2026-05-01

• #0010

  #0010Models & APIr/LocalLLaMA3 weeks ago

  `gemma-4-31B-it-DFlash` released, but blocked on `llama.cpp` support

  80radar

  gemma-4-31B-it-DFlashModel — a 31B Gemma variant tuned with DFlash

  Weights are up on Hugging Face, but local testing is still blocked by unmerged llama.cpp PR #22105. Useful only for tracking right now; wait for merge before judging real usability.

  • —The model is already published at huggingface.co/z-lab/gemma-4-31B-it-DFlash, so distribution started before runtime support landed.
  • —Testing is gated by ggml-org/llama.cpp PR #22105; without that merge, local inference flow is effectively blocked.
  • —This is a release you bookmark, not deploy today. The next real checkpoint is PR merge, then compatibility and performance checks.

  Source: www.reddit.com/r/LocalLLaMA/comments/1t0s4qv/gemma431bitRead original →
• #0009

  #0009Agents & toolsSimon Willison3 weeks ago

  `Codex CLI` 0.128.0 adds the `/goal` autonomous loop

  80radar

  Codex CLICLI coding agent — /goal enables autonomous runs

  Codex CLI 0.128.0 introduces /goal, a loop that keeps working until it judges the task complete or a configured token budget runs out. For indie developers, this is meaningful because it pushes Codex CLI closer to a lightweight autonomous coding agent without requiring extra orchestration.

  • —The new /goal command gives Codex CLI a self-continuation loop, similar in spirit to the Ralph loop mentioned in the article.
  • —The article suggests the feature is driven mainly by injected prompt templates such as goals/continuation.md and goals/budget_limit.md, implying prompt-layer agent behavior rather than a deeply new runtime architecture.
  • —From an indie developer perspective, this can reduce manual prompting for longer coding tasks, but the token budget limit means cost control and stopping conditions remain important.

  Source: simonwillison.net/2026/Apr/30/codex-goals/#atom-everythiRead original →
• #0008

  #0008Agents & toolsHermes Agent Releases3 weeks ago

  `Hermes Agent` `v0.12.0` adds self-maintaining `Curator` and faster startup

  80radar

  Hermes AgentAgent tool — self-maintaining work stack with Curator

  A background Curator now grades, merges, and prunes skills on a 7-day cycle, pushing Hermes closer to self-maintenance. With broader integrations and a ~57% TUI cold-start cut, it looks more usable as a daily agent stack now.

  • —The release bundles an autonomous hermes curator that ranks usage, archives weak skills, and writes reports to logs/curator/run.json and REPORT.md; skill sprawl gets managed without manual cleanup.
  • —The self-improvement fork moved to a rubric-first flow, prefers updating the actively used skill, inherits live provider/model/credentials, and stays limited to memory plus skills; fewer flaky saves, less agent drift.
  • —LM Studio became a native provider with auth, hermes doctor, reasoning transport, and live /models listing; local-model setups got much less hacky.
  • —ComfyUI v5 and TouchDesigner-MCP are now bundled by default, while Spotify and Google Meet integrations landed too; the default install covers more real workflows out of the box.

  Source: github.com/NousResearch/hermes-agent/releases/tag/v2026.Read original →
• #0007

  #0007Agents & toolsOpenAI Codex Releases3 weeks ago

  `Codex CLI` `0.128.0`, persisted `/goal` flows and tighter agent controls

  80radar

  Codex CLICoding CLI — persistent /goal and agent controls

  Long-running agent work is now first-class with persisted /goal, pause/resume, and runtime continuation. Stronger permission profiles and plugin lifecycle controls make unattended CLI automation more usable now.

  • —Persisted /goal adds create, pause, resume, clear, app-server APIs, and runtime continuation; longer jobs no longer need ad hoc orchestration.
  • —Permission handling got sharper with built-in profiles, sandbox profile selection, cwd controls, and active-profile metadata; safer defaults for automation work.
  • —Plugin ops expanded to marketplace install, remote uninstall, bundle caching, bundled hooks, and external-agent config import; setup friction drops noticeably.
  • —codex update, configurable TUI keymaps, /statusline and /title edits, plus resume/interruption fixes reduce day-to-day CLI rough edges.

  Source: github.com/openai/codex/releases/tag/rust-v0.128.0Read original →
• #0006

  #0006Agents & toolsClaude Code Releases3 weeks ago

  `Claude Code` `v2.1.126` adds gateway model picker, project purge, broader permission bypass

  70radar

  Claude CodeCoding CLI — stronger gateway sync and project reset

  Model selection now reads compatible gateway /v1/models, and project state can be fully cleared with claude project purge. The release also loosens --dangerously-skip-permissions and fixes enough auth, timeout, and Windows issues to make daily use smoother right now.

