Skills
Rewrite noisy tool output before the model sees it using PostToolUse output replacement
Because PostToolUse can now replace the output of any tool (not just MCP tools) via `hookSpecificOutput.updatedToolOutput`, you can intercept verbose or low-signal tool results — giant test logs, raw HTTP dumps, paginated API noise — and hand the model a trimmed, structured summary instead. PostToolUse inputs also now carry `duration_ms`, so you can branch on how long a tool took. Add a PostToolUse hook that collapses known-noisy commands to their salient lines, cutting tokens and keeping the agent's context focused without changing the underlying tool.
↳ Follow the thread