2026-05-15

• #0023

  #0023Agents & toolsGeekNews5 days ago

  `Codex` Lands in the `ChatGPT` Mobile App

  80radar

  Active coding sessions now carry over to the phone with thread state, approvals, plugins, and project context intact. This makes long-running agent work less desktop-bound and worth trying for on-call fixes or quick intervention.

  • —The app pulls in active threads from laptop, devbox, and remote environments, so work can continue across devices without restarting context.
  • —Approvals, plugins, project context, and live status are exposed on mobile; that turns the phone into a control surface, not just a viewer.
  • —Fast check-ins become practical when an agent is mid-run and needs a small decision, reducing the need to sit in front of the main machine.
  • —The value is highest for long-running coding tasks and remote sessions; short local edits still feel faster on desktop.

  Source: news.hada.io/topic?id=29516Read original →
• #0022

  #0022Agents & toolsGeekNews5 days ago

  `Codex` lands in the `ChatGPT` mobile app for remote agent control

  80radar

  CodexCoding agent — runs and reviews tasks remotely

  You can now approve, monitor, and manage coding runs from your phone. Strong quality-of-life upgrade if Codex is already part of your workflow.

  • —Mobile support centers on remote control: approve actions, check progress, and manage runs away from the desktop.
  • —This extends existing agent sessions instead of replacing desktop use, so handoff during commute or downtime gets much smoother.
  • —The practical gain is faster unblock cycles when long-running tasks need a quick human approval.
  • —Teams are not the story here; solo operators get the clearest benefit from being able to keep jobs moving anywhere.

  Source: news.hada.io/topic?id=29514Read original →
• #0021

  #0021Infra & SaaSvercel_blog5 days ago

  `AI Gateway` adds request-time provider ranking controls

  80radar

  AI GatewayAI routing service — failover across model providers

  Routing can now optimize on price, first-token latency, or throughput at request time instead of Vercel's blended default. Useful when one model has many providers and the cheapest or fastest route materially changes margin or UX.

  • —Set sort on providerOptions.gateway to 'cost', 'ttft', or 'tps' depending on whether margin, snappiness, or long-output speed matters most.
  • —Ranking is computed at request time, so newly added providers, price changes, and observed latency shifts flow through without code changes.
  • —Fallback is strict: providers are attempted in sorted order, and the next one is used only if the higher-ranked provider is unavailable.
  • —sort works with Zero Data Retention filtering and with order; pinned providers stay first, then the rest follow the chosen ranking.
  • —Each response exposes routing metadata with a sort block showing candidates, metric values, attempt order, and health-based deprioritization for debugging.

  Source: vercel.com/changelog/sort-providers-by-cost-latency-or-tRead original →
• #0020

  #0020Agents & toolsClaude Code Releases5 days ago

  `Claude Code` `v2.1.142`, more controllable background agents and much better daemon reliability

  80radar

  Background sessions are now configurable from the CLI and far less fragile across sleep, upgrades, worktrees, and MCP timeouts. If you rely on dispatched agents, this is a practical update worth taking now.

  • —claude agents adds --add-dir, --settings, --mcp-config, --plugin-dir, --model, and --effort. Background runs are easier to pin inside scripts and repeatable workflows.
  • —Fast mode now defaults to Opus 4.7 instead of 4.6, with CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1 as an escape hatch. Speed-oriented usage changes immediately without prompt edits.
  • —MCP_TOOL_TIMEOUT now correctly affects remote HTTP and SSE MCP servers, removing the old 60-second ceiling. Long-running tool calls are less likely to fail for the wrong reason.
  • —Sleep/wake reconnects, pre-existing git worktrees, binary-upgrade daemon exits, Chrome-extension crash loops, and Windows network-drive deadlocks were fixed together. Dispatched agents should disappear or wedge far less often.

  Source: github.com/anthropics/claude-code/releases/tag/v2.1.142Read original →
• #0019

  #0019Agents & toolsopencode_releases5 days ago

  `OpenCode` `v1.14.51`, experimental background subagents added

  70radar

  OpenCodeOpen-source coding agent — ships as CLI and desktop app

  You can now leave subagents running while continuing the main session. That makes longer coding tasks less blocking, and the release also cleans up several GPT-5, LiteLLM, Azure, and shell-output edge cases.