  • —/model now lists models from a gateway's /v1/models when ANTHROPIC_BASE_URL targets an Anthropic-compatible endpoint; easier to swap self-hosted or routed backends.
  • —claude project purge [path] deletes transcripts, tasks, file history, and config state, with --dry-run, --yes, --interactive, and --all; useful when resetting broken local context.
  • —--dangerously-skip-permissions now skips prompts for writes to .claude/, .git/, .vscode/, and shell config files, while catastrophic delete commands still require confirmation.
  • —OAuth login got a fallback for pasted codes in terminal plus fixes for localhost callback failures, slow proxies, IPv6-only devcontainers, and rare refresh-token races.
  • —Windows handling improved: more PowerShell 7 install paths are detected, PowerShell becomes the primary shell when enabled, and CJK text rendering bugs were fixed.

  Source: github.com/anthropics/claude-code/releases/tag/v2.1.126Read original →
• #0005

  #0005Agents & toolsGitHub Changelog3 weeks ago

  `GitHub Copilot` for `Visual Studio` adds cloud agents, user-level custom agents, and a debugger agent

  70radar

  Visual Studio is moving from inline assist to full agent workflows inside the IDE. Useful if you already live in Microsoft tooling; less urgent if your stack is centered on lighter editors.

  • —Cloud agent sessions now start directly from the IDE, cutting friction for remote agent runs and longer tasks.
  • —Custom agents gain user-level support, so personal workflows no longer need team-wide setup or admin overhead.
  • —A new Debugger agent can validate issues during debugging, pushing Copilot closer to task execution instead of suggestion-only use.

  Source: github.blog/changelog/2026-04-30-github-copilot-in-visuaRead original →
• #0004

  #0004Indie businesspractical_founders_podcast3 weeks ago

  `ProSpend`: Selling as a Growth Move, Not an Exit

  60radar

  ProSpendSpend management SaaS — mid-market focused, expanded into AP

  A 10-year bootstrap reached 1,000 customers and sold to a strategic buyer to speed UK expansion. The useful lesson: acquisition can replace VC when distribution, market access, and control matter more than pure ownership.

  • —Started with expense management, then expanded into AP, POs, and budgets; adjacent finance workflows lifted ACV instead of chasing new personas.
  • —ACVs sat around $15K-$40K with about 1,000 customers and 50 employees; a focused mid-market wedge scaled into a solid SaaS base.
  • —Channel partnerships with MYOB and resellers now drive about 50% of new customers; indirect distribution became a major growth lever.
  • —Meaningful founder income took 5-6 years; bootstrapping preserved control, but demanded long tolerance for cash-flow pressure.
  • —The 2025 sale to ISH kept Sharon operating the company while adding UK market access; selling worked as a growth financing choice, not a finish line.

  Source: practicalfounders.com/podcast/why-selling-your-company-cRead original →
• #0003

  #0003Otherr/ClaudeAI3 weeks ago

  Practical `Claude` business playbook built from **35+** founder stories

  50radar

  A GitHub repo packages repeatable patterns for turning Claude into daily business ops, not demos. Useful if you want tested workflow scaffolding without digging through dozens of threads.

  • —The repo condenses 35+ Reddit founder threads into one operating manual, cutting discovery time fast.
  • —Coverage spans service businesses, micro-SaaS, persistent memory, and daily workflows, so it is broader than prompt tips.
  • —It includes an Infinity Barrier guardrail pattern, signaling that failure containment is treated as a first-class design concern.
  • —The Vibe → Value framework and revenue-making archetypes make it more actionable than generic AI-builder inspiration.

  Source: www.reddit.com/r/ClaudeAI/comments/1t0in32/i_built_a_praRead original →
• #0002

  #0002Models & APIAnthropic TypeScript SDK3 weeks ago

  `anthropic-sdk-typescript` `v0.92.0` improves Managed Agents and env-based headers

  50radar

  anthropic-sdk-typescriptTypeScript SDK — official client for Anthropic APIs

  Managed Agents got API polish, and headers can now be set via env without extra wiring. Useful if you script multi-env deployments; otherwise this is a small but practical SDK update.

  • —Managed Agents APIs were improved in v0.92.0, which lowers friction if you already build agent flows on Anthropic's SDK.
  • —Headers can now be configured via environment variables, making staging, tenant, or proxy-specific setup easier to automate.
  • —Bedrock error events delivered in chunk frames now raise APIError, reducing silent failure handling edge cases in streaming paths.

  Source: github.com/anthropics/anthropic-sdk-typescript/releases/Read original →
• #0001

  #0001Models & APIAnthropic TypeScript SDK3 weeks ago

  `Anthropic Bedrock SDK` `v0.29.1` normalizes chunk-frame errors to `APIError`

  40radar

  Anthropic Bedrock SDKSDK — Anthropic client tooling for Amazon Bedrock

  Streaming failures from Bedrock chunk frames now surface as APIError instead of irregular error events. Small release, but it removes edge-case handling in stream consumers and lowers retry/debug friction.

  • —Released on 2026-04-30 as a patch over v0.29.0 with a single Bedrock fix, so upgrade scope is narrow.
  • —Error events delivered inside chunk frames now throw APIError; downstream code can handle stream failures with one error path.
  • —If you pipe Bedrock streams into logs, retries, or UI state, this cuts custom guards for malformed event shapes.
  • —No new features or pricing changes here; value is operational stability, not new capability.

  Source: github.com/anthropics/anthropic-sdk-typescript/releases/Read original →