A layered testing system for measuring Cline's performance at different levels.

Note: Smoke tests (Layer 2) are partially disabled while the eval framework is repointed at the new SDK CLI. The scenarios under evals/smoke-tests/ are preserved and npm run eval:smoke:run still works against whatever cline is on $PATH (install with npm i -g cline). The old build-and-link helpers and the auto-running cline-evals-regression.yml workflow are off until someone wires the build step at the new SDK CLI.

evals/
├── smoke-tests/           # Quick provider validation (minutes)
│   ├── run-smoke-tests.ts
│   └── scenarios/         # 5 curated test scenarios
│
├── e2e/                   # Full E2E with cline-bench (hours)
│   └── run-cline-bench.ts
│
├── cline-bench/           # Real-world tasks (git submodule)
│   └── tasks/             # 12 production bug fixes
│
├── analysis/              # Metrics and reporting framework
│   ├── src/
│   │   ├── metrics.ts     # pass@k, pass^k calculations
│   │   ├── classifier.ts  # Failure pattern matching
│   │   └── reporters/     # Markdown, JSON output
│   └── patterns/
│       └── cline-failures.yaml
│
└── baselines/             # Performance baselines for regression detection

Location: src/core/api/transform/__tests__/

Tests API transform logic without LLM calls:

• Thinking trace preservation
• Tool call parsing (XML, native formats)
• Provider format conversions

npm run test:unit -- --grep "Thinking\|Tool Call"

Location: evals/smoke-tests/

Quick validation across providers with real LLM calls:

• 5 curated scenarios
• 3 trials per test for pass@k metrics
• Runs the cline CLI with --config, -y, -t, and -m

# Set API key (Cline provider)
export CLINE_API_KEY=sk-...

# Run smoke tests
npm run eval:smoke:run

# Run specific scenario
npm run eval:smoke:run -- --scenario 01-create-file

# Run with specific model (overrides per-scenario models)
npm run eval:smoke:run -- --model anthropic/claude-sonnet-4.5

Location: evals/e2e/ + evals/cline-bench/

Full agent tests on production-grade tasks:

• 12 real-world coding problems
• Docker/Daytona execution via Harbor
• Nightly CI runs

# Prerequisites: Python 3.13, Harbor, Docker
npm run eval:e2e

# Specific task
npm run eval:e2e -- --tasks discord

# Different provider
npm run eval:e2e -- --provider openai --model gpt-4o

The framework calculates:

With 3 trials:

• All pass → pass (reliable)
• All fail → fail (broken)
• Mixed → flaky (needs investigation)

• Current PR gate: contract tests only
• Smoke test CI: temporarily disabled while the workflow is repointed at the SDK CLI
• Nightly: E2E tests with cline-bench are not yet implemented, see TODO

# Run all fast tests
npm run test:unit
npm run eval:smoke:run

# Run E2E (requires setup)
cd evals/cline-bench
# Follow README.md for Harbor setup
npm run eval:e2e

1. Create evals/smoke-tests/scenarios/<name>/config.json
2. Add optional template/ directory with starting files
3. Run to verify: npm run eval:smoke:run -- --scenario <name>

1. Add to src/core/api/transform/__tests__/
2. Run: npm run test:unit -- --grep "YourTest"

Contribute to cline/cline-bench

• cline-bench tasks
• Smoke test scenarios

• Nightly E2E CI: Add scheduled workflow for cline-bench tests
  • Requires: Docker runner, Harbor setup, ~1-2 hour timeout
  • Should run on schedule (e.g., nightly) not per-PR
  • Separate secrets for E2E environment
• Native tool calling smoke tests: Add CLI support for native_tool_call_enabled setting to test Claude 4 with native tools