Skip to content

layer-10-plugins.md

Latest commit

 

History

History
388 lines (317 loc) · 20.6 KB

layer-10-plugins.md

File metadata and controls

388 lines (317 loc) · 20.6 KB

Layer 10 — eliza/plugins/* (sample-driven survey)

Files: 2,575 TypeScript across 99 plugin/app dirs under eliza/plugins/. Surveyed: 99 / 99 dirs (sample-driven; per-dir, not per-file). Refactored: 0 / 99.

CAVEAT — survey only. A true file-by-file audit at this scale is multi-week work. This pass is bulk-driven: caller-grep, package metadata, registry diff, dist/test split, top-K LOC sample. Per-dir status is at the dir level. Most dirs are flagged [?] (deferred deep audit) by design — the survey's job is to identify the dozen-ish dirs worth full audits, not to audit 99 dirs.

Walking-order note (why Layer 10 is the bulk sweep)

eliza/ is a gitignored, separate git checkout (CLAUDE.md §"Dependencies on elizaOS"). In the default packages mode Milady resolves @elizaos/* from npm — none of the source files under eliza/plugins/ are in the production bundle. They only matter when:

  1. The user runs bun run eliza:local to switch to source-priority mode.
  2. A Milady contributor patches an upstream plugin.

So Layer 10 is upstream code Milady consumes, not Milady's own code. Every deletion candidate below requires a parallel decision upstream in the elizaOS/eliza monorepo, not just here. That's why this is the bulk sweep saved for last per the AUDIT.md walking order.

Dir-level status conventions (survey-mode)

Status Meaning
[!] Surveyed; findings recorded; live and used
[x] Surveyed; clean / no concerns at the dir level
[?] Surveyed superficially; needs deep file-by-file audit
[-] Deletion candidate (no callers, museum piece, stub, or build leak)

Cross-cutting findings (apply to whole layer)

F1. Build leak: eliza/plugins/dist/ (113 files, untracked)

There is a stray dist/ directory at the plugins root containing built artifacts (index.js, init.js, models/, providers/, etc.) whose code imports @elizaos/core and @google/genai. Reads as plugin-google-genai build output that escaped one level too high. Untracked in the eliza repo. Safe to rm -rf locally; no upstream PR needed since it's not in git.

  • Path: eliza/plugins/dist/
  • File count: 113 (all generated)
  • Action: delete locally; investigate the bundler config in the affected plugin that wrote here (likely a outDir: "../dist" mistake).

F2. Empty plugin dirs (3) — abandoned scaffolds

Three dirs have only a dist/ and node_modules/, zero source, no package.json, untracked in eliza's git:

  • eliza/plugins/app-form/ — empty (compare to live eliza/plugins/plugin-form/ — 16 src files; suspicious naming overlap)
  • eliza/plugins/plugin-plugin-manager/ — empty (registry entry plugin-manager.json references @elizaos/plugin-plugin-manager from npm)
  • eliza/plugins/plugin-robot-voice/ — empty

All three are local artifacts of a prior bun install that hydrated node_modules for npm-only packages. Safe to delete locally. None should be in the source tree.

F3. 91 nested dist/ directories — local build output

Every plugin has its own dist/ after bun run build. Total disk: ~150 MB. None are tracked in git. The largest:

dist size plugin
51 MB plugin-vision
36 MB plugin-elevenlabs
25 MB plugin-n8n-workflow
15 MB app-lifeops
10 MB plugin-pdf

plugin-vision's 51 MB dist with only 40 source files suggests model binaries or tokenizers bundled into output. Worth investigating in the deep audit of that plugin.

F4. Registry-only plugins (27) — npm-resolved, source not in checkout

Plugins listed under eliza/packages/app-core/src/registry/entries/plugins/ but without a sibling dir in eliza/plugins/:

auto-trader, blooio, browser, clipboard, code, copilot-proxy, directives,
eliza-classic, evm (→ plugin-wallet), experience, gmail-watch, goals,
hedera, localdb, memory, moltbook, n8n, prose, rss, s3-storage, scheduling,
trajectory-logger, trust, tts, twilio, vercel-ai-gateway, webhooks

These are not dead — they're upstream npm packages Milady consumes via @elizaos/plugin-<name>. They live in separate elizaos-plugins/* repos. Out of Layer 10 scope; flagged here only to clarify the registry/dir gap.

