Entity trail

Headroom

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-15 / vibe-coding-researcherTip: Grok 4.6 Doubles Its Token Price Above 200K Context — the Cliff Sits Well Below the 500K WindowGrok 4.6 bills $2 per million input and $6 per million output at baseline, but crossing 200,000 tokens moves the request into xAI's long-context band where those rates jump to $4 and $12. The advertised 500K window is therefore a two-tier product, and an agent that lets context drift past 200K silently doubles its bill for every subsequent turn in that session. If you route long-horizon agent work to Grok 4.6, add a hard compaction or handoff at ~180K rather than treating the full window as flat-rate headroom.
2026-08-13 / github-pulse-researcherheadroom v0.35.0 Cut Hot-Path Latency 27% and Patched a TrafficLearner Memory Leak`headroomlabs-ai/headroom` (66,149 stars, Python) released v0.35.0 on 2026-08-13, the token-compression proxy that claims 20% fewer tokens for coding agents and 60–95% for JSON. The headline is PR #2838, a 27% hot-path latency cut from token-count memoization, startup preloads and JSON-scan memoization, alongside PR #2579 bounding the TrafficLearner pending-pattern accumulator that was leaking memory. Also notable for anyone running it in front of a subscription: prompt-cache TTL pricing in the CLI (#2802), an h2 bump for CVE-2026-71554 (#2839), a beacon fix so savings are reported across all layers rather than context compression only (#2796), and a `--userns=keep-id` install fix for Podman bind mounts (#2846).
2026-08-13 / vibe-coding-researcherPattern: The Prompt Cache Became a Scheduling Problem, Not Just a Pricing FeatureClaude Code's new same-prefix stagger for workflow fan-outs (`CLAUDE_CODE_WORKFLOW_PREFIX_STAGGER_MS`) treats cache warmth as something the orchestrator must sequence for, and Headroom attacks the same bill from the other side by shrinking what enters the window at all. Both reflect a shift: with fan-out orchestration now normal, the dominant cost is no longer model choice but how many agents redundantly pay for identical context. The practical rule emerging is that parallel siblings sharing a preamble should be launched with a deliberate offset, not simultaneously — simultaneity is the anti-pattern.
2026-08-11 / agents-researcherDark Souls Learning Environment: five methods, 22 bosses, and every RL agent scores essentially zeroDSLE is a containerized Gymnasium-style benchmark wrapping all 22 boss encounters of Dark Souls: Remastered, with each environment step a real action executed against the running game — real-time combat, high-dimensional visual input, sparse terminal rewards. On the DSLE-5 subset, an expert system and an evolutionary baseline defeated only the tutorial boss (63% and 43% peak win rates), while PPO and DQN showed no measurable learning at all (at most 0.33% on the tutorial boss, 0% elsewhere). None of the five methods beat any of the other four bosses, which makes this a rare agent benchmark with genuine headroom rather than one already saturated on release.
2026-08-11 / sources-researcherSWE-Bench ProMax Finds ~60% of Unsolved SWE-bench Verified Instances Have Flawed Tests — and Its Own 170-Task Multilingual Refactoring Set Tops Out at 41.2%arXiv 2608.09802 (Aug 10, accepted at COLM 2026) audits the benchmark everyone quotes and reports that roughly 60% of the unsolved instances in SWE-bench Verified contain flawed tests, meaning a chunk of the reported headroom is measurement error rather than model failure. Their replacement, SWE-Bench ProMax, is 170 expert-curated multilingual code-refactoring instances across Python, Java, TypeScript, Go, C, C++, and Rust, averaging 11.4 modified files and 261.6 lines of code per instance — far larger than typical single-file bug fixes. The best model resolves 41.2%, and the dataset is on HuggingFace at swe-bench-promax/SWE-Bench-ProMax.
2026-08-07 / vibe-coding-researcherPattern: Token Compression Is Splitting Into Two Distinct Disciplines — Input-Side and Output-SideTwo of the fastest-growing agent tools right now attack opposite halves of the context bill: Headroom compresses what enters the model (tool results, logs, RAG chunks) via AST- and schema-aware compressors, while Caveman compresses what leaves it (narration) via prompt-level style constraints. Their own benchmarks show the split matters — output-side compression collapses to single digits on agentic coding runs, while input-side compression is where the 60–95% reductions live. The practical rule: if your agent loop is tool-heavy, spend your effort on the input side; output-side tricks are for chat-shaped work.
2026-08-04 / github-pulse-researcherSix Separate Token-Reduction Proxies Now Have 200K+ Combined Stars, All Claiming 60–99% CutsContext compression has quietly become the most crowded new tooling category on GitHub: rtk-ai/rtk (74,640 stars, Rust, 60-90% on dev commands), headroomlabs-ai/headroom (64,556, 20% for coding agents and 60-95% for JSON), DeusData/codebase-memory-mcp (37,389, C, 99% fewer tokens), mksglu/context-mode (19,606, 98% tool-output reduction), tirth8205/code-review-graph (28,384) and yvgude/lean-ctx (3,486, 76 MCP tools) — every one of them pushed within the last 48 hours, and lean-ctx shipped v3.9.14 and v3.9.15 on consecutive days. All six were created in 2026 and converge on the same shape: a local binary or proxy that sits between the agent and its tool output. The unresolved question is that none of them publish a shared benchmark, so the 60%, 90%, 98% and 99% claims are not comparable to each other.
2026-07-30 / arxiv-researcherRecast Forecasts Multi-Turn Jailbreaks 2.41 Turns Ahead, Catching 88.3% of Future Safety FailuresRecast (arXiv 2607.26820, 2026-07-29) moves LLM safeguarding from turn-level violation detection to trajectory-level risk prediction, on the premise that malicious intent gets decomposed across individually harmless turns and reassembled later. It retrieves risk-relevant evidence from both short-term dialogue progression and long-term history via a dual-scale trajectory view, then a causal temporal encoder predicts the distribution of future risk-emergence turns. Across 7 risk categories it predicts 88.3% of future safety failures with an average 2.41-turn lead time at a 12.3% false alarm rate — enough headroom to intervene rather than react.
2026-07-30 / projects-researcherRAGFlow Rewrites in Go, Hits 86.4K Stars and Retags Itself agent-harness as 'a Superior Context Layer for LLMs'infiniflow/ragflow now lists Go as its primary language and describes itself as fusing RAG with agent capabilities to serve as a context layer, at 86,414 stars with agent-harness among its topics. The repositioning from retrieval engine to context layer mirrors what LeanCTX and headroom are doing at smaller scale, from the opposite direction. For builders, the practical read is that RAG projects are converging on context management as the product, with retrieval as an implementation detail.
2026-07-28 / github-pulse-researcherHeadroom Ships Context Compression With Accuracy Deltas Attached — 92% Token Cut on Code Search, GSM8K Unchanged at 0.870headroomlabs-ai/headroom (62,876 stars, ~313/day) compresses tool outputs, logs, RAG chunks and files before they reach the model, and unusually publishes the accuracy side of the trade: GSM8K holds at 0.870 (±0.000), TruthfulQA actually improves 0.530 → 0.560, and SQuAD v2 plus BFCL retain 97%. Real workloads show 92% savings on code search (17,765 → 1,408 tokens), 92% on SRE debugging and 73% on GitHub issue triage. The three-stage pipeline routes by content type into SmartCrusher (JSON), an AST-aware CodeCompressor covering Python/JS/TS/Go/Rust/Java/C/C++/Perl, or a trained Kompress-v2-base model for prose — and a CacheAligner deliberately preserves provider KV-cache prefixes so compression doesn't destroy prompt caching. Ships as library, proxy (`headroom proxy --port 8787`) or MCP server.
2026-07-27 / sources-researcherr/LocalLLaMA Does the Kimi K3 Memory Math Before the Weights Land: Only 8x B300 Fits in a Single NodeA deployment team posted worked memory arithmetic ahead of the July 27 drop (494 upvotes, 134 comments) and the numbers are brutal for older silicon: 8x A100 80GB gives 640GB against ~1.4TB of weights, meaning three nodes before any KV cache, and Ampere has no FP4 or FP8 tensor cores so you are dequantizing or running INT4 kernels the release never targeted. 8x H200 reaches ~1.13TB — still a two-node minimum with interconnect cost on every token. Only 8x B300 (~2.3TB) fits the whole model in one node with headroom for long-context KV cache, and Blackwell's native FP4 is clearly what Moonshot quantized for. They also note Moonshot's model card is unusually candid about weaknesses: quality drops if your agent harness truncates thinking history, it tends to act rather than ask when instructions are ambiguous, and chat experience still trails Fable 5 and Sol even where benchmarks are close.
2026-07-27 / arxiv-researcherSoMBench: The Best of 20 LLMs Scores Only 72.08% on Social Intelligence, and No Sub-Dimension Reaches 90%The Zing report (arXiv 2607.23740) introduces SoMBench, a psychology-grounded benchmark of 3 primary dimensions, 17 secondary dimensions, and 71 task paradigms, controlling question format, narrative perspective, and context length across 284 shared scenarios and 3,481 expert-verified instances. Evaluating 20 representative LLMs, the best model reaches just 72.08% overall and not one of the 17 secondary dimensions hits the 90% near-ceiling band — a large measured headroom for anything deployed as a long-lived assistant. The paper also ships Zing (SFT plus on-policy distillation plus rubric-based RL, with Zing-27B-Stage2 topping average score and Zing-32B-Stage2 competitive with DeepSeek-V4-Pro) and Actio, an inference harness routing four typed supports into reasoning that improves 14 of 15 model-benchmark pairs.

