🧭 Claude for Government Launches in Beta

Anthropic today opened Claude for Government in beta — a dedicated offering that lets federal, state, and local agencies access Claude's most capable models without needing a separate cloud-provider relationship. Anthropic remains the contracted and billing party; agencies request access directly at claude.com/solutions/government.

What sets it apart from Claude Enterprise

Claude for Government is built on the same foundation as Claude Enterprise but operates inside an isolated, dedicated FedRAMP boundary with additional compliance guarantees:

• FedRAMP High authorization — the highest tier, covering workloads that handle sensitive but unclassified federal data.
• Hash-chained audit log — every administrative action is recorded in a tamper-evident log that organisation administrators can review directly in the product. No Anthropic support ticket required to pull audit data.
• Two-person approval for sensitive operations — Anthropic's own access to customer data requires a second internal approver, closing a class of insider-threat scenarios that single-admin-approval architectures leave open.
• Metering-only usage exports — usage reports contain only counts and timestamps, not prompt content, so agencies can respond to ATO and IG audit requests without moving sensitive material outside the FedRAMP boundary.
• FedRAMP Secure Configuration Guide — Anthropic is publishing this as a public-facing document, giving agency security teams a concrete baseline for hardening their Claude for Government deployments.

Scope: all three branches, direct billing

A companion announcement, Offering Expanded Claude Access Across All Three Branches of Government, extends coverage to the legislative and judiciary branches alongside the federal civilian executive. A limited-time programme makes Claude for Government (Enterprise tier, unlimited seats) available to federal agencies at $1 through August 2026 — the intent being to eliminate procurement friction so agencies can pilot and generate the ATO evidence needed for a full deployment.

🧭 Claude Code August Patch Notes: Five Stability Fixes Worth Knowing

The Claude Code changelog for this week carries five fixes that are easy to miss but meaningful for daily use — particularly for teams running long Claude Code sessions or using the memory system heavily.

Fix-by-fix breakdown

• Memory frontmatter truncated at inline # — If a memory file's frontmatter contained a value like description: Track cache-hit rates # updated, everything after the # was silently dropped on save. The bug affected any memory whose description used an inline comment. Fixed: values are now written verbatim and the inline comment is preserved.
• Feature flags going stale after OAuth token rotation — In sessions longer than the OAuth access-token lifetime, GrowthBook feature flags cached on startup were not refreshed when the token rotated. Features that changed mid-session (such as new model rollout toggles) would stay in their at-launch state for the rest of the session. Fixed: flag cache is invalidated on token rotation.
• GrowthBook null-feature crash — A separate GrowthBook bug: if a feature evaluated to null (e.g. a flag targeting no users in the current environment), Claude Code would crash rather than treating it as a disabled flag. Fixed with a null-guard.
• Unbounded memory growth from oversized --settings files — Pointing --settings at a device file, a named pipe, or a multi-gigabyte file caused Claude Code to grow without bound as it attempted to read the settings. Files larger than 2 MiB now fail at startup with a clear error message rather than consuming all available memory.
• Bash tool killing the Claude session via pkill pattern match — On Linux, a pkill -f command whose pattern accidentally matched the Claude Code CLI process would terminate the CLI itself. Claude Code now spawns commands in a process group that is explicitly excluded from the pattern match scope.

New: docker and Podman permission prompts

Claude Code now shows permission prompts for docker commands (including the Podman docker shim) that carry daemon-redirect flags: --url, --connection, --identity, and Podman's --remote mode. These flags can redirect the docker client to an arbitrary daemon — including one outside the project's intended scope — so they now require explicit user approval before execution.

New: --forward-subagent-text flag

The new --forward-subagent-text CLI flag (and its environment-variable equivalent CLAUDE_CODE_FORWARD_SUBAGENT_TEXT) includes subagent text and thinking in stream-json output. This is useful for building observability tooling around multi-agent runs where you want to stream subagent reasoning to a log aggregator or custom UI.

# Update Claude Code to pick up the August patch set
npm install -g @anthropic-ai/claude-code@latest
claude --version

# Verify memory files are no longer truncating inline # values
# (open any .claude/memory/*.md and check the description field)