F5. Dir-only plugins (34) — connector / provider plugins not in registry

Plugins with a checkout dir but no <name>.json registry entry. These are the chat / messaging connectors and a few utility plugins:

action-bench, bluebubbles, bluesky, browser-bridge, calendly,
claude-code-workbench, discord, executecode, farcaster, feishu, google-chat,
google-meet-cute, imessage, instagram, line, matrix, nostr, nvidiacloud,
robot-voice, shopify, signal, slack, sql, streaming, suno, telegram,
twitch, vertex, wallet, web-search, wechat, whatsapp, x, xmtp

Connectors (discord, telegram, x, slack, etc.) are wired through the channel-keyed auto-enable map at eliza/packages/agent/src/config/plugin-auto-enable.ts, not the dashboard registry. That's correct — the registry is for user-toggleable plugins; connectors enable per-onboarding.

The handful of utility plugins in this list (action-bench, executecode, browser-bridge, etc.) do have callers and load through other paths. Not dead.

F6. eliza/cloud overlap (informational)

eliza/cloud/ has parallel implementations for the connector-style plugins (discord, telegram, slack, x, mcp, wallet) under cloud/services/gateway-*, cloud/apps/api/v1/*, and cloud/packages/lib/services/*. This is by design:

  • eliza/plugins/plugin-discord — runtime plugin loaded into a local Eliza agent (DM/channel handling).
  • eliza/cloud/services/gateway-discord — cloud-hosted webhook gateway that bridges Discord events into hosted agents.

Different deployment surfaces, different code, same external API. Not collisions, not dedup candidates. Flag noted to prevent a future agent from "consolidating" them.

F7. app-lifeops test-fixture mass — not runtime code

app-lifeops total file count is 400, but 84 of those are test files under test/scenarios/ and scenarios/. The actual runtime source is ~316 files / ~200 K LOC — still the largest single plugin/app by an order of magnitude. Treat as a deep-audit candidate in its own right.

F8. No v1/v2/legacy subdir splits

find -name v1 -o -name v2 -o -name legacy -o -name old across Layer 10 returns nothing. Versioning happens via version in package.json (2.0.0-alpha.536 is the modal version), not via in-tree branches. Good.

F9. No *.generated.ts files

Zero generated source files. All .d.ts and .js build output is properly contained under dist/. Good.

Sample LOC table (top 25 by source LOC)

Computed wc -l on *.ts/*.tsx, excluding node_modules, dist, .d.ts, .test.*. Calls "loc" loosely — includes blank/comment lines. For top 10, the test split is also shown.

Rank Plugin Source files Test files LOC Cross-monorepo callers
1 app-lifeops 363 84 200,826 50
2 plugin-wallet 187 20 49,236 21
3 plugin-agent-orchestrator 79 3 35,994 38
4 plugin-discord 68 9 28,797 64
5 plugin-sql 85 69 19,422 160
6 app-steward 56 5 17,795 30
7 plugin-x 53 2 13,924 54
8 plugin-n8n-workflow 60 25 13,797 25
9 app-companion 60 1 11,187 43
10 app-2004scape 58 1 8,709 16

Notable: app-lifeops is 5× the next-largest. plugin-sql is the canonical bridge into the runtime (160 callers, mostly internal eliza test suites). app-2004scape has 8.7 K LOC for a museum-piece game demo — see deletion candidates below.

Deletion candidates (top 10)

Hard rule: all of these are decisions for the upstream elizaOS/eliza repo, not Milady. Milady's choice is whether to stop tracking them in the registry / docs / scaffolds. Local deletion of eliza/plugins/<name>/ is fine in either mode.

Confirmed-dead (zero monorepo callers + non-essential)

  1. plugin-action-bench — 0 callers, 28 src files, "action benchmarking" harness. Not in registry, not in auto-enable, not imported. Likely an internal eliza eval tool that doesn't belong in the plugin tree.
  2. plugin-calendly — 0 callers, 9 src files. No registry entry, no auto-enable. Likely a hackathon/half-built integration.
  3. plugin-google-meet-cute — 0 callers, 16 src files. Joke name, no wiring. Pure dead weight.
  4. plugin-nvidiacloud — 0 callers, 11 src files. Not in registry, not in auto-enable. Superseded by plugin-openrouter / plugin-vertex.
  5. plugin-vertex — 0 callers, 13 src files. Same story; not in registry. Superseded by plugin-google-genai.
  6. plugin-web-search — 0 callers, 5 src files. Package name uses old scope @elizaos-plugins/* (not @elizaos/*) — clearly an artifact from pre-migration era.
  7. plugin-xmtp — 0 callers, 5 src files. XMTP messaging connector; neither in registry nor in any auto-enable map.

