Trigix v1.2.0

Deeper AI-native capabilities — the agent runs on Chinese / self-hosted models,
richer RAG retrieval, and acting tools — plus a deployment and distribution
chain that actually builds and deploys. Backward compatible with 1.1.0.

Added

AI-native

• The Agent node runs on OpenAI-compatible models (Qwen, DeepSeek, Zhipu,
  Moonshot, a self-hosted vLLM/Ollama gateway) in addition to Anthropic, so the
  tool-use agent works in a self-hosted / China deployment where the Anthropic
  API is unreachable.
• Agent tools: a sandboxed http_request (default-deny egress, SSRF guard,
  DNS-rebinding-safe IP pinning, response size cap) and custom-node tools that
  let the agent call the tenant's own registered nodes.
• Hybrid RAG retrieval — vector + full-text fused with Reciprocal Rank Fusion —
  and an optional min_score floor; helps queries that hinge on exact tokens
  (codes, identifiers, English terms inside CJK text).
• Optional cross-encoder reranking over a Cohere/Jina/BGE-compatible endpoint
  (e.g. a self-hosted bge-reranker), with a dependency-free local fallback.
• An HNSW vector index for retrieval at scale, and CJK tokenization for the
  hybrid keyword side when a Chinese FTS config (pg_jieba / zhparser) is present.
• The Agent node reports token usage (_agent_usage).

Deployment & CI

• CI: a Docker image build smoke job (builds the platform and AI runtime images
  and checks they run) and a Helm job (lint + render across value permutations
  • kubeconform schema validation).
• A Dockerfile for the AI runtime (the repo had a platform image but none for
  the AI runtime).
• Helm chart: an ai-runtime Deployment + Service, so the chart deploys the
  full stack — platform + AI runtime + PostgreSQL/pgvector + Redis; the AI
  runtime and Redis were also added to docker-compose.prod.yml.
• The Helm chart is published to GHCR (oci://ghcr.io/bj-qizhi/charts/trigix)
  and attached to GitHub Releases; a workflow auto-publishes on a chart-v*
  tag, syncs both channels, and bumps the README install version.

Changed

• Both Agent LLM backends run the (synchronous) model call off the event loop.
• The platform image tracks the latest stable Rust (rust:1-slim), matching CI,
  so it no longer rots when a dependency raises its edition / MSRV.
• Helm chart 0.3.2, appVersion 1.2.0.

Fixed

• The platform Docker image could not build: a stale crate name left over from
  the agentflow→trigix rename, a Rust base image too old for the dependency
  tree (a transitive crate now requires edition 2024), and the migrations
  directory missing from the build stage (sqlx::migrate! reads it at compile
  time). The runtime image also lacked curl, which its healthcheck used.
• Helm chart could not deploy: the platform DATABASE_URL was a password-less
  placeholder so it could not authenticate, and the platform Service/PDB
  selectors also matched the Redis (and now AI runtime) pods. Both the platform
  and AI runtime now build the DSN from the postgres secret, and each component
  has a scoped selector.
• docker-compose.prod.yml: dropped a migrations initdb mount that conflicted
  with the app's own sqlx::migrate! step on a fresh database.
• The condition node now evaluates operators (gt/lt/contains/…) and a
  source path instead of silently falling back to an existence check, and the
  bundled gallery templates were corrected to read values from the right paths.