Unverified50% confidenceFactExact time
After setting DFLASH_TOKENS=15, in citation requests with a 20k context it causes the engine to crash, reporting EngineDeadError and a marlin_gemm allocation failure, after which every request returns 500 and it cannot self-recover.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/16/2026
First Seen
Valid until: 10/30/2026
Sources
Related Entities
Related Claims
UnverifiedIn the ticket service failure demonstration, Catalyst reproduced the failure from a 96-request run and reduced it to a 16-request reproducer; the candidate fix achieved 0 errors out of 16, p99 of approximately 16ms, and passed all 3 holdout scenarios.70% similarUnverifiedIn LLM calls, client errors such as 400 (malformed request) and 401 (invalid key) should not trigger switching to a backup provider, because resending will only get the same error; only transient provider-side failures such as connection failures, timeouts, 429, and 5xx should trigger fallback.69% similarUnverifiedThis version fixes an issue where excessively large failure output from background tasks caused API request bodies to exceed size limits and fail63% similarUnverifiedv2.1.276 fixes an issue where every request would fail with a 400 error (Input tag 'advisor_20260301') when ANTHROPIC_BASE_URL pointed to a proxy or gateway.61% similarUnverifiedThe new version fixes an issue where sessions would fall into an infinite retry loop of unexpected tool_use_id 400 errors, and corrupted transcripts will now self-heal as much as possible.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1043371API
curl https://kongchang.com/api/v1/knowledge/claims/1043371MCP
get_claim(id=1043371)