AI API receipts

Upstream gateway and parser bugs, replayed through Badgr’s outcome contract. Each page shows what Badgr decided and the receipt it recorded. Where the contract cannot tell a response is wrong, the page says so.

Badgr rejects these (19)

parse_arguments turns comma-containing values into Python lists
sgl-project/sglang#43147 · receipt replay-sglang-43147-number-comma · failed
PoolsideV1Detector silently retypes nullable/union string arguments
sgl-project/sglang#43149 · receipt replay-sglang-43149-nullable-string · failed
Qwen3CoderDetector streaming emits unterminated tool JSON
sgl-project/sglang#43150 · receipt replay-sglang-43150-unterminated-json · failed
Qwen3CoderDetector streaming parser fabricates tool calls from prose
sgl-project/sglang#43151 · receipt replay-sglang-43151-prose-becomes-tool-call · failed
Parallel tool-call arguments dropped when streaming /v1/messages
BerriAI/litellm#45348 · receipt replay-litellm-45348-parallel-tool-args-dropped · failed
Proxy emits a non-Responses error frame mid-stream
BerriAI/litellm#45406 · receipt replay-litellm-45406-invalid-error-frame · failed
google-genai streamGenerateContent: a stream dropped before the first chunk is never retried
BerriAI/litellm#45457 · receipt replay-litellm-45457-stream-dropped-before-first-chunk · failed
glm47 streaming parser silently swallows the real tool call after a stray <tool_call> in prose
vllm-project/vllm#60703 · receipt replay-vllm-60703-swallowed-tool-call · failed
glm47 streaming parser emits tool-call arguments that are not valid JSON
sgl-project/sglang#43273 · receipt replay-sglang-43273-glm47-invalid-stream-json · failed
Top-level import soundfile in the Transcribe handler breaks /v1/messages streaming without the proxy extra
BerriAI/litellm#45379 · receipt replay-litellm-45379-soundfile-import-breaks-stream · failed
Claude Code to Bedrock: per-turn-control beta is dropped but the per-message output_config is still sent
BerriAI/litellm#45411 · receipt replay-litellm-45411-bedrock-per-turn-control-400 · failed
Concurrent /v1/messages calls sometimes answer 500 "dictionary changed size during iteration" while the spend row records success
BerriAI/litellm#44748 · receipt replay-litellm-44748-500-while-spend-records-success · failed
/v1/completions returns HTTP 500 for use_beam_search + echo + logprobs
vllm-project/vllm#60875 · receipt replay-vllm-60875-beam-search-echo-logprobs-500 · failed
Responses API mid-stream 429 can be recorded against the fallback deployment and cool it down
BerriAI/litellm#45645 · receipt replay-litellm-45645-fallback-cooled-by-primary-429 · failed
routing_strategy "cost-based-routing" breaks every sync Router method with RouterRateLimitError
BerriAI/litellm#45718 · receipt replay-litellm-45718-cost-routing-sync-raises · failed
chat->Responses->chat bridge converts string user content to text blocks and breaks Databricks json_object
BerriAI/litellm#45618 · receipt replay-litellm-45618-bridge-string-content · failed
DotsToolParser drops tool calls when request.tools is None
vllm-project/vllm#60865 · receipt replay-vllm-60865-dots-parser-drops-tool-call · failed
SamplingParams accepts float for int fields, then LLM.generate() hangs on an EngineCore msgspec decode error
vllm-project/vllm#60998 · receipt replay-vllm-60998-float-max-tokens-hangs · failed
dots and Muse Glimmer detectors rewrite string-typed argument values (null becomes None, JSON-looking strings coerced)
sgl-project/sglang#43017 · receipt replay-sglang-43017-string-args-rewritten · failed

Badgr cannot detect these (27)

