Files
nexu-io--open-design/specs/current/plugin-driven-flow-plan.md
T
wehub-resource-sync 070959e133
landing-page-staging / Deploy landing page to staging (push) Has been skipped
landing-page-ci / Validate landing page (push) Failing after 4s
visual-baseline / Capture visual baselines (push) Has been cancelled
bake-plugin-previews / Bake plugin previews (push) Has been cancelled
chore: import upstream snapshot with attribution
2026-07-13 12:00:47 +08:00

12 KiB
Raw Blame History

Plugin-driven flow plan

Parent: spec.md · Siblings: architecture-boundaries.md · maintainability-roadmap.md · run.md

Purpose

Make plugins 1:1 drive a complete run, give bare query input a deterministic fallback, and align Home entry with the NewProject taxonomy plus migration shortcuts (Figma / folder / repo). The current shape lets users pick a plugin in the rail but the actual run still falls back to a generic agent loop because (a) plugin-local SKILL.md is never injected, (b) Home direct-query never pins a plugin, and (c) the daemon stage runner is still a stub.

Goals / non-goals

  • Goals
    • Every plugin reaches the agent prompt with its own SKILL.md body and pipeline contract intact, so the same Use plugin action yields the same artifact across runs.
    • "Naked" Home query input gets auto-bound to a default scenario plugin keyed off the project kind so there is no zero-plugin path.
    • Home input card surfaces an intent rail (Prototype / Slide deck / Image / Video / Audio / From Figma / From folder / Template / Other) modelled on the Lovart entry pattern.
  • Non-goals
    • Marketplace / third-party plugin distribution changes.
    • Transport changes between web and daemon.
    • Replacing the agent CLI loop. We keep the LLM as the primary worker; we only formalise stage orchestration around it.

Current gaps (referenced to code)

# Where Symptom
G1 apps/daemon/src/server.ts#firePipelineForRun runStage stub returns synthetic critique.score: 4; pipeline events are emitted but no real per-atom work runs.
G2 apps/daemon/src/plugins/apply.ts#pickFirstSkillId ./SKILL.md is treated as a global skill id and never matches the registry, so a plugin's own SKILL.md is silently dropped from composeSystemPrompt.
G3 apps/web/src/components/HomeView.tsx#submit When no plugin is active, pluginId is null; resolvePluginSnapshot then returns null and the run goes through the legacy non-plugin code path.
G4 apps/web/src/components/EntryShell.tsx#DEFAULT_SCENARIO_PLUGIN_BY_KIND Every kind maps to od-new-generation, so image/video/audio and migration kinds never reach their bundled scenarios (od-figma-migration, od-code-migration).
G5 apps/web/src/components/HomeHero.tsx No category / model chips below the input card; first-touch users have to know which plugin they want.

Target architecture

plugin manifest (od.pipeline.stages[])
  └─ daemon stage runner          (real workers; replaces stub in G1)
       └─ atom executor registry  (per-atom handler)
            └─ agent CLI process  (LLM + tool calls, prompt carries plugin skill + active stage)
  • The agent stays the primary "worker"; the new layer is the daemon-side enforcement of stage order, atom-prompt fragments, and exit signals.
  • Plugin-local SKILL.md flows into the same ## Active skill slot as global skills, just sourced from plugin.fsPath/<path> instead of SKILLS_DIR.
  • The default-scenario binding moves to a single resolver shared by /api/projects, /api/runs, and EntryShell so the client and daemon never disagree.

