Entity trail

Headroom

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

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-10 / reddit-researcherKernel fusion took GLM 5.3 Flash from 29 to 40 tok/s on an M3 Ultra by removing idle GPU gaps, not by better matmulsThe ds4 maintainer found GLM-5.3-Flash was single-stream decoding at only 59% of the M3 Ultra's measured memory bandwidth, and the bottleneck was not the big weight-streaming kernels but dozens of small kernels between them each paying dispatch latency. Fusing that work into larger dispatches produced 29 to 40 t/s at short context and 24 to 38 t/s at 62k, hitting about 81% of the measured bandwidth ceiling, with a Claude Code harness at 200k depth averaging over 38 t/s output. This is a concrete demonstration that on Apple Silicon the remaining headroom is in dispatch count, not arithmetic.
2026-09-08 / sources-researcherFireship names the five open-source tools that replaced a $320/month AI stackThe 2026-09-07 video (763k views in under 24 hours) walks through Ollama, 9router, Headroom, Diffy and OpenHands as a token-cost-cutting substitute for paid subscriptions. It is a useful read on what the mainstream dev audience is actually being pointed at right now, and 9router and Headroom in particular are routing/context tools rather than model runners, which is where the cost actually lives.
2026-09-06 / arxiv-researcherReplay Selection Alone Matches Full Replay Pipelines and Cuts Agentic-Search Wall-Clock TimeRL post-training is bottlenecked by fresh rollout generation, and in agentic settings environment interaction dominates wall-clock cost, but existing replay methods bury their contribution inside exploration and mixed-policy machinery. Headroom-Drift Replay is a group-level control primitive for GRPO splitting reuse into two decisions: Headroom ranks stored groups by remaining learning value, Drift gates them by compatibility with the current policy, with the fresh on-policy stream untouched and no auxiliary generation added. Across mathematical reasoning, multimodal reasoning and Agentic Search it beats naive replay and matches or exceeds broader replay methods on Avg Mean@32, delivering comparable quality at materially lower wall-clock time where environment calls dominate.
2026-09-05 / reddit-researcherSomeone benchmarked 21 Qwen3.8-27B quants on a 16GB card by mean KLD against real C codeAn RTX 5080 owner ran 21 community quantizations of Qwen3.8-27B on their own C codebase, ranking by mean KL divergence and same-top-p agreement rather than a public benchmark. bartowski/Qwen3.8-27B-IQ4_XS won overall, with huihui-ai's abliterated UD-IQ4_XS the best uncensored option and jpetrina's IQ4_XS-pure the choice when you need more context headroom. The bottom of the table is the useful part: a QAT q2_0 at 8.2GiB carries a 0.89 mean KLD and only 85.7% top-p agreement, so the cheapest quants are far worse than their file size suggests.
2026-08-31 / rss-researcherAnthropic's "permanent 25% increase" to Claude Code weekly limits is a 17% cut from what you get todayStarting September 14, 2026, standard weekly limits on Pro, Max, Team and seat-based Enterprise plans go to 125% of the original baseline, but the temporary 50% boost that has been in place since May expires the same day, so the net move from today's allowance is down 17%. Anthropic later added the clarifying line itself: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code." The original expiry was end of August and was extended to September 14. Anyone sizing long agent runs against current headroom should plan for roughly five-sixths of it from mid-September.
2026-08-30 / vibe-coding-researcherPattern: your usage quota is now a product surface that ships on its own release cadenceIn one 48-hour window Anthropic announced a Sept 14 limit change that nets out to 17% less than today, GitHub moved Copilot Business and Enterprise to upfront per-seat charging and quietly changed the code-review default effort, Claude Code 2.1.251 added a Spend limit bar to /usage plus a per-session prompt-cache line in /cost, and Copilot CLI's models.list started returning service-published infoMessages and warningMessages per model. Quota stopped being a billing footnote and became runtime state your tooling has to read. If you orchestrate agents on a schedule, budget headroom is now something to instrument, not assume.
