promptkit/docs/policy/testing.md

# Testing Policy

## Purpose

Our tests exist to make **incorrect changes expensive and correct changes cheap**.

We do not optimize for test count, line coverage, exhaustive isolation, or the fewest possible tests. We optimize for sufficient confidence in important behavior while imposing as little unnecessary friction as possible on future development.

## Every test has a cost

Testing is not an unqualified good. Every test imposes both an immediate cost and a continuing lifetime cost.

A test must be:

- written and reviewed;
- understood by future maintainers and coding agents;
- executed in maintainer-run validation;
- diagnosed when it fails;
- updated when legitimate behavior changes;
- maintained as fixtures, APIs, and dependencies evolve; and
- removed or rewritten when it becomes redundant, brittle, misleading, or obsolete.

Tests also create cognitive and architectural friction. They can constrain refactoring, duplicate policy, slow feedback loops, add noise to failures, and cause harmless implementation changes to require unrelated edits across the suite.

A test is warranted only when the confidence it provides justifies these costs.

Apply this cost-benefit analysis at two levels:

1. **Per test:** What realistic defect does this test detect, how consequential would that defect be, and is that protection worth the test's lifetime cost?
2. **Across the suite:** Does this collection provide materially more confidence than a smaller, simpler suite would?

The preferred test suite is a **lean suite that provides sufficient confidence in the risks that matter, without redundant or low-value tests**. We seek sufficient confidence with the least unnecessary testing friction, not the fewest possible tests.

Some friction is intentional. Tests should make dangerous changes—such as breaking compatibility, corrupting data, violating security boundaries, or reintroducing subtle bugs—require deliberate review. They should not make ordinary internal changes needlessly expensive.

The cost of a test is not a reason to omit testing by default. Do not cite maintenance cost abstractly. When omitting a plausible test, be able to state why the protected failure is low-risk, already covered, obvious, reversible, or cheaper to detect elsewhere. For consequential, subtle, or difficult-to-observe behavior, the presumption should favor testing.

## Default testing style

Use a **classical/Detroit-style** approach:

- Test observable behavior, resulting state, contracts, and invariants.
- Use real internal collaborators when they are fast and deterministic.
- Use fakes, stubs, or mocks primarily at expensive, nondeterministic, destructive, or external boundaries.
- Prefer package-level behavioral tests over tests coupled to private helpers or internal call sequences.
- Treat exact collaborator interactions as testable behavior only when the interaction itself is a requirement.

Examples of appropriate seams include clocks, randomness, subprocesses, remote APIs, object storage, email, and paid LLM calls.

## Test execution requirements

Promptkit currently uses maintainer-run validation rather than hosted CI.
Maintainers run the complete local workflow in the
[development guide](../development.md#maintainer-validation) before accepting
changes. That guide is the canonical owner of exact commands, formatting,
documentation-link validation, and repository-hygiene checks.
Introducing hosted CI later would supplement, not silently redefine, this
documented validation model.

Maintainer validation must include ordinary and race-enabled package tests,
static analysis, a complete build, and execution of both maintained offline
consumer examples. The preparation example protects assembled preparation and
inspection behavior. The execution example separately protects assembled
`Run`, injected-client, validation, usage, and result behavior.

Tests in the default suite must be deterministic, offline, and independent of
real credentials. They must not invoke paid APIs, use live network
dependencies, or depend on mutable external services. Tests that require live
infrastructure must be explicitly opt-in and clearly separated from the
default suite.

Control clocks, randomness, environment variables, and other process-global or
machine-specific state when they affect behavior. Tests must be parallel-safe:
they should run repeatedly and alongside other tests without depending on
execution order, shared mutable state, fixed ports, or state left by an earlier
test.

## Test types and assets

Use each test type where it protects a distinct risk:

- Unit and package tests protect focused behavior and invariants through the
  narrowest stable boundary.
- Contract tests protect exported behavior, compatibility, and error identity
  relied upon by consumers.
- Integration tests use real collaborators when correctness depends on their
  interaction, while replacing live or nondeterministic external boundaries.
- External-package root tests exercise the public facade as a Go consumer,
  while internal package tests own focused implementation behavior.
- The maintained offline preparation and execution examples protect distinct
  representative assembled consumer workflows without contacting a model
  provider.
- Fixtures should be minimal, synthetic, versioned with the behavior they
  exercise, and free of credentials or private data.
- Golden files are appropriate only when the complete output is intentionally
  stable and semantic review of updates is practical.
- Failure-path tests should cover consequential malformed input, dependency
  failure, cancellation, partial results, and recovery behavior.

## What deserves tests

Prioritize tests for:

1. Public and package-level contracts.
2. Domain rules and important invariants.
3. Boundary conditions and malformed input.
4. Failure handling, cancellation, retries, recovery, and partial success.
5. Serialization, schemas, compatibility, and round trips.
6. Previously observed or plausible regressions.
7. Representative integration and consumer workflows.

