Tool Result Limit

ToolResultLimitMiddleware truncates oversized tool-result text in after_tool_call, before CubeLoop publishes ToolExecutionEndEvent. The model, the checkpointer, and any host event consumer all see the truncated result.

Use it when a tool can return unbounded output — shell commands, web fetches, MCP tools — and a huge result would blow the context window or a host event-size budget.

Basic setup

from cubeloop import Agent
from cubeloop.middleware import ToolResultLimitMiddleware

agent = Agent(
    model=provider.model("claude-sonnet-4-6"),
    middleware=[
        ToolResultLimitMiddleware(),
    ],
)

The default cap is 20,000 characters of text content. Results at or under the cap pass through unchanged (after_tool_call returns None). Over the cap, text is cut to the first max_chars characters and a notice is appended:

[truncated: tool result was 1082017 characters; showing first 20000. Narrow the tool call and retry.]

Image blocks are kept. Only TextContent counts toward the cap.

Options

ToolResultLimitMiddleware(
    max_chars=20_000,
    exclude_tool_names={"load_skill"},
)

• max_chars — maximum characters of text kept from the original result. Must be >= 1.
• exclude_tool_names — tool names that skip truncation. Use this for tools whose full payload is load-bearing (for example a skill body returned by load_skill).

Placement

Put this after other after_tool_call middleware so they rewrite the full result first, and the truncated text is what ToolExecutionEndEvent carries.

CubeLoop's default composer is last-non-None wins. A later middleware that returns an AfterToolCallResult without content would drop this rewrite. Keep ToolResultLimitMiddleware last, or use a composer that merges content and details.

Subagent child agents do not inherit parent middleware. Pass the same instance (or a second one) via SubagentMiddleware(inherited_middleware=...) if child tool results should be capped too.