ToolResultEvictionMiddleware

Middleware that evicts oversized tool results to the AbstractFilesystem before downstream reasoning sees the bloated message list.

When the text content of a ToolResultBlock in canonical agent state or the current reasoning view exceeds ToolResultEvictionConfig#getMaxResultChars(), this middleware:

• Writes the full result to {evictionPath}/{agentName}/{sanitized-toolCallId}-{contentHash} in the filesystem.

• Compacts AgentState#contextMutable() first, so legacy pre-reasoning hooks cannot hide an original result from persistence cleanup.

• Rebuilds the current ReasoningInput, preserving hook deletions and rewrites while replacing results that remain visible to the immediate model call.

• Marks compacted results in metadata so they are not recursively evicted on later turns.

Tools listed in ToolResultEvictionConfig#getExcludedToolNames() are never evicted (e.g. readFile — evicting would cause re-read loops).