report-filter

# report-filter ## When to invoke Immediately after **any** subagent or tool chain returns output. Applies to: - Return value of `Task` / `Agent` calls - Pipeline stage outputs - Long tool outputs (test runner, build log) where only signal matters - Chained skill handoffs where stage N+1 parses stage N Every dispatch that went out through an RTCO brief must come back through this filter. ## Why Unfiltered agent output contains: - Pleasantries ("I found that...", "Hope this helps!", "Let me know if...") - Restatement of the original task ("You asked me to find X. I looked for X.") - Meta-narration ("I'll now explain my process in detail...") - Hedging and self-assessment ("I believe this might be...") None of this is signal. It costs tokens on the return trip and dilutes the signal when the caller reads or feeds the output into the next stage. This skill extracts the signal, validates against the brief's declared OUTPUT schema, and either emits clean structured output or **rejects** the report back to the subagent with a fix request. ## Rules ### Drop (zero preservation) - Pleasantries, greetings, sign-offs. - Restatement of the task. - Process narration ("First I did X, then Y"). Exception: keep only if the brief's OUTPUT schema explicitly requested a process log. - Hedging phrases with no added information ("I think", "it seems", "possibly"). Keep hedging **only** when it encodes genuine uncertainty about a technical claim. - Apologies. ### Keep verbatim (no paraph...