Claude Code Daily Briefing - 2026-06-25

Release Summary

The CLI shipped two releases on 6/24 (v2.1.190 and v2.1.191) — and v2.1.188/189 were skipped, in the same vein as v2.1.180/182/184 that appear to have been pulled just before publishing. But today’s headline isn’t the CLI; it’s Claude Tag, the new product Anthropic launched in beta on 6/23 that puts Claude into Slack channels as a shared AI teammate (see New Features). The second axis is another incident on 6/24 — the fourth stability wobble this week (Security & Limitations).

Full release notes

New Features & Practical Usage

Claude Tag — Claude joins your Slack channel as a shared AI teammate (6/23, beta)

Anthropic launched Claude Tag (6/23, beta). The idea is to make Claude a persistent, shared AI teammate inside a Slack channel — anyone can @Claude to hand off a task, and Claude works through it asynchronously over hours or days, in stages, posting results back to the thread. Everyone in the channel can see what it’s working on and pick up where the last person left off.

• Channel-level shared memory: Claude Tag gives a whole channel one shared Claude that builds memory from channel activity — and, with permission, automatically learns from other Slack channels and data sources.
• Ambient mode: when enabled, Claude proactively keeps you updated about whatever it thinks you need — surfacing things before you ask.
• Separate identities and access: admins can create distinct Claude identities per use, so a sales Claude won’t share memories with an engineering Claude. Admins control which tools, data, and codebases it connects to, plus spend limits, at a granular level.
• Model and availability: it runs on Opus 4.8 and is in beta for Claude Enterprise and Team plans.

# Delegate right in the Slack channel
@Claude clean up this incident thread into a timeline and follow-up action items
@Claude review this PR, and ping me in ambient mode if you get stuck

The practical takeaway: Claude is moving past the terminal, IDE, and web and into the Slack channel where teams actually congregate. If the 6/22 Artifacts feature (turning session work into an internal shared page) was about sharing the output, Claude Tag is the next step up — stationing the agent itself in the collaboration space, shifting the center of gravity from a tool one person runs locally to a coworker the whole channel shares. Note that the existing Claude in Slack app retires 8/3 and Claude Tag replaces it, so teams on that app have migration to handle (see Ecosystem). Anthropic — Introducing Claude Tag · TechCrunch

/rewind can now reach back past a /clear (v2.1.191)

A small but genuinely mistake-reversing change. /rewind (which rewinds the conversation and code state to an earlier point) can now go back to before you ran /clear.

# When you accidentally /clear and lose all your context
/rewind   # can now return to a point before /clear and recover the conversation

The point is that /clear is no longer an irreversible cliff — if you hit /clear to tidy up a long session and it took the context you actually needed with it, you can now /rewind past it. This is a safety net aimed squarely at people running long interactive sessions, more than unattended automation. GitHub v2.1.191

Developer Workflow Tips

Comma-separated hook matchers were silently dead — re-verify your guard hooks

If you enforce CLAUDE.md rules with a hook, do not skip this fix. v2.1.191 fixed a bug where comma-separated hook matchers (e.g. "Bash,PowerShell") silently never fired. In other words, if you bundled multiple tools into one matcher with a comma to guard them, that hook may actually have been dead.

// .claude/settings.json — a comma-bundled matcher may not have fired before v2.1.191
{
  "hooks": {
    "PreToolUse": [
      { "matcher": "Bash,PowerShell", "hooks": [ /* ... */ ] }
    ]
  }
}

This is the exact counterpart to the 6/22 point — enforce must-follow rules with a hook, not a prompt — because without a guarantee that the hook actually fires, there is no enforcement. If you use comma-bundled matchers, (1) update and (2) reproduce a dangerous command to confirm the hook actually blocks it. (If in doubt, splitting one matcher per tool is the surest option.) GitHub v2.1.191

MCP reliability stepped up for unattended runs — retries + headless OAuth

If you run Claude Code unattended in CI/cron with MCP servers attached, v2.1.191’s MCP work pays off immediately. (1) Capability discovery now retries transient network errors with a short backoff, (2) MCP OAuth discovery and token requests also retry once after a transient error, and (3) headless environments skip the browser popup and go straight to the paste-the-URL prompt. On top of that, MCP 404 errors now show the URL and point to your MCP config, so root-causing is faster.

