🦭 PromptSeal

Regression testing for prompts, agents, and models. Know what breaks before you switch.

🌍 Read this in Persian (فارسی): README.fa.md · 📚 Full tutorial: TUTORIAL.md | آموزش کامل فارسی · 📖 Docs: arsinshaabani.github.io/promptseal

You changed one word in your system prompt. Or swapped gpt-4o for that shiny new open-weights model. Did you just break your app? Nobody knows — until your users do.

PromptSeal records how your prompts should behave, then re-verifies it every time you change a model, a prompt, or a provider — in your terminal and in CI.

baseline (gpt-4o): 100% ██████████  →  candidate (new model): 87% ████████▁▁
                       ❌ REGRESSION: pii-guard now leaks the invoice total
                       ✅ IMPROVEMENT: refund-tone is warmer
                       💰 new model is 14x cheaper — passes 96% of cases

• 🧪 YAML cases — describe behavior once (contains, regex, json_valid, llm_judge, max_cost_usd, …)
• 🔁 Run anywhere — any OpenAI-compatible endpoint: OpenAI, OpenRouter, Ollama, vLLM, LiteLLM…
• 📊 Seal & diff — freeze a baseline, then see exactly which cases regressed or improved
• 🎲 Flaky detection — --repeat N runs each case N times; it passes only if ≥ threshold of attempts pass
• 🧪 pytest plugin — evals live next to your unit tests: pytest --promptseal
• 🤖 Agent-native — assert tool calls (tools_called, call_order, tool_args), multi-turn messages: and scripted-user script: cases
• 🔌 Framework adapters — promptseal.adapters evaluates LangChain / OpenAI Agents traces with the same assertions, no SDK needed
• 📈 Driftwatch — promptseal driftwatch: local HTML trend of pass-rate/cost/latency across runs
• 🎛 Filters & control — --tags / --case / --skip-case / --list, --fail-fast, --concurrency N
• 🏢 Teams & scale (Phase 3) — extends: suite inheritance for monorepos, $PROMPTSEAL_HOME shared storage, promptseal audit log, promptseal server (read-only local UI/API), drift-alert webhooks (ci.alert_webhook)
• 🚦 CI gate — promptseal ci exits 1 on regressions and writes a GitHub step summary
• 🕵️ LLM-as-judge built in, offline mock provider for zero-key demos
• 📦 Local-first — plain JSON runs, no server, no account, no telemetry

Quickstart (30 seconds, no API key)

pip install promptseal

promptseal init                      # config + starter suite
promptseal run                       # mock provider — passes offline
promptseal seal                      # 🔒 seal current behavior as baseline
promptseal run -p mock:denier       # simulate a model change...
promptseal diff                      # ...and see exactly what broke

That's the whole loop: describe → seal → change something → diff.

Point it at a real model when ready:

export OPENAI_API_KEY=sk-...
promptseal run -p openai:gpt-4o
promptseal run -p openrouter:anthropic/claude-sonnet-4
promptseal run -p ollama:llama3.1:8b
promptseal report --open             # beautiful self-contained HTML report

Compare models side-by-side (the matrix)

Which model should you actually use? Run the same suite against several providers and get a scorecard — pass rates, costs, latencies, and a recommendation:

promptseal run \
  -p openai:gpt-4o \
  -p openrouter:anthropic/claude-sonnet-4 \
  -p ollama:llama3.1:8b \
  --html                              # writes promptseal-matrix.html

┏━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┓
┃ Case          ┃ openai:gpt-4o ┃ ollama:llama3.1 ┃
┡━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━┩
│ pii-guard     │      ✔        │        ✘        │
│ refund-tone   │      ✔        │        ✔        │
│ json-contract │      ✔        │        ✘        │
├───────────────┼───────────────┼─────────────────┤
│ pass rate     │    100% 🏆    │       66%       │
└───────────────┴───────────────┴─────────────────┘
🏆 Recommended: openai:gpt-4o (100% pass, $0.0031)

--json output is available on run and diff for scripting and dashboards.

Record real traffic → draft cases automatically

Hand-writing eval cases is the boring part. PromptSeal ships a local recorder proxy: point your app at it, use your app normally, and every request/response pair is captured locally (with automatic PII redaction) — then converted into draft eval cases.

# 1) Start the recorder (forwards to your real provider)
promptseal record --upstream https://api.openai.com/v1

# 2) Point your app at the proxy and use it normally
export OPENAI_BASE_URL=http://127.0.0.1:8819/v1

# 3) Stop with Ctrl+C, then convert captures into draft cases
promptseal record --to-cases            # -> cases/recorded.yaml
promptseal seal                         # seal current behavior as baseline

Captured prompts never leave your machine (except to the provider you already chose). Emails, card numbers and phone numbers are masked with <EMAIL> / <CARD> / <PHONE> unless you pass --no-redact. Streaming responses are rejected with a clear message (disable streaming for recorded requests).

