Aller au contenu principal

Task Breakdown: Agent Plugins Standard Interop

Feature ID: agent-plugins Plan: ./plan.md Status: Draft Last updated: 2026-08-09

How to use​

Phases map to PRs (one or two PRs per phase, stacked on develop). Tasks sequential unless (parallel). Every phase ends with the same standing DoD gate:

DoD gate (every phase): fresh full pnpm build green (catches TS2305 cross-package export misses); pnpm test green in touched packages; node dist/openapi/generate-openapi.js full-app bootstrap green (DI crash-loop gate) whenever a module/provider changed; migration included in the SAME PR as any entity change, entity added to packages/agent/src/database/_entities-inventory.ts; module-shape pin specs (apps/api/src/agents/agents.module.spec.ts, apps/api/src/trigger/trigger-internal.module.spec.ts) updated with any provider add; pnpm format:check with the ROOT glob; grep apps/web/e2e for every touched endpoint/DTO; branch e2e dispatched and green before merge; PR reviewed to a clean bot pass (NN #14/#18).

Nothing-changes invariant (every phase): git diff must show zero behavioral edits to: plugin-loader.service.ts, plugin-manifest-validator.service.ts, plugin-installer.service.ts (native paths), everworks-skills plugin, Skill/ SkillBinding entities' existing columns, existing skills routes/DTOs, apps/mcp, .agents/.claude/.cursor skill dirs, skills-lock.json. Additive-only files (capability interface additions, _entities-inventory.ts appends, module provider appends, locale key appends) are expected and fine.

Phase 0 — Conformance library (PR-1)​

  • T1. Scaffold packages/agent-plugins (@ever-works/agent-plugins): tsup ESM build, Vitest, MIT, zero NestJS deps. Own deps: ajv@^8, gray-matter@^4, yaml if needed.
  • T2. Vendor canonical schemas under src/schemas/1.0.0/ (plugin.schema.json, mcp.schema.json) imported via resolveJsonModule so they ship inside dist/.
  • T3. src/manifest.ts: closed manifest validator with exact AP-4 severity split. Unit tests per AP-1…AP-6 (incl. MUST-NOT-reject cases: non-semver version, weird URLs, 1-char and dotted names).
  • T4 (parallel). src/skills.ts: discovery walker (immediate children, regular-file check) + Agent Skills frontmatter validator (name=dir, charset, lengths, allowed-tools string→array tokenizer, metadata string-map check). Tests per AP-7…AP-9.
  • T5 (parallel). src/mcp.ts: closed mcp.json validator, closed server union, URL/header rules, reserved env keys, version-match with plugin.json. Tests per AP-11…AP-15.
  • T6 (parallel). src/paths.ts containment (realpath symlink resolution) + src/expand.ts placeholder expansion. Tests per AP-10, AP-16…AP-17 (expansion in args/env-values/cwd only; non-recursive; unknown placeholders literal).
  • T7. fixtures/: full conformance corpus (valid minimal, valid full, every fatal/non-fatal manifest case, skip-one-skill, disable-MCP-only, per-server skip, containment escapes, dir-name mismatch, oversized description). Add fixtures path to .prettierignore.
  • T8. src/serialize.ts: plugin.json + SKILL.md emitters; slug→name guard; round-trip property test (emit → parse → deep-equal).
  • T9. Optional CI oracle job: skills-ref validate (Python) over skill fixtures, non-blocking.

