docs(llms): setup agents documentation

This commit is contained in:
Carlos Valente
2026-08-25 21:26:39 +02:00
committed by Carlos Valente
parent a84b8b5530
commit 5acfc4a122
10 changed files with 485 additions and 0 deletions
+52
View File
@@ -0,0 +1,52 @@
# Testing strategy
Use for changed behaviour or tests.
## Layers
### Pure unit
Put detailed business-rule coverage on pure functions. Test public inputs/outputs with Vitest: relevant boundaries, invalid input, ordering, rollover, errors.
Every bug fix needs a regression test that fails before the fix. Prefer behaviour over mocks/implementation assertions.
### Service and state
Use service, DAO, store, or hook tests for orchestration: transitions, transactions, persistence, cache, notifications, external effects.
Do not repeat all pure cases here. Prove delegation and sequencing.
### End-to-end
Reserve Playwright for key journeys and high-risk cross-layer integrations: edit/run rundown, playback, imports, rundown switching, cloud-prefixed navigation.
No E2E for edges already proven in lower layers. Add E2E only when lower layers cannot prove the user-facing integration.
## Compact before handoff
Development harnesses may be broad, repetitive, diagnostic, temporary. Before handoff:
1. Identify distinct required behaviours/regressions.
2. Keep the smallest readable set that catches them.
3. Parameterise repetition only when the table reads better.
4. Remove diagnostic assertions, redundant permutations, temporary fixtures, private-detail coupling.
5. Keep rare cases that encode real domain rules.
Optimise for future readers, not minimum line count.
## Quality
- Name tests by observable behaviour.
- Keep setup local/explicit unless a fixture improves comprehension.
- Avoid arbitrary waits, wall-clock dependence, cross-test state, weak assertions.
- Prefer realistic typed fixtures over large snapshots or masking casts.
- Test non-mutation when promised.
- No tests for trivial type/format changes or framework behaviour Ontime does not own.
## Review prompts
- Business logic directly unit-testable?
- Test catches the reported bug/regression?
- Cases distinct, not repeated path?
- Temporary harness leaked?
- E2E justified by cross-layer risk?