🚀 llmhq-promptops

Git blame for prompts. When prod breaks, know exactly which prompt version was running, in seconds. No SaaS, no vendor lock-in — all in your own git.

llmhq-promptops is a prompt management framework that records every deploy, builds an immutable snapshot at CI time, and lets you ask promptops blame --at <timestamp> to find out what was running when an incident happened. Built for teams who already use git for everything else.

✨ Key Features

• 🔎 Incident archaeology — promptops blame --at <ts> resolves "what prompt was running in prod at that moment" by composing the deploy log with git history.
• 🐳 Production runtime without .git/ — promptops snapshot build writes a self-contained .promptops/snapshot.json. Ship that in your Docker image; AutoResolver picks it up automatically.
• 📒 Deploy event log — append-only .promptops/deploys.jsonl records every deploy with provenance (env, commit, actor, metadata). Committed to git alongside your prompts.
• 🔄 Automated git versioning (opt-in) — after promptops hooks install, a pre-commit hook detects prompt changes, bumps semver in YAML metadata, and re-stages. promptops init repo never touches .git/hooks/ on its own.
• 📝 Version-aware testing — :unstaged, :working, :latest, commit-<sha>, or any tag.
• 🐍 Python SDK — PromptManager(resolver=AutoResolver()) works the same in dev (uses git) and prod (uses snapshot).

⚡ Quick Start

Two paths depending on what you want to do.

Path A — 60-second demo on a sample repo

Walk through the incident-archaeology flow on a pre-baked repo, no setup of your own.

pip install llmhq-promptops

git clone https://github.com/llmhq-hub/promptops.git
cd promptops/examples/incident-archaeology-demo
./setup.sh                       # builds a tmp git repo with pre-baked history
cd /tmp/promptops-demo

# 1. Look at what's been deployed
promptops deploy list

# 2. "Production broke at 10:00 UTC. What was running?"
promptops blame --at 2026-05-20T10:00:00Z

# 3. Same question, full text of one prompt
promptops blame --at 2026-05-20T10:00:00Z --prompt summarizer

That's the hero use case. Three commands, one provable answer.

Full walkthrough + screencast script: examples/incident-archaeology-demo/README.md

Path B — Try it on your own repo

For when you want to add this to a real project.

pip install llmhq-promptops

cd <your-project>                # any git repo
promptops init repo              # creates .promptops/ (does NOT touch git hooks)

# Scaffold your first prompt → .promptops/prompts/user-onboarding.yaml
promptops create prompt user-onboarding

# Render it (the starter template uses {{ user_input }}; :unstaged reads
# your uncommitted working-tree edits)
python -c "from llmhq_promptops import get_prompt; print(get_prompt('user-onboarding:unstaged', {'user_input': 'World'}))"

# Record a deploy and look it up later
promptops deploy event --env prod -m release=v0.1.0
promptops blame --at now

# Optional: enable auto-versioning on commit (opt-in since v0.4.0)
promptops hooks install

📖 Usage Examples

Recommended production setup

from llmhq_promptops import PromptManager, AutoResolver

# Same code in dev (uses .git/) and prod (uses .promptops/snapshot.json)
manager = PromptManager(resolver=AutoResolver(repo_path="."))

prompt = manager.get_prompt("user-onboarding", {"name": "Alice"})

Resolving with full provenance

When you need to know which version actually ran (for logging, observability):

resolved = manager.resolve("user-onboarding")
# resolved.text       — raw YAML
# resolved.version    — e.g. "v1.2.3" or "commit-abc12345"
# resolved.commit     — full 40-char SHA
# resolved.source     — "git" or "snapshot"
# resolved.resolved_at — tz-aware UTC datetime

log_to_observability({
    "prompt": resolved.prompt_id,
    "version": resolved.version,
    "commit": resolved.commit,
    "source": resolved.source,
})

Specific versions

from llmhq_promptops import get_prompt

prompt = get_prompt("user-onboarding")                    # smart default
prompt = get_prompt("user-onboarding:v1.2.1")             # tagged version
prompt = get_prompt("user-onboarding:commit-abc12345")    # untagged commit
prompt = get_prompt("user-onboarding:unstaged")           # working tree
prompt = get_prompt("user-onboarding:working")            # HEAD (committed)

Using with LLM frameworks

from llmhq_promptops import get_prompt

prompt = get_prompt("user-onboarding:working", {"name": "John"})

# OpenAI
import openai
openai.chat.completions.create(
    model="gpt-4",
    messages=[{"role": "user", "content": prompt}],
)

# Anthropic
import anthropic
anthropic.Anthropic().messages.create(
    model="claude-3-sonnet-20240229",
    messages=[{"role": "user", "content": prompt}],
)

🔧 CLI Commands

Full help: promptops --help or promptops <command> --help.

