Telexed

AI news through a solo-operator lens — only what changes your day8 of 412

FILTER[All][Agents & tools][Models & API][Generative media][Infra & SaaS][ASO & growth][Indie business][Idea signals][Other][★6+ high-signal]

clear filters

Sat, May 91 dispatches

#0412
#0412Agents & tools Claude Code Releases2 weeks ago
`Claude Code` `v2.1.136`, fixes MCP auth churn and resume/path edge cases
60radar
Claude CodeCoding agent — AI dev tool for terminal and IDE workflows
Remote MCP setups got more reliable: token refresh races, login loops, and lost sessions were fixed. If your workflow leans on multiple tools and IDE integration, this patch removes daily friction.
- Concurrent MCP token refresh no longer drops refresh tokens, which should stop repeated re-auth across several remote servers.
- /resume and --continue now work when project paths contain underscores, fixing a small but nasty session-recovery edge case.
- IDE integrations no longer lose .mcp.json, plugin, or claude.ai connector servers after /clear, which mattered for long-lived editor sessions.
- settings.autoMode.hard_deny adds unconditional block rules in auto mode, giving tighter guardrails for risky file or tool actions.
Source: github.com/anthropics/claude-code/releases/tag/v2.1.136Read original →
FIG-4121:1
60radar
FIG-4121:1

Fri, May 83 dispatches

#0411
#0411Agents & tools Claude Code Releases2 weeks ago
`Claude Code` `v2.1.133`: worktree base control, effort propagation, multi-session fixes
70radar
Claude CodeCoding tool — an agentic CLI for worktree-based dev
Worktree branching is now explicit, and effort state finally propagates cleanly through hooks and Bash. The bigger win is stability: concurrent sessions, MCP OAuth, subagents, and remote interrupts all got fixes that remove real workflow breakage.
- New worktree.baseRef picks fresh or head, so new worktrees can start from origin/<default> or local HEAD; safer when juggling unpushed changes.
- Hooks now receive effort.level plus $CLAUDE_EFFORT, which makes effort-aware scripts and tool routing much easier to implement.
- A refresh-token race that pushed parallel sessions into 401 failures was fixed; this directly reduces flaky multi-session agent runs.
- MCP OAuth now respects HTTP(S)_PROXY, NO_PROXY, and mTLS across discovery, registration, token exchange, and refresh; enterprise network setups stop randomly breaking.
- Subagents can now discover project, user, and plugin skills again, removing a nasty limitation for delegated workflows.
Source: github.com/anthropics/claude-code/releases/tag/v2.1.133Read original →
FIG-4111:1
70radar
FIG-4111:1
#0410
#0410Agents & tools GitHub Changelog2 weeks ago
`GitHub Copilot` deprecates `Claude Sonnet 4` across all experiences
60radar
One model option disappeared across chat, inline edits, agent mode, and completions on May 6, 2026. If your Copilot output quality depended on it, prompt behavior and fallback model choice need a quick recheck.
- Coverage is broad: Copilot Chat, inline edits, ask, agent mode, and code completions all lose the same model path.
- The change landed on May 6, 2026, so teams using fixed Copilot habits may see response style shifts immediately.
- This is a tooling workflow issue, not just model news; saved prompts and comparison baselines can drift after removal.
Source: github.blog/changelog/2026-05-07-claude-sonnet-4-deprecaRead original →
FIG-4101:1
60radar
FIG-4101:1
#0409
#0409Agents & tools Hermes Agent Releases2 weeks ago
`Hermes Agent` `v0.13.0` turns long-running agent work into a durable board
80radar
Hermes AgentCoding agent — resilient long-running task execution
Long-running agent runs now survive interrupts, recover from failures, and stay pinned to a goal. If you automate multi-step coding or ops flows, this is a real upgrade worth testing now.
- Kanban adds durable multi-agent execution with heartbeat, reclaim, zombie detection, retries, and hallucination recovery.
- /goal keeps the agent locked on one target across turns, cutting drift in longer coding and automation sessions.
- Security tightened hard: 8 P0s closed, redaction defaults ON, stranger DMs blocked on WhatsApp, TOCTOU gaps patched.
- Checkpoints v2 rewrites persistence with pruning and disk guardrails, while gateway auto-resume restores sessions after restarts.
Source: github.com/NousResearch/hermes-agent/releases/tag/v2026.Read original →
FIG-4091:1
80radar
FIG-4091:1