Qwen3CoderDetector silently flips boolean parameters to false
sgl-project/sglang#43153 · receipt replay-sglang-43153-boolean-flip · success
HermesDetector reorders streamed normal text across a tool call
sgl-project/sglang#43148 · receipt replay-sglang-43148-reordered-text · success
HarmonyParser silently deletes streamed content that follows a tool call
sgl-project/sglang#43156 · receipt replay-sglang-43156-deleted-text · success
/v1/messages to Responses: assistant text blocks are reordered around function calls
BerriAI/litellm#45470 · receipt replay-litellm-45470-reordered-assistant-turn · success
Mistral content returned as a list of chunks is truncated to the last text chunk
BerriAI/litellm#45378 · receipt replay-litellm-45378-truncated-text · success
Consecutive user messages are mishandled when rendering the chat template
sgl-project/sglang#43155 · receipt replay-sglang-43155-consecutive-user-roles · success
Inkling "reasoning off" is ignored, so a request that disabled thinking is processed with reasoning on
sgl-project/sglang#43145 · receipt replay-sglang-43145-reasoning-off-ignored · success
get_hf_text_config erases the text sub-config dtype on the thinker_config path
sgl-project/sglang#43157 · receipt replay-sglang-43157-thinker-dtype-erased · success
get_rope_config fabricates rope_theta=10000 for transformers-v5 per-layer-type rope configs
sgl-project/sglang#43158 · receipt replay-sglang-43158-rope-theta-fabricated · success
grouped_topk selects wrong experts with more than eight groups
vllm-project/vllm#60752 · receipt replay-vllm-60752-grouped-topk-wrong-experts · success
/v1/messages streaming needs undeclared redis; Bedrock streams skip success callbacks without an error
BerriAI/litellm#45380 · receipt replay-litellm-45380-redis-skips-success-callbacks · success
apply_reasoning_enabled does not mirror the minimax-m3 read-side rule
sgl-project/sglang#43146 · receipt replay-sglang-43146-minimax-m3-reasoning-toggle · success
Azure SDK Responses calls (/openai/responses) bypass the router and go to api.openai.com
BerriAI/litellm#45033 · receipt replay-litellm-45033-azure-responses-wrong-provider · success
tool_choice="required" with the hermes structural tag drops parallel tool calls
vllm-project/vllm#60769 · receipt replay-vllm-60769-required-drops-parallel-tool-calls · success
A 10-minute Mistral transcription is billed as $28.8B because a FLAC with no sample count is read as 2^63 frames
BerriAI/litellm#45357 · receipt replay-litellm-45357-flac-transcription-cost-explodes · success
A pre_call guardrail adds litellm_metadata, so tag-based routing ignores the request tags
BerriAI/litellm#45543 · receipt replay-litellm-45543-guardrail-breaks-tag-routing · success
Mistral replayed reasoning_content is dropped instead of being sent as a thinking chunk
BerriAI/litellm#45546 · receipt replay-litellm-45546-mistral-reasoning-not-replayed · success
post_call keeps a full serialized copy of every response for the life of the request
BerriAI/litellm#45541 · receipt replay-litellm-45541-post-call-holds-full-response-copy · success
UnsupportedParamsError on the primary model is hidden behind a 200 from the fallback model
BerriAI/litellm#44815 · receipt replay-litellm-44815-config-error-hidden-by-fallback · success
Responses streaming bridge omits sequence_number on most events and hardcodes output_item.done to 1
BerriAI/litellm#44923 · receipt replay-litellm-44923-responses-stream-missing-sequence-numbers · success
Python SDK loses the first SSE event type when the stream starts with a UTF-8 BOM
langchain-ai/langgraph#9225 · receipt replay-langgraph-9225-sse-leading-bom-loses-event-type · success
CuTe DSL GDN decode reads the state pool transposed, producing garbage output
sgl-project/sglang#43331 · receipt replay-sglang-43331-cutedsl-gdn-decode-garbage · success
DeepSeekV3Detector streaming drops the first tool call when two complete calls arrive in one increment
sgl-project/sglang#43523 · receipt replay-sglang-43523-dropped-first-tool-call · success
/v1/messages does not accumulate deployment-level tpm for native Anthropic routes
BerriAI/litellm#45702 · receipt replay-litellm-45702-messages-tpm-not-enforced · success
Missing usage in a provider response is reported as 0 tokens, which looks identical to a real zero
BerriAI/litellm#45586 · receipt replay-litellm-45586-missing-usage-reads-zero · success
JSON flow result with a "data" or "result" key reads back wrong from the API
PrefectHQ/prefect#23372 · receipt replay-prefect-23372-result-key-reads-back-wrong · success
Olmo3 tool parser corrupts string arguments with newlines and drops calls split across lines
vllm-project/vllm#60421 · receipt replay-vllm-60421-olmo3-newline-corrupts-args · success