feat: shared assertion templates (include_assertions)

[christso]

Problem

Users frequently reuse the same assertion sets across multiple tests and eval files — safety checks, format validation, tone requirements. Currently they must copy-paste assertion blocks, leading to:

• Duplication across EVAL.yaml files
• Drift when updating shared criteria
• Verbose eval files that obscure test-specific logic

Proposed Design

Reusable assertion template files

# .agentv/templates/safe-response.yaml
assertions:
  - type: llm-grader
    prompt: ./graders/no-hallucination.md
    required: true
  - type: llm-grader
    prompt: ./graders/safe-content.md
    required: true
    min_score: 0.9
  - type: contains
    value: "disclaimer"
    negate: true

Reference from EVAL.yaml

# evals/EVAL.yaml
tests:
  - id: refund-request
    input: "I want a refund"
    assertions:
      - include: safe-response          # resolves .agentv/templates/safe-response.yaml
      - type: llm-grader
        prompt: ./graders/refund-quality.md   # test-specific assertion
  - id: greeting
    input: "Hello"
    assertions:
      - include: safe-response          # same shared template
      - type: contains
        value: "hello"

Suite-level includes

# Apply to all tests
assertions:
  - include: safe-response

tests:
  - id: test-1
    input: "..."
    # inherits safe-response assertions + can add test-specific ones

Resolution rules

1. include: name resolves to .agentv/templates/{name}.yaml
2. Relative paths also work: include: ./my-templates/safety.yaml
3. Templates can include other templates (max depth 3 to prevent cycles)
4. Test-level assertions merge with included assertions (not replace)
5. skip_defaults: true on a test still skips suite-level includes

Template Location

Two resolution mechanisms, no config needed:

1. Convention directory: include: safe-response resolves to .agentv/templates/safe-response.yaml
2. Relative path: include: ./my-templates/safety.yaml resolves relative to the eval file

Use case	How
Shared across repo	.agentv/templates/ (convention)
Co-located with evals	include: ./shared/safety.yaml
Shared across eval suites	include: ../../common/safety.yaml
Monorepo shared	include: ../../../packages/evals/templates/safety.yaml

Relative paths are explicit and traceable — no ambiguity about which directory a template resolved from.

Implementation

Files to modify

1. packages/core/src/evaluation/validation/eval-file.schema.ts — add include variant to EvaluatorSchema
2. packages/core/src/evaluation/loaders/evaluator-parser.ts — resolve include references, load template files, flatten into assertion list
3. packages/core/src/evaluation/yaml-parser.ts — handle include resolution during test loading

Template discovery

.agentv/templates/           # convention directory
  safe-response.yaml
  json-output.yaml
  professional-tone.yaml

Discovery chain (closest wins): {eval-dir}/.agentv/templates/ → {repo-root}/.agentv/templates/

Template file format

Same as an assertion block — a YAML file with a top-level assertions array. Each entry is a standard evaluator config. This keeps templates authorable by AI agents (same schema as inline assertions).

Research Context

Inspired by RSpec's shared_examples pattern — define reusable test behaviors that can be included with context. See test-framework-assertion-patterns research.

Dependencies

• feat: configurable threshold + naming taxonomy cleanup #925 must land first (introduces min_score field used in template examples)

Acceptance Signals

1. include: name resolves correctly — loads from .agentv/templates/{name}.yaml — verified by test
2. include: ./path resolves relative paths — verified by test
3. Suite-level includes apply to all tests unless skip_defaults: true — verified by test
4. Test-level includes merge with test-specific assertions — verified by test
5. Nested includes work up to depth 3; depth > 3 produces a clear error — verified by test
6. Missing template produces a clear error with the resolved path — verified by test
7. Schema validation accepts include entries in assertion arrays — verified by test
8. eval-schema.json regenerated with include support
9. All existing tests pass — no regressions

Non-goals

• Template parameters/variables (keep it simple — templates are static assertion sets)
• Template versioning
• Remote template registries
• Cross-repo template sharing (use relative paths with monorepo layout)
• Configurable template_dirs in config.yaml (relative paths already cover custom locations — a config-based search path creates "which directory did this resolve from?" debugging headaches)
• Template inheritance/override (templates are flat assertion lists, not class hierarchies)