  • —Background subagents are the only real feature jump here: parallel task flow inside OpenCode gets more practical for longer refactors and research runs.
  • —LiteLLM now needs `v1.85.0-rc.2+` for current GPT-5 tool-call behavior, so older proxy setups can break quietly until upgraded.
  • —Azure gpt-5.5 completions requests were fixed, plus NVIDIA endpoints now send the required billing origin header for provider compatibility.
  • —Session reliability improved in several failure paths: cancelled assistant messages, repeated auto-compaction, truncated shell streams, and missing worktree POST bodies.
  • —Desktop app now shows MCP registration/auth status with direct re-auth flows, reducing connector debugging friction inside the client.

  Source: github.com/anomalyco/opencode/releases/tag/v1.14.51Read original →
• #0018

  #0018Agents & toolsCline Releases5 days ago

  `Cline CLI` `v3.0.3` adds isolated runs with `--worktree`

  70radar

  ClineOpen-source coding agent — works in CLI and editors

  Safer branch-level experimentation lands without manual repo setup. The new --worktree flow plus resume support makes parallel agent attempts much easier, and the auth/filepath fixes remove several sharp edges.

  • —--worktree auto-creates a fresh git worktree under ~/.cline/worktrees/; paired with --taskId and --continue, it enables isolated retries of the same task.
  • —The history UI now shows live session status and refreshes rows in place, which matters when several long-running CLI jobs are open at once.
  • —OpenAI-compatible auth was restored and model metadata now survives config or migration, reducing provider setup churn.
  • —macOS screenshot paste/read failures caused by Unicode path variants like U+202F are fixed, removing a subtle TUI reliability bug.
  • —--thinking none now works for reasoning-enabled Ollama models, and detached hub event errors no longer crash sessions.

  Source: github.com/cline/cline/releases/tag/cli-v3.0.3Read original →
• #0017

  #0017Agents & toolsGitHub Changelog5 days ago

  `GitHub Copilot app` enters technical preview as a desktop agent workspace

  70radar

  GitHub Copilot appDesktop coding agent app — starts from GitHub context

  GitHub is turning Copilot into a GitHub-native desktop workspace for isolated agentic coding. If the handoff from issue to execution is smooth, this could become a serious alternative to browser-first agent flows.

  • —Positioning is clear: a native desktop surface tied to GitHub work, not just an editor add-on.
  • —The promise is isolated execution and steering, which matters when testing agent runs without polluting the main repo state.
  • —Technical preview means the direction is notable now, but reliability and workflow depth still need proof.
  • —If it reduces friction from issue, PR, or repo context into action, it can compete with standalone coding agents.

  Source: github.blog/changelog/2026-05-14-github-copilot-app-is-nRead original →
• #0016

  #0016Agents & toolsopencode_releases5 days ago

  `OpenCode` `v1.15.0`, core event system and JS SDK event fixes

  60radar

  OpenCodeOpen-source coding agent — built for extensible sessions and events

  Event delivery is now more reliable across sessions and integrations, and missing JavaScript SDK event types are back. Useful if you build tooling on top of OpenCode; desktop users get a small quality bump, but this is mainly an infrastructure-stability release.

  • —The core now uses an Effect-based event system, aiming for more complete event delivery across sessions and integrations.
  • —Custom tool loading no longer fails on invalid exports, so one bad module is less likely to break the whole setup.
  • —Project instruction lookup errors are ignored during load, which keeps sessions opening even when discovery is flaky.
  • —Versioned event projector lookup was fixed, reducing the risk of replaying events with the wrong handlers.
  • —The JavaScript SDK restores missing session and message event types, which matters if you consume OpenCode events programmatically.

  Source: github.com/anomalyco/opencode/releases/tag/v1.15.0Read original →
• #0015

  #0015Generative mediaGeekNews5 days ago

  `Supertonic 3` launches ultra-light on-device TTS with 31 languages and emotion tags

  60radar

  Supertonic 3On-device TTS engine — multilingual with emotion tags

  A small-footprint TTS now handles expressive cues like laugh and scream, while improving pronunciation and voice cloning. Strong fit for offline narration, in-app voice UX, and cost-sensitive shipping.