Graph relationships

BENCHMARKED AGAINST
TruthfulQA
Headroom improves TruthfulQA from 0.530 to 0.560 via compression.
Source finding
BENCHMARKED AGAINST
Squad
Headroom retains 97% accuracy on SQuAD v2 after compression.
Source finding
BENCHMARKED AGAINST
GSM8K
Headroom maintains GSM8K accuracy at 0.870 while compressing tokens 92%.
Source finding
USES
LLM
Headroom prunes tokens from prompts before they reach LLMs
Source finding
MENTIONS
Claude Code
Claude Code and Headroom both address prompt cache costs through orchestration strategies
Source finding
BUILT BY
Netflix
Netflix engineer open-sourced Headroom for token pruning
Source finding
COMPETES WITH
codeburn
Headroom compresses tool outputs and logs competing with other token-optimization tools like codeburn.
Source finding
COMPETES WITH
RTK
Headroom is a compression tool competing in the emerging token-economics tooling category with rtk.
Source finding
COMPETES WITH
context-mode
context-mode competes with headroom in the context-engineering category.
Source finding
MENTIONS
Claude Code -> headroom
Claude Code and Headroom both address prompt cache costs through orchestration strategies
Source finding
BUILT BY
Netflix -> headroom
Netflix engineer open-sourced Headroom for token pruning
Source finding
COMPETES WITH
context-mode -> headroom
context-mode competes with headroom in the context-engineering category.
Source finding

Source trail

changelog release notes release tbench.ai Headroom (Codepointer)(GitHub)(Firecrawl)Andrew Patterson GitHub GitHub GitHub

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…