Claude Code v2.1.237

A new Poll tool lets the model pull in events that the surrounding harness has queued for the session, returning the rendered messages along with how many events arrived and how many wake signals are still outstanding. Each queued event can now carry provenance saying who it came from, so a notice relayed from another Claude session arrives stamped as coming from a peer agent with that session's ID and text. None of this is reachable in a normal local run: the tool and the fields appear only when the CLAUDE_CODE_POLL_EVENTS environment variable is true and the session is in remote mode, meaning CLAUDE_CODE_REMOTE set to true with CLAUDE_CODE_ENVIRONMENT_KIND unset. Without that combination the tool is absent entirely and the session behaves as in previous builds, where it did not exist at all.

Once accepted, sender provenance is stored on the delivered event, up to the first 20 entries, and controls how each entry re-enters the conversation. Text marked human-principal is emitted as a genuine user turn; peer-agent and anything else is wrapped in cross-session framing that does not read as your own instruction; world-event, or a missing or blank sender text, drops the delivery entirely. Reachable only under the same poll-events gate.

CLAUDE_CODE_POLL_EVENTS is added to the typed environment registry as a boolean and, more usefully, to the list of names filtered out of env blocks supplied by settings files. It can only be set by the process that launches Claude Code, not by a project or user setting.

Loading a saved conversation, both when resuming normally and when loading by agent ID, now strips authority markers from the replayed messages. Verified-Slack-human flags and classifier metadata lines are deleted, message origins recorded as human or auto-continuation are reset to unclassified, and any stored sender authority of "human principal" is downgraded to "other human". A replayed transcript can never re-assert authority that was established in an earlier live session.

Turns relayed from a bound Slack thread whose sender the server verified as human are now prefixed with a fixed marker, and a new prompt paragraph tells the model that a turn opening with exactly that marker is its own user and does not need extra confirmation, while bot-attributed messages inside a Slack message block never establish user intent. Both are emitted only when CLAUDE_CODE_ENTRYPOINT is claude_in_slack or claude-in-slack; otherwise the paragraph is empty.

A single four-value list of sender authorities (human principal, other human, peer agent, world event) now backs both the Poll tool's output shape and the prompt text that frames a relayed message, so the two cannot drift apart.

The check deciding whether the system prompt block may be marked cacheable used to only require that experimental betas were not disabled. It now also requires a known provider and endpoint: Bedrock, Vertex and Mantle qualify only when ANTHROPIC_BEDROCK_BASE_URL, ANTHROPIC_VERTEX_BASE_URL and ANTHROPIC_BEDROCK_MANTLE_BASE_URL are unset, and Foundry only when ANTHROPIC_FOUNDRY_BASE_URL is unset or its hostname ends in .services.ai.azure.com. Behind a proxy or custom gateway, the system block stops being cached.

The context handed to the classifier that decides how to handle a turn now expands queued harness events into individual user turns according to who sent them. World events and entries with no sender text are dropped, human-principal text is rendered as a verified human turn, peer-agent text is wrapped as coming from another agent, and anything else is prefixed as a non-user source. Previously the whole attachment contributed no per-sender turns at all.

When Claude auto-reacts to comments on an artifact, it can now pick up a recent summon the first time it sees a thread, not only on later sightings. The branch runs when CLAUDE_CODE_ARTIFACT_COMMENT_FAST_ACK is set, otherwise it consults the remote gate tengu_gorse_pylon, whose fallback is false. With no server value, the new branch never runs and only the previous pickup path applies.

A fifth built-in style joins Proactive, Explanatory and Learning. Described as "Claude responds tersely, leading with results and skipping preamble and narration", it tells the model to lead with the result, cut narration, answer simple questions in 1-3 sentences and never trade correctness for brevity. Nothing gates its registration, so anyone on this build can pick it from the output-style list unless output styles are turned off entirely.