# hallbayes

**Repository Path**: mirrors_trending/hallbayes

## Basic Information

- **Project Name**: hallbayes
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2025-09-15
- **Last Updated**: 2026-03-21

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Berry

## Workflow verification playbooks

Want the “with vs without hallucination detector” experience? Start here:

- `docs/workflows/README.md` — index of workflow playbooks
  - Search & Learn
  - Generate Boilerplate/Content
  - Inline Completions
  - Greenfield Prototyping
  - RCA Fix Agent

Each playbook includes a **maximally contrasting** worked example (❌ vibes vs ✅ evidence + verifier).

Berry runs a local MCP server with a safe, repo‑scoped toolpack plus verification tools (`detect_hallucination`, `audit_trace_budget`).

Berry ships a single MCP surface: **classic**.

Classic includes:
- Verification tools (`detect_hallucination`, `audit_trace_budget`)
- Run & evidence notebook tools (start/load runs, add/list/search spans)

See `docs/MCP.md` and `docs/workflows/README.md`.

Berry integrates with Cursor, Codex, Claude Code, and Gemini CLI via config files committed to your repo.

## Supported verifier backends

Berry’s current verification method requires token logprobs (Chat Completions-style `logprobs` + `top_logprobs`).

Supported today:
- `openai` (default): OpenAI-compatible Chat Completions endpoints with logprobs (OpenAI, OpenRouter, local vLLM, or any compatible `base_url`)
- `gemini`: Gemini Developer API `generateContent` with token logprobs via `logprobsResult` (when enabled for the model)
- `vertex`: Vertex AI `generateContent` (Gemini models) with token logprobs via `logprobsResult`
- `dummy`: deterministic offline backend for tests/dev

Not supported yet:
- Anthropic (OpenAI-compat layer ignores `logprobs`)

## Quickstart

1) Install (from this repo):

```bash
pipx install -e .
```

Fallback:

```bash
pip install -e .
```

2) In the repo you want to use:

First, set your verifier API key (recommended):

```bash
berry setup
```

Then install repo-scoped MCP config files:

```bash
berry init
```

Optional: enable strict verification gates for that repo:

```bash
berry init --strict
```

3) Reload MCP servers in your client.

Optional: register Berry globally (user-level configs) so you don't have to commit repo files:

```bash
berry integrate

# macOS .pkg installers may also deploy system-managed configs:
#   berry integrate --managed-only
```

4) Use a prompt/workflow (Search & Learn (verified), Generate Boilerplate/Content (verified), Inline completion guard, Greenfield prototyping, RCA Fix Agent).

## Docs

- `docs/USAGE.md` — task‑oriented guides
- `docs/CLI.md` — command reference
- `docs/CONFIGURATION.md` — config files, defaults, and env vars
- `docs/MCP.md` — tools/prompts and transport details
- `docs/PACKAGING.md` — release pipeline (macOS pkg + Homebrew cask)

## Tests

```bash
pytest
```