Browser access for coding assistants
- OpenCode Browse lets coding assistants use your signed-in browser — Developers can automate existing web sessions and take back control for logins, authentication prompts or payments.
- Browse pairs a Chromium extension with a local server. It supports Chrome, Edge, Brave and other Chromium browsers, alongside OpenCode, Claude Code, Codex and other clients.
- An agent is a model running a multi-step task with tools. Browse connects through MCP, the standard for connecting agents to tools, and provides background tabs, screenshots and network capture.
- The MIT-licensed repository documents lending an existing tab and handing control back to a person. Only a real user click completes a handoff.
- Its site allowlist restricts navigation, but requests made by pages remain unrestricted. The read-only option blocks ordinary interactions, while JavaScript execution can still submit forms or change pages.
- A remote browser can connect over Tailscale or WireGuard. Browse itself leaves that connection unencrypted, so its documentation warns against exposing it to the open internet.
- Caveat: These restrictions are convenience controls, not a security boundary protecting your signed-in accounts from an assistant’s actions. (HN 281 points · 159 comments · product documentation and source repository)
Desktop privacy and permissions
-
A researcher finds Windows recording which coding assistants you use — Windows users have a specific diagnostic-data setting to review even when their coding tools run locally.
- Xusheng Li’s November 16 investigation follows executable-name checks through
MissionControlAggregator.dllinto saved usage bits and a diagnostic report. - On Windows 11 ARM64 build
26200.9974, renaming a copy ofcmd.exetocodex.exeset the Codex usage bit. The test records recognition by name. - Li found the assistant-usage event queued for upload when optional diagnostics were allowed. A clean US-region Windows Home installation enabled optional diagnostics by default in his test.
- The inspected report records which tools were observed, rather than prompts, source code or time spent using each tool.
- Microsoft’s settings guide explains diagnostic controls. Review Settings → Privacy & security → Diagnostics & feedback → Send optional diagnostic data.
- Caveat: The investigation demonstrates collection and upload queuing on tested configurations, not a verified rollout to every Windows installation or receipt by Microsoft. (HN 633 points · 487 comments · original reverse-engineering account)
- Xusheng Li’s November 16 investigation follows executable-name checks through
-
Apple adds narrower permissions for apps controlling your Mac — Developers of desktop assistants need to handle users granting access to particular folders or applications.
- Apple’s November 16 announcement introduces the changes in macOS 27.3 beta 2. Device Control and Data Access splits into System & App Protection Override and Accessibility.
- Users can choose specific protected folders or apps and configure access to ask each time. A visible indicator shows when one app controls another.
- Apple’s new support guide explains how to remove broad access and add the app again with narrower permissions.
- The changes also affect ordinary automation and utility apps. Backup and assistive-technology developers can request entitlements, with further details still pending.
- Caveat: Previously granted permissions remain in effect, so installing the beta does not automatically narrow existing access. (HN 144 points · 84 comments · Techmeme · Apple developer announcement and support guide)
Writing requirements that generated code can follow
- Explicit behavior rules reduce coding failures in a controlled study paper notes → — Writing down forbidden behavior, boundary cases and rules that must always hold gives developers checks beyond a tidy specification template.
- Elmira Onagh and Maleknaz Nayebi studied 72 participants, with 24 in each specification condition. Participants used ChatGPT-4o for two small Python tasks, graded against 33 hidden tests each.
- Mean failures fell from 21.54 without a required specification to 19.67 with a lightweight template and 11.38 with explicit behavioral obligations. The authors report reductions of 8.7% and 47.2%, respectively.
- Complete correctness remained rare: one of 24 submissions in the most explicit group passed every test. Neither other group produced a completely correct submission.
- After requirements changed, four of 24 explicit-group submissions regressed, versus two of 24 in each other group. The explicit group’s regressions lost less previously passing behavior.
- The replication package includes data, tests and analysis. The adjusted failure-count comparison lost statistical significance after correcting for multiple comparisons.
- Caveat: This tests specification work, including planning and reflection, on small ChatGPT-4o tasks rather than autonomous coding in production repositories. (arXiv 2611.20195 · submitted November 15)
Checking memory scores and recovery advice
-
Memory-test scores change when prompts and graders change paper notes → — Compare memory systems under matching answer instructions and grading rules before treating a higher score as better recall.