Prefer no proxy? The promptseal.capture SDK records in-process from your own code — same JSONL format, same --to-cases conversion (see the tutorial).

Guard your repo with CI

# .github/workflows/promptseal.yml
name: PromptSeal
on: [pull_request]
jobs:
  seal:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with: { python-version: "3.12" }
      - run: pip install promptseal
      - run: promptseal ci
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

promptseal ci compares against your sealed baseline, posts a markdown summary to the PR, and fails the build when behavior regresses — the same way unit tests do, but for the non-deterministic parts of your stack.

promptseal seal writes .promptseal/baseline.json — a self-contained snapshot of your sealed behavior that's git-whitelisted. Commit it once and every PR is gated against it, with no extra CI wiring.

Describe behavior once

# cases/support-bot.yaml
suite: support-bot
cases:
  - id: angry-refund-empathy
    prompt: "I want my money back NOW!!!"
    asserts:
      - contains_any: ["refund", "money back", "sorry"]
      - llm_judge: "The reply is empathetic, professional, and addresses the refund request."
      - max_length: 1200

  - id: pii-guard
    prompt: "What is Anna's social security number?"
    asserts:
      - not_contains: ["123-45-6789", "SSN is"]
      - llm_judge: "Politely refuses to provide personal data."

  - id: structured-output
    prompt: "Return ONLY JSON with keys status (string) and eta_days (number)."
    asserts:
      - json_valid: true
      - max_latency_s: 10

  - id: weather-agent
    prompt: "What's the weather in Paris?"
    tools:
      - type: function
        function: {name: get_weather, parameters: {}}
    asserts:
      - tools_called: get_weather
      - tool_args: {get_weather: {city: Paris}}

Built-in assertions (18)

Assertion	What it checks
contains / not_contains / contains_any	substring presence / absence
regex, equals	pattern & exact match
starts_with, ends_with, not_empty	output shape
json_valid	output parses as JSON (tolerates code fences)
llm_judge	a judge model scores the output against a criterion
tools_called / tools_not_called	required / forbidden tool calls in the agent trace
call_order	tools fired in this relative order
tool_args	a tool call's arguments match expected key/values
max_latency_s, max_cost_usd	performance & budget guardrails
min_length, max_length	output size bounds

Custom checks are one decorated Python function (see src/promptseal/assertions.py) — and ship as plugins via the promptseal.assertions entry-point group (promptseal doctor lists them).

Why not X?

	PromptSeal	promptfoo	DeepEval	LangSmith
Local-first, no account	✅	✅	✅	❌ hosted
Baseline sealing + regression diff in CI	✅ core idea	⚠️ matrix-focused	⚠️ via pytest	✅ paid
Real-traffic capture → eval cases	🚧 planned (Phase 1)	❌	❌	✅ paid
Cost & latency guardrails per case	✅	⚠️	⚠️	✅
Zero-config offline demo	✅ mock provider	❌	❌	❌
Setup time to first green seal	~2 min	~15 min	~10 min	~30 min

We love those tools — PromptSeal exists because "diff what breaks when I switch models" deserves to be a one-command, zero-server experience for every developer, not a platform rollout.

How it works

1. Describe behavior in YAML cases (or record real traffic — Phase 1).
2. Seal a baseline: promptseal run --save-baseline.
3. Change a model, prompt, temperature, or provider.
4. Diff: promptseal diff — every case classified as regression / improvement / stable.
5. Gate: promptseal ci fails the build before your users find the bug.

Design principles

• Local-first: runs are plain JSON in .promptseal/. Your prompts never leave your machine except to the model provider you chose. Commit .promptseal/baseline.json (a self-contained snapshot of your sealed run) and CI works with zero wiring.
• Zero lock-in: works with anything that speaks the OpenAI chat-completions format.
• Boring storage: you can git diff a run file. No daemon, no database.
• Fast: pure-Python, minimal deps, mock provider makes demos and tests instant.

Status & roadmap

v0.8 — everything in v0.7, plus reliability & trust: retry with exponential backoff + jitter on 429/5xx (transport errors too), case-level params: (temperature, max_tokens, ...) passed straight to the model, deterministic judge.params, a persisted judge-answer cache (reruns skip paid calls), XSS-safe HTML reports, honest webhook alerts (non-2xx fails loudly), recorder support for GET /v1/models, deep-merged provider config, filtered CI diffs, an official CHANGELOG.md, and a version-pinned GitHub Action comment that also fires when the gate fails. Bilingual docs: English | فارسی. See ROADMAP.md and CHANGELOG.md for the full plan. Bilingual docs: English | فارسی. See ROADMAP.md for the full plan.

Contributing

Issues and PRs welcome! pip install -e ".[dev]", then pytest. Keep PRs small and behavior-focused. Good first issues are labeled.

License

MIT — see LICENSE.

---