Build/scaffold leaks (not source code at all)

  1. eliza/plugins/dist/ — F1 above. 113 generated files. Delete locally.
  2. eliza/plugins/app-form/ — F2 above. Empty hydration shell.
  3. eliza/plugins/plugin-plugin-manager/ + plugin-robot-voice/ — F2 above. Empty hydration shells.

Museum-piece "demo apps" (need product call before deleting)

app-2004scape, app-defense-of-the-agents, app-clawville, app-hyperscape, app-babylon, app-screenshare — these look like elizaOS demo / game / hackathon apps. They have low caller counts (7-18, mostly each other or scaffolds) and high LOC for what they do. Defer to product: are these shipped, are they marketing demos, are they being used to validate the orchestrator?

If the answer is "no" for any of them, they're 50-100 K LOC of dead weight each with a dist/ to match.

Deep-audit candidates (top 10)

Live, complex, high caller-count, and exposed to Milady's runtime contract. These warrant a real file-by-file audit before any further refactor of the layers above touches them.

  1. plugin-sql (85 src / 19 K LOC / 160 callers) — the SQL/Drizzle bridge. Used by every test suite and many runtime paths. Schema changes here cascade everywhere.
  2. plugin-anthropic (21 src / 66 callers) — primary LLM provider for Milady (Opus 4.7 default per CLAUDE.md). Any change to model defaults, prompt-cache headers, or provider switch lives here.
  3. plugin-discord (68 src / 28 K LOC / 64 callers) — largest connector; the cloud/local-runtime overlap (F6) means contract changes need both surfaces to agree.
  4. plugin-x (53 src / 13 K LOC / 54 callers) — second-largest social connector; high churn historically.
  5. app-lifeops (363 src / 200 K LOC / 50 callers) — the scaffolded "personal OS" app; its sheer size means it could hide architectural violations that bleed into shared infra.
  6. app-companion (60 src / 11 K LOC / 43 callers) — the desktop companion app entrypoint.
  7. plugin-agent-orchestrator (79 src / 36 K LOC / 38 callers) — spawns Codex/Claude/OpenCode sub-agents (per CLAUDE.md). Critical path for the coding-agent product surface.
  8. plugin-elizacloud (49 src / 45 callers) — the Cloud SDK integration. Auth, billing, app-domain, monetization touch this.
  9. plugin-telegram (15 src / 48 callers) — high callers for a small plugin suggests heavy boundary surface; quick win to deep-audit.
  10. plugin-openai (25 src / 114 callers) — second LLM provider by default. Mirror of plugin-anthropic concerns.

Per-dir status table

Format: [status] dir-name src-files | LOC | callers | one-liner

(For dirs marked [?] the survey did NOT inspect source files. Bulk metadata only. Promote to [!] after a deep audit.)

Apps (app-*, 25 dirs)

  • [?] app-2004scape 58 | 8709 | 16 | Old-school RuneScape clone demo. Museum-piece candidate.
  • [?] app-babylon 7 | – | 18 | Babylon.js sample app. Demo.
  • [?] app-browser 9 | – | 7 | In-app browser surface. Low callers.
  • [?] app-clawville 6 | – | 14 | Game demo (@clawville/app-clawville). Museum.
  • [!] app-companion 60 | 11187 | 43 | Desktop companion app. Deep-audit.
  • [?] app-contacts 8 | – | 12 | Contacts surface.
  • [?] app-defense-of-the-agents 6 | – | 16 | Game demo. Museum candidate.
  • [?] app-elizamaker 11 | – | 20 | Eliza-maker scaffold UI.
  • [-] app-form 0 | 0 | – | Empty hydration shell. Delete.
  • [?] app-hyperliquid 11 | – | 14 | Hyperliquid trading UI.
  • [?] app-hyperscape 5 | – | 18 | 3D scape demo.
  • [?] app-knowledge 6 | – | 11 | Knowledge-base app.
  • [!] app-lifeops 363 | 200826 | 50 | Personal-OS app. Deep-audit (largest).
  • [?] app-phone 24 | – | 15 | Phone surface.
  • [?] app-polymarket 15 | – | 16 | Polymarket UI.
  • [?] app-scape 43 | – | 14 | Scape app variant.
  • [?] app-screenshare 5 | – | 9 | Screen share. Museum candidate.
  • [?] app-shopify 13 | – | 25 | Shopify storefront app.
  • [!] app-steward 56 | 17795 | 30 | Property steward. Deep-audit.
  • [?] app-task-coordinator 25 | – | 19 | Task coordination.
  • [?] app-training 52 | – | 34 | Training UI; pairs with Layer 8 trajectory persistence.
  • [?] app-vincent 16 | – | 25 | Vincent surface.
  • [?] app-wallet 17 | – | 20 | Wallet UI app.
  • [?] app-wifi 8 | – | 13 | WiFi surface.
  • [?] app-workflow-builder 3 | – | 7 | Workflow builder stub.