The point is a guard against a single network blip breaking the whole unattended run — the same direction as 6/24’s CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT (which aborts a hung remote MCP call instead of letting it block forever). Together, a remote MCP server that wobbles transiently in CI will either recover via retry or fail cleanly. If you run MCP in an unattended pipeline, update. GitHub v2.1.191

Security & Limitations

Another 6/24 incident — elevated error rates on Opus 4.8 (the fourth this week)

On 6/24, Claude hit another incident. Per aggregator records, an elevated error rate on Opus 4.8 began in the afternoon (around 13:18 UTC) and lasted roughly 1h40m before recovering (the official status page didn’t load at check time, so the timing/duration is from aggregators).

Each incident recovered, but the pattern is the signal — after the 6/18 full outage, the 6/20 ‘Opus 4.8 errors’, and the 6/23 multi-model incident, this is the fourth stability wobble this week alone. Claude Code in the terminal is tied to the same backend, so even a short outage can break an unattended run wholesale. The unchanging practical answer is to put retries, backoff, and failure alerts in the pipeline (review them alongside the 6/22 CLAUDE_CODE_MAX_RETRIES cap of 15 and RETRY_WATCHDOG). And Fable 5 / Mythos 5 remain offline for everyone on day 13+ of the export-control suspension — for non-US developers, the response is unchanged: keep a fallbackModel like Opus 4.8 in place. Claude Status — Incident History · StatusGator

Ecosystem & Plugins

The existing Claude in Slack app retires 8/3 — a 30-day migration window (admin action required)

The counterpart to the Claude Tag launch is a support-end notice. The existing Claude in Slack app retires on 8/3, and Claude Tag replaces it entirely. It does not migrate automatically, so admins must act.

• 30-day window: admins must opt into Claude Tag during the migration window (roughly 30 days) and configure channel access, tool connections, and memory scope before the 8/3 deadline.
• Why it matters: Claude Tag’s identity, memory, and tool-access model differs from the old app, so settings do not carry over — skip the migration and the old app’s Slack integration goes dark after 8/3.

The practical read is simple: if your team used Claude in Slack, this isn’t a feature to admire — it’s a migration with a deadline. Like the 6/20 managed MCP connectors and the 6/23 org-level model restrictions, it’s another scene in Anthropic steadily widening the admin-console-governed enterprise surface — both onboarding and offboarding are now tied to admin settings. Anthropic — Introducing Claude Tag

Community News

• Anthropic: its own team generated ~65% of its code with an internal Claude Tag (Anthropic’s own figure): In introducing Claude Tag, Anthropic said its product team generated about 65% of its code with an internal version of Claude Tag — and that most of the code that built Claude Tag itself was written that way (these are Anthropic’s own numbers). Precision aside, the message is clear: Anthropic is positioning the agent not as a personal assistant but as a coworker stationed in the channel, and offering a self-referential proof that it built its own product with that model. Read alongside the 6/18 ‘Addy Osmani — 4x the code, +12% productivity’ and the 6/23 ‘The Untrainable’ discussion of benchmark/output volume ≠ business value, treat ‘65%’ as a signal of delegation scope, not a productivity guarantee. Anthropic — Introducing Claude Tag
• TechCrunch: ‘Claude Tag is learning your company, one Slack message at a time’: TechCrunch framed the heart of Claude Tag as its learning memory — Claude accumulates how your org works from channel activity and, with permission, absorbs other channels and data sources. The flip side of that convenience is governance: unless admins separate and limit what it remembers, which department’s Claude can access what, and how far it’s allowed to learn from the start, internal context pools in one place more broadly than intended. The recommended design is to scope memory, tool access, and spend limits per identity before rollout. TechCrunch

Minor Changes

Practical changes and fixes in v2.1.191 not covered above.

