Active

PR Review Recipes

Automating code quality with deterministic workflows

January 2026

The Goal

Automate PR review for amplifier-app-cli using Amplifier's recipe system.

Version 1

Built confidence with foundational checks

Ruff
Linting & formatting
Pyright
Type checking
CLI
--help, --version

Solid foundation. But was it enough?

The Gap

A routine PR merged. All checks passed.

Then it failed at runtime.

The issue only appeared when users created actual sessions — something our recipe never tested.

Root Cause

We tested that it launched. We didn't test what users actually do.

✗ Tested

  • amplifier --help
  • amplifier --version
  • Static code analysis

✓ Needed

  • Real session creation
  • Tool invocations
  • Provider connections
Same Day

Discovered → root-caused → improved → validated

Discovered
Runtime failure
Root Cause
Gap identified
Improved
Recipe updated
Validated
Re-tested original PR

How the Recipe Works Now

Seven phases, fully automated

1 FETCH
2 CLONE
3 ANALYZE
4 SHADOW
5 SMOKE
6 REVIEW
7 REPORT
# PR metadata from GitHub  FETCH
# Clone PR branch + amplifier-core  CLONE
# Cross-repo contract checking  ANALYZE
# Create isolated container  SHADOW
# Real sessions, tools, providers  SMOKE
# Code quality + security audit  REVIEW
# Generate findings report  REPORT

Two Layers of Detection

🔍 Static Analysis

  • Contract checking — API signatures between repos
  • Breaking changes — Removed params, changed types
  • Import validation — Module dependencies

Catches: What could break

Dynamic Testing

  • Shadow environment — Isolated container
  • Real user flows — Session creation, tool calls
  • Runtime validation — Module loading, execution

Catches: What does break

Shadow Environment

Test what users actually do, in complete isolation

Install from PR code
Create real sessions
Invoke tools
Connect providers
Run smoke tests

Safe. Repeatable. Tests the full user experience before merge.

Validation

Tested the improved recipe against the PR that got through

✓ Static
Contract analysis caught the API signature mismatch between repos
✓ Dynamic
Shadow smoke tests caught the runtime session-creation failure

Both layers now catch it. The gap is closed.

The Pattern

Each gap becomes the next improvement

🔨
Build
🧪
Test
💡
Learn
🔧
Improve
🔄
Repeat

Not a failure. A feedback loop.

Velocity

74 commits across 44 days — Dec 15, 2025 to Jan 28, 2026

Brian Krabach
68 (92%)
Marc Goodner
3
Mollie Munoz
1
Sam Schillace
1
momuno
1
6 stages
code-review recipe with model routing: haiku for classification, sonnet for analysis, opus for strategy
2-in-1
comprehensive-review.yaml composes code-review + security-audit as sub-recipes

Sources & Methodology

Data Collection

  • Recipe file analysis: find + cat on review recipe YAML files
  • Git history: git log --oneline (74 commits found)
  • Contributors: git log --format="%an" | sort | uniq -c
  • Timeline: First commit Dec 15, 2025; latest Jan 28, 2026

Repository

microsoft/amplifier-bundle-recipes

Key files analyzed:

  • examples/code-review-recipe.yaml
  • examples/comprehensive-review.yaml

Attribution

Primary contributor: Brian Krabach (~92% of 74 commits)

Additional contributors: Marc Goodner (3), Mollie Munoz (1), Sam Schillace (1), momuno (1)

Known Gaps

  • Exact runtime-failure PR not identified by SHA
  • V1 → V2 timeline is narrative from existing deck, not commit-verified

Data as of

February 20, 2026

Feature status

Active

Ship with Confidence

PR Review Recipes — Now in Amplifier

1 / 14