Skip to content

Threads

github-actions[bot] edited this page Oct 4, 2026 · 35 revisions

Wiki source: docs/02-business-product/clawai-threads-product-spec.md

ClawAI Threads

Threads turns private chat into sourced publications through asynchronous multi-model review. Generation records public intent after a clear disclosure; the draft remains private until its owner approves publication.

Product choices

  • Article, research article, guide, and technical explanation formats.
  • Three to five authors; unanimous agreement on the exact draft hash; Judge at least 80; Critic at least 75; up to three rounds.
  • A user-selected spend cap is required. Existing plan and credit rules apply; there is no new PAYG price.
  • Authenticated users can comment, react, and request changes under existing moderation controls.
  • Approved publications are independently indexable. Chat-share lockdown stays.
  • Approved revisions remain public anonymously after account deletion; private snapshots and generation artifacts are erased.

Architecture

claw-threads-service owns publications and community state. claw-thread-generation-service owns isolated jobs. Chat provides filtered immutable snapshots and research-service provides evidence. Read the architecture and ADR-159.

Generation uses Auth's aggregate cap ledger. Existing research, Judge, and Critic plan uses are reserved once at enqueue; each provider wallet hold is sub-held before the call. Settlement counts measured cost, and a job cannot close with an unresolved sub-hold. The cap introduces no new PAYG price.

The existing /api/v1/threads route remains the chat-thread alias. Publication APIs use /api/v1/thread-publications.

Generation requests a versioned snapshot from Chat's internal endpoint. Ownership is checked in Chat by matching the requested owner and thread in one serializable read. Snapshots include ordered user/assistant text and a SHA-256 digest, while excluding system/tool rows, failures, placeholders, duplicate chunks, attachment metadata, and messages matching known secret patterns. An over-limit transcript (2,000 messages or 2 MiB) is rejected whole. JSON and Markdown exports are available; TOON stays unavailable until a verified codec shows semantic round-trip and useful token savings.

The two service containers probe /api/v1/health over loopback HTTPS with Node's built-in HTTPS client; see the deployment architecture.

Clone this wiki locally