- The November 14 study uses LoCoMo, a benchmark, a fixed set of tasks for comparing models. It covers ten conversations and 1,540 questions, with two answering models and three graders.
- The authors report score changes up to 7.2 percentage points from switching the judge model, a model grading another model’s output. Changing answer instructions moved scores up to 9.3 points.
- Their Becos memory system scored 79.7% versus 73.0% for the complete conversation with GPT-4o mini and short answers. With GPT-4.1 mini, scores were nearly identical: 83.2% versus 83.1%.
- Becos supplied about 90% fewer context characters. These savings exclude the cost of storing and searching memories, and the longer-answer comparison changes several instructions at once.
- The released answers, verdicts and scoring script support rescoring and baseline reruns. Becos itself and its retrieved contexts are absent from that release.
- Caveat: Both authors work for Becos’s developer, and the tested configuration was developed on LoCoMo without a fully unseen evaluation set. (arXiv 2611.19199)
-
Recovery instructions sometimes make assistants stop after temporary errors paper notes → — Compare detailed advice against a neutral extra opportunity to continue, and verify the resulting application state separately.
- Sachin Gupta’s November 16 paper tests LangGraph and CrewAI against an isolated GitLab installation. Tasks supply the exact issue and report text, separating execution from writing quality.
- Across 1,944 planned primary runs, guided GPT-4o mini completed 57.4% of observed tasks versus 65.7% with neutral feedback. Missing outcomes leave the other two model comparisons inconclusive.
- All twenty pairs where neutral feedback succeeded and advice failed involved temporary rate limits. The guided assistants received feedback but made no further program requests and reported incompletion.
- A separate 624-run study of a revised policy found 161 correct states plus one unknown in 192 guided runs, versus 142 in 192 neutral runs. The policies combine execution information and advice.
- Final-state correctness, required source reads, verification and final reporting are scored separately. The paper describes its methods, but complete execution records and code are unavailable publicly.
- Caveat: These are constructed tasks within one GitLab workflow, with no independent human validation or evidence that one advice policy works generally. (arXiv 2611.20711)
Exporting notes before a service closes
- Mem sets shutdown dates and provides a notes migration guide — Users of the AI notes service have a concrete deadline to save their library outside the application.
- Notes become read-only on December 18, 2026. Access ends when Mem shuts down on January 15, 2027.
- Mem says charging has stopped. Previously paid features and limits remain available through the remaining service period, subject to the read-only date.
- The migration guide starts with downloading a ZIP from Settings → Exports → My Computer. It includes an Obsidian route and advises checking links, attachments, transcripts and selected recordings afterward.
- Caveat: A downloaded archive still needs checking in the destination app before you can rely on it as a complete migration. (HN 70 points · 105 comments · company notice and migration instructions)
All gathered items - what was cut and why (9)
- California’s lawsuit against OpenAI → DEDUP: yesterday Afternoon’s lead already covers the filing and company response. No further court development was verified.
- Google’s automated-intrusion report discussion → DEDUP: yesterday Afternoon’s attack-workflow item already covers the same report. A new discussion URL adds no new finding.
- Formal operations specifications → DEDUP: yesterday Afternoon’s paper item already explains the management-program results and proposed AI application.
- Dan Luu’s Astra testing comparison → STALE: the substantive account is worth reading, but verified HN submissions already link it on November 3. Its renewed front-page appearance is not a newly performed comparison.
- Nemotron 3.5 Bolt’s technical report → LOW_UTILITY: the November 16 paper is newly listed, but the model card dates the model release October 23. No separately verified new release or outside evaluation justifies replaying its vendor comparisons this edition.
- Anthropic’s reported three-day AMD migration → UNVERIFIABLE: the collected investor-facing claim lacks a checked public implementation and comparable measurement in this run. Its sweeping productivity comparison is withheld.
- A study of repeated agent handoffs → UNVERIFIABLE: the November-format identifier is paired with an October 14 submission and no later version on the inspected page. Its release timing could not be reconciled.
- Model ML’s funding report → LOW_UTILITY: the collected funding and valuation headline supplies no checked new tool release or engineering method for this edition.
- FlashMap’s speculative decoding announcement → STALE: the collected post remains dated October 28, with no verified material update since earlier cuts.