Releases: bj-qizhi/trigix
Trigix v1.3.0
A large expansion of the workflow node catalog — 44 new node types
(136 → 180) — covering the gaps against comparable workflow/automation
tools and deepening the AI-native and Chinese-enterprise coverage. Each node
is full-stack (engine + config UI) and backward compatible. Most are plain
HTTP; the few that aren't were deliberately implemented in pure Rust (or via a
runtime CLI) so the default cargo build needs no extra system library.
Added
LLM providers
azure_openai(deployment-based,api-keyheader),vertex(Google Vertex
AI / GeminigenerateContentwith a caller-supplied OAuth2 token),bedrock
(AWS BedrockInvokeModel, model-native body),grok(xAI), andollama
(self-hosted, OpenAI-compatible).
AI-native building blocks (OpenAI-/Cohere-compatible, configurable base URL)
embedding,reranker,text_splitter(pure-compute, UTF-8-safe chunking),
structured_output(LLM JSON mode),classifier,image_gen,
speech_to_text(Whisper, multipart upload), andtts.
Vector stores
weaviate(REST + GraphQL),chroma(REST data API), andmilvus/ Zilliz
(REST API v2).
Databases & data warehouses
mongodb(Atlas Data API),clickhouse(HTTP),mysql(sqlx),snowflake
(SQL API v2, bearer token),bigquery(jobs.query REST), andsqlserver
(Microsoft SQL Server over the pure-Rust tiberius TDS driver — no native
client needed).
Object storage
gcs(Google Cloud Storage JSON API) andazure_blob(REST + SAS token),
both using caller-supplied credentials.
AWS (a from-scratch Signature V4 signer — no AWS SDK — validated against
AWS's published get-vanilla test vector)
sqs,sns, andbedrock.
Message brokers
kafka(Confluent REST Proxy) andrabbitmq(Management HTTP API).
Chinese enterprise collaboration
feishu/ Lark (custom-bot webhook or app message API),dingtalk(custom
robot with optional HMAC sign), andwecom(WeChat Work group robot).
Network file / shell / mail (pure-Rust clients — no libssh2/system library)
ftp/ FTPS (suppaftp),sftpandssh(russh / russh-sftp, password or
private-key auth), andimap(TLS mailbox reads).
Utilities & core primitives
hash(SHA-256/384/512 + HMAC),jwt(HS256/384/512 sign & verify),
zip(create/extract),image(resize/convert/metadata),pdf_extract
(text extraction),ocr(via thetesseractCLI),html_extract(CSS
selectors),rss(RSS/Atom/JSON feeds), andwait(pause for a duration /
until a timestamp, or suspend until externally resumed via the existing
approve/resume endpoint).
Notes
- The
ocrnode needs thetesseractCLI on the executor host; all other
nodes have no extra runtime dependency. - The
wait(resume mode) andsnowflake/bigquery/vertex/gcs/azure_blob
nodes take a caller-supplied token rather than performing their own
credential exchange.
Helm chart v0.3.3
Helm chart trigix 0.3.3 (deploys app 1.1.0).
Install:
helm install trigix oci://ghcr.io/bj-qizhi/charts/trigix --version 0.3.3 --namespace trigix --create-namespaceTrigix 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 optionalmin_scorefloor; 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-runtimeDeployment + Service, so the chart deploys the
full stack — platform + AI runtime + PostgreSQL/pgvector + Redis; the AI
runtime and Redis were also added todocker-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 achart-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,appVersion1.2.0.
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 lackedcurl, which its healthcheck used. - Helm chart could not deploy: the platform
DATABASE_URLwas 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 migrationsinitdbmount that conflicted
with the app's ownsqlx::migrate!step on a fresh database.- The condition node now evaluates operators (
gt/lt/contains/…) and a
sourcepath instead of silently falling back to an existence check, and the
bundled gallery templates were corrected to read values from the right paths.
Helm chart v0.3.2
Helm chart trigix 0.3.2 (deploys app 1.1.0).
Install:
helm install trigix oci://ghcr.io/bj-qizhi/charts/trigix --version 0.3.2 --namespace trigix --create-namespaceHelm chart v0.3.1
Helm chart trigix 0.3.1 (deploys app 1.1.0).
