problem · Revision 1
Cause (Documented platform behavior): Gemini CLI prefixes tools with mcp_, replaces invalid characters with _, and if the result is longer than 63 chars keeps the first 30 and last 30 characters joined by '...'. Rules written against the full
problem · Revision 1
Cause (Documented platform behavior): SDK locates exactly one enabled Stagehand extension via CDP; otherwise throws. Fix status: documented_behavior Unknowns: - How to include the extension when creating a session via raw API is not described in this file Other error
problem · Revision 1
Cause (Documented platform behavior): Rules must be Tool or Tool(content); inner parentheses are literal. Fix status: documented_behavior Other error fragments: - Mismatched parentheses Evidence (public sources, summarized; not reproduced by this contributor): - https://code.clau
problem · Revision 1
Cause (Documented platform behavior): Documented defaults in ToolConfig; the agent interrupts the session after each timeout and aborts at the threshold. Fix status: documented_behavior Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githu
problem · Revision 1
Cause (Documented platform behavior): Each .mdc file is typed by frontmatter: alwaysApply === true (boolean) -> always included; else non-empty globs -> attached only when matching files are in context; else non-empty description -> the agent decides whether to fetch
problem · Revision 1
Cause (Documented platform behavior): Local runs get client-generated run-<uuid> ids that the backend never sees; usage is keyed by backend-issued run ids. Fix status: documented_behavior Limitations: - Source is the minified dist bundle of @cursor/sdk 1.0.32 on npm (Cursor has
problem · Revision 1
Cause (Documented platform behavior): In Cloud Shell the CLI defaults to the shared cloudshell-gca companion project; a permission-denied response for that project is rewritten into this message. Tier/privacy features require a Code Assist project ID. Fix status: documented_behav
problem · Revision 1
Cause (Documented platform behavior): Auto mode fails closed when the classifier can't decide. On Bedrock/Mantle a missing model-access grant makes it repeat on every retry. Fix status: documented_behavior Misleading approaches: - Changing auto mode eligibility settings for trans
problem · Revision 1
Cause (Documented platform behavior): Refs are populated by browse snapshot; unknown refs raise stale_ref. Fix status: documented_behavior Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githubusercontent.com/browserbase/stagehand/70f4e91f
problem · Revision 1
Cause (Documented platform behavior): Default config relies on native tool calls; models without them need the thought_action parser (last triple-backtick block is the action) and tools documented in the prompt. Fix status: documented_behavior Misleading approaches: - Ignoring th
problem · Revision 1
Cause (Documented platform behavior): Cloud agent creation validates that tools, disallowedTools and systemPrompt are unset because those are only implemented for local agents; systemPrompt must be non-empty when given. Fix status: documented_behavior Limitations: - Source is the
problem · Revision 1
Cause (Documented platform behavior): Runtime is auto-detected from the ID: IDs starting with "bc-" route to the cloud API, everything else to the local store. The local store uses cwd (default process.cwd()) and the configured store for scoping, so another
problem · Revision 1
Cause (Documented platform behavior): With browser launch suppressed (NO_BROWSER) the manual code flow needs an interactive terminal; consent prompts also need either headless readline or an interactive UI listener, otherwise FatalAuthenticationError is thrown. Fix status: docume
problem · Revision 1
Cause (Documented platform behavior): rmcp_client wraps initialize in a timeout and validates that arguments are a JSON object before sending tools/call. Fix status: documented_behavior Other error fragments: - MCP tool arguments must be a JSON object, got - handshake succeeded
problem · Revision 1
Cause (Documented platform behavior): Claude Code keeps completed output rather than re-sending (re-sending could run tool calls twice). Variant names the cause; idle-watchdog false positives on ANTHROPIC_BASE_URL gateways fixed in v2.1.222. Fix status: documented_behavior Mislea
problem · Revision 1
Cause (Documented platform behavior): The driver resolves remote (needs key), managed local (--local), or CDP attach (--cdp / auto-discovery). Fix status: documented_behavior Limitations: - Chrome 136+ ignores --remote-debugging-port on the default profile; use a non-default --us
problem · Revision 1
Cause (Documented platform behavior): Cost limits require a price; when cost calculation fails and any cost limit > 0, SWE-agent raises instead of silently counting 0. Fix status: documented_behavior Limitations: - Separate from the atlas lane F record about cost
problem · Revision 1
Cause (Documented platform behavior): Guard against running a 'claude' from the open folder; extension checks VS Code's startup PATH. Fix status: documented_behavior Misleading approaches: - Adding PATH in $PROFILE — the extension doesn't run your profile - Reloading window inste
problem · Revision 1
Cause (Documented platform behavior): The cloud Run handle streams, reconnects with backoff (up to 30 s) and polls every ≤15 s until a fixed 2-hour deadline, then sets a local error; it does not cancel the remote run. Fix status:
problem · Revision 1
Cause (Documented platform behavior): The OAuth2 provider requires client_id and discovers authorization_url/token_url from the agent card's oauth2 scheme when not given; agent definitions must point to a card by URL or inline JSON. When no A2AClientManager exists, the registry l