Thu, May 72 dispatches

#0408
#0408Agents & tools Cursor Changelog2 weeks ago
`Cursor` adds PR review, parallel plan execution, and split PR workflows
80radar
Code review and execution are getting compressed into one loop. Cursor is moving past autocomplete toward workflow orchestration, and this release looks immediately useful for faster shipping.
- The release bundles a new PR review flow with faster plan execution, tying review and implementation closer together.
- Parallel agents are positioned as the speed lever for plans; multi-step work should finish in less wall-clock time.
- Quick-action pills target common workflows, which cuts prompt friction on repetitive IDE actions.
- Split PRs in the title suggests smaller, cleaner change sets are becoming a first-class workflow inside Cursor.
Source: cursor.com/changelog/05-07-26Read original →
FIG-4081:1
80radar
FIG-4081:1
#0407
#0407Agents & tools Claude Code Releases2 weeks ago
`Claude Code` `v2.1.132` adds session env vars and hardens terminal/resume flows
60radar
Claude CodeCoding agent CLI — stronger terminal recovery and session stability
More of a stability drop than a feature release: terminal recovery, paste handling, resume state, and MCP failure cases all got cleaned up. Worth updating if Claude Code is part of your daily shell or IDE loop.
- CLAUDE_CODE_SESSION_ID now reaches Bash subprocesses, aligning tool env with hook session_id and making per-session logging easier.
- CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN=1 keeps output in native scrollback, useful if you prefer normal terminal history over fullscreen rendering.
- Resume and plan-mode edge cases were fixed, including emoji-corrupted sessions, ignored --permission-mode, and lost plan mode after ExitPlanMode.
- A severe MCP issue was patched: malformed stdio server output could drive 10GB+ RSS memory growth, and failed tools/list now retries with clearer status.
Source: github.com/anthropics/claude-code/releases/tag/v2.1.132Read original →
FIG-4071:1
60radar
FIG-4071:1

Wed, May 62 dispatches

#0406
#0406Agents & tools Claude Code Releases2 weeks ago
`Claude Code` `v2.1.129` adds session plugin fetch and background auto-update
70radar
Claude CodeAI coding tool — terminal-first dev workflow
Plugins are easier to load on the fly, and self-updates now run in the background on package installs. Small release, but the workflow friction drop is real enough to use now.
- --plugin-url <url> pulls a plugin .zip into the current session, which lowers the setup cost for trying custom tooling fast.
- CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE lets Homebrew and WinGet installs upgrade in the background, then asks for a restart instead of blocking work.
- Gateway model discovery is now opt-in via CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1, reducing surprise model list changes on custom gateways.
- skillOverrides now behaves correctly across off, user-invocable-only, and name-only, which makes shared agent setups easier to control.
- Several regressions were fixed around /clear, subagent UI, prompt cache TTL, and MCP allow rules, so day-to-day CLI reliability improved.
Source: github.com/anthropics/claude-code/releases/tag/v2.1.129Read original →
FIG-4061:1
70radar
FIG-4061:1
#0405
#0405Agents & tools GitHub Changelog2 weeks ago
`GitHub MCP Server` secret scanning is now generally available
60radar
GitHub MCP ServerMCP server — GitHub workflows with built-in secret scanning
Secret scanning now lands directly inside MCP-based coding workflows. If you rely on agentic IDEs or CLIs, this is an immediate security upgrade with almost no adoption friction.
- GitHub MCP Server now supports secret scanning in GA, so the feature is no longer experimental.
- The target surface is MCP-compatible AI tools, including examples like GitHub Copilot CLI and agent-friendly IDE flows.
- This pushes security checks closer to where code is generated, which matters more as more edits originate from agents.
Source: github.blog/changelog/2026-05-05-secret-scanning-with-giRead original →
FIG-4051:1
60radar
FIG-4051:1

`Claude Code` `v2.1.136`, fixes MCP auth churn and resume/path edge cases

`Claude Code` `v2.1.133`: worktree base control, effort propagation, multi-session fixes

`GitHub Copilot` deprecates `Claude Sonnet 4` across all experiences

`Hermes Agent` `v0.13.0` turns long-running agent work into a durable board

`Cursor` adds PR review, parallel plan execution, and split PR workflows

`Claude Code` `v2.1.132` adds session env vars and hardens terminal/resume flows

`Claude Code` `v2.1.129` adds session plugin fetch and background auto-update

`GitHub MCP Server` secret scanning is now generally available