Kind → default scenario plugin

metadata.kind / taskKind Bundled scenario
prototype, other, template od-new-generation
deck od-new-generation (until a deck-specialised scenario lands)
image, video, audio od-media-generation (Stage C)
figma-migration taskKind od-figma-migration
code-migration taskKind od-code-migration
tune-collab taskKind od-tune-collab

Stages

Stage Status Notes
A shipped Plugin-local SKILL.md reaches ## Active skill; Home query auto-binds default scenario per kind.
B shipped (MVP) Chip rail mirrors NewProject taxonomy + adds Figma / folder / template shortcuts. Secondary chip rows (model picker for image, inline figmaUrl input) deferred.
C shipped (MVP) Bundled od-media-generation scenario for image/video/audio; uses existing media-image / media-video / media-audio atoms rather than a new wrapper atom.
D shipped (MVP) apps/daemon/src/plugins/atoms/registry.ts + built-ins.ts introduce an atom worker registry; firePipelineForRun now drives stages through runStageWithRegistry. Set OD_PIPELINE_RUNNER=stub to fall back to the v1 canned-signal runner. Real workers only ship for critique-theater today (reads run_devloop_iterations.critique_summary for the latest parseable score); every other FIRST_PARTY_ATOM registers a permissive worker so happy-path convergence matches v1.
E shipped (MVP) pnpm guard + pnpm typecheck green; daemon plugin suites (atom-registry, pipeline-runner, local-skill, apply, bundled-scenarios, headless-run, bundled, pipeline, simulate) all green; web 720/720 green; full daemon suite 1847/1852 with the remaining 5 failures all reproduced on HEAD without these changes (3× finalize-design macOS-tmpdir symlink issue; 2× chat-route/origin-validation order-dependent flakes that pass in isolation). Cross-app browser e2e covering the bare-query / chip-click flows is deferred to a follow-up since the existing Playwright harness only ships packaged-app smoke.

Stage A — Plugin actually injects, Home never runs naked

Smallest change with the largest stability win.

# File Change
A1 apps/daemon/src/plugins/apply.ts Replace pickFirstSkillId with pickFirstSkillBinding returning `{ kind: 'global', id }
A2 apps/daemon/src/plugins/snapshots.ts + contracts AppliedPluginSnapshot Add optional pluginSkillBody. Persist alongside other JSON fields.
A3 apps/daemon/src/server.ts#composeDaemonSystemPrompt When snapshot carries pluginSkillBody, prefer it over the global skill body for the ## Active skill block.
A4 apps/web/src/components/EntryShell.tsx Split DEFAULT_SCENARIO_PLUGIN_BY_KIND to map per kind; add migration kinds.
A5 apps/daemon/src/server.ts (/api/projects, /api/runs) When the body carries no pluginId/appliedPluginSnapshotId and no project pin exists, look up the bundled scenario for taskKind and apply it.
A6 apps/daemon/tests New tests cover local skill injection + default-by-kind binding.

Exit criteria

  • Picking a plugin → run logs show ## Active skill — <plugin> block with the plugin's SKILL.md contents.
  • Submitting a Home query with no plugin → snapshot row exists and pluginId matches the kind mapping.
  • Existing plugin / non-plugin tests still pass.

Stage B — Home intent rail (Lovart-style)

  • HomeHero adds a home-hero__rail row.
  • Primary chips: Prototype · Slide deck · Image · Video · Audio · From template · From Figma · From folder · Other.
  • Secondary row appears after primary selection where it makes sense (e.g. Image surfaces a model chip: GPT Image 2 / Nano Banana Pro / Seedance 2.0).
  • Selecting a chip pre-applies the matching scenario plugin via applyPlugin() so Enter behaves the same as the explicit Use plugin path.

Stage C — Media + migration scenario fill-in

  • Add bundled scenario od-media-generation. The pipeline reuses the already-shipped media-image / media-video / media-audio atoms; no dedicated media-generate wrapper is needed and the original plan's mention of a separate atom is superseded by this note.
  • The scenario shares taskKind: 'new-generation' with od-new-generation. The daemon's collectBundledScenarios dedupes by taskKind, preferring the canonical od-<taskKind> id so the pipeline-fallback stays deterministic.
  • Surface "From Figma" / "From folder" chips on the Home rail.
    • "From Figma" applies the od-figma-migration plugin (which carries the figmaUrl input). A dedicated inline figmaUrl field is deferred to a follow-up; the chip's prompt-template substitution still surfaces {{figmaUrl}} so the user can edit before submit.
    • "From folder" prefers the Electron native picker (when available) and falls back to opening the existing modal-based import form.

