish that the 3 MB target or the newly clarified consumer-management acceptance is met.

    The current CI core inventory totals 6,098,557 B: JavaScript 4,550,807 B, declarations 362,035 B, native source/assets 1,070,584 B, other 115,131 B. This is a file-category breakdown, not plugin attribution or final release-artifact proof. The remaining work is to complete the provider extractions, attribute the retained package closures, remove proven plugin-only delivery, and demonstrate approximately 3 MB unpacked with the real package/consumer checks. Do not delete shared networking/capture code just because a plugin also uses it.

    #4583’s previously shown local .tgz install command is not supported by the current npm-name parser; its evidence instructions now require a supported managed install or an explicit management-owner extension, rather than a private-store bypass.

  5. changed the title [-]0.22: finish plugin-free core bundling and reach ~3 MB unpacked[/-] [+]0.22: ship a mobile-first, plugin-free core and measure size reduction[/+] on Nov 16, 2026
  6. thymikee commented on Nov 16, 2026

    @thymikee
    MemberAuthor

    Decision: size direction: approximately 3 MB unpacked is aspirational, not a hard release/acceptance ceiling. The maintainer expects roughly 30–50% overall reduction by moving default-plugin and non-mobile implementation out of the core delivery. Report achieved before/after bytes and percentages against a declared pre-extraction baseline using equivalent package preparation; do not treat this expectation as a measured result or weaken functionality/checks to hit a number. The zero-plugin-implementation-in-core and agent-device-managed-installation contracts remain required. Web-only runtime extraction is now explicitly in scope under #4597; its external backend/browser was already managed separately and is not a new core-byte saving. This supersedes earlier hard “reach 3 MB” wording. Existing extraction sequencing and owners remain unchanged.

  7. thymikee commented on Nov 18, 2026

    @thymikee
    MemberAuthor

    Decision: optional platforms: HarmonyOS, Vega OS, Linux desktop and macOS desktop will be separately installed, agent-device-managed plugins. Web remains tracked in #4597. Use the existing plugin lifecycle for installation, selection and typed recovery. Core keeps platform metadata and shared contracts. Each plugin delivers its platform runtime, setup and diagnostics, dedicated dependencies, and platform-only native code and assets. Existing command behavior and explicit install guidance remain available through the managed plugin.

    Finalize the platform registration and lifecycle contract in the Web work. Then split the platform moves into independently reviewable changes. Move the macOS Swift helper and its build and discovery logic with macOS. Move the AT-SPI Python helper and its discovery logic with Linux. Keep the Apple runner and snapshot bridge code that iOS and tvOS still use. Verify packaged contents, fresh managed installs, and representative platform commands. Report exact packed byte changes per layer. The current archive contains 115,366 bytes under apple/macos-helper/ and 10,645 bytes under linux/, separate from desktop JavaScript and the shared Apple assets.

    Keep Undici with Limrun through the plugin's @limrun/api dependency and with Maestro through its runScript HTTP implementation. Core provision-kit uses Agent, ProxyAgent and request to download app install sources. After Limrun extraction, measure a narrow transport build against the existing 518,980-byte Undici chunk. Require parity for address pinning, proxy and NO_PROXY, TLS identity, per-hop redirect checks, streaming limits, cancellation, and cleanup. Compare a smaller upstream entry or an existing Node transport before taking on a maintained vendor fork. Measure the complete installed footprint and verify the packed runtime.

  8. thymikee commented on Nov 18, 2026

    @thymikee
    MemberAuthor

    Current state: The BrowserStack and AWS extractions landed in #4579 and #4617. Limrun #3311 is in progress. The optional platform work is tracked separately in #4597 (web and shared registration), #4317 (HarmonyOS), #4672 (Vega OS), #4676 (Linux desktop and AT-SPI), and #4675 (macOS desktop and its Swift helper). The platform contract review precedes the individual migrations and will identify shared files that need one owner. #4674 tracks the core app downloader footprint separately. Its transport analysis can run alongside the contract review. Its final package comparison requires the Limrun SDK to be outside the core archive.

  9. thymikee commented on Nov 18, 2026

    @thymikee
    MemberAuthor

    Current state: BrowserStack and AWS source and packed live tests landed through #4579 and #4617. Their registry nightlies at 0.22.0-nightly-0-22.20261118.187 contain real plugin factories and match registry integrity. Clean managed registry-install and shared consumer acceptance remain. #3311 owns Limrun and #3759 section 4. #4597 owns the reviewed platform contract before #4317, #4672, #4676 and #4675. #4674 owns the core downloader footprint. #4583 now points at the supported manager tarball path. #4162/#4585 continue through #4539 and its follow-up. The milestone will use exact source and hosted evidence for PR readiness, separately from live, publisher and ADR prerequisites.

  10. thymikee commented on Nov 18, 2026

    @thymikee
    MemberAuthor

    Current state: the published agent-device, @agent-device/browserstack, and @agent-device/aws-device-farm nightlies at 0.22.0-nightly-0-22.20261118.187 match registry integrity. A clean consumer ran the published CLI and installed both providers by exact version through the shared manager. Both selected versions were compatible, both became active in an isolated daemon, and remove/reinstall verified the selected-versus-active distinction and restart recovery. The preinstall control returned typed missing-plugin guidance. This advances #3309 and #3310 and the registry consumer portion of #4591.

    The test found that AWS still says the plugin is private and unpublished, and the unversioned recovery commands target npm latest, which points to the empty placeholders for both providers. The follow-up aligns install guidance with the matching 0.22 nightly and leaves stable publication separate. Actual T3/e2e consumer integration and the other package and live gates remain tracked.

  11. thymikee commented on Nov 20, 2026

    @thymikee
    MemberAuthor

    Status 2026-11-20. Extractions merged into 0.22 since the last note: web (#4792, about 49 kB unpacked), Vega (#4828, about 10 kB), plugin contract v3 (#4784), optional platform families (#4781). Open PRs: Limrun (#4846, about 461 kB), HarmonyOS (#4857, about 39 kB). Not started: Linux desktop (#4676), macOS desktop (#4675), downloader footprint (#4674). Sizes are the per-PR size-report deltas against each PR's base, not a combined before/after measurement; the final measurement at named SHAs is still owed here.

    Publication prerequisites that the closed extraction issues hand to this one (npm placeholder plus trusted publisher, then drop private): @agent-device/web (#4597), @agent-device/vega (#4672), and later @agent-device/limrun, @agent-device/harmonyos. A 0.22 release whose core lacks these families must not ship before the matching package is published.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions