promptcanary 0.2.2

Detect silent behavioral drift in LLM providers and custom models — before it breaks production.

🐦 PromptCanary

Detect silent behavioral drift in LLM providers — before it breaks production.

---

LLM providers frequently update their models without changing the model string or announcing behavioral shifts. These changes cause silent regressions that are expensive to debug. PromptCanary catches them automatically.

---

The Problem

You ship a production AI assistant. Everything works. Then one day — with no API change, no model rename, no announcement — the behavior shifts. JSON keys reorder. Step-by-step reasoning appears where it didn't before. Refusals trigger on edge cases that passed last week. Your downstream parser breaks. Your agent loop fails.

Silent model drift is real, common, and expensive.

The Solution

pip install promptcanary
promptcanary init my-suite
promptcanary run --provider openai/gpt-4o --save-baseline

# One week later... run again and compare:
promptcanary compare --provider openai/gpt-4o --fail-on-drift

⚠️  HIGH drift in 'my-suite': 3 regression(s) detected
    Score: 94.0% → 71.0% (Δ -23.0%)

┌─ Regressions ────────────────────────────────────────────────────────────────┐
│ JSON Validity    │ format   │ prompt_3 │ 1.00 → 0.00 │ Δ -1.00 │ JSON parse  │
│ Direct Answer    │ reason   │ prompt_1 │ 1.00 → 0.00 │ Δ -1.00 │ Preamble    │
│ Response Length  │ format   │ prompt_5 │ 0.92 → 0.51 │ Δ -0.41 │ 4x longer   │
└──────────────────────────────────────────────────────────────────────────────┘

---

Features

• 🔌 Works with any provider — OpenAI, Anthropic, Google, Ollama, vLLM, and 100+ more via LiteLLM
• 📋 15 built-in probes across format, reasoning, safety, and factual categories
• 🔧 Custom probes — one decorator, zero boilerplate
• 📊 Rich reports — terminal, Markdown, HTML, and JSON
• 🤖 GitHub Actions native — scheduled checks, PR comments, auto-issue on drift
• ⚡ Fast setup — working canary in under 10 minutes
• 🧪 CI-ready — --fail-on-drift exits non-zero for automated gating
• 🏗️ Clean architecture — Pydantic v2, fully typed, 80%+ test coverage

---

Quick Start

Install

pip install promptcanary

1. Scaffold a suite

promptcanary init my-suite
cd my-suite

Edit canary.yaml to describe the prompts and behaviors that matter to you:

name: my-production-suite
probes:
  - type: json_validity
  - type: direct_answer
    expect_direct: true
  - type: refusal
    expect_refusal: false
prompts:
  - text: "Return JSON: {name: 'Alice', role: 'engineer'}"
    description: "Core JSON format canary"
  - text: "What is the capital of France? One sentence."
    expected_keywords: ["Paris"]

2. Run and save baseline

export OPENAI_API_KEY=sk-...
promptcanary run --provider openai/gpt-5.4 --save-baseline

┌─ PromptCanary Run Report ────────────────────────────────────────────────────┐
│ my-production-suite  ·  Score: 100.0%  ·  Pass rate: 100.0%  ·  openai/gpt-5.4
└──────────────────────────────────────────────────────────────────────────────┘
✅ All probes passed.
✅ Baseline saved: baselines/my-suite__openai-gpt-5.4__20260629T090000_abc12345.json

3. Detect drift

# Run whenever you want to check — daily, weekly, or in CI:
promptcanary compare --provider openai/gpt-5.4 --fail-on-drift

Prefer a free, local model? Use Ollama instead — no API key needed:

ollama pull qwen3.6:27b
promptcanary run --provider ollama/qwen3.6:27b --save-baseline
promptcanary compare --provider ollama/qwen3.6:27b --fail-on-drift

---

Python SDK

from promptcanary import (
    CanarySuite, LiteLLMProvider, FileBaselineStore,
    CanaryPrompt, JsonValidityProbe, StepByStepProbe,
    KeywordPresenceProbe, compare,
)
from promptcanary.core.reporter import Reporter, DriftReporter

# ── Build suite ───────────────────────────────────────────────────────────────
suite = CanarySuite(
    name="production-agent",
    prompts=[
        CanaryPrompt(
            text='Return JSON: {"action": "search", "query": "Paris weather"}',
            expected_keywords=["action", "query"],
        ),
        CanaryPrompt(
            text="What is the capital of France? One sentence.",
            expected_keywords=["Paris"],
        ),
    ],
    probes=[
        JsonValidityProbe(),
        KeywordPresenceProbe(required_keywords=["Paris"]),
        StepByStepProbe(expect_steps=False),
    ],
)

provider = LiteLLMProvider("openai/gpt-5.4", temperature=0.0)

# ── Run ───────────────────────────────────────────────────────────────────────
result = suite.run(provider)
Reporter(result).print_terminal()

# ── Save baseline ─────────────────────────────────────────────────────────────
store = FileBaselineStore("baselines/")
snapshot = store.save(result)

# ── Compare ───────────────────────────────────────────────────────────────────
new_result = suite.run(provider)
drift = compare(snapshot, new_result)
DriftReporter(drift).print_terminal()

if drift.has_drift:
    print(drift.summary)
    # ⚠️ HIGH drift in 'production-agent': 2 regression(s) detected ...