Stage D — Real stage / atom workers (replaces the stub)

As shipped (MVP):

  • apps/daemon/src/plugins/atoms/registry.ts owns the atom worker registry: registerAtomWorker, runStageWithRegistry, and the frozen PERMISSIVE_DEFAULT_SIGNALS table. Real-worker outputs replace permissive defaults wholesale so a real score of 5 never gets clipped to 4; cross-worker conflicts inside a single stage still pessimistically merge (false-wins / lowest-number-wins).
  • apps/daemon/src/plugins/atoms/built-ins.ts registers a worker for every FIRST_PARTY_ATOMS entry on first use. Only critique-theater ships a real watcher today — it reads run_devloop_iterations.critique_summary for the latest parseable score=N token. Every other atom registers a permissive worker that returns no signals (so the defaults flow through) — that keeps backwards-compat with the v1 stub while documenting the worker surface for future migrations.
  • apps/daemon/src/server.ts#firePipelineForRun now branches on OD_PIPELINE_RUNNER: default (registry) drives stages through runStageWithRegistry; OD_PIPELINE_RUNNER=stub falls back to the canned signal pump for diagnostic bisection.

Deferred to a follow-up:

  • Real watchers for file-write, live-artifact, direction-picker, discovery-question-form, todo-write, etc. These require either an agent-write protocol that touches DB rows the daemon can observe, or a side-channel (artifacts table / genui surface responses) wired into worker reads. The registry shape is ready; only the workers themselves remain.
  • Strengthen ## Active stage block so the agent has to acknowledge each atom's outputs. (Today the contracts-side renderActiveStageBlock already exists; we still need to gate stage progression on the agent's structured acknowledgement.)

Stage E — Verification gate

  • pnpm guard + pnpm typecheck green.
  • Daemon + web package tests green.
  • New e2e covering: bare query, plugin pick, Figma chip, folder chip — all four end with a pipeline_stage_completed SSE.

File map

New (planned across stages)

  • plugins/_official/scenarios/od-media-generation/{open-design.json,SKILL.md}
  • plugins/_official/atoms/media-generate/{open-design.json,SKILL.md}
  • apps/daemon/src/plugins/atoms/media-generate.ts
  • apps/daemon/src/plugins/atoms/registry.ts
  • apps/web/src/components/home-hero/chips.ts
  • apps/web/tests/components/home-hero-chips.test.tsx
  • apps/daemon/tests/plugins-default-binding.test.ts
  • apps/daemon/tests/plugins-local-skill.test.ts

Modified

  • apps/daemon/src/plugins/apply.ts
  • apps/daemon/src/plugins/snapshots.ts
  • apps/daemon/src/server.ts
  • apps/daemon/src/plugins/pipeline.ts
  • apps/daemon/src/prompts/system.ts
  • apps/web/src/components/EntryShell.tsx
  • apps/web/src/components/HomeHero.tsx
  • apps/web/src/components/HomeView.tsx
  • apps/web/src/styles/home/index.css
  • packages/contracts/src/plugins/apply.ts (AppliedPluginSnapshot adds optional pluginSkillBody)

Risks

  • R1 — Plugin SKILL.md may conflict with a project-pinned skill. Resolution order: plugin > project skill > kind default.
  • R2 — Media surface already drives prompts through the media generate CLI; od-media-generation must not double-inject.
  • R3 — Stage D changes runtime behaviour. Keep OD_BUNDLED_ATOM_PROMPTS=0 and an explicit OD_PIPELINE_RUNNER=stub escape hatch until e2e stabilises.

Open questions

  • Q1 — Persist the Home secondary chip selection (model picker) to user config or keep session-only? Session-only for v1.
  • Q2 — Should "From folder" be a top-level chip or live behind a "More" submenu? Top-level for first release; revisit after telemetry.

Verification commands

pnpm guard
pnpm typecheck
pnpm --filter @open-design/daemon test
pnpm --filter @open-design/web test