Claude Code. Any Model. The most powerful AI coding agent now speaks every language.
Test cull — sacrificed coverage record
This records the behavior guards deliberately removed in the aggressive test cull (per explicit instruction to "cut harder" past the adversarial verifier's protection). These were the sole automated guard for the listed behaviors. Deleting them means a regression in these areas will not be caught by the test suite. Recorded here so the loss is explicit and can be reversed if any of these regress in the wild.
Highest-risk removals (real shipped bugs these guarded)
provider-routing.test.ts— model→dialect routing matrix, including the #102 regression guard ([email protected] → DefaultAPIFormat, not GLMModelDialect). #102 was a real shipped bug. Also removed: gemini/gpt/o-series/codex/qwen/deepseek/minimax/xiaomi dialect selection, vendor-prefix routing, mid-string false-positive guards, and the PROVIDER_PROFILES↔BUILTIN_PROVIDERS consistency checks.tool-count-cap.test.ts— OpenAI 128-tool cap (and Codex-uncapped divergence). This exact value was silently deleted once (commit 3edc60f) and broke every OpenAI run with128 tools. No other checked-in test guards it; the head-slice is only exercised on the live
oai@path.authority.test.ts— credential/provider registration: local providers (ollama/lmstudio/vllm/mlx), kimi/codex/vertex/native-anthropic registration, composite invalidate() fan-out (both halves), login/logout routing to primary. Dropping these means--model ollama@…/cx@…/ vertex routing can break undetected, and hydrate-on-add (TUI) key application relies on the invalidate fan-out.
Other sole-guard removals
model-selector.test.ts— picker→Firebase slug mapping, user-deployed-provider classification, fixedModel (kimi-coding single-model) declarations.provider-definitions.test.ts— duplicate name/shortcut/prefix detection, local vs remote transport classification, direct-API classification, display names.routing-rules.test.ts— loadRoutingRules non-null + catch-all guarantee, displayName.format-translation.test.ts— MiniMax context-window/vision, Gemini stream format, Codex shouldHandle, ProviderProfile registration completeness.- transport oauth tests — kimi-coding catalog-name delegation, Gemini CodeAssist envelope
- headers, codex payload model normalization under OAuth.
mcp-server.test.ts— async getRecommendedModels (not the stale sync path).op-source.test.ts— sniff memoization;model-loader.test.ts— sync freshness gate.
If any of these behaviors regress, restore the corresponding test from git history (the commit immediately before the aggressive cull).