Using Google Gemini instead

provider = LiteLLMProvider("gemini/gemini-3.5-flash", temperature=0.0)
result = suite.run(provider)

Using a free, local model via Ollama (no API key, zero cost)

# 1. Install Ollama and pull a model: `ollama pull qwen3.6:27b`
provider = LiteLLMProvider("ollama/qwen3.6:27b", temperature=0.0)
result = suite.run(provider)

Load from YAML

from promptcanary import CanarySuite, LiteLLMProvider

suite = CanarySuite.from_yaml("canary.yaml")
provider = LiteLLMProvider("anthropic/claude-sonnet-4-6")
result = suite.run(provider)

---

Built-in Probes

Format & Structure

Probe	Detects
JsonValidityProbe	Invalid JSON output
JsonSchemaProbe(required_keys=[...])	Missing or forbidden JSON keys
JsonKeyOrderProbe(expected_order=[...])	Key reordering in JSON output
ResponseLengthProbe(min=10, max=4000)	Length explosions or sudden brevity
MarkdownHeaderProbe(expected_headers=[...])	Missing section headers
KeywordPresenceProbe(required=[...], forbidden=[...])	Keyword drift
ExpectedKeywordsProbe	Keywords declared on CanaryPrompt.expected_keywords

Reasoning Style

Probe	Detects
StepByStepProbe(expect_steps=True)	Loss or gain of chain-of-thought reasoning
VerbosityProbe(expected_words=200)	Word count drift
ConfidenceLanguageProbe(expect_hedging=False)	Hedging vs. confident language shifts
DirectAnswerProbe(expect_direct=True)	"Sure!", "Great question!" preamble

Safety & Refusal

Probe	Detects
RefusalProbe(expect_refusal=False)	Unexpected refusals (or missing ones)
SafetyLanguageProbe(expect_safety_language=False)	New disclaimer injection

Tool Use (Agent Workflows)

Probe	Detects
ToolCallPresenceProbe(expect_tool_call=True)	Missing or unexpected function calls
ToolCallNameProbe("search_web", allow_aliases=[...])	Wrong function called
ToolCallArgsProbe(required_args=[...], forbidden_args=[...])	Missing/leaked arguments
ToolCallSchemaProbe(schema={...})	Full structural validation (name + args + types)

Factual

Probe	Detects
FactualConsistencyProbe("Paris")	Drift from known-correct answer
SentimentProbe(expect_positive=None)	Tone shifts

---

Custom Probes

from promptcanary.core.probes.base import probe
from promptcanary.core.models import CanaryPrompt, LLMResponse, ProbeCategory, ProbeResult

@probe("tool_call_format", name="Tool Call Format", category=ProbeCategory.CUSTOM)
def check_tool_call(prompt: CanaryPrompt, response: LLMResponse) -> ProbeResult:
    """Verify the model always calls the search tool when asked to search."""
    has_tool_call = '"function": "search"' in response.content
    return ProbeResult(
        probe_id="tool_call_format",
        probe_name="Tool Call Format",
        category=ProbeCategory.CUSTOM,
        prompt_id=prompt.id,
        passed=has_tool_call,
        score=1.0 if has_tool_call else 0.0,
        details="Tool call found." if has_tool_call else "Expected tool call missing.",
    )

Custom probes are auto-registered and can be used in YAML configs by their probe_id.

---

GitHub Actions Integration

Add to .github/workflows/promptcanary.yml:

name: PromptCanary Drift Check
on:
  schedule:
    - cron: "0 9 * * 1"    # Every Monday
  workflow_dispatch:

jobs:
  canary:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: pip install promptcanary
      - name: Run and compare
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        run: |
          promptcanary run --provider openai/gpt-5.4 --output-json results.json
          promptcanary compare --current results.json --baseline baselines/latest.json --fail-on-drift

On drift, PromptCanary will:

• Print a detailed terminal report
• Exit with code 1 (fails the job)
• Optionally post a PR comment or open a GitHub issue (see .github/workflows/promptcanary.yml)

---

Supported Providers

PromptCanary works with any provider supported by LiteLLM — cloud or local, paid or free.

