Add antoinezambelli/forge-guardrails

agents/antoinezambelli__forge-guardrails/README.md

@@ -0,0 +1,61 @@
+# forge-guardrails
+
+**A reliability layer for self-hosted LLM tool-calling.**
+
+Forge takes an 8B local model from single-digit pass rates to **84% across a
+26-scenario eval suite** — and lifts Claude Sonnet from 85% to 98% on the same
+workload — by applying composable guardrails without changing a single line of
+your agent code.
+
+## What it does
+
+Forge intercepts model responses before they reach your client or downstream
+logic and applies (in order):
+
+1. **Response validation** — every tool call is checked against the declared `tools` array; unknown names and malformed shapes are caught before surfacing.
+2. **Rescue parsing** — Mistral `[TOOL_CALLS]`, Qwen `<tool_call>` XML, and fenced JSON are normalised to the canonical OpenAI `tool_calls` schema.
+3. **Retry with error tracking** — on validation failure, forge retries inference with a corrective tool-result message (up to `max_retries`, default 3).
+4. **Synthetic `respond` tool injection** — prevents small models from breaking the tool-calling loop by routing plain-text responses through a structured call (stripped from the outbound response so the client never sees it).
+
+## Three usage modes
+
+| Mode | How | Best for |
+|------|-----|----------|
+| **Proxy server** | `python -m forge.proxy` | Drop-in for existing clients (opencode, aider, Claude Code, Continue) |
+| **WorkflowRunner** | Python API | Building new agentic loops directly on forge |
+| **Guardrails middleware** | `Guardrails` facade | Adding forge's stack inside your own orchestration loop |
+
+## Supported backends
+
+Ollama · llama-server (llama.cpp) · Llamafile · vLLM · Anthropic API
+
+## Quick install
+
+```bash
+pip install forge-guardrails                    # core
+pip install "forge-guardrails[anthropic]"       # + Anthropic client
+```
+
+## Claude Code integration
+
+Point Claude Code at a forge-guarded local model:
+
+```bash
+python -m forge.proxy --backend llamaserver --gguf path/to/model.gguf --port 8081
+# then:
+ANTHROPIC_BASE_URL=http://localhost:8081 ANTHROPIC_AUTH_TOKEN=anything claude
+```
+
+## Eval harness
+
+```bash
+python -m tests.eval.eval_runner --backend llamafile --gguf path/to/model.gguf --runs 10
+python -m tests.eval.report eval_results.jsonl
+```
+
+## Links
+
+- Repository: https://github.com/antoinezambelli/forge
+- PyPI: https://pypi.org/project/forge-guardrails/
+- Paper: https://doi.org/10.1145/3786335.3813193
+- Docs: https://github.com/antoinezambelli/forge/tree/main/docs

agents/antoinezambelli__forge-guardrails/metadata.json

@@ -0,0 +1,14 @@
+{
+  "name": "forge-guardrails",
+  "author": "antoinezambelli",
+  "description": "A reliability layer for self-hosted LLM tool-calling — guardrails, rescue parsing, retry loops, and proxy mode for Ollama, llama.cpp, vLLM, and Anthropic.",
+  "repository": "https://github.com/antoinezambelli/forge",
+  "version": "0.7.2",
+  "category": "developer-tools",
+  "tags": ["llm", "tool-calling", "guardrails", "self-hosted", "ollama", "llama-cpp", "proxy", "agentic", "reliability", "python"],
+  "license": "MIT",
+  "model": "anthropic:claude-sonnet-4-6",
+  "adapters": ["system-prompt", "claude-code"],
+  "icon": false,
+  "banner": false
+}