autoresearch-prompt-manager 0.1.0

Versioned prompt CRUD, A/B experiments, metrics, and LLM optimization API

prompt-manager

Versioned prompt CRUD, A/B experiments, metric collection, and LLM-driven optimization API.

Install

# Just the client SDK (for services that fetch prompts)
pip install prompt-manager[client]

# Full API server
pip install prompt-manager[api]

# Client + metric reporting
pip install prompt-manager[client,metric]

# Everything
pip install prompt-manager[all]

Quick start

Start the API

PM_DATABASE_URL=postgresql://user:pass@localhost:5432/prompts \
  prompt-manager serve

Use the client SDK

from prompt_manager.client import PromptManagerClient

client = PromptManagerClient(base_url="http://localhost:8910")

# Resolve a prompt (returns latest version, experiment-aware)
prompt = await client.resolve("welcome-email", session_id="user-123")
print(prompt.body)       # "Hi {name}, welcome to {company}!"
print(prompt.version)    # 2

# Report a quality metric
await client.report_metric("welcome-email", str(prompt.version_id), "quality", 8.5)

API endpoints

Method	Path	Description
POST	/prompts	Create prompt (auto-creates v1)
GET	/prompts	List all prompts
GET	/prompts/{slug}	Get prompt by slug
POST	/prompts/{slug}/versions	Create new version
GET	/prompts/{slug}/versions	List versions
GET	/resolve/{slug}	Resolve prompt (experiment-aware)
POST	/experiments	Create A/B experiment
PATCH	/experiments/{id}/status	Start/pause/conclude
POST	/metrics	Report metric signal
GET	/metrics/aggregate	Aggregated metrics per version
POST	/optimize	Trigger LLM optimization
GET	/health	Health check

Key features

• Slug-based addressing -- resolve("welcome-email") not UUIDs
• Immutable versions -- append-only, SHA-256 dedup, full audit trail
• Experiment routing -- MurmurHash3 deterministic + Thompson Sampling (auto_optimize)
• Sticky sessions -- same user always sees same variant
• Metric collection -- quality signals per version, batch ingestion
• MCP server -- expose all tools via Model Context Protocol
• CLI -- prompt-manager serve, prompt-manager migrate, prompt-manager health

Database

Requires PostgreSQL 14+. Migrations run automatically on startup.

# Docker
docker compose up -d

# Or point to existing Postgres
PM_DATABASE_URL=postgresql://user:pass@host:5432/dbname prompt-manager migrate

Configuration

All settings via PM_-prefixed environment variables:

Variable	Default	Description
PM_DATABASE_URL	postgresql://localhost:5432/prompt_manager	PostgreSQL DSN
PM_HOST	0.0.0.0	Bind host
PM_PORT	8910	Bind port
PM_LLM_PROVIDER	anthropic	LLM for optimization
PM_LLM_MODEL	claude-sonnet-4-20250514	Model ID
PM_LLM_API_KEY	--	API key

Part of autoresearch-prompt-manager

autoresearch-prompt-manager  (this package -- API, client, metrics)
  -> autoresearcher-shonku   (optimization agents)
  -> shonku                  (agent framework)
  -> agno                    (runtime -- https://agno.com)

LLM-driven optimization is powered by agno and AgentOS.

Contributing

1. Fork autoresearch-prompt-manager
2. cd packages/prompt_manager && pip install -e '.[dev,api,client,metric]'
3. Make changes, pytest, ruff check src/
4. Integration tests: PM_DATABASE_URL=... python3 -m pytest tests/integration/
5. Submit a PR

License

MIT