Plugins (plugin-*, 73 dirs)

  • [-] plugin-action-bench 28 | – | 0 | 0 callers. Delete candidate.
  • [!] plugin-agent-orchestrator 79 | 35994 | 38 | Sub-agent spawner. Deep-audit.
  • [!] plugin-agent-skills 28 | – | 36 | USE_SKILL mechanism (per CLAUDE.md).
  • [!] plugin-anthropic 21 | – | 66 | Default LLM provider. Deep-audit.
  • [!] plugin-app-control 26 | – | 6 | APP create/load. Auditable today (used by orchestrator).
  • [?] plugin-bluebubbles 16 | – | 11 | iMessage bridge.
  • [?] plugin-bluesky 17 | – | 16 | Bluesky connector.
  • [!] plugin-browser-bridge 8 | – | 32 | High caller count for 8 files.
  • [-] plugin-calendly 9 | – | 0 | 0 callers. Delete candidate.
  • [?] plugin-claude-code-workbench 10 | – | 16 | Claude Code workbench surface.
  • [?] plugin-cli 6 | – | 17 | CLI plugin.
  • [?] plugin-commands 12 | – | 19 | Commands plugin.
  • [?] plugin-computeruse 35 | – | 24 | Computer-use plugin.
  • [!] plugin-discord 68 | 28797 | 64 | Largest connector. Deep-audit.
  • [?] plugin-edge-tts 9 | – | 24 | Edge TTS.
  • [?] plugin-elevenlabs 4 | – | 22 | ElevenLabs TTS. 51 MB dist — investigate.
  • [!] plugin-elizacloud 49 | – | 45 | Cloud SDK. Deep-audit.
  • [?] plugin-executecode 5 | – | 2 | EXECUTE_CODE action.
  • [?] plugin-farcaster 35 | – | 20 | Farcaster connector.
  • [?] plugin-feishu 15 | – | 17 | Feishu connector.
  • [?] plugin-form 16 | – | 11 | Form runtime (compare to dead app-form/).
  • [?] plugin-github 14 | – | 15 | GitHub plugin.
  • [?] plugin-google-chat 11 | – | 14 | Google Chat connector.
  • [!] plugin-google-genai 17 | – | 45 | Gemini provider.
  • [-] plugin-google-meet-cute 16 | – | 0 | 0 callers. Delete candidate.
  • [!] plugin-groq 7 | – | 47 | Groq provider.
  • [?] plugin-imessage 16 | – | 22 | iMessage plugin.
  • [?] plugin-inmemorydb 9 | – | 10 | In-memory DB; check vs plugin-sql.
  • [?] plugin-instagram 15 | – | 13 | Instagram connector.
  • [?] plugin-line 13 | – | 15 | LINE connector.
  • [?] plugin-linear 29 | – | 3 | Linear plugin. Low callers.
  • [?] plugin-local-ai 15 | – | 10 | Local AI provider.
  • [!] plugin-local-embedding 9 | – | 47 | Embedding provider. High caller count.
  • [?] plugin-local-storage 6 | – | 1 | Local storage. Low callers.
  • [?] plugin-matrix 9 | – | 17 | Matrix connector.
  • [!] plugin-mcp 34 | – | 18 | MCP server plugin. Cloud overlap.
  • [?] plugin-minecraft 20 | – | 3 | Minecraft plugin.
  • [?] plugin-music-library 47 | – | 9 | Music library.
  • [?] plugin-music-player 39 | – | 8 | Music player.
  • [?] plugin-mysticism 32 | – | 3 | Mysticism plugin (low callers).
  • [!] plugin-n8n-workflow 60 | 13797 | 25 | n8n workflow integration.
  • [?] plugin-nostr 11 | – | 17 | Nostr connector.
  • [-] plugin-nvidiacloud 11 | – | 0 | 0 callers. Delete candidate.
  • [!] plugin-ollama 16 | – | 35 | Ollama local provider.
  • [!] plugin-openai 25 | – | 114 | Second-most-cited plugin. Deep-audit.
