forgo.cloud
Sign in
Repo workspace

forkjoin-ai/gnosis

protocol69 `runHiddenEmbed` — distributed hidden-state embedding over SkyMesh

distributed-inference/PROTOCOL69_HIDDEN_EMBED.md
forkjoin-ai/gnosis
Repo overview Open in workbench

protocol69 runHiddenEmbed — distributed hidden-state embedding over SkyMesh

Make hidden-state embedding a first-class protocol69 work type so the native node fleet (gaming PC, R1s, future boxes) produces the contextual flux-vector DB via the sovereign dispatch protocol instead of ad-hoc fat-station HTTP. Browser-WebGPU agents run TinyLlama generation and CANNOT produce the qwen-896 embeddings the DB lives in, so embedding work is executed by a native agent that calls a local qwen fat-station /hidden-embed.

The contract (all three layers MUST agree on this)

Gateway base: https://skymesh.forkjoin.ai (prod) / http://127.0.0.1:8788 (dev). NEW routes are flag-gated by SKYMESH_EMBED_ENABLED (default false) so deploying the prod worker is safe.

1. ENQUEUE POST {gateway}/protocol69/embed/offer

Body: { "batchId": string, "items": [{ "key": string, "tokens": number[] }], "model": "qwen2.5-0.5b", "dim": 896 }

  • Creates one workflow per item, single leg: workKind:"runHiddenEmbed", executor:"nativeEmbed", state:"pending", carrying {key, tokens, model, dim, certificate} (the constant antiqueue cert).
  • Returns { ok:true, batchId, enqueued: number }. Idempotent on key (skip keys already filled).

2. CLAIM POST {gateway}/protocol69/work/claim (existing route, extend matching)

Body: { "agentId": string, "skills": ["runHiddenEmbed"] }

  • Returns { status:"claimed", work:{ workflowId, dispatchId, legId, cacheKey, key, tokens, model, dim, certificate, workKind:"runHiddenEmbed" } } or { status:"idle", work:null }.

3. ACK POST {gateway}/protocol69/work/ack (existing route, store projection)

Body: { "ackKind":"legAck", "legAck":{ schemaVersion, contractId, certificate, workflowId, legId, dispatchId, cacheKey, workKind:"runHiddenEmbed", resultHash, provenance, projection:{ key, dim, embedB64 } } }

  • embedB64 = base64 of the little-endian f32 embedding (dim floats). Gateway stores it keyed by key, marks the workflow filled.
  • Certificate fields MUST equal the constant antiqueue control-plane cert (validated exact, MeshCacheNodeDO ~509-516).

4. RESULT GET {gateway}/protocol69/embed/result?batchId=...&since=...

Returns { results: [{ key, dim, embedB64 }], done: number, total: number } (paged). Lets the dispatcher collect fills without re-claiming.

Layers

  • Gateway (apps/skymesh/src/durables/MeshCacheNodeDO.ts): the 4 routes above, flag-gated.
  • Native agent (distributed-inference/src/bin/protocol69-embed-agent.rs): claim→POST local /hidden-embed→ack loop. Args: --skymesh-url, --fat-station-url (local qwen), --agent-id. Reuse the claim/ack structs + cert constant from protocol69-monster-coordinator.rs.
  • Dispatcher (open-source/flux-vector/scripts/embed-via-protocol69.ts): read a word list + vocab, tokenize, enqueue in batches, poll RESULT, write decoded f32 → ~/.flux-embed/ctxemb.json (resume-aware; skip keys already present). Mirrors embed-via-mesh.ts but over the gateway.

Consistency invariant

The native agent's local fat-station MUST load the SAME knot/model (qwen, 896-dim) as the existing R1-built DB, or embeddings won't share a space. The gaming PC redeploy = native qwen fat-station + this agent pointed at it.