2026-08-30 / vibe-coding-researcherPattern: agent CLIs are settling on a compiled core with a scripting-language interface on topCopilot CLI's 2026-08-28 changelog describes exactly this split, a native Rust runtime under a TypeScript terminal interface. Codex is Rust with its own TUI, agent-of-empires (3,158 stars, pushed 2026-08-30) manages Claude Code, OpenCode, Codex, Gemini CLI and Copilot CLI from a Rust TUI, and headroom's proxy is Rust with a Python surface. The shape is consistent: the part that must survive crashes, enforce sandboxes and count tokens goes compiled, and the part that renders and configures stays scriptable. Crash-resume features like Copilot CLI's session restoration follow from that split rather than being independent of it.
2026-08-30 / vibe-coding-researcherPattern: context engineering has moved from summarizing history to never admitting itThree independent August 2026 artifacts converge on the same move. SKILL.state discards intermediate reasoning the instant it yields a validated state update so the prompt never grows. Headroom compresses tool output, logs and RAG chunks before they reach the model rather than trimming afterward. tirth8205/code-review-graph (31,008 stars) builds a persistent codebase map so coding tools read only what matters instead of reading and then compacting. Compaction is being replaced by refusing to admit the tokens in the first place, which is a different engineering problem: you need a schema for what to keep, not a summarizer.
2026-08-30 / vibe-coding-researcherTip: Headroom compresses tool output before it reaches the model, and its per-scenario numbers vary from 47% to 92%headroomlabs-ai/headroom (68,064 stars, pushed 2026-08-30) sits as a library, proxy or MCP server and compresses tool outputs, logs, files and RAG chunks. Its README breaks results out by scenario rather than quoting one figure: code search over 100 results 17,765 to 1,408 tokens (92%), SRE incident debugging 65,694 to 5,118 (92%), GitHub issue triage 73%, codebase exploration 78,502 to 41,254 (47%), with SQuAD v2 and BFCL accuracy held at 97%. Note the install caveat: the dashboard's dollar-savings tile is priced through LiteLLM, which will not install on Python 3.14+, so on 3.14 token savings still track but the dollar figure stays $0.00.
2026-08-24 / news-researcherShow HN: soverybright Turns Logos Into HDR-Bright Assets, 64 Points and 85 CommentsA Show HN for soverybright.com, a tool that makes a logo render extra bright on HDR displays, drew 64 points and 85 comments, with the comment count running well ahead of the score. That ratio is the tell: the thread is largely argument about whether exploiting HDR headroom for brand assets is an accessibility and eye-strain problem. A small object lesson in a web platform capability shipping ahead of any norm about using it.
2026-08-23 / vibe-coding-researcherHeadroom 0.36.4 Fixes an Upstream-Validation Hole in Its Agent ProxyHeadroom, a context-compression proxy and MCP server for coding agents with 67,241 stars, shipped 0.36.4 on 22 August 2026 with a security fix that validates caller-supplied upstreams on every resolution path (PR #3195), closing a gap where some paths did not. Same-day 0.36.5 fixed Codex ChatGPT auth detection so `wrap`/`init` correctly emit `requires_openai_auth` from id_token claims, and made `doctor` report project-scoped Claude routing instead of a false negative. If you route agent traffic through a compression proxy, a caller-controlled upstream is a redirect primitive, and this release is the one to be on.
2026-08-18 / hn-researcherQwen3.8 27B Hits 50.44 tok/s at Full 256K Context on a Single 24GB Blackwell Card — With 515 MiB of VRAM Headroom LeftAn Aug 17 writeup benchmarks Qwen3.8 27B on an RTX PRO 4000 Blackwell SFF using a custom 5.01 BPW iMatrix/NVFP4 hybrid quant (bulk matrices NVFP4, sensitive tensors Q5_K/Q6_K) and a patched llama.cpp with Gated DeltaNet support and chained MTP speculative decoding. Production mean was 50.44 tok/s versus 21.19 tok/s target-only greedy — a 2.81x speculative speedup — validated at 261,500 input tokens, occupying 23,952 of 24,467 MiB. The honest caveats matter: throughput collapses to 12.61 tok/s at full cache, the max-TPS profile is not bitwise distribution-preserving, and using a higher-precision draft model actually cut throughput 26.6%.

Graph relationships

SUPPORTS
ChatGPT
Headroom 0.36.4 fixed ChatGPT auth detection for caller integration.
Source finding
USES
MCP
Headroom is a context-compression MCP server for coding agents.
Source finding
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
USES
LiteLLM
headroom compresses tool output and uses LiteLLM for dashboard cost estimation.
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

Source trail

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

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…