• ~37% less CPU during streaming: by coalescing text updates to 100ms, CPU use during streaming responses dropped about 37%, and long-session memory growth (the terminal output cache) was reduced — a noticeable change for long sessions that had your laptop fans spinning (v2.1.191)
• Stopping a background agent is now permanent: stopping an agent from the tasks panel now sticks — it no longer resurrects (it used to come back after being stopped) (v2.1.191)
• /voice explains restrictions: when /voice is disabled by org policy, it now explains the restriction instead of a generic ‘not available’ (v2.1.191)
• /permissions Recently-denied approvals persist: in the ‘Recently-denied’ tab, approving a denial now persists on close (it used to be silently discarded) (v2.1.191)
• Sandbox network permission remembers hosts: hosts you allow with ‘Yes’ in the network permission dialog are now remembered for the session, so it stops re-prompting on every connection (v2.1.191)
• Managed-settings forced refresh fix: forceRemoteSettingsRefresh now takes effect when set via MDM or file policy, and the fetch sends Cache-Control: no-cache to stop proxies from serving stale responses (v2.1.191)
• Many terminal-UI fixes: the welcome splash art overflowing an 80×24 macOS Terminal, the /login URL truncating across lines in Windows Terminal, Cmd+click not working in Ghostty over ssh/tmux, and the scroll position jumping to the bottom during streaming were all fixed (v2.1.191)

Recommended Reads

• “Jobs and Software is Fucked”: A 10-year engineer (ex-Blizzard) takes on the collapse of the hiring market, amplified by AI, head-on. The sharp diagnosis: the structure now disadvantages candidates who follow the rules — those who quietly use AI in coding assessments gain an edge, while companies respond with keyword-based resume filters and AI-proctored tests that block documentation access yet can’t stop phone-based AI use. Heavier still is the trend of kicking away the junior ladder entirely — entry paths close amid the expectation that AI will eliminate junior roles. The author frames AI coding adoption as abandoning the dignity of craftspeople, testers, and engineers and refuses to surrender to it — in the same vein as the 6/21 ‘Is AI eroding our skills’ and the 6/23 ‘The Untrainable’, this is the rawest version of the question of what’s left for people and professions when tools compress the work. urflow
• “Programmers in the AI era: from code generation to controlling nondeterminism”: As AI automates code generation, the developer’s core role shifts toward design, verification, and control. The central thesis is crisp: the essence of programming isn’t typing a lot of code — it’s filling in the details of vague requirements and abstracting them into reusable forms. So developers aren’t replaced but evolve into system designers and harness engineers, taming AI’s nondeterminism with deterministic checks — tests, compilers, linters, approval gates — while owning the strategic details of problem definition, architecture, and risk. Read alongside 6/21’s ‘surfingcomplexity — keep synthesis and verification human’ and 6/23’s ‘Who Will Price Intelligence’, it puts what to automate vs. what to hold onto into the language of harness engineering. velog (Korean)

Interesting Projects & Tools

• ponytail — make AI agents think like the laziest senior developer: An open-source tool that injects the lazy-senior mindset of “the best code is code you never write” into AI coding agents. It enforces a 6-step check before generating code — is it really needed → can a standard library do it → is there a native platform feature → can an existing dependency do it → does it fit in one line — pruning over-engineering without compromising security, data integrity, or accessibility. Adjust rigor with /ponytail [lite|full|ultra|off], flag over-engineering in a diff with /ponytail-review, and analyze a whole repo with /ponytail-audit. One case reported 54% less code, 100% safety retained, ~20% lower cost, and ~27% faster; it’s compatible with Claude Code, Cursor, and 14 agents total, under MIT — a tool that turns the ‘abstract carefully’ principle from the read above into agent guardrails. GitHub
• Monet — local memory for AI coding agents: A memory system aimed at the chronic problem that an agent is “a blank slate every time you open a new session.” It works on three principles — (1) autonomous note-taking (the agent records decisions and patterns on its own), (2) relevance-ranked retrieval (frequently referenced memories surface first), and (3) progressive improvement (the more it accumulates, the more project-specific it gets). It evolved from plain markdown files into a structured memory index, runs 100% locally (no data leaves your device), stores in SQLite (~/.monet), is free, and exposes an open MCP interface to work with Claude, Cursor, and other MCP-supporting agents — where 6/23’s Agent-Blackbox measured where an agent leaks tokens, Monet is the counterpart that handles what an agent should remember. Monet