Topics with suffixes like "use cases", "workflows", "review",
"examples" were previously echoed near-verbatim into search_query,
returning near-zero matches because nobody posts the literal phrase
(2026-04-19 Hermes Agent Use Cases failure).
Unit 2 — planner breadth:
1. Planner prompt rule: STRIP intent-modifier phrases from search_query
(keep them in ranking_query). Paraphrase across 4-5 subqueries that
each express the intent differently.
2. Planner prompt rule: quote only multi-word proper nouns like
"Hermes Agent", not the user's full topic.
3. Raise _max_subqueries cap from 3 to 5 for how_to / opinion / product /
breaking_news / prediction. Comparison stays at 4; factual / concept
stay at 2 unless the topic carries an intent modifier.
4. Deterministic fallback: when intent is non-{comparison,prediction}
and topic contains an intent modifier, append 3 paraphrased
subqueries (workflows, production, experience).
Unit 3 — deterministic fallback defaults:
5. _infer_intent default changed from "breaking_news" to "concept".
Prior default forced strict_recent freshness on unclassified topics,
biasing against older relevant material. Recency-signal regexes
("trending", "this week", etc.) added above the default so genuinely
time-sensitive topics still classify correctly.
6. _keyword_query now quotes only title-cased multi-word proper nouns
("Hermes Agent", "Claude Code"), not the user's full typed topic.
Hyphenated compounds and lowercase terms are left as bare keywords
so platform tokenizers broaden rather than narrow retrieval.
7. New stderr warning when plan_query runs with no --plan and no LLM
provider: surfaces that the deterministic fallback path is weaker
than the --plan-from-Claude-Code path, so callers know to generate
and pass a plan.
Tests: 37 planner tests pass, including 11 intent-modifier and 7
fallback-defaults tests.
Matt Van Horn
4d9f29d2ed