Dreamer (@leumas/dreamer)

The ecosystem's multi-stage ideation pipeline: a data-defined chain of LLM stages — the default is idea → expansion → feasibility → selection → scaffold → notes — where each stage can read earlier stages' outputs and, critically, run on any configured model (ported from the legacy Dreamer tool, which was hardwired to a single local Ollama model in five places).

How agents call it

Args: topic (required) · depth standard|deep (default standard) · mode standard|wild (wild favors unconventional-but-plausible ideas) · provider/model (any @leumas/providers brain, default auto) · pipeline (a custom definition; defaults to the 6-stage pipeline) · stageModels (per-stage model overrides) · inputValues (values for a pipeline's declared {{input.*}} tokens).

Any model — and mixable per stage (the whole point)

This is the headline capability. provider/model set the run default, but stageModels overrides the model per stage by key, so a single run can brainstorm cheaply and decide expensively:

POST /api/dreamer/run
{
  "topic": "carbon-negative concrete",
  "depth": "deep",
  "provider": "auto",
  "stageModels": { "idea": "llama3.1", "selection": "claude-opus-4-8" }
}

Resolution per stage: stageModels[key] → stage.provider/stage.model → the run's provider/model → 'auto'. 'auto' always resolves to whatever provider is available, so a call with no model set still runs. Never call a raw provider — always go through the injected registry.

The pipeline is DATA (add a stage with no code change)

A pipeline is portable JSON: { name, version, stages: [{ key, label, task, outputFormat, priorOutputs }] }. Studio can edit it, an action/HTTP caller can pass one in, and a tenant can store their own. Validate any custom definition first — GET /api/dreamer/pipeline returns the default; POST /api/dreamer/pipeline/validate returns { ok, errors }. The one validator rejects duplicate keys, self-references, unknown priorOutputs, and forward references (a stage consuming a later stage's output).

SSE vocabulary

start · stage_start · stage_done · done · error (+ : ping heartbeat every 15s; client disconnect aborts the run; read POST-SSE with the shared admin/_shared/sse.js streamPostSse — EventSource is GET-only). Every stage checkpoints to /db/dreamer_runs, so a dropped stream loses nothing and a partial/failed run is still readable (the legacy tool left orphan folders with no status).

Persistence

dreamer_runs dynamic collection (auto-written when a connector is injected): the doc is created running up-front and updated after each stage with steps, outputs, usage, then finalized done/aborted/error. No separate list route — query /db/dreamer_runs.

Gating

requireAuth + requireMembership('intelligence') (admins bypass) + PassNode guard feature:dreamer.run — author a passnode_rules doc with that resource id to meter/charge runs; no rule = free pass. LLM calls cost real tokens, so meter before white-labeling.