Migrating from Other Systems

This guide helps teams migrate from custom eval scripts or other frameworks to Vally. The core idea: your existing concepts map cleanly to Vally’s model.

Concept mapping

Your concept	Vally equivalent	Notes
Test case / example	Stimulus	A prompt + grader config in eval.yaml
Assertion / check	Grader	Implements the Grader interface
Test run / execution	Trajectory	Flat event log captured during agent execution
Test result	GraderResult	{ passed, score, evidence }
Test suite / harness	eval.yaml	Collection of stimuli with scoring config
Test config / profile	eval.yaml config	Execution settings (runs, timeout, model, executor, judge_model)
CI check	vally lint or eval	CLI commands with exit codes
Gold reference / expected output	reference-based grader input	Passed as trajectoryB in pairwise graders
Test report	LintReporter / EvalReporter	LintConsoleReporter, EvalConsoleReporter built-in, extensible

Migration steps

1. Inventory your existing checks

   List every assertion your current system makes. For each one, categorize it:

   Check	Type	Vally grader
   ”Output contains X”	String match	output-contains (built-in)
   “Output does NOT contain X”	Negated string match	output-not-contains (built-in)
   “File was created”	File check	file-exists (built-in)
   “File was NOT created”	Negated file check	file-not-exists (built-in)
   “File contains pattern”	Content check	file-contains (built-in)
   “File does NOT contain text”	Negated content check	file-not-contains (built-in)
   “Output matches regex”	Pattern match	output-matches (built-in)
   “Output does NOT match regex”	Negated pattern match	output-not-matches (built-in)
   “Command exits 0”	Script check	run-command (built-in)
   “Agent produced output”	Session check	exit-success (built-in)
   “Session completed cleanly”	Session health	completed (built-in)
   “LLM says output is good”	LLM judge	prompt (built-in)
   “Compare A vs B”	A/B comparison	pairwise (built-in)

2. Convert test cases to stimuli

   Each test case becomes a stimulus in eval.yaml:

   Before: custom script

   # run_test("Write tests for add()", expect_file="add.test.js")

   After: eval.yaml

   stimuli:
     - name: test-generation
       prompt: "Write tests for the add() function"
       graders:
         - type: file-exists
           config:
             path: "add.test.js"

3. Convert custom assertions to graders

   For checks that don’t map to built-in graders, implement the Grader interface:

   import type { Grader, GraderInput, GraderResult, GraderMetadata } from "@microsoft/vally";
   
   export class MyCustomCheck implements Grader {
     metadata: GraderMetadata = {
       name: "my-custom-check",
       description: "Checks something specific to my domain",
       determinism: "complex-static",
       costProfile: "low",
       portability: "t2-domain",
       reference: "reference-free",
       temporalScope: "trajectory-level",
     };
   
     async grade(input: GraderInput): Promise<GraderResult> {
       // Port your existing assertion logic here
       const passed = /* your check */;
       return {
         name: this.metadata.name,
         kind: "code",
         passed,
         score: passed ? 1 : 0,
         evidence: passed ? "Check passed" : "Check failed because...",
       };
     }
   }

4. Set up your eval.yaml

   Combine your stimuli and configure scoring:

   eval.yaml

   name: migrated-eval-suite
   type: capability
   config:
     runs: 3
     timeout: 120s
     model: gpt-5.5
   stimuli:
     # All your converted test cases
   scoring:
     threshold: 0.7

5. Update CI

   Replace your existing eval CI step with Vally:

   .github/workflows/eval.yml

   - run: npm install -g @microsoft/vally-cli
   - run: vally lint .
   - run: vally eval --eval-spec eval.yaml --output-dir ./results

What you gain

After migration, you get several things that are hard to build with custom scripts:

Feature	Custom scripts	Vally
Shared grader taxonomy	❌ Ad-hoc	✅ Declared per grader
Multi-trial metrics (pass@k)	❌ Build yourself	✅ Built-in
Trajectory capture + replay	❌ Build yourself	✅ Built-in
Re-grade without re-running agent	❌ Not possible	✅ vally grade
Shared graders across teams	❌ Copy-paste	✅ Plugin registry

Tip

You don’t have to migrate everything at once. Start with lint (free, instant) and add eval stimuli incrementally.

Next steps

• Writing eval specs — advanced patterns
• Writing custom graders — full guide
• Add to CI — GitHub Actions setup