[contexto]: Implement LocalBackend #116

Problem

@ekai/contexto only ships with RemoteBackend, which calls an external api for summarization, ingestion and search. Users who want to run fully local have no option today. The ContextoBackend interface was designed to be pluggable, but there's no local implementation.

Goal

Implement a LocalBackend class that satisfies the ContextoBackend interface by running the full pipeline locally: extract, summarize, embed, cluster, persist.

Interface to implement

interface ContextoBackend {

  ingest(payload: WebhookPayload | WebhookPayload[]): Promise<void>;

  search(query: string, maxResults: number, filter?: Record<string, unknown>, minScore?: number): Promise<SearchResult | null>;

}

Ingestion pipeline

The remote API runs this pipeline on every episode/combined event. The LocalBackend needs to replicate it:

WebhookPayload
  → Extract (strip OpenClaw metadata envelope, extract user + assistant + tool text)
  → Summarize via LLM → EpisodeSummary { summary, key_findings, status, confidence, evidence_refs }
  → Combine into content: summary + key_findings as bullets
  → Embed the combined content
  → addToMindmap() (AGNES clustering)
  → Persist via MindmapStorage

1. Extract

Strip the OpenClaw metadata envelope from user messages (Sender (untrusted metadata): ...) and extract text from ContentBlock[] arrays. Handle both string and ContentBlock[] content formats.

2. Summarize

Call an LLM to produce a structured EpisodeSummary from the raw episode (user message + assistant response + tool outputs):

interface EpisodeSummary {

  summary: string;              // 1-3 sentence concise summary

  key_findings: string[];       // at least one entry

  status: 'complete' | 'partial' | 'blocked';

  confidence: number;           // 0.0 to 1.0

  evidence_refs: EvidenceRef[];

  open_questions?: string[];

}

The content stored in the mindmap is summary + "\n" + key_findings joined as bullets. The full episode is preserved in metadata.episode for context injection at retrieval time. The remote API uses openai/gpt-4o-mini via OpenRouter with temperature: 0.2 and response_format: { type: 'json_object' }. Validation should degrade gracefully (mark as partial if fields are missing, don't fail).

3. Cluster + persist and search

Call Mindmap.add() which handles addToMindmap() and storage.save. Make the storage layer pluggable

Reference

• RemoteBackend (pattern to follow): packages/contexto/src/client.ts

• ContextoBackend interface: packages/contexto/src/types.ts

• Mindmap class: packages/mindmap/src/mindmap.ts

Scope

• Episode summarizer (port LLM call + validation from api-server)

• LocalBackend class implementing ContextoBackend

• Wire into plugin config (backend: 'local' option in PluginConfig)

• Support pluggable storage

• Tests for extract → summarize → embed → ingest → search round-trip