A package-level contract is behavior relied upon by another package or major collaborator, not every observable detail of a package implementation.

For behavior involving **data integrity, destructive operations, compatibility, security, concurrency, idempotency, or recovery**, presume that durable tests are required unless the behavior is already credibly protected at another layer.

Do not add tests merely because a function, branch, or line exists. Do not add a test when the same meaningful risk is already adequately protected elsewhere.

## Choose the right test boundary

Test through the narrowest stable boundary that expresses the behavior clearly.

This is often the package API, but it may instead be:

- a smaller pure function when dense domain logic is most clearly isolated there;
- a package-level operation when several internal collaborators jointly produce the behavior; or
- a larger integration boundary when correctness emerges from interaction with a real dependency.

Do not force all behavior through oversized consumer-workflow tests. Do not
test every private helper merely because it exists. Choose the boundary that
gives durable confidence with the least incidental coupling.

## Test behavior, not implementation

A test should protect a decision, contract, or invariant—not memorialize the current implementation.

Before adding or retaining a test, ask:

> What realistic defect would this test catch?

A test is suspect when its main purpose is to detect that someone:

- changed an internal constant;
- renamed or split a private helper;
- reordered equivalent internal operations;
- changed incidental formatting;
- replaced one correct algorithm with another; or
- refactored internal object structure without changing behavior.

Refactoring should normally require no test edits unless the refactored structure is itself part of the contract.

A test can be factually correct and still have negative value. Accurately describing current behavior is not enough; the protected behavior must be important enough to justify the future friction.

## Expected effects of different changes

Use the following expectations when evaluating test failures and test maintenance:

| Change | Expected effect on tests |
|---|---|
| Internal refactor that preserves behavior | Existing tests should normally remain unchanged and continue to pass. |
| Change to an internal default with no contractual significance | Behavioral tests should normally remain unchanged; tests should derive expectations from configuration or relationships rather than duplicate the old value. |
| Intentional change to public behavior, policy, schema, or compatibility guarantees | The relevant tests should be reviewed and changed deliberately. |
| Accidental violation of a contract or invariant | Tests should fail; fix the production code rather than rewriting the tests to accept the defect. |

A test failing is not the same as a test needing to be edited. Many tests may correctly fail because of one production defect. The maintenance smell is a correct internal change that requires unrelated expectation updates throughout the suite.

## Separate mechanism from policy

Configurable thresholds and defaults must not be duplicated throughout the test suite.

The following fragments are illustrative rather than standalone Go programs.
Do not encode an internal concurrency limit indirectly:

```go
// Production policy:
const maxConcurrency = 4

// Brittle test:
err := startProcesses(5)
require.Error(t, err)
```

Instead, test the mechanism relationally:

```go
const limit = 2
runner := NewRunner(limit)

require.NoError(t, runner.Start(limit))
require.ErrorIs(t, runner.Start(limit+1), ErrTooMuchConcurrency)
```

The test should prove:

- the configured limit is accepted; and
- one beyond the configured limit is rejected.

The production default should be tested exactly only when its literal value is itself a public, operational, safety, protocol, or compatibility requirement.

Apply the same rule to limits, timeouts, capacities, retry counts, and ranges: test relationships and behavior, not duplicated literals.

For concurrency limits, test both kinds of behavior when relevant:

1. **Configuration enforcement:** invalid or excessive requested values are handled correctly.
2. **Runtime enforcement:** observed peak concurrency never exceeds the configured limit.

Use a test-controlled limit and measure the behavior relative to that limit. Do not merely assert today's default value.

## Avoid semantic duplication across layers

Each behavior should have a clear test owner.

- Parser tests own parsing cases.
- Validator tests own validation rules.
- Domain tests own transformations and invariants.
- Boundary tests own external integration behavior.
- Orchestrator tests own coordination and failure propagation.
- Consumer-workflow tests prove that representative assembled library use
  works.

Higher-level tests should not repeat every lower-level case. A single intentional policy change should not require unrelated edits across many test files.

Tests that are individually reasonable may still be collectively redundant. Evaluate the marginal value of each additional test in light of the protection already provided by the rest of the suite.

## Use test doubles deliberately

Choose the least elaborate test double that provides the required control or observation.

As a default:

1. Prefer real collaborators when they are fast and deterministic.
2. Use small in-memory fakes when realistic stateful behavior is helpful.
3. Use stubs when a dependency only needs to provide controlled responses.
4. Use mocks when the interaction itself is contractual.

Mocks are appropriate when the contract includes facts such as:

- a notification is sent exactly once;
- a transaction is committed only after successful writes;
- cancellation reaches a subprocess;
- an expensive API is called no more than once; or
- a security audit event is emitted.