Phase 1 — Package registry + local-dir source + skills read-side (PR-2, PR-3)​

  • T10. Entities agent_plugin_packages (Tier A, incl. dataManifest) + agent_plugin_package_allowlist (Tier D) per plan §2.5; append to _entities-inventory.ts; migration CreateAgentPluginPackages. Read the generated SQL before commit (NN #16). Account-transfer whitelist entries for packages land HERE, same PR (types.ts + export literal + both import paths + SyncPushOptions) — reference-only export (name+version+source), never package content; deferring this to Phase 3 would reopen the silent-no-round-trip bug class.
  • T11. packages/agent/src/agent-plugins/ module: AgentPluginPackageService (register/validate/list/remove; findings persisted per package), local-dir acquirer (AGENT_PLUGINS_DIR scan, optional-with-default env per constants.ts pattern). Barrel export entry in packages/agent/package.json export map.
  • T12. Feature flag FEATURE_AGENT_PLUGINS default false in apps/api/src/config/constants.ts + .env.example + compose env files.
  • T13. AgentPluginCatalogService (plan §2.2) in packages/agent/src/agent-plugins/: entries from registered package rows + conformance lib; version synthesis per AP-21; pre-flight findings by calling the existing gates directly (MAX_BODY_BYTES/assertNoSecrets/ assertNoInjectionTokens — reachable because this is platform code, not a plugin). Wire into SkillsFacadeService.listEntries as an @Optional()-injected additive source merged LAST (existing behavior bit-identical when absent/flag-off). Add optional provenance fields (packageName?/packageVersion?/sourceKind?) to SkillCatalogEntry (additive interface change). Barrel-export the new module in packages/agent/package.json export map.
  • T14. API controller apps/api/src/agent-plugins/ (list/install local/uninstall/findings), throttled 5/min, admin allowlist CRUD parallel to admin/plugins/allowlist. DTO validation + validation-authz-matrix e2e specs for every new route.
  • T15. CLI verbs (agent-plugins list/install/status/uninstall) following dynamic.command.ts pattern.
  • T16. Integration test (Jest, in packages/agent — the facade + services live there; conformance-lib unit tests stay Vitest in packages/agent-plugins): fixture package on disk → register → catalog entries appear via SkillsFacadeService → installFromCatalog → identical Skill rows; assert existing catalog e2e pins unaffected (flag off by default; flow-skills-catalog-pagination.spec.ts untouched).

Phase 2 — git + npm sources + updates (PR-4)​

  • T17. Git acquirer: URL-first shallow clone (isomorphic-git primitives per plan §2.3), ref-pin + resolved-SHA recording, containment validation post-clone, re-fetch/update path.
  • T18. npm acquirer: pacote manifest→allowlist→extract (no symlink, no scripts, no dep install); registry config reuse; 409/424/502/504 error mapping parity with the code-plugin installer.
  • T19. Boot re-materialization (warmupFromDb pattern) for git/npm packages; startupProbe budget check.
  • T20. Update-available end-to-end: catalog service computes updates for git/npm packages; facade surfaces a badge + explicit re-sync action; the same facade wiring finally calls the existing provider checkForUpdates seam (zero non-test callers today) so everworks-skills updates surface too.
  • T21. Web UI: Sources + Packages pages under settings; install flow with per-component findings display; i18n keys in ALL apps/web/messages/*.json; PostHog events.
  • T22. Env wiring per 2026-06-12 checklist (k8s manifests dev/stage/ prod + deploy workflows + compose; defaults make it non-fatal) + explicit operator note re ArgoCD-managed live env (outside this repo).

Phase 3 — MCP client: remote transports + bindings (PR-5, PR-6)​

  • T23. McpServerConfigService (plan §2.2) in packages/agent/src/agent-plugins/: resolves bound servers → validated configs + provenance via the conformance lib. No plugin capability, no facade-capabilities.ts append. Barrel-export from packages/agent/package.json export map (TS2305 bug class).
  • T24. Entity agent_mcp_server_bindings (Tier C — skill_bindings template PLUS net-new columns serverName/enabled/settings/ secretSettings/credentialsSecretRef, per plan §2.5) + migration + inventory append. Account-transfer whitelist entries for bindings (same 4-place set as T10) — masked secrets, credentialsSecretRef pointers deliberately omitted — implementing the ordered package-reference import contract (plan §2.5): resolve/remap destination packageId FIRST; absent package → pending-reference row + bindings imported enabled: false; disallowed source → skip-and-report; never auto-fetch content on import; never fail the whole import.
  • T25. packages/agent/src/mcp/: McpClientService (streamable-http + sse via @modelcontextprotocol/sdk — new dep of packages/agent), per-server failure isolation, connect-time credential injection (client-generated headers; masked reads), SSRF egress policy, redirect policy per AP-15 (no package-header cross-origin forwarding without explicit authorization; no client-generated credential forwarding cross-origin at all — with tests), run-end disconnect.
  • T26. McpToolSource injected into AgentToolService. resolveAllowedTools as a new optional source (domain-tool-source pattern); mcp__<server>__<tool> naming; name/description sanitization; builtin-collision drop + WARN; run-log WARNs for skipped servers. Update module pin specs.
  • T27. Bindings API (/api/agents/:id/mcp-servers CRUD + standalone delete; agent-scoped targets only in v1) + MCP Servers tab on /agents/[id] + i18n + e2e specs (binding CRUD, authz matrix, cross-tenant isolation). Domain chat tools per the in-code DoD (agent-tool.service.ts:317-319, "every new entity"): packages + bindings get tool sources here; agent_plugin_package_allowlist is exempt (admin-only surface, like plugin_allowlist which has no chat tools) and skill_files is exempt (child rows surfaced through existing skill tools) — exemptions stated so the DoD check is explicit.
  • T28. Decision + implementation: MCP tool invocations → PluginUsageEvent usage accounting (recommended: on).

