Fetching from the wire…
01
02
03
04
05
06
07
08
Infra2026-09-16 · source-backed
PR #28869 forces \n</think> rather than bare </think> when llama.cpp terminates a reasoning budget for qwen3-coder templates. The old string didn't match what the template renders on subsequent requests, so every follow-up missed the prompt cache. One character, large cost. The general rule for anyone injecting forced tokens into a chat template: the injected text has to be byte-identical to what the template itself emits.
Each link below shares sources, entities, or timing with this story.
agentic-kv-cache simulates cross-request prefix caching against real traces, not synthetic ones: 68,266 requests across 393 Claude Code sessions at 64-token blocks, plus Mooncake traces at 512-token blocks. It models prefix-contiguous hits, radix eviction constraints and pinne...
frontier-harness-eval/eval ran Kimi K3 through nine coding-agent harnesses in 12 configurations over 30 identical software-engineering tasks, 360 runs. Codex led on pass rate at 66.7% with $3.47 median cost per pass. Claude Code matched DSH Creator at 63.3% but cost $18.34 per...
rust-v0.150.1, published August 27, fixes remote compaction ignoring retained images when computing token budget; it now trims older images as needed (GitHub). Sessions passing screenshots to a vision-capable agent were budgeting as though those images cost nothing, which prod...
Created July 31, Ratchet's argument is that every "prefer the stdlib, keep the diff small" ruleset is an open loop: the model reads it and nothing verifies compliance. So a PostToolUse hook measures every edit and reports back into the live session: net line delta against a bu...
vercel-labs/deepsec (6,383 stars, +1,082 this week) runs a three-stage pipeline: scan with regex matchers and no AI, process for AI investigation emitting findings plus a recommendation, and revalidate re-checking existing findings against git history to see if they were alrea...
2.1.265 fixes resuming a foreground-spawned subagent changing its tool list and system prompt prefix, and agent teammates plus resumed subagents moving SubagentStart hook context and preloaded skills out of the prompt prefix on later turns. Both are called out explicitly as br...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.