The prior pipeline.py only logged the planner outcome when an external
--plan was passed ("[Planner] Using external plan (N subqueries)").
The internal LLM planner and the deterministic fallback ran silently,
so retrieval-breadth failures were invisible without --debug.
After plan finalization, emit a unified trace:
[Planner] Plan: intent=X, freshness=Y, cluster_mode=Z, subqueries=N, source=external|llm|deterministic
[Planner] sq1 label=... search="..." sources=[...]
[Planner] sq2 ...
Stderr only; does not touch the user-facing stdout synthesis. The
source= annotation distinguishes --plan (external), provider-backed
(llm), and deterministic paths — so when the 2026-04-19 Hermes Agent
Use Cases failure mode recurs, the trace tells the user which path ran
and what subqueries it produced.
Tests: added test_planner_trace_always_fires_on_mock_run which captures
stderr on a mock pipeline run and asserts the summary + per-subquery
lines appear.
Matt Van Horn
5f218aaac5