Do not use mocks merely to isolate every object or reproduce the implementation's call graph.

## Go-specific guidance

Use:

- table-driven tests for meaningful behavioral categories and boundaries;
- `t.TempDir()` for real filesystem behavior;
- `httptest.Server` for realistic HTTP interactions;
- fuzz tests for parsers, normalization, path handling, and broad input spaces;
- golden files only when the complete output is intentionally stable;
- integration tests where correctness depends on component interaction; and
- a small number of representative consumer-workflow tests.

Avoid exact error-string assertions unless the wording is itself contractual. Prefer `errors.Is`, `errors.As`, typed errors, or structured error fields.

At public API boundaries, prefer stable error identity, structured values, and
the smallest semantic fragment needed to identify the failure. Do not snapshot
complete diagnostic wording unless it is contractual.

Golden-file updates must require an explicit local flag. Ordinary validation
runs must never update golden files automatically, and maintainers must inspect
the semantic diff before accepting an update.

Keep tests readable and direct. Test helpers and fixture frameworks must earn their own maintenance cost; do not build elaborate test infrastructure for small or isolated needs.

## Coverage

Coverage is a diagnostic, not a target.

Use it to find untested critical branches and unexpectedly weak packages. Do not write low-value tests solely to increase a percentage, and do not infer test quality from coverage alone.

Pure domain logic will often warrant higher coverage than facade wiring or
external adapters. Uneven coverage is acceptable when it reflects risk.

Increasing coverage is valuable only when the newly covered behavior protects a meaningful risk at an acceptable cost.

## Regression tests

A bug fix should normally include a regression test that fails before the fix and passes afterward.

Retain the test when the defect could realistically recur and its consequences justify the ongoing cost. Prefer the narrowest durable test of the violated contract or invariant; do not preserve accidental implementation details from the original bug.

Not every historical bug requires a permanent test. If the underlying design has made recurrence impossible, the test has become redundant, or a stronger invariant test now subsumes it, remove or consolidate it.

## Deleting or rewriting tests

Tests are maintained code, not permanent historical artifacts.

Delete or rewrite a test when its maintenance cost exceeds the confidence it provides.

Strong candidates include tests that:

- require updates after harmless internal changes;
- directly assert private constants without protecting a real contract;
- duplicate the same policy across several layers;
- verify mock choreography rather than outcomes;
- snapshot large amounts of incidental output;
- test trivial private helpers already exercised through stable package behavior;
- protect risks already covered more effectively elsewhere;
- are flaky, misleading, obsolete, or disproportionately expensive to diagnose; or
- no longer correspond to a plausible failure mode.

Several brittle tests may encode one genuine requirement. Replace them with one durable behavior-level or invariant test rather than preserving all of them.

Deleting a low-value test can improve the quality of the suite by reducing noise, maintenance burden, and friction around legitimate change.

## Reviewing a proposed test

Use the following questions when the value, boundary, or durability of a proposed test is not self-evident. Significant test additions should be reviewable against them, but written answers are not required for every routine test.

1. What realistic defect would it catch?
2. How likely is that defect?
3. How consequential would it be?
4. Is the behavior already protected elsewhere?
5. At which layer should this behavior be owned?
6. Does the test assert a durable contract or an incidental implementation detail?
7. Could the implementation be refactored without changing the behavior and without editing this test?
8. What should cause this test to fail?
9. What legitimate changes should not cause this test to fail?
10. What ongoing maintenance, execution, and diagnostic cost will the test impose?
11. Is there a smaller or more direct test that protects the same risk?

Do not add the test when its expected lifetime cost exceeds its expected protective value.

When deciding not to test plausible behavior, record or be able to explain why the risk is low, already protected, obvious, reversible, or cheaper to detect elsewhere.

## Definition of sufficient

A test suite is sufficient when:

- important contracts and invariants are protected;
- meaningful boundaries and failure modes are exercised;
- realistic and consequential regressions are credibly protected against silent recurrence;
- behavior involving data integrity, destructive operations, compatibility, security, concurrency, idempotency, and recovery is credibly protected;
- important external boundaries have realistic integration coverage;
- representative complete workflows are tested;
- failures provide useful signal rather than redundant noise;
- legitimate internal changes usually do not require test edits; and
- additional tests would mostly repeat existing protection or preserve inconsequential implementation details.

Sufficiency is a risk judgment, not a coverage percentage or test count.
Reassess it as the library, its consumers, and the consequences of failure
evolve.

The governing rule is:

> Test heavily where failure is consequential, subtle, or difficult to detect after the fact. Test lightly where failure is obvious, reversible, and inexpensive—and retain no test whose lifetime cost exceeds the confidence it provides.