Repository navigation
Releases
0.5.1 — Proxy API documentation
Highlights
Release 0.5.1 completes the documentation for the Proxy API introduced in 0.5.0.
Proxy API documentation
- New Proxy API docs page (GitHub Pages → Running → Proxy API):
- What the proxy does, launch flags and the library API (
create_proxy_app/run_proxy) - Endpoint reference:
POST /v1/chat/completions(full request schema + theminicsoptions object),GET /v1/models,GET /health - The complete
minicsmetadata block field-by-field reference (version, model, latency, documents used, citations, diagnostics) - Streaming SSE format, error/degradation matrix
- Client examples: curl, Python and TypeScript OpenAI SDKs, external GUIs (Open WebUI, LibreChat, ChatBox)
- Configuration integration (retrieval/rerank), deployment notes (auth, nginx, Docker) and a Proxy vs studio API comparison
- What the proxy does, launch flags and the library API (
- Registered in the docs site: sidebar, pager, homepage card and search index
- Cross-linked from
docs/index.md, the README anddocs/api.md
No code changes — drop-in upgrade.
Install / upgrade: pip install --upgrade minichat-studio==0.5.1
Docker: docker pull ghcr.io/jasonjimnz/minics:0.5.1
0.5.0 — Pi-safe skills, rerank pipeline, OpenAI RAG proxy
Highlights
Release 0.5.0 ships three major additions: Pi-compatible agent skills, an optional rerank pipeline, and the OpenAI-compatible RAG proxy server.
Pi compatibility (agent skills)
- Fixed silent skill drops: YAML descriptions containing
:are now quoted — Pi's strict frontmatter parser rejected the parentminicsskill - Fixed
minics --home X <cmd>being ignored (subparser default clobbered the main parser value; both flag positions now work) minics skills installauto-saves the Pi project-trust decision (--pi-trust), warns about frontmatter issues that would make Pi drop a skill- New
minics-proxyskill; 12 skills ship in the package
Rerank (optional, disabled by default)
- New
retrieval.rerankconfig: Cohere-compatiblePOST /rerank(Cohere, Jina, TEI, vLLM) with URL/key fallback to the embeddings endpoint - The hybrid retriever reorders the fused pool (
rerank_score,ranks,sources) and degrades gracefully to the fused order on failure - Setup flags (
--rerank-model,--no-rerank,--rerank-pool, …), Settings UI card,test_connectionrerank report
Proxy server (new)
minics proxy— FastAPI/uvicorn OpenAI-compatible endpoint (/v1/chat/completions,/v1/models,/health)- Automatic grounding: retrieval runs first, the context is injected as
[n]references, streaming supported - Responses carry a
minicsmetadata block: version, model, latency, documents used, citations, retrieval diagnostics - Per-request grounding options via the
minicsbody field (use_rag,use_graph,use_grounding,top_k,document_ids)
Also
- FastAPI + uvicorn + requests added as dependencies
- 58 unit tests + full E2E verification against a live LLM/embeddings stack
Install / upgrade: pip install --upgrade minichat-studio==0.5.0
Docker: docker pull ghcr.io/jasonjimnz/minics:0.5.0
0.4.0
MiniCS Lite 0.4.0
This release makes MiniCS far easier to feed and far easier to drive: a new
bulk document importer for the terminal, and a complete, precise agent skillset
that ships inside the package — installable for Pi, OpenCode, Codex and any
skill-compatible agent with a single command.
New: minics scan — bulk document import
Scan a directory (absolute or relative to the current working directory) and
add every compatible document to the store one by one, with live terminal
progress (convert → embed → graph, per file):
minics scan ./docs # all compatible formats, recursive
minics scan /data/papers --pdf # only PDFs
minics scan ./notes --markdown --txt # only Markdown and plain text
minics scan ./docs --tags manuals,v2 # tag every imported document
minics scan ./docs --dry-run # preview what would be imported
minics scan ./docs --no-index # import only; index later with reindex- Format flags
--markdown,--txt,--pdf,--docx,--latex— with
no flag, all compatible formats are used (PDF, TXT, Markdown, DOCX,
LaTeX). - Idempotent: files already in the store (same SHA-256) are skipped and
reported — re-running after new documents is the intended workflow. - Hidden folders and vendor directories (
node_modules,.venv,
__pycache__,dist, …) are skipped automatically. - Per-file failures are reported and the scan continues; a summary
(imported / duplicates skipped / failed / total) is printed at the end.
New: minics skills — the bundled agent skillset
MiniCS now ships a complete agent skillset inside the PyPI package —
nothing is fetched from GitHub. Install it for your coding agents:
minics skills install # project-local: ./.agents/skills/
minics skills install --global # user-wide: ~/.agents/skills/
minics skills install --interactive # pick individual skills
minics skills install --force # overwrite an existing installation
minics skills list # bundled skills + install statusOne parent skill plus ten focused sub-skills:
| Skill | Covers |
|---|---|
minics |
Parent map: concepts, workflow, quick CLI reference, store layout. |
minics-setup |
LLM/embedding endpoints, wizard, flags, troubleshooting. |
minics-documents |
Import pipeline, minics scan, indexing, reindexing. |
minics-datasets |
Datasets, ChatML entries, lifecycle, versioning. |
minics-authoring |
Enhance, suggest tags, evaluate (6 dimensions), grounded generation. |
minics-retrieval |
Hybrid RAG + graph, RRF fusion, tuning knobs. |
minics-chat |
Grounded chat, RAG/graph/grounding toggles, [n] citations. |
minics-export |
Collections; ChatML, JSONL, Alpaca, ShareGPT, Markdown. |
minics-python |
The AppContext library API for scripts and notebooks. |
minics-server |
minics serve, the GHCR image, volumes, reverse proxies. |
minics-skills |
The installer itself: local vs global, update flows. |
The SKILL.md + frontmatter format is the emerging cross-agent standard, so a
single install serves Pi, OpenCode, Codex and any other skill-compatible
agent — local installs can be committed so the whole team shares them.
Documentation
- New: CLI reference — every
minicssubcommand, with a
deep dive onscan(flags, sample output, duplicate handling). - New: Agent skills — installation guide and a dedicated
view for every skill in the catalogue, plus agent-compatibility notes. - New: Skill usage examples — 34 deep, worked
examples (real commands, expected output, agent tasks) across all 11 skills,
plus 3 end-to-end playbooks: paper archive → fine-tuning dataset,
nightly knowledge refresh, and new machine setup. - README:
scanandskillsadded to the quick start and entry points; the
skillset joins the highlights table.
Fixes & internals
- The CLI no longer crashes on legacy Windows consoles (cp1252): output
streams are reconfigured witherrors="replace", so Unicode arrows and
dashes in help, skill descriptions and progress lines are safe. - New
minics.skills_installmodule: bundles the skillset as package data and
copies it locally (with per-skill skip/force semantics). - New tests for
scan(import, dedup, dry-run, format filters, missing dir)
and the skills installer (list, local install, skip/force, interactive
selection, unknown names).
Artifacts
| Artifact | Where |
|---|---|
| Python package | PyPI — minichat-studio 0.4.0 |
| Server image | ghcr.io/jasonjimnz/minics:0.4.0 (and :latest) |
| Docs site | https://jasonjimnz.github.io/minics/ |
Upgrading
pip install -U minichat-studio
minics skills install --global --force # refresh the bundled skillsNo schema changes — existing ~/.minics stores work as-is.
Full changelog: 0.3.3...0.4.0
0.3.2
Initial working state with all sections, small issues can be hidden that hasn't been triggered on tests