  • —Supports 31 languages including Korean, which lowers the friction for multilingual voice features without a cloud dependency.
  • —Adds 10 emotion tags such as laugh, breath, and scream, so scripted dialogue can sound less flat with simple text markup.
  • —Pronunciation accuracy improved, and repetition or omission failures were reduced; this matters more than flashy demos in production TTS.
  • —Voice cloning quality also improved, making it more usable for character voices, guided audio, or branded app narration on-device.

  Source: news.hada.io/topic?id=29522Read original →
• #0014

  #0014Agents & toolsGeekNews5 days ago

  `DeerFlow` `2.0`, ByteDance’s Long-Running SuperAgent Harness

  60radar

  DeerFlowOpen-source agent harness — parallel orchestration for long tasks

  Breaks multi-hour work into parallel sub-agents for research, coding, and content generation. Worth tracking if you want a self-hosted orchestration layer instead of bolting long-running loops onto a coding agent.

  • —Handles research, coding, and content jobs in one harness, aiming at workflows that run for minutes to hours rather than single prompts.
  • —v2 is a full rewrite with no shared code from the Deep Research-style v1, so expect architecture changes instead of an incremental patch.
  • —Built on LangGraph and LangChain, which lowers adoption cost for teams already using that stack but adds framework dependency.
  • —The useful angle is orchestration, not model novelty. Strong fit when one agent is no longer enough and ad-hoc scripts start to sprawl.

  Source: news.hada.io/topic?id=29518Read original →
• #0013

  #0013Agents & toolsr/ClaudeAI4 days ago

  Claude UI Cloning Works Better With Exact Design Specs, Not Longer Prompts

  50radar

  Predictable UI generation came from exact tokens, full states, spacing scales, and navigation graphs. Useful as a reusable spec checklist before asking Claude to build screens.

  • —Exact values beat fuzzy labels: #1A1A1A is stable, while “dark gray” creates inconsistent screens.
  • —State coverage matters early: empty, loading, error, and filled states reduce model-made UI decisions.
  • —A 4/8/16/24 spacing scale produced steadier layouts than annotating every per-element gap.
  • —Navigation written as a graph removes button-destination guessing, especially in multi-screen clones.

  Source: www.reddit.com/r/ClaudeAI/comments/1tdwau0/i_turned_50_pRead original →
• #0012

  #0012OtherGitHub Changelog4 days ago

  `GitHub Projects` adds built-in timestamp fields

  50radar

  Project views can now sort and filter by Created, Updated, and Closed without manual columns. Small change, but it removes admin work and makes backlog aging and closure tracking much easier.

  • —Three built-in fields are now available: Created, Updated, Closed. No custom date-property setup needed.
  • —These fields can be added to any project view, so stale items and recently touched work become easier to spot.
  • —Closed is useful for measuring cycle endings and reviewing what actually shipped in a given window.

  Source: github.blog/changelog/2026-05-15-timestamp-fields-in-gitRead original →
• #0011

  #0011Idea signalsr/SaaS4 days ago

  `Stripe Billing` fee pain is turning into a switch-or-build signal

  50radar

  Once recurring revenue grows, billing logic can feel expensive enough to trigger migration math. A complaint at $2,500/month suggests room for leaner billing stacks, migration tools, or hybrid in-house setups.

  • —Monthly spend hit $2,500 for billing logic alone, separate from card processing fees; the pain starts after early-stage convenience stops being cheap.
  • —The extra cost is framed as nearly 1% on top of payment fees, which makes Stripe Billing feel overpriced once volume rises.
  • —The decision set is already clear: stay, switch, or build internally. That usually means the product category has weak lock-in at this price point.
  • —A narrow product opportunity sits in migration tooling, subscription logic primitives, or cost dashboards that show when hosted billing stops making sense.

  Source: www.reddit.com/r/SaaS/comments/1tduh1u/stripes_billing_fRead original →
• #0010

  #0010Indie businessGeekNews5 days ago

  `Claude for Small Business`, workflow pack for SMB tools

  50radar

  Instead of a new model, this packages approvals around finance and ops actions. Useful if you want Claude to sit on top of back-office tools without giving it fully autonomous control.