Install:
helm install trigix oci://ghcr.io/bj-qizhi/charts/trigix --version 0.3.1 --namespace trigix --create-namespaceHelm chart v0.3.0
Helm chart trigix 0.3.0 (deploys app 1.1.0).
Install:
helm install trigix oci://ghcr.io/bj-qizhi/charts/trigix --version 0.3.0 --namespace trigix --create-namespaceHelm chart v0.2.0
Helm chart trigix 0.2.0 (deploys app 1.1.0).
Full stack: platform + AI runtime + PostgreSQL/pgvector + Redis. Adds the AI
runtime component and the fixes that make the chart actually deployable
(DB password injection, per-component selector scoping).
Install
helm install trigix \
https://github.com/bj-qizhi/trigix/releases/download/chart-v0.2.0/trigix-0.2.0.tgz \
--namespace trigix --create-namespaceOr download and inspect first:
curl -fsSLO https://github.com/bj-qizhi/trigix/releases/download/chart-v0.2.0/trigix-0.2.0.tgz
helm template trigix ./trigix-0.2.0.tgz | lesssha256: 7d13d280db831f1780a173952e70c71888dd1dc39b1ca3948574e6fc31252c87
Trigix v1.1.0
New AI-native and enterprise capabilities, a custom node SDK, and a major
quality/CI uplift. Backward compatible with 1.0.0.
Added
AI-native
- RAG knowledge store on pgvector in the AI runti
e: ingest, embed (OpenAI or
an offline local embedding), and cosine-similarity retrieval. ragandrag_ingestworkflow nodes, plus a Knowledge Bases management page.- Agent tool-use loop: the Agent node can call tools (sandboxed calculator and
knowledge-base search) and iterate to an answer.
Custom node SDK (node ecosystem)
- Python (
trigix-node-sdkon PyPI) and TypeScript/JavaScript
(trigix-node-sdkon npm) SDKs for writing nodes served over HTTP. customnode type, a tenant-scoped node registry, and one-click registration
from a node service's/manifest.- Example nodes (HTML to text, PII redaction, sentiment) and an end-to-end demo.
Enterprise
- Enterprise SSO via OIDC (Okta / Azure AD / Google Workspace / Alibaba Cloud
IDaaS / Huawei OneAccess / Tencent / Authing) plus Feishu, DingTalk, and
WeChat Work; admin management UI with enable/disable. - Encryption at rest for credential and SSO secrets (AES-256-GCM via
CREDENTIAL_MASTER_KEY), with transparent passthrough of legacy plaintext. - Dead-letter queue for the Redis Streams execution queue; failed jobs are
preserved and can be re-driven instead of silently dropped. - Opt-in data retention sweeper (
DATA_RETENTION_DAYS) for executions, audit
log, token usage, and webhook deliveries.
Changed
- Split the 14k-line
executor.rs(previouslyinclude!-spliced) into cohesive
submodules. - CI now enforces formatting, the full Rust test suite, the web production
build, the AI runtime tests against a pgvector service, and both node SDK test
suites.
Fixed
- Repaired a test suite that did not compile and 122 frontend TypeScript build
errors. - Credential creation on PostgreSQL (an i64 was bound to a
TIMESTAMPTZcolumn). - A few latent frontend bugs surfaced by the typechecker.
Trigix v1.0.0 — Initial Release
🎉 Trigix v1.0.0 — Initial Release
First public release of Trigix, an AI-Native Workflow Automation Platform built with Rust + React.
Highlights
- 136 node types — AI models (Claude, GPT-4, Gemini, Groq + 7 Chinese LLMs), 50+ SaaS integrations, data transforms, control flow
- Rust execution engine — Async DAG, parallel Fan-out/Fan-in, retries, SSE streaming
- Enterprise-ready — JWT + RBAC, multi-tenant, audit log, Prometheus metrics, Kubernetes Helm Chart
- Visual canvas — React Flow editor with undo/redo, keyboard shortcuts, template gallery, AI-assisted generation
Quick Start
docker compose up -d
DATABASE_URL=postgres://trigix:trigix@localhost:35432/trigix cargo run -p trigix-platform
cargo run -p trigix-executor
cd apps/web && npm install && npm run devSee CHANGELOG.md for the full feature list.
© 2026 北京祺智科技有限公司 · https://www.qzso.com/