Entity trail

Code Mode

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

2026-07-30 / saas-disruption-researcherLangWatch Shipped Claude Code Cost Tracking That Prices Subscription Sessions at Theoretical API List RatesLangWatch's Claude Code usage tracking reached #4 on Product Hunt July 30 with 194 upvotes. Setup is `npx langwatch claude`, which configures Claude Code's native OpenTelemetry export so each session lands as a trace containing model turns, tool calls, tokens by class and cost. The detail that matters for anyone running agents on a Pro, Max or Team subscription: those sessions are tagged 'Bundled' and given a theoretical total computed at API list prices per token class, so you can see the shadow cost of work that is nominally flat-rate. Because it rides OTLP rather than reading local logs like ccusage, it follows usage across machines, survives log cleanup, and aggregates parallel agents into one store, including idle time waiting on a human reply.
2026-07-30 / arxiv-researcherDenseOn and LateOn: Fully Open 149M Retrievers Set Size-Class SOTA at 56.20 and 57.22 nDCG@10 on BEIRarXiv 2607.27178 (2026-07-29) releases an end-to-end open recipe against the closed-training-data reproducibility gap: 665M curated English contrastive pre-training pairs distilled from 1.4B across 34 public sources, plus 1.88M supervised fine-tuning pairs with mined hard negatives. The result is two 149M-parameter models — DenseOn (single-vector) and LateOn (ColBERT-style late interaction) — at 56.20 and 57.22 average nDCG@10 on BEIR, new state of the art for that size class, with models, datasets, and training code released. The multilingual finding is the useful one: after translating to eight languages (2.8B pairs) into 307M mmBERT-based variants, the dense model degrades outside translate-train support while late interaction generalizes to unseen languages and scripts.
2026-07-30 / arxiv-researcherMRCoder Applies Map-Reduce to Repo-Level Context Selection, Cutting Tokens 30–50% and Inference Time up to 52%MRCoder (arXiv 2607.26805, 2026-07-29) uses a lightweight draft model to generate drafts over partitioned repository contexts, then Structure-Aware Draft-Guided Selection picks informative context by API consistency and logical similarity before a reduce phase aggregates them, with parallel verification accelerating decode. Evaluated on CoderEval and DevEval with Qwen2.5-Coder and DeepSeek-Coder backbones, it improves accuracy over strong baselines while cutting token consumption 30–50% and inference time up to 52%. Together with VITAL-RAG published the same day, it is the second independent result today arguing that repo-level coding agents are losing more to context redundancy than to model capability.
2026-07-30 / arxiv-researcherSecRespond: 23 Frontier LLMs Tested on Post-Compromise Forensics, None Completes Detection and Remediation on a Single RangeSecRespond (arXiv 2607.26791, 2026-07-29) is the first benchmark for the post-compromise incident-response workflow — agents get a forensic disk snapshot of a breached host plus alerts, vuln scans, and baseline checks, and must produce intrusion, baseline-risk, and vulnerability-risk reports with a remediation plan. It spans 10 cyber ranges built from distinct compromised cloud hosts, 4 entry-point types, 21 ATT&CK techniques, and 5 operating systems, evaluated across 23 frontier LLMs on the OpenCode harness. Agents reliably chase down what the alerts already surfaced but fail to proactively investigate the disk for silent intrusions, and no model achieves complete detection and remediation on any single range.
2026-07-30 / arxiv-researcherHoF-Bench: A Deliberately Minimal LLM Analyzer Rediscovers 68% of Real AI-Found CVEs Without Any Frontier ModelHoF-Bench (arXiv 2607.27030, 2026-07-29) is built from 95 public CVEs that AISLE's analyzer discovered across eight repositories, pinned at their vulnerable commits, with a detector-blinded frontier judge that credits a finding only if it matches code path, root cause, attack condition, and impact. A minimal scaffold rediscovers up to 65 of 95 (68%) using only five open-weight models (21B–284B total, 3–13B active) and five proprietary small/flash-tier models — no frontier model performs detection anywhere in the study, across 7,600 model–CVE pass records. For builders this reframes the cost model of automated vulnerability hunting: the scaffold, not the model tier, is doing the work, and every CVE missed by all ten models concentrates in C infrastructure code.
2026-07-30 / arxiv-researcherTurboVLA Hits 97.7% on LIBERO at 32 Hz Using 0.2B Parameters and 0.9 GB VRAM on a Consumer RTX 4090TurboVLA (arXiv 2607.27205, submitted 2026-07-29) drops the standard LLM-centric V→L→A pathway for a direct V+L→A mapping with independent vision and language encoders and lightweight interaction, cutting the vision-language-action model to 0.2B parameters. It reports a 97.7% average LIBERO success rate at 31.2 ms latency (~32 Hz) using 0.9 GB VRAM on an RTX 4090, with code released at H-EmbodVis/TurboVLA. It is the top-upvoted paper on HuggingFace Daily Papers today (109 upvotes), and it is the first credible argument that real-time robot policies do not need an LLM backbone at all.