  • [!] plugin-openrouter 20 | – | 40 | OpenRouter provider.
  • [?] plugin-pdf 10 | – | 26 | PDF plugin. 10 MB dist — fonts? investigate.
  • [-] plugin-plugin-manager 0 | 0 | – | Empty hydration shell. Delete.
  • [?] plugin-rlm 8 | – | 3 | RLM plugin. Low callers.
  • [?] plugin-roblox 16 | – | 7 | Roblox plugin.
  • [-] plugin-robot-voice 0 | 0 | – | Empty hydration shell. Delete.
  • [!] plugin-shell 28 | – | 23 | Shell action.
  • [?] plugin-shopify 14 | – | 7 | Shopify plugin.
  • [!] plugin-signal 16 | – | 29 | Signal connector.
  • [?] plugin-simple-voice 11 | – | 6 | Voice plugin.
  • [!] plugin-slack 22 | – | 21 | Slack connector.
  • [?] plugin-social-alpha 43 | – | 3 | Social-alpha plugin (low callers for 43 files).
  • [!] plugin-sql 85 | 19422 | 160 | Drizzle/SQL bridge. Deep-audit (highest callers).
  • [?] plugin-streaming 4 | – | 20 | Streaming plugin.
  • [?] plugin-suno 6 | – | 6 | Suno music plugin.
  • [?] plugin-tailscale 12 | – | 1 | Tailscale plugin (1 caller).
  • [?] plugin-tee 18 | – | 3 | TEE plugin.
  • [!] plugin-telegram 15 | – | 48 | Telegram connector. Deep-audit (high callers / small).
  • [?] plugin-twitch 11 | – | 16 | Twitch connector.
  • [-] plugin-vertex 13 | – | 0 | 0 callers. Delete candidate.
  • [?] plugin-video 5 | – | 7 | Video plugin.
  • [?] plugin-vision 40 | – | 17 | Vision plugin. 51 MB dist (largest) — investigate.
  • [!] plugin-wallet 187 | 49236 | 21 | Wallet plugin (#2 by LOC).
  • [-] plugin-web-search 5 | – | 0 | 0 callers, old @elizaos-plugins/* scope. Delete.
  • [?] plugin-wechat 10 | – | 18 | WeChat connector.
  • [?] plugin-whatsapp 32 | – | 24 | WhatsApp connector.
  • [!] plugin-x 53 | 13924 | 54 | X (Twitter) connector. Deep-audit.
  • [!] plugin-xai 7 | – | 33 | xAI provider.
  • [-] plugin-xmtp 5 | – | 0 | 0 callers. Delete candidate.

Top-level

  • [-] dist/ 113 | – | – | Build leak. Delete locally.

Promotion criteria

A [?] dir promotes to [!] (deep-audited) only after:

  1. Reading every source file in the dir.
  2. Mapping its inbound and outbound dependency edges to other layers.
  3. Applying the eight AGENTS.md axes per file.
  4. Recording per-file findings in this document (would expand it 50-100x).

A [?] dir promotes to [-] (deletion) only after:

  1. Verifying the cross-monorepo grep is correct (no dynamic imports under enabled_plugins config or registry strings missed by package-name grep).
  2. Confirming no scaffolding template (in eliza/templates/) or test fixture references it.
  3. Getting upstream sign-off (the file lives in elizaOS/eliza, not Milady).

What this enables

Phase 3 (Electrobun decomposition) and Phase 4 (chat fallback honesty) do not touch this layer — they're confined to Layers 1–9. Layer 10 is inventory + cleanup-debt tracking. The 7 hard deletion candidates and 4 build leaks are safe Milady-side cleanup independent of upstream coordination.