Claude Code updates 2.1.80 with Channels and full --resume restores

• Anthropic shipped Claude Code 2.1.80 with a broad CLI update that restores full --resume state for parallel tool calls, adds a rate_limits field for statusline scripts, and cuts startup memory by about 80 MB on 250k-file repos, according to the 2.1.80 changelog.
• The release also tightens memory handling: Claude is now told to treat stored memories as historical context, verify them against current files before acting, and if they conflict, to “trust what you observe now,” per the system prompt diff.
• A new research-preview --channels feature lets select MCP servers push messages into a running Claude Code session, with Telegram and Discord as the first official integrations, as shown in the channels launch and described in the channels docs.
• The update also restores “many previously blocked SQL analysis functions” and fixes proxy-related 400s for fine-grained tool streaming on API proxies, Bedrock, and Vertex, which makes this release more than a chat-control demo for teams running longer or more complex sessions release thread.

Claude Code 2.1.80 is a reliability-heavy release disguised as a feature drop. The biggest operational fix is that resumed sessions now restore all parallel tool_use and tool_result pairs instead of surfacing the old “[Tool result missing]” failure mode, according to Anthropic’s release thread. The same changelog says the build also fixes 400s in fine-grained tool streaming when Claude Code is routed through API proxies, Bedrock, or Vertex, and improves @ file autocomplete responsiveness in large repos.

The release adds a few concrete control surfaces for power users. Anthropic’s CLI changes says statusline scripts now get a rate_limits field covering 5-hour and 7-day windows, plugin entries can now be declared inline in settings.json via a new source: "settings", and skills or slash commands can set effort frontmatter to override model effort on invocation. The public changelog also lists --channels as a research-preview flag.

SQL workflows that regressed earlier are back. The prompt and surface diff notes that SQL analysis functions no longer show the old “cannot construct due to unbound types” error, restoring prior analysis behavior for users who depended on those tool paths.

Channels is the new headline feature: Anthropic says it lets “select MCPs” control a live Claude Code session, starting with Telegram and Discord, so you can message the agent from a phone while the session stays attached to your local environment launch post. That is materially different from a mobile chat wrapper; the feature is designed to push events into an already running coding session.

The docs frame this as a “research preview” and make the scope clear. Anthropic’s channels docs says Channels requires Claude Code 2.1.80 or later, works with claude.ai login rather than console or API-key auth, and is meant for external events like chat messages, alerts, and webhooks entering a live session. The same docs describe security controls including sender allowlists and access policies, plus persistent or background deployment modes.

Telegram and Discord are the first official plugins, with setup guides linked from the launch thread through the Telegram README and Discord README. Supporting posts from Wes Roth and TestingCatalog echo the same implementation detail: this is remote control over MCP, not a separate hosted Claude endpoint.

The system-prompt change is small but important for engineers who rely on Claude Code memory across long-lived projects. The new guidance says memory “can become stale over time” and Claude should verify current files or resources before answering, then update or delete the stale memory if reality has changed memory guidance. That reduces a specific failure mode where an agent confidently acts on old repo state.

The release also closes a few sharp edges that show up in real usage rather than demos. Anthropic’s fix list includes a voice mode WebSocket fix for Cloudflare bot detection on non-browser TLS fingerprints, a fix for managed settings not applying at startup when remote-settings.json was cached, and the removal of /remote-control from deployments where it cannot actually function. Together with full --resume restoration, those changes make long-running and remotely supervised sessions more predictable than the previous build, which only landed about 23 hours earlier according to the release metadata.