Stop the current clip and emit it via onClip. Resolves when finalized.

Escape the current React commit phase entirely.

Upload files via the Files API; return {mount_name: file_id}.

Poll /state until busy is False. Raises BotError on timeout. Default matches the bot's action_timeout (600s / 10 min): pathfinding pl

* Recursively collect every DrawingML text run (<a:r>) under `node`. * Each run contributes its literal text (<a:t>) and optional font size * (<a:rP

Call reset_run and verify the start_kit landed (inventory has iron_pickaxe, position is at depth). Retries the whole thing — covers the bot re

Update `agent_id` to declare `callable_agents` (Multiagent sessions). Shape per platform.claude.com/docs/en/managed-agents/multi-agent: c

Resolve the URL CMA uses to reach this participant's bot. Preferred (event mode): a header-authenticated endpoint. We store the relay key as

Assemble the full CMA agent spec from the AGENT dict.

Expand templates × variants, then cycle if needed to reach n items.

Preflight for the bot's MCP URL. Two things must be true for the cloud agent to play: (a) the bot is alive on localhost, and (b) the bot's out

Create/refresh the static_bearer vault credential binding `token` to `mcp_server_url`. Credentials live inside a vault: find-or-create a vault

GET the bot's /state and sum diamond counts in inventory. Returns None on any failure — this is pure feedback, not correctness.

Read the bot's running diamond counter for this run. Best-effort; returns 0 on any failure so cost reporting never blocks on the bot.

Create a fresh worker session for one task and return its final text.

{item: delta_count}, only items that changed.

Yield every skill in the org, paginating explicitly. The API's default page size can claim has_more=false prematurely, so we request big pages

Emit a minimal `turn` event matching harness/logging_.py so harness/verify.py can consume this trace unchanged. We don't have a full GameState

Show what the participant's AGENT dict became as a Managed Agent — the 'this is a real object in your account' moment.

Turn AGENT['skills'] entries into CMA skill params. A string entry is a local skill directory — upload it once, cache the skill_id, reuse ther

The assistant content from the Anthropic SDK is a list of typed blocks (TextBlock, ToolUseBlock). Convert them to plain dicts so JSON can enco

The `reasoning` field is for logs only; bot.js doesn't accept it.

Pull sink JSONLs from the session sandbox into local data/.runs/<run_id>/. The CMA beta has no file-download endpoint for agent-written outputs,

Create/update a static_bearer credential, tolerating both live API schemas: some deployments take mcp_server_url INSIDE auth (the SDK's typed

Unrelated auth ERRORs around 14:15 that self-resolve — a red herring.

Healthchecks, GC pauses, cache misses — realistic chatter.

* Prepare everything the graders need for one task — pure data, no scoring. * The harness has no idea what the individual graders measure; it just ha

Return the active system prompt. With narrate=True, append a demo-mode instruction telling Claude to chat() a one-line plan before each action.

* Flatten each text shape (<p:sp>) on the slide to {placeholder type, full * text, largest font size} — the three signals the title heuristic needs.

Upload everything and create/update the `slot` agent. Prints summary.

Stop everything and release the capture.

* Format one cell using the grader's own formatter (e.g. "100%", "4.2/5"), * appending a signed delta vs the previous run when both values are numeri