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):

pipx install -e .

Fallback:

pip install -e .

2. In the repo you want to use:

First, set your verifier API key (recommended):

berry setup

Then install repo-scoped MCP config files:

berry init

Optional: enable strict verification gates for that repo:

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:

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

pytest