Open Fusion Test Strategy

Open Fusion Test Strategy

Required Context

Read the owning spec before designing tests:

• API behavior: docs/specs/001-openai-compatible-api.md
• Orchestration: docs/specs/002-llm-orchestration-routing.md
• Config: docs/specs/003-single-json-configuration.md
• Providers: docs/specs/004-provider-adapters-openrouter.md
• Streaming/tools: docs/specs/005-streaming-tools-response-normalization.md
• Ops/security: docs/specs/007-observability-resilience-security.md

Test Pyramid

Prefer this order:

1. Unit tests for pure validation, mapping, policy, and normalization logic.
2. Integration tests for NestJS modules, guards, filters, config loading, and service wiring.
3. Contract tests for public OpenAI-compatible envelopes, provider adapter behavior, and streaming chunks.
4. E2E tests for critical client flows through HTTP.
5. Live provider tests only when explicitly configured and cost-controlled.

What To Mock

Mock or fake by default:

• Vercel AI SDK model calls;
• OpenRouter/network calls;
• time, ids, and environment variables;
• streaming token sources;
• provider failures and timeouts.

Do not mock the code under test. Mock at external boundaries.

Required Coverage Areas

API:

• valid chat completion request;
• invalid request envelope;
• non-finite numeric fields (NaN, Infinity, -Infinity);
• message count, message content, and payload limit failures;
• unknown public model;
• auth failure;
• OpenAI-style error body;
• streaming chunk format and [DONE].

Config:

• valid minimal config;
• unknown provider type;
• missing env secret;
• invalid model/provider reference;
• invalid route/orchestrator reference;
• secret redaction.

Orchestration:

• direct orchestrator response;
• allowed delegation;
• blocked delegation;
• max delegation limit;
• blocked, failed, and timed out attempts counted against maxDelegations;
• delegate timeout;
• finish reason propagation;
• depends_on: [] does not become a pre-final task by itself;
• parallel pre-final task failure cancels in-flight work or explicitly ignores late results;
• final-only streaming.

Providers:

• non-streaming success;
• streaming success;
• tool-capable path;
• malformed provider tool payloads;
• malformed provider-supplied delegate messages;
• unsupported capability;
• unknown internal model id maps to internal error without exposing the unresolved id in the public error body;
• provider timeout;
• provider error normalization.

Ops:

• request id propagation;
• structured log fields;
• chat_completion.failed for validation/model/tools policy errors before SSE;
• redaction;
• generic public internal_error messages for internal configuration failures, with diagnostic identifiers kept out of the OpenAI-compatible response body;
• payload limits;
• health checks without paid provider calls.

Test hygiene:

• Restore process.env by mutating the existing object, not by assigning a new object to process.env.
• Assert behavior and public contracts first; use mock call assertions only to clarify boundary interactions.

Test Quality

• Assert observable behavior, not private implementation.
• Keep fixtures small and named.
• Avoid snapshots for dynamic OpenAI-compatible envelopes unless normalized.
• Use builders/helpers only after duplication is real.
• Make failures diagnostic: include model/route/status expectations clearly.