📗 Guides

Runnable examples live in examples/, including Docker, GitHub Actions, hook lifecycle, error handling, and FastAPI.

📁 Project Structure

.promptops/
├── prompts/           # YAML prompt templates (auto-versioned)
├── deploys.jsonl      # Append-only deploy event log (committed to git)
├── snapshot.json      # Production-runtime snapshot (built at CI time)
├── config.yaml        # Hook + tool configuration
├── tests/             # Test datasets
├── results/           # Test reports
├── logs/              # Audit logs
└── reports/           # Auto-generated version change notes

📋 Prompt Schema

PromptOps reads two interchangeable layouts. promptops create prompt scaffolds the compact prompt: form; the richer metadata: form below adds tags, multiple models, and inline tests. Both are parsed the same way.

Extended (metadata:) format:

# .promptops/prompts/user-onboarding.yaml
metadata:
  id: user-onboarding
  version: "1.2.0"            # auto-incremented once you run `promptops hooks install`
  description: "User onboarding welcome message"
  tags: ["onboarding", "welcome"]
  models:                    # NOTE: nested under metadata
    default: gpt-4-turbo
    supported: [gpt-4-turbo, claude-3-sonnet, llama2-70b]

template: |
  Welcome {{ user_name }}!
  Available features:
  {% for feature in features %}
  - {{ feature }}
  {% endfor %}

variables:
  user_name: {type: string, required: true}
  features:  {type: list,   default: ["Browse", "Purchase"]}

tests:
  - dataset: .promptops/tests/onboarding-data.json
    metrics: {max_tokens: 150, min_relevance: 0.8}

Compact (prompt:) format — what promptops create prompt generates:

# .promptops/prompts/user-onboarding.yaml
prompt:
  id: user-onboarding
  description: "User onboarding welcome message"
  model: gpt-4-turbo
  template: "Welcome {{ user_name }}!"
variables:
  user_name: {type: string, required: true}

🔄 Automated Versioning

Automated versioning is opt-in — enable it once per repo with promptops hooks install. (Since v0.4.0, promptops init repo never modifies .git/hooks/ on its own.) After that, semantic version rules applied by the pre-commit hook:

• PATCH (1.0.0 → 1.0.1) — template content changes only
• MINOR (1.0.0 → 1.1.0) — new variables added (backwards compatible)
• MAJOR (1.0.0 → 2.0.0) — required variables removed (breaking change)

Workflow (after promptops hooks install):

1. Edit a prompt → changes in working tree
2. promptops test runtest --prompt name:unstaged (test before commit)
3. git add + git commit → pre-commit hook bumps version and re-stages
4. Post-commit hook tags the new version and generates a changelog entry

Zero manual version management.

🌟 Version References

🛠️ Requirements

• Python 3.10+
• Git (for versioning at dev time — not required at production runtime if you ship snapshot.json)
• YAML + Jinja2 (auto-installed)

📚 Dependencies

• Core: Typer (CLI), Jinja2 (templates), PyYAML (parsing), GitPython (git access)

🤝 Contributing

See CONTRIBUTING.md.

git clone https://github.com/llmhq-hub/promptops.git
cd promptops
python -m venv venv
source venv/bin/activate
pip install -e .
pip install pytest
pytest tests/

📄 License

MIT — see LICENSE.

📞 Support

• Issues: github.com/llmhq-hub/promptops/issues
• Discussions: github.com/llmhq-hub/promptops/discussions
• Releases & changelog: github.com/llmhq-hub/promptops/releases

Made with ❤️ for teams shipping LLMs in production.