fix: provider spec conformance audit — 29 violations fixed, 50+ drift tests added

[jpr5]

Summary

Systematic spec conformance audit of all 21 aimock providers via 60+ specialized agents (MSAL). Found and fixed 29 spec violations across 10 providers. Added 50+ new drift tests covering 9 previously untested providers plus error shapes and reasoning/thinking for 5 existing providers.

Phase 0: Drift test gap closure (18 agents)

• 9 new drift test files: images, speech, transcription, moderation, ElevenLabs, fal.ai (shapes + queue lifecycle), video, rerank
• Error shape tests for OpenAI Chat, Responses, Claude, Gemini, Cohere
• Reasoning/thinking drift tests for OpenAI Chat, Responses, Claude, Gemini

Phase 1-3: Per-provider conformance audit (33 agents)

Every handler audited for: request conversion, non-streaming shape, streaming shape, tool calls, error format.

Phase 4: Fix round (7 agents)

Request conversion fixes:

• Responses API: max_output_tokens and response_format silently dropped
• Gemini: maxOutputTokens, topP, topK dropped; spurious functionCall.id
• Cohere: structured content, native v2 tool format, temperature, max_tokens
• Ollama: tool_calls on assistant messages, images, system on generate

Response shape fixes:

• OpenAI Chat: error responses missing param: null, leaking internal status
• Moderation: missing illicit/illicit/violent categories
• Transcription: verbose response always emitting empty words/segments
• Search: missing Tavily fields (query, images, response_time, answer)
• Rerank: spurious document field

WebSocket fixes:

• Realtime WS: 9 violations — ID prefixes (evt-→event_), missing session/response fields, previous_item_id tracking, output item status
• Gemini Live WS: 4 violations — config alias, standalone turnComplete, complete httpToGrpc mapper

Anthropic thinking:

• signature field + signature_delta event added to Claude Messages and Bedrock invoke thinking blocks

CR: 13-agent MSAL review

Found 4 additional bugs (previous_item_id tracking, incomplete gRPC mapper, test false positive, Anthropic signature). All fixed.

Closes audit plan: https://www.notion.so/35a3aa38185281cc8001c5e863a098dd

Test plan

• pnpm test — 2873 passed
• npx tsc --noEmit — clean
• 50+ new drift tests with SDK shape comparisons
• Negative assertions throughout
• 13-agent CR converged (4 findings fixed, confirmation implicit in combined verify)

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@168

commit: f5ad360