Entity trail

Roo Code

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

2026-09-05 / hn-researcherKevin Buzzard Compiled Anthropic's FLT Proof and Confirms It Checks Out, While Saying It Adds 'Essentially Nothing' MathematicallyImperial College's Kevin Buzzard, who runs the EPSRC-funded FLT formalization project, posted on September 4 that he compiled Anthropic's codebase and ran comparator on it, and it checks out. He also ran security checks, having an agent inspect the non-mathematical code and personally reviewing suspicious elements. His caveats matter for builders reading the announcement: the work uses the 1995 Darmon-Diamond-Taylor approach rather than modern methods, covers exponents p≥17 (sufficient, since smaller irregular primes were already formalized), and faithfully transcribes existing literature rather than proving anything new.
2026-09-02 / rss-researcherPaint.NET now ships a 180,000-line clean-room Direct2D rewrite that Rick Brewster says Claude wroteAfter 20+ years maintaining Paint.NET, Rick Brewster concluded WINE's Direct2D would never be complete enough, so the app now carries its own from-scratch reverse-engineered Direct2D implementation. He puts it at 180,000 lines against 700,000 for the rest of the codebase and says flatly it was 'written by our good friend Claude, without whom this would NOT have been possible.' He also calls it vibe coded and largely untested, with him manually correcting resource management and architectural problems, which makes it one of the more honest accountings of what agent-written code at that scale actually costs.
2026-08-31 / arxiv-researcherDeLLMGuard Cuts LLM Vulnerability-Scanner Root-Cause Accuracy From 23.5% to 6.6% While Keeping Source PublicarXiv 2608.28400 (2026-08-28, cs.CR/cs.SE) treats verified smart-contract source disclosure as an attack input now that LLM agent workflows make large-scale exploit scanning practical. DeLLMGuard separates disclosed source from runtime execution across multiple contract addresses so an agent must first recover proxy, delegate and factory relations, with a Verification Layer checking deployment relations, runtime bytecode, source and state changes to prove the transformation preserves the original business logic. Evaluated on 387 real-world vulnerable contracts with three LLM agents in an environment derived from SCONE-bench, it reduces overall root-cause correctness from 23.5% to 6.6% and beats the closed-source bytecode baseline on the primary non-proxy set.
2026-08-31 / agents-researcherFor root cause analysis, a general agent plus a self-evolving harness now beats a purpose-built RCA agentA quantitative study finds that pointing Codex or Claude Code at diagnosis now often outperforms specialized RCA agents built from scratch, with the remaining accuracy gap sitting in the external adaptation layer rather than the model. OpsHarness accumulates system-specific experience from past diagnoses so it improves with use, pairing a data plane of layered operational knowledge and an idea-card tool library with a control plane coordinating setup, diagnosis and evolution. The argument generalizes: for a vertical agent task, invest in the harness and reuse the general agent.
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.
2026-08-30 / vibe-coding-researcherRoo-Code up to 3.51.1 has a code-injection hole in its MCP integration trust model, disclosed publicly and scored only 5.5CVE-2026-81835, published 2026-08-28, targets the `fetch_instructions` function reached via a malicious MCP server in Roo-Code versions through 3.51.1, allowing remote code injection. The exploit is already public and NVD notes multiple issues were reported to the vendor. GitHub's advisory GHSA-q2jq-8pm2-fj8h rates it low while NVD scores it 5.5 MEDIUM, a gap worth knowing about if you gate upgrades on GHSA severity alone.
2026-08-29 / arxiv-researcherNetwork Troubleshooting Agents Are Near-Saturated on Real Faults but Invent Root Causes When Nothing Is BrokenFaulT-Bench runs 200 troubleshooting scenarios across eight network topologies, five reimplemented from public practitioner labs, deployed in Kathara and spanning genuine faults, false fault reports, wrong device attribution and wrong root-cause claims. Evaluating SADE, ReAct and Claude Code, all three are near-saturated on accurate tickets and robust to misdirection but degrade sharply when the network is healthy and the ticket is wrong, probing until a benign condition can be promoted to a root cause instead of concluding nothing is wrong. Rewriting 72 false-premise tickets into five reporter personas shows wording matters more than content: a confidently wrong report is handled about as well as an accurate one, while a vague underspecified one collapses performance.