Phase 4 — stdio transport + PLUGIN_DATA (PR-7)​

  • T29. AGENT_PLUGINS_DATA_DIR (default /app/agent-plugins-data); per-(tenant, package) data dirs; create-before-launch; delete on uninstall; SaaS write-through sync via boot-selected IStoragePlugin with DB-side key manifest.
  • T30. Stdio launcher: from-scratch env (buildSubprocessEnv pattern) + expanded package env overlay + PLUGIN_ROOT/PLUGIN_DATA last; single- token command resolution (bare vs ./-relative) with containment; cwd rules per AP-13; process supervision + teardown at run end.
  • T31. Triple gate: feature flag + AGENT_PLUGINS_STDIO deployment setting (default off; v1 SaaS keeps it off — no sandbox route is built in this feature) + per-binding explicit enable; disabled stdio surfaces as "present, disabled by policy" (AP-19).
  • T31b. Phase-4 env wiring repeat of the 2026-06-12 checklist for AGENT_PLUGINS_DATA_DIR + AGENT_PLUGINS_STDIO: k8s manifests (dev/stage/prod) + deploy workflow env blocks + .env.example + compose files + explicit operator note for the ArgoCD-managed live env (outside this repo). Defaults keep boot non-fatal; the wiring is still required for the values to be settable.
  • T32. Full-spec conformance run: fixture packages exercising every §10.1 checklist row against the live loader; publish the checklist result as a doc page.

Phase 5 — Sidecars + export (PR-8, PR-9)​

  • T33. skill_files entity + migration + storage via IStoragePlugin (uploads-stack validation: sha256 naming, MIME sniff, per-file + per-skill size caps). NOTE: uploads stack does NOT secret-scan — wire assertNoSecrets (packages/agent/src/utils/secret-scan.ts) as a NEW check on text-like files only. Package service ingests sidecars; findings for oversized/rejected files.
  • T34. Materialization into claude-code workspace staging (claude-code.plugin.ts:542-558 seam): .claude/skills/<name>/… + .mcp.json next to seedMetadata writes — with the execution-gate split (plan §4.5/§5, Greptile P1 on PR #2000): references/+assets/ only by default (executable bits stripped); scripts/ only for packages passing the stdio-grade triple gate; .mcp.json only for bind-enabled servers. Tests assert a non-gated package's scripts NEVER reach the workspace. Explicitly does NOT touch the Wave-2 Task-workspace path (workspaceCwd has no consumer — see plan §4.5).
  • T35. Export API + CLI + UI: scope-selected skills → conformant package (zip); round-trip gate (AP-22/23); slug guard enforcing the FULL spec name rule (≤64, no --, no leading/trailing hyphen — the DTO accepts all three) with rename prompt.
  • T36. ever-works-mcp package descriptor export (streamable-http https://mcp.ever.works/mcp, credential-free; docs updated by addition — pointer from docs/features/mcp-server.md, no removal).

Phase 6 — Hardening + docs + conformance claim (PR-10)​

  • T37. User-facing docs page docs/features/agent-plugins.md + sidebar entry in apps/docs/sidebarsPlatform.ts; internal docs stay off-nav.
  • T38. Desktop staging (prepare-bundle.js + <userData>/agent-plugins default source) + headless-node XDG dir.
  • T39. apps/mcp whitelist decision: expose agent-plugins management tools or document out-of-scope.
  • T40. Security review pass (stdio gate, SSRF policy, containment fuzzing over fixture escapes); Sentry tags; rate-limit re-check.
  • T41. Conformance statement doc: map every §10.1 row + AP-1…AP-23 to test evidence; announce "Agent Plugins v1.0.0 compatible (client: skills + MCP; producer: skills packages, plus the Ever Works MCP-server package descriptor)" — the single canonical claim wording, used verbatim in spec §1.3, the ADR, and marketing. MUST document client-side policy refusals explicitly (64KB body cap, secret/injection gates, stdio-off deployments, SSRF egress denies): spec-valid packages a deployment may decline to load, stated as policy, not as conformance gaps.

Effort estimate (engineer-days, honest)​

PhaseDays
0 — conformance lib4–5
1 — registry + local + skills5–6
2 — git/npm + updates + UI5–6
3 — MCP remote + bindings6–8
4 — stdio + PLUGIN_DATA5–7
5 — sidecars + export5–6
6 — hardening + docs3–4
Total33–42 (plus review/CI wall-clock)

Skills-only value ships at end of Phase 2 (~14–17 d); the full-standard claim lands with Phase 4; export completes the producer story in Phase 5.