  • —Connects QuickBooks, PayPal, and HubSpot into task-specific workflows, so the product is aimed at real back-office execution, not just chat.
  • —Users pick tools and a job inside Claude Cowork, then Claude runs the flow. That lowers setup friction compared with building custom agent glue first.
  • —Sending, publishing, and payment actions still need human approval, which makes it more practical for revenue and ops tasks than fully hands-off automation.

  Source: news.hada.io/topic?id=29525Read original →
• #0009

  #0009Idea signalsOpenAI5 days ago

  `ChatGPT` adds a personal finance preview for U.S. `Pro` users

  50radar

  Account-linked financial guidance is moving into mainstream assistants. That raises the bar for finance UX, but the immediate opening is narrower workflow tools, not a full-stack consumer fintech clone.

  • —Access is limited to U.S. `Pro` users in preview, with no API or team angle mentioned, so this is more product-direction signal than an immediate builder tool.
  • —The differentiator is secure financial account connection plus advice grounded in goals and priorities; context-rich vertical agents are becoming the expected UX.
  • —Finance is a hard market on trust, compliance, and data access. A focused layer like budgeting explainers, cash-flow summaries, or decision support is more buildable than a broad robo-advisor copy.

  Source: openai.com/index/personal-finance-chatgptRead original →
• #0008

  #0008Infra & SaaSvercel_blog5 days ago

  `Vercel CLI` adds native `curl` syntax

  50radar

  Protected deployments can now be hit with plain curl-style requests using your Vercel auth. It removes header and token boilerplate for preview smoke tests, so updating the CLI is worth it.

  • —vercel curl now accepts full URLs, bare hostnames, and the --url flag, so existing curl habits mostly transfer over.
  • —Vercel auth is reused to bypass Deployment Protection, which cuts the usual preview-token and header setup.
  • —Linked projects can pass just a path instead of a full URL, making repeated endpoint checks much faster.
  • —This is a small feature, but it sharpens local-to-preview debugging loops where protected environments slow teams down.

  Source: vercel.com/changelog/use-native-curl-syntax-with-vercel-Read original →
• #0007

  #0007Otherr/LocalLLaMA5 days ago

  Self-Training With Verifiable Rewards Pushes `Qwen 2.5` 7B to **112/164** on HumanEval

  50radar

  A self-generated code-and-tests loop produced a large jump without human-written training pairs. Cheap enough to replicate, but still a one-off experiment rather than a product-ready recipe.

  • —The loop is simple: generate problems, sample multiple solutions, keep (failed attempt, fixed attempt) pairs, and let a Python interpreter score them.
  • —After fixing a grading bug, Qwen 2.5 7B moved from 25 to 112/164 on HumanEval; that is a big enough jump to treat as a real benchmark signal.
  • —A Qwen 2.5 14B run used 100 mined pairs, took 95 minutes on an H100, and cost $3.50; the barrier here is much lower than typical RL folklore.
  • —Control training on fake pairs gave 25/164, identical to base, which suggests the lift came from correction data rather than format imitation.

  Source: www.reddit.com/r/LocalLLaMA/comments/1tde3m1/i_let_a_smaRead original →
• #0006

  #0006OtherGeekNews5 days ago

  `scrcpy` `4.0` adds flexible virtual display resizing

  50radar

  scrcpyAndroid mirroring tool — low-latency control over USB or Wi-Fi

  Virtual displays can now resize to match the client window with --flex-display. The SDL3 migration should reduce resize-related friction, so it's a solid quality-of-life update for Android testing setups.

  • —--flex-display or -x lets the virtual display track the client window size dynamically, which cuts friction when testing multiple Android layouts.
  • —The move from SDL2 to SDL3 shifts the app onto an actively maintained graphics layer, improving upstream support and bug-fix velocity.
  • —Window resizing gets safer operationally because display and client size stay aligned instead of drifting during interactive mirroring.
  • —This is not a workflow reset, but teams using scrcpy daily for QA, demos, or device debugging will feel the upgrade immediately.

