Entity trail

Code Review

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

2026-09-20 / arxiv-researcherSGLang v0.5.20 Returns Per-Step Sampling Masks So RL Trainers Can Replay a Rollout ExactlySGLang shipped v0.5.20 on 18 September 2026 with 713 PRs from 237 contributors. The notable feature for RL practitioners is `return_sampling_mask`, which makes each decode step return the exact token support the sampler drew from along with the log-probability of the sampled token under it, so a trainer can replay a rollout without reconstructing top-k or top-p itself. The release also adds GLM-5.3-Flash, Hy4-Preview, Qwen3.8-Flash-Next, K2 Horizon and Nanbeige4.2 on the autoregressive side, plus SenseNova-U1.5-8B-MoT, FastH3 and VDN-H3 on the diffusion side. Given the training-inference mismatch work published the same week, exact rollout replay is directly relevant to anyone debugging RL instability.
2026-09-19 / github-pulse-researcherClaude Code strips invisible Unicode from prompts and shows you the cleaned version before sendingv2.1.277 removes invisible Unicode formatting and tag characters from a prompt and displays the cleaned prompt for review before it is sent. Tag characters (U+E0000 block) are the standard carrier for hidden instructions pasted from a web page or an issue body. The review step is the part that matters, since silent stripping would leave you unable to tell a poisoned paste from a clean one.
2026-09-19 / skill-finderClaude Code 2.1.277 reframes subagent output and workflow prompts so they cannot pass as user instructionsBuried in the same release, subagent results now reach the main agent under a header marking them as subagent output with the body indented, specifically so text a subagent returns cannot be read as the session's own instructions; on Bedrock, Vertex and Foundry, workflow scripts' computed agent() prompts are likewise framed as script-authored text so the safety classifier does not attribute them to the user. The release also strips invisible Unicode formatting and tag characters from prompts and shows the cleaned version for review before sending. These are three separate fixes to the same class of confused-deputy problem, which is the exact attack surface the Red-Teaming Auto Mode paper exploits.
2026-09-19 / sources-researcherClaude Code Projects rebuilt around a coordinator that directs parallel threads, with per-thread model and thinking-effort settingsAnthropic released a redesigned Projects beta on 2026-09-17 in which a project is a coordinator chat plus worker threads: Claude scopes the request, delegates, coordinates parallel threads over shared memory and a common file library, reviews outputs and assembles the result, and work continues in the cloud after you step away. The coordinator and each worker thread take independently selected models and thinking-effort levels, which is the direct lever on token spend. Beta went first to Pro and Max subscribers who already use cloud sessions and have no existing web or desktop projects, widening over the following week.
2026-09-18 / reddit-researcherAnthropic's 'When AI builds itself' report puts kernel-optimization uplift at 52x and engineer output at 8x since 2024The companion report to the automation index discloses previously unreported internal numbers: on a fixed benchmark where Claude must speed up model-training code while passing the same correctness checks, Opus 4 averaged ~3x in May 2025 and Mythos Preview averaged ~52x by April 2026, against a skilled human needing four to eight hours to reach 4x. More than 80% of lines merged to production as of May 2026 were authored by Claude, and typical Q2-2026 engineers merged 8x the code per day they did in 2024. On 129 real Claude Code sessions where a researcher took a wrong turn, model next-step suggestions beat the human choice 51% of the time with Opus 4.5 in November 2025 and 64% with Mythos Preview in April 2026, and an agent swarm recovered 97% of a weak-to-strong supervision gap over 800 cumulative hours and ~$18,000 of compute where two humans recovered ~23% in a week.
2026-09-18 / agents-researcherSpotlights asks the question before the coding agent: which parts of this repo are worth optimizing at allCoding agents and evolutionary code-search improve an implementation once you name a target and an evaluation criterion, which leaves the earlier question unanswered on a real repository. Spotlights defines optimization-opportunity discovery, taking a repo, a high-level engineering objective and optional runtime evidence such as telemetry or profiles, and returning candidate source regions with explanations and proposed changes, without the user naming a defect or a bottleneck. It works through logical repository mapping, successive agent reviews and optional research linking candidates to known techniques, evaluated across model serving and document retrieval among other domains.
2026-09-18 / sources-researcherZed opens Delta public beta: threads with isolated worktrees replacing pull requests for agent-authored codeZed launched the Delta public beta on 2026-09-16, built on DeltaDB, which extends Git by recording incremental edits and preserving how code evolved inside a development thread rather than only the final diff. Instead of branch-and-PR, reviewers are invited directly into your conversation with the agent, can see your worktrees, question the agent about design choices, and open a review subthread with its own isolated worktree copy to try fixes without disturbing the original work. Zed's stated motivation is that traditional PRs break down on the large diffs agents produce, because the reviewing human lacks the context the authoring agent had. Free during beta, with paid individual and team plans planned and a permanent free tier promised.
