Summary

• Add a transform_tool_result hook to ContextEngine so context engines can rewrite tool output before it is written into session history.
• Wire the hook into the agent tool-call path after the existing HookEvent::TransformToolResult hook and before artifact spill, sanitization, injection guarding, and session commit.
• Expose the hook through ContextEngineHooks for scriptable/manual/plugin configuration.
• Include transform_tool_result in scriptable hook validation, metrics, traces, plugin hook path checks, and plugin integrity checks.
• Make build_context_engine create a scriptable layer when any declared scriptable hook exists, including transform_tool_result, assemble, compact, subagent hooks, and on_event.

Hook Contract

The script receives:

{
  "type": "transform_tool_result",
  "agent_id": "...",
  "tool_name": "...",
  "tool_use_id": "...",
  "input": {},
  "content": "...",
  "is_error": false,
  "status": "completed"
}

To rewrite the model-visible tool result content, it returns:

{
  "content": "rewritten tool output"
}

If content is missing or not a string, the original tool result is kept.

If the hook fails under the default warning policy, the original tool result is kept.

Why

This gives per-agent context engines a deterministic pre-ingest cleanup point for noisy tool outputs.

For example, a Rust coding agent can trim cargo check or cargo test output with Rust-specific rules before that output enters conversation history, while an Android/Kotlin agent can use different Gradle/logcat cleanup rules. Agents that do not configure the hook keep the existing behavior.

Verification

• cargo test -p librefang-runtime transform_tool_result
• cargo test -p librefang-runtime hook_integrity_all_hook_fields_checked
• cargo check -p librefang-types -p librefang-runtime
• cargo clippy -p librefang-types -p librefang-runtime --all-targets -- -D warnings
• git diff --check

Out Of Scope

• No changes to the existing Rust HookEvent::TransformToolResult contract.
• No LLM summary request protocol changes.
• No built-in language-specific output cleaners.