智能工作流

Build a minimal skill

Create files under reyn/local/my_skill/:

```yaml

Quick Start

Requirements: Python 3.11+, a LiteLLM-compatible model endpoint.

pip install -e .               # local install; web UI: pip install -e ".[web]"
export OPENAI_API_KEY=sk-...   # or set the key for your LiteLLM proxy
reyn init                      # creates reyn.yaml + .reyn/config.yaml

A note on weak default models. Reyn's default models.standard points at a low-cost LLM. Two phenomena are normal at the weak tier and dissolve on stronger models: - Occasional empty replies on tool-heavy queries (e.g. "list available skills" / "explain how X works"). Measured ~15% rate on gemini-2.5-flash-lite. Tracked as G12 in docs/deep-dives/journal/dogfood/giveup-tracker.md. - Capability questions leak router-internal vocabulary in non-English replies (e.g. asking 「何ができる?」 may return text that mentions invoke_action / list_actions / skill__X / etc. verbatim). Trace-driven A/B at N=10: weak gemini-2.5-flash-lite clean rate 20%, strong gemini-2.5-flash 87.5% with no prompt change. Tracked as G31 in the same file with the full matrix. Edit reyn.yaml's models.standard to point at a stronger model if either rate matters for your use.

Wire it up (Claude Desktop example)

Two transports; pick whichever fits your setup.

Recommended — SSE (shared with the web UI, dev-loop friendly):

```bash reyn web --port 8080 # leave running in any terminal

and auto-wraps it with the artifact envelope.

reyn run index_docs '{"type":"index_docs_input","data":{"source":"my_docs","path":"docs/**/*.md","description":"Project documentation"}}'

The CLI also accepts the bare data dict

How Reyn compares

Reyn is more constrained. If you want maximum LLM autonomy and creative agent behavior, LangGraph or AutoGen will feel less restrictive.

Reyn ships a RAG framework foundation, not a mature RAG product. The differentiator is that you write your indexing strategy as a skill.md — LLM-driven adaptive chunking with a deterministic postprocessor chain — not a Python pipeline. Override the chunker per-source by swapping a single python step. End-to-end smoke (= reyn run index_docs against docs/concepts/*.md → 418 chunks via real gemini-embedding-001 → reyn chat with natural concept queries) returned indexed semantic answers in 3/3 runs (batch 22, 2026-05-10). Maturity gaps (rerank / HyDE / contextual retrieval / RAG eval framework / IDE integration / vector store variety beyond SQLite) live downstream — see Project Status and docs/concepts/data-retrieval/rag.md.

Reyn is smaller. No chain abstractions, no rich vector store ecosystem — those live downstream (see care-boundary.md).

Reyn is opinionated about state. The Workspace is the only inter-phase data channel; Events are the only audit log. Other frameworks let you pass state in-memory or through callbacks — convenient, but invisible to crash recovery and audit trails.

Time-travel debugging. Reyn ships a replay CLI that walks any past run step by step (--mode replay), and a compare CLI that diffs two runs side by side (--mode compare). See docs/reference/dogfood-tracing.md.