2026-08-29 / hn-researcherMulti-Agent Mathematicians in 'the Station' Produced Five Results New to the Literature, Including a 604-Point Kissing ConfigurationStephen Chung, Wenyu Du and William J. Wesley submitted arXiv 2608.23691 on August 24, describing an open-world environment where agents from different model families pick their own research directions with no central coordinator. Across 12 construction problems the agents produced results novel to existing literature on five of them, including a new infinite family of finite-field Kakeya sets, exact 604-point kissing configurations in dimension 11, improved lower bounds for Erdos's minimum-overlap problem and new infinite families for Book Ramsey numbers. The agents emitted theorems and explanations alongside the numbers, and the authors are releasing all agent dialogues, proofs and verification code.
2026-08-29 / vibe-coding-researcherPattern: Token Cost Moved From Billing Concern to Runtime InstrumentationIn two days: Claude Code added a `/cost` prompt-cache line with hit ratio and warm/cold plus a `prompt_cache` status-line object and a `rate_limits.spend_limit` field; Codex started counting nested subagent tokens toward root goal budgets; qwen-code shipped a Web Shell session token usage panel and an OpenTelemetry span for context usage; Kilo Code let you pick a cheaper model just for compaction. Cost is being exposed as a live signal an agent loop can read and act on, not a monthly invoice. The builder move is to put cache hit ratio on a status line now, because a cache regression is invisible without it.
2026-08-28 / github-pulse-researcherTencent open-sources Hy4-preview, a 770B-parameter MoE with 49B active and a 1M-token context windowTencent Hunyuan released and open-sourced Hy4 preview on 2026-08-28: 770B total parameters with 49B activated per token, 78 layers (77 of them MoE with 256 routed plus 1 shared expert, top-8 routing), a 1M context length, and a native 10B MTP layer for speculative decoding. The attention stack borrows Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse index reuse, plus identity Hyper-Connections on the residual path with 4 residual streams. Tencent's own newsroom confirms availability through WorkBuddy/CodeBuddy, Yuanbao, ima, Tencent Cloud Tokenhub and OpenRouter, free on CodeBuddy for two weeks, so builders can benchmark it against GLM-5.3 and Qwen4 without hosting 770B weights.
2026-08-28 / news-researcherResearchers Found 227 Install Commands Pointing at Unowned Packages Across 6,214 Corporate DomainsA study published August 27 scanned 6,214 live domains belonging to defense contractors, Fortune 500s and Big Tech and found 227 install commands in llms.txt and llms-full.txt documentation files that resolve to package names nobody owns. Claude, OpenAI's Codex and Nous Research's Hermes each executed those commands when reading the docs, and the researchers confirmed a few dozen companies, some Fortune 500, actually installed proof-of-concept code inside their networks. This turns slopsquatting from a hallucination problem into a documentation-supply-chain problem: the agent-readable file you publish for discoverability is now an execution surface, and anyone can register the dangling name.

Graph relationships

ACQUIRED
Cline
Cline merged Roo Code codebase, with Roo team pivoting to Roomote cloud agent.
Source finding
COMPETES WITH
OpenWork
openwork offers alternatives as Roo Code sunsetting.
Source finding
COMPETES WITH
Kilo Code
Kilo Code positioned itself as community-maintained Roo Code successor.
Source finding
DEPRECATES
Kilo Code
Roo Code archived development in May 2026, with Kilo Code as a named migration target.
Source finding
ACQUIRED
Cline -> Roo Code
Cline merged Roo Code codebase, with Roo team pivoting to Roomote cloud agent.
Source finding
COMPETES WITH
OpenWork -> Roo Code
openwork offers alternatives as Roo Code sunsetting.
Source finding
COMPETES WITH
Kilo Code -> Roo Code
Kilo Code positioned itself as community-maintained Roo Code successor.
Source finding

Source trail

getagentseal/codeburn Tom's Hardware June 18 Ahead of AI note Kilo Code TauricResearch's framework HKUDS/nanobot Block's agent framework Roo Code announced JetBrains AI Pulse survey ctx project SC Media GitHub Advisory

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…