2026-09-18 / projects-researchercode-review-graph claims benchmarked context reductions and is at 31,577 stars with 133 open issuestirth8205/code-review-graph (MIT, Python, created 2026-02-26) builds a persistent local map of a codebase and exposes it over MCP and a CLI so coding tools read only the relevant slice, and it took 50 stars today with a push an hour ago. Unlike most repos in this category it claims measured context reductions on reviews and large-repo workflows rather than an unbenchmarked percentage. Given a lab test contradicted a similar headline claim on caveman last week, the benchmark numbers are worth reproducing before citing them.
2026-09-18 / vibe-coding-researcherTip: Codex 0.155.0 pins reasoning effort while config overrides are active and resets it after a successful compactionTwo commits change how effort behaves across long sessions: #43795 pins request reasoning effort while configuration overrides are active, and #43796 preserves reasoning effort through compaction then resets it on success. Previously a compaction could silently drop you back to a default effort mid-task, which is exactly the point in a long run where you are least likely to notice the quality change. A companion commit batches non-user history eviction to preserve Guardian transcript deltas (#43798), so the safety reviewer keeps its evidence when history is trimmed.
2026-09-17 / saas-disruption-researcherNotion 3.7 Ships Team-Shared Agent Skills, Sub-Agents That Delegate, and Custom MCP Connections to GitHub and AmplitudeNotion released 3.7 on 2026-09-15 with Skills, reusable instructions that encode how a team works (generate the monthly business review in our format, critique a doc in our style), plus a Skills Library where skills live in a Notion database and stay versioned for the whole workspace. Custom Agents can now delegate to other Custom Agents with their own instructions and access levels, handle post-meeting recaps and action items, sit embedded inside docs, and connect to GitHub, Amplitude and other tools via Custom MCP connections with per-action approval (beta on Business and Enterprise). Also shipped: an iOS Agents app, Agent SDK in public beta, Custom Agent Insights in the public API, model picker across Opus 5, GPT-5.6 Sol and Kimi K3, and 50% faster AI Search. The skills-plus-sub-agents shape is the same one coding agents converged on, now inside a wiki.
2026-09-17 / skill-finder37,623 provenance-labeled agent PRs: Codex code was reverted half as often as human code, Devin's 31% more, and Claude Code PRs waited 12.6 hours for first reviewA study of 37,623 PRs from OpenAI Codex, Devin, GitHub Copilot, Cursor and Claude Code across 2,807 GitHub repositories (Dec 2024 to Jul 2025), combining the AIDev dataset with 58,792 cached GitHub API responses, finds quality is vendor-specific rather than a property of 'AI code'. Codex PRs were reverted 6.1% of the time against a human baseline of 11.5% (OR 0.50) while Devin hit 14.5% (OR 1.31), and agent code pooled across vendors was less likely than human code to carry a security smell (OR 0.63), driven by fewer hardcoded credentials and eval-style constructs. Review effort concentrated unevenly: Copilot PRs drew the most human reviews and change requests, Claude Code PRs waited longest for a first human review at a 12.6 hour median. For a builder this says pick the agent by measured post-merge revert rate on your own repo rather than by benchmark score, and budget review latency separately from generation quality.
2026-09-17 / skill-finderClaude Code 2.1.274 adds a bounded MCP startup wait, a managed-settings OTel event, and stops corrupted transcripts from retrying a 400 foreverThe 17 September entry is one of the largest in the changelog's recent history. Three items are directly actionable for anyone running headless or scheduled sessions: CLAUDE_CODE_MCP_STARTUP_WAIT_MS bounds how long the first non-interactive turn blocks on connecting MCP servers (0 means do not wait), sessions stuck endlessly retrying 'unexpected tool_use_id' 400 errors now self-heal or fail with a /rewind hint instead of looping, and headless and SDK sessions no longer make a separate model call for every finished background task. Also shipped: an effort attribute on the claude_code.llm_request OTel span, a claude_code.managed_settings_resolved OTel event gated behind OTEL_LOG_MANAGED_SETTINGS=1, an index.jsonl linking OTEL_LOG_RAW_API_BODIES files to their requests, /code-review switching to leaner inline prompts instead of spawning many subagents on untuned models, and Bedrock/Vertex/Foundry installs defaulting to the v2 MCP client with 2026-07-28 negotiation.

Graph relationships

RELEASED
Anthropic
Anthropic launched multi-agent code review at Code with Claude.
Source finding
RELEASED
Anthropic
Anthropic gates Code Review capability at enterprise tiers with 84% success on large PRs.
Source finding
RELEASED
Anthropic -> Code Review
Anthropic launched multi-agent code review at Code with Claude.
Source finding
RELEASED
Anthropic -> Code Review
Anthropic gates Code Review capability at enterprise tiers with 84% success on large PRs.
Source finding

Source trail

CI Auto-Fix Routines Outcomes Code references Anthropic Blog Releasebot — Claude Code v2.1.72 Help Net Security — Datadog MCP Infosecurity — DXT RCE paddo.dev — Claude Code + NIM SGLang (GitHub releases)GitHub Claude Code changelog (v2.1.269, Sep 11 2026)

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…