Provider	Example model string	API key env var	Cost
OpenAI	openai/gpt-5.5 (flagship), openai/gpt-5.4 (balanced), openai/gpt-5.4-mini (fast)	OPENAI_API_KEY	Paid
Anthropic	anthropic/claude-opus-4-8 (flagship), anthropic/claude-sonnet-4-6 (balanced)	ANTHROPIC_API_KEY	Paid
Google Gemini	gemini/gemini-3.1-pro (flagship), gemini/gemini-3.5-flash (balanced), gemini/gemini-3.1-flash-lite (fast)	GEMINI_API_KEY	Paid
xAI	xai/grok-4	XAI_API_KEY	Paid
Ollama (local, free)	ollama/qwen3.6:27b, ollama/deepseek-r1:14b, ollama/gpt-oss:20b, ollama/llama3.3:8b	(none)	Free
vLLM (self-hosted)	hosted_vllm/<org>/<model>	(none)	Free (self-hosted compute)

Model availability changes quickly. Run litellm --test or check LiteLLM's provider docs for the latest model strings before relying on any example here.

# Cloud providers
LiteLLMProvider("openai/gpt-5.4")
LiteLLMProvider("anthropic/claude-sonnet-4-6")
LiteLLMProvider("gemini/gemini-3.5-flash")
LiteLLMProvider("xai/grok-4")

# Local models — zero cost, full privacy, no API key required
LiteLLMProvider("ollama/qwen3.6:27b")        # strong general-purpose, Apache 2.0
LiteLLMProvider("ollama/deepseek-r1:14b")    # reasoning-focused, MIT licensed
LiteLLMProvider("ollama/gpt-oss:20b")        # OpenAI's open-weight release
LiteLLMProvider("hosted_vllm/meta-llama/Llama-3.3-8B-Instruct")

Why test local models too?

Local, open-weight models (via Ollama or vLLM) make excellent zero-cost canaries: running them hourly costs nothing and catches infrastructure-level regressions (prompt template bugs, parser issues) independent of any vendor's API changes. See notebooks/ci_integration.ipynb for a cost-aware multi-provider scheduling strategy that mixes free local models with periodic paid-provider checks.

---

Report Formats

Every run produces multiple output formats:

promptcanary run --provider openai/gpt-5.4 \
  --output-json results.json \
  --output-md report.md \
  --output-html report.html

• Terminal — colour-coded table with scores and details
• Markdown — GitHub-flavoured, ideal for PR comments
• HTML — self-contained dark-theme interactive report
• JSON — machine-readable for downstream automation

---

Trend Visualization

Track score history, per-probe heatmaps, and drift timelines across multiple runs:

from promptcanary.storage.file import FileBaselineStore
from promptcanary.utils.visualization import plot_score_history, plot_probe_heatmap

store = FileBaselineStore("baselines/")
snapshots = [store.load_from_path(p) for p in sorted(Path("baselines").glob("*.json"))]

# Works everywhere — zero dependencies, ASCII sparkline in any terminal
plot_score_history(snapshots, mode="ascii")

# Interactive HTML — requires: pip install promptcanary[viz]
plot_score_history(snapshots, output_path="trend.html")
plot_probe_heatmap(snapshots, output_path="heatmap.html")

See notebooks/analyzing_drift_trends.ipynb for a full walkthrough of identifying which probe regresses first during gradual drift.

Architecture

promptcanary/
├── core/
│   ├── models.py       # Pydantic v2 domain types
│   ├── suite.py        # CanarySuite orchestrator
│   ├── comparator.py   # Drift comparison engine
│   ├── reporter.py     # Terminal/MD/HTML/JSON output
│   └── probes/         # 19 built-in probes + registry
├── providers/
│   ├── base.py         # BaseLLMProvider ABC
│   └── litellm.py      # LiteLLM adapter
├── storage/
│   └── file.py         # Local JSON baseline storage
├── utils/
│   └── visualization.py # Trend charts (ASCII + optional Plotly)
└── cli.py              # Typer CLI

Key design choices:

• All models are Pydantic v2 with full type hints and JSON serialisation
• Probes are stateless, auto-registered, and composable
• No network calls in the core — only in provider adapters
• BaselineStore and BaseLLMProvider are ABCs enabling custom backends

---

Contributing

We welcome contributions of all kinds! The highest-value contributions are:

• New probes — especially for specific domains (tool use, agents, legal, medical)
• Community canary suites — canary.yaml examples for specific use cases
• Storage backends — S3, GCS, database adapters
• Bug reports with reproduction cases

See CONTRIBUTING.md for setup, conventions, and the PR checklist.

---

Roadmap

Done in v0.1.x:

• ToolCallProbe family — function name + argument schema stability for agent workflows
• Trend visualization — score history, probe heatmaps, drift timelines (ASCII + Plotly)
• Property-based tests (Hypothesis) for comparator and scoring invariants
• Full CLI test coverage

Planned:

• suite.arun() — async parallel execution
• Export connectors — Langfuse, Phoenix/Arize
• SemanticSimilarityProbe — embedding-based semantic drift
• S3 / GCS baseline storage backends
• Optional web dashboard

---

License

MIT — see LICENSE.

---

Built with ❤️ for AI engineers who care about production reliability.