  Source: news.hada.io/topic?id=29505Read original →
• #0005

  #0005Agents & toolsGeekNews5 days ago

  `Claude Code`/`Codex` skill adds post-task learning drills

  40radar

  Turns agentic coding into guided practice, not just output. After refactors or schema work, it offers 10-15 minute optional drills, so the tool can train judgment while shipping.

  • —After tasks like new file creation, schema changes, and refactors, it suggests a short follow-up exercise instead of ending at code output.
  • —The 10-15 minute format is small enough to fit real workflow, which makes deliberate practice more likely to stick.
  • —Useful signal for agent UX: copilots can improve operator skill, not only automate execution.
  • —Best read as a reusable interaction pattern for coding agents, not a platform-level product launch.

  Source: news.hada.io/topic?id=29531Read original →
• #0004

  #0004OtherSimon Willison5 days ago

  `datasette-llm-limits` `0.1a0` adds per-user LLM spend caps

  40radar

  DatasetteOpen-source data publishing tool — strong plugin extensibility

  LLM usage inside Datasette can now be hard-capped by user or globally with rolling windows and USD limits. Useful if you expose prompts to users; otherwise the impact is narrow but the cost-control pattern is worth copying.

  • —Works with datasette-llm and datasette-llm-accountant, so enforcement sits on top of existing usage tracking.
  • —Config supports per-user or global limits, with windows like rolling-24h and explicit amount_usd caps.
  • —Example policy sets $1.00 per actor over 24 hours, which is a clean template for metered AI features.
  • —Best fit is multi-user internal tools or lightweight SaaS surfaces where prompt access needs a hard budget guardrail.

  Source: simonwillison.net/2026/May/15/datasette-llm-limits/#atomRead original →
• #0003

  #0003Agents & toolsOpenAI5 days ago

  `Codex` Workflows for Sales Ops: Briefs, Prep, Forecasts

  40radar

  Sales artifacts become agent jobs: briefs, prep packets, reviews, plans, and deal diagnosis. Useful as a template for turning messy business inputs into repeatable Codex tasks.

  • —Codex is positioned beyond code: pipeline briefs, meeting prep, forecast reviews, account plans, and stalled-deal analysis from real work inputs.
  • —The reusable pattern is artifact generation from scattered context. That maps cleanly to founder sales, support triage, churn review, and weekly ops.
  • —No new product capability, pricing, or benchmark is included. Value sits in workflow design, not platform news.

  Source: openai.com/academy/codex-for-work/how-sales-teams-use-coRead original →
• #0002

  #0002Agents & toolsOpenAI5 days ago

  `Codex` workflows for data science team outputs

  40radar

  CodexCoding agent — turns work inputs into code and analysis outputs

  A practical prompt-pattern set: turn work inputs into briefs, KPI memos, scoped analyses, and dashboard specs. Useful if analytics work keeps blocking product decisions.

  • —Covers five output types: root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specs.
  • —The angle is not code generation; Codex is positioned as an analyst agent that converts raw work context into decision artifacts.
  • —Best reuse: standardize recurring metric investigations into templates, then feed tickets, logs, and dashboards as inputs.

  Source: openai.com/academy/codex-for-work/how-data-science-teamsRead original →
• #0001

  #0001Agents & toolsGoogle AI Forum5 days ago

  `Antigravity` hangs when Git repo uses `extensions.worktreeConfig=true`

  40radar

  AntigravityCoding agent tool — automates work inside the IDE

  A Git config edge case can stall agent sessions completely while chat still accepts prompts. If your repo enables extensions.worktreeConfig, unset it for now; otherwise Antigravity is unreliable in affected workspaces.

  • —The failure is reproducible on macOS with Antigravity 1.107.0 when opening a nested folder inside a repo using extensions.worktreeConfig=true.
  • —Logs show workspace resolution breaking on worktreeconfig, then workspace infos is nil; the chat UI looks alive but the agent stops responding.
  • —Restarting or signing in again does nothing until running git config --unset extensions.worktreeConfig and reloading the language server or window.
  • —This matters if you use Git worktrees or per-worktree config: repository features, not auth or network, can silently brick the coding agent.

  Source: discuss.ai.google.dev/t/antigravity-agent-becomes-unrespRead original →