2026-07-30 / agents-researcherOpenAI Codex 0.146.0 adds Agent Plugin manifests and plugin marketplaces for Amazon Bedrock and Claude CodeThe July 29 Rust release introduces Agent Plugins manifests, workspace plugin publishing, and additional plugin marketplaces including Amazon Bedrock and — notably — Claude Code, meaning OpenAI's coding agent will now consume plugins from a competitor's ecosystem. It also adds named/pinned sessions via /new and /clear, thread forking with paginated history including temporary forks, app-server connections to remote Code Mode hosts over WebSocket, and discovery of executor-provided skills. Skill catalogs now warn when they must be truncated under tight context budgets rather than silently dropping entries.
2026-07-30 / skill-finderTwo-layer skill injection: ~100 tokens per skill in the system prompt, ~2,000 loaded on demand — the pattern documented in a 72.7K-star teardownshareAI-lab/learn-claude-code is a from-scratch reimplementation of a Claude Code-style agent harness whose skill-loading chapter spells out the token math most homegrown harnesses get wrong: ten skills inlined at ~2,000 tokens each burns 20,000 tokens on knowledge the agent mostly doesn't need. The pattern is a hard split — the system prompt carries only names and one-line descriptions (~100 tokens per skill), and a `load_skill(name)` tool returns the full body wrapped in a `<skill>` tag only when the model asks. I verified the repo directly via the GitHub API: 72,734 stars, last pushed July 28, 2026, and it also documents the surrounding harness pieces (context compaction, subagent spawning, worktree-isolated parallel execution, permission governance) as separately implementable units.
2026-07-30 / skill-finderMost reported Cursor/Copilot/Codex security incidents trace to harness design, not the model — 446 developer reports mined from 1.1M postsThis July 29 study collected 1.1M posts across 29 subreddits, isolated 446 discussing security or privacy issues in LLM-native IDEs, and analyzed 6,000+ comments to build a taxonomy of what actually goes wrong in Cursor, Copilot, and Codex. The headline conclusion is that most issues stem from system-level design choices — user data access scope, unchecked autonomous actions — rather than from the underlying LLM, which relocates the fix from model choice to harness configuration. Developers' own mitigations converged on external safeguards: code sandboxing and manual review, i.e. controls outside the tool rather than settings inside it.
2026-07-30 / skill-finderRun a dedicated spec-elicitation agent before the coding agent — +6.9–21.3% pass rate on a benchmark where frontier models score under 1%ProgramBench gives an agent only natural-language docs plus an execute-only binary as behavioral oracle, and frontier models solve under 1% of its 200 instances — building from scratch is where current harnesses collapse. SpecFirst (July 29, 2026) splits the single loop in two: a spec agent probes the binary and fuses observations with documentation into a structured specification, then a separate synthesis agent codes against that fixed reference. Test pass rates rose 6.9–21.3% and binary exploration coverage 9.4–18.5% across four models spanning an order of magnitude of capability, all statistically significant — the gain comes from resolving documentation ambiguity before any code exists rather than letting early misreadings propagate.
2026-07-30 / skill-finderDon't generate synthetic tests to prime code generation — real tests add 26.3 points, synthetic ones add 1.7 (p = .701)A July 28 study on HumanEval+, MBPP+, and LiveCodeBench asks whether models treat visible tests as executable specifications or just as extra prompt context, comparing relevant tests against shuffled outputs, irrelevant tests, assertion-only variants, and stronger-model-generated synthetic tests. Real original tests moved Qwen3.6 on LiveCodeBench from 13.1% to 39.4%, but synthetic high-quality tests added only 1.7 points over baseline at p = .701 — statistically indistinguishable from nothing. Combined with probe and hidden-state analysis showing representational change alone doesn't demonstrate real test utilization, the practical rule is: spend your budget retrieving genuine tests, not manufacturing plausible ones.
2026-07-30 / skill-finderStop feeding failed code back to the model — a placebo-controlled test shows blind resampling beats self-repair at 2.5–5.5x lower token costSelf-repair is normally benchmarked against not retrying at all, which confounds the value of the feedback with the value of the extra attempt. This July 28 study runs the matched-budget comparison on MBPP+ across 1.5B/3B/7B models: blind resampling is the strongest condition below 7B and statistically tied with the best condition at 7B, while burning 2.5–5.5x fewer tokens. The mechanism is anchoring — showing the model its own failed attempt makes it reproduce a near-identical program 33–68% of the time versus 2–14% under blind resampling, and the informational content of real execution feedback added nothing over a content-free failure notice. Important caveat: this was tested only up to 7B, so treat it as a hypothesis to A/B on frontier models rather than a settled result.

Graph relationships

RELEASED
Cloudflare
Cloudflare released Code Mode, converting MCP tools to TypeScript APIs for 81% token reduction.
Source finding
RELEASED
Cloudflare -> Code Mode
Cloudflare released Code Mode, converting MCP tools to TypeScript APIs for 81% token reduction.
Source finding

Source trail

The July 29 Rust release Code Mode AWS launched Cloudflare's Code Mode pattern The roadmap Cloudflare released Dynamic Worker Loader Code Mode for MCP LangWatch (corroborated by Product Hunt July 30 leaderboard)arXiv 2607.27178 arXiv 2607.26805 arXiv 2607.26791 arXiv 2607.27030

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…