The 2026-04-19 Hermes Agent Use Cases run had a Nate Herk YouTube video
titled "I Tested Claude's New Managed Agents" score 51 and rank #2
with zero Hermes content. The reranker had intent-specific scoring hints
but no entity-grounding check, so topic-vicinity matches (one offhand
OpenClaw mention) drifted to the top.
Add _primary_entity(topic) that strips intent-modifier suffixes ("use
cases", "workflows", etc.) so "Hermes Agent use cases" yields
primary_entity="Hermes Agent". Pass the entity through to both the LLM
and fallback scoring paths.
Fallback path: if primary_entity is not found (case-insensitive) in
title + snippet, subtract ENTITY_MISS_PENALTY (25 pts). Skip the
demotion for candidates with no text at all (image-only TikToks etc.)
to avoid false negatives on thin-text sources.
LLM path: add a "Primary entity grounding" hint to _build_prompt when
primary_entity is non-empty. Instructs the LLM to score candidates
without the entity at <=30.
Tests: 24 rerank tests pass, including 8 new entity-grounding tests.
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.
The engine's ## Ranked Evidence Clusters block is a scratchpad for the
model to read, not user-facing output. Two consecutive /last30days runs
on 2026-04-19 (Hermes Agent Use Cases) dumped it verbatim as user output
because the prior canonical-boundary text (Pass through the lines ABOVE
this boundary verbatim) was ambiguous about scope.
Split render_compact stdout into two bounded blocks:
- <!-- EVIDENCE FOR SYNTHESIS: ... --> wraps Ranked Evidence Clusters,
Stats, and Source Coverage. Transform into prose per LAW 2.
- <!-- PASS-THROUGH FOOTER: ... --> wraps the emoji-tree footer only.
Emit verbatim per LAW 5.
Rewrite _render_canonical_boundary to scope pass-through to the footer
block explicitly and give the model a concrete self-check string
(### 1. followed by a score tuple) as the named LAW 6 failure signal.
Add LAW 6 to SKILL.md OUTPUT CONTRACT with the observed violation
(2026-04-19 Hermes Agent Use Cases) and a worked transformation example.
Adds docs/releases/v3.0.9.md as the GitHub Release body and appends
the matching CHANGELOG.md entry.
Covers what shipped in v3.0.9 (Class 1 refuse-gate, LAW 1 WebSearch
precedence, END-boundary, stale SKILL.md deletion) plus the community
contributions that landed across 3.0.1-3.0.8 that had never been
announced (TikTok + YouTube top comments, Hermes support, multi-key
rotation, cross-platform fixes, HTTP layer consolidation, eval
fixtures).
Contributors credited: @j-sperling, @stephenmcconnachie, @zaydiscold,
@iliaal, @Chelebii, @Gujiassh, @hnshah, @george231224, @shalomma,
@BryanTegomoh, @uppinote20, @zerone0x, @thinkun, @thomasmktong,
@fanispoulinakisai-boop, @pejmanjohn, @zl190, @Jah-yee, @dannyshmueli,
@Cody-Coyote.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Five Opus 4.7 self-debugs on v3.0.8 (3 passing, 2 failing runs) converged
on four fixes:
1. Engine refuses Class 1 demographic-shopping queries at main() front-door.
Birthday-gift failure mode becomes structurally impossible - the pipeline
never runs on a doomed query. Exit code 2 with a REFUSE message on stderr
pointing the model to ask for hobbies/relationship/budget. Escape hatch:
LAST30DAYS_SKIP_PREFLIGHT=1 for "just run it" overrides.
2. Delete stale `.agents/skills/last30days/SKILL.md` (1382 lines, April 13
snapshot) and `.hermes-plugin/SKILL.md` (269 lines, April 13 snapshot).
Peter Steinberger's self-debug named the first file as the one it read
instead of the real SKILL.md. One SKILL.md per plugin, at the plugin root.
Sync script simplified: Hermes now always uses main SKILL.md.
3. render_compact() appends an explicit END-OF-CANONICAL-OUTPUT boundary
with pass-through instruction. The model had the canonical body in its
buffer on the Peter run and discarded it; the boundary makes pass-through
the path of least resistance.
4. LAW 1 gains a verbatim-pattern override clause naming the exact WebSearch
tool-result reminder ("CRITICAL REQUIREMENT: MUST include Sources:
section") that caused Peter's trailing Sources leak. No more ambiguity
at synthesis time.
Tests: tests/test_preflight.py, 29 scenarios covering Class 1 matches
(birthday gift, best-for-demographic, what-to-buy-relationship), qualifier
skips (budget, hobbies, activity after year-old), and the REFUSE message
shape.
Validation gate before merging to main: re-run the 5 debug topics
(Peter Steinberger, birthday gift for 40 year old, Kanye West, Garry Tan,
OpenClaw vs Paperclip vs Hermes) on v3.0.9 and confirm 5/5 canonical
compliance. Rollback to v3.0.8 if any previously-passing topic regresses.
Plan: docs/plans/2026-04-18-015-fix-engine-refuse-keyword-traps-delete-stale-skillmd-files-plan.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- scripts/compare.sh now runs /last30days vs /last30days-beta (was
/last30days vs /last30days-3:last30days-skill-private which was a stale
private install name that no longer works)
- CLAUDE.md adds a Beta channel section pointing at mvanhorn/last30days-skill-private
so future agent sessions discover the two-skill layout on project load
No runtime impact on /last30days. Engine code unchanged.
Plan: docs/plans/2026-04-17-005-feat-beta-skill-from-private-repo-plan.md
(plan file is gitignored per PR #259, not included in this diff)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- install matrix now leads with claude.ai (widest audience, one-click path)
- direct download link to GitHub release's 'latest' asset URL
- 3-step UI walkthrough with link to Settings > Capabilities > Skills
- Claude Code / OpenClaw / Gemini / manual paths still documented, collapsed
- removes the bash scripts/build-skill.sh requirement from end-user flow
- description leads with imperative 'Research' + 'what people actually say' (strong trigger signal for community/social-research prompts)
- argument-hint shows 3 concrete user phrasings instead of marketing copy
- 176 chars, well under Anthropic's 200-char cap
- preserves all source coverage (Reddit, X, YouTube, TikTok, Hacker News, Polymarket, GitHub, web)
Per ecosystem research (April 2026), trigger description quality is the single
biggest lever separating 500-install skills from 350k-install skills.
Atomic bump across all four manifests:
- SKILL.md (root)
- skills/last30days/SKILL.md (internal spec)
- .claude-plugin/plugin.json
- gemini-extension.json
CHANGELOG entry documents the skill-upload packaging fix, vendor/ removal,
legacy plans/ removal, and the new scripts/build-skill.sh builder.
Wraps git archive with --prefix=last30days/ so the zip contains a single
top-level skill folder matching SKILL.md's name: frontmatter. Enforces:
- refuses to build with a dirty working tree (prevents shipping untracked changes)
- fails if zip exceeds 200 files (claude.ai's empirical upload cap)
- fails if zip contains more than one SKILL.md (avoids name: confusion)
Output at dist/last30days.skill (gitignored).
Both plans describe work that was already shipped:
- feat-add-websearch-source.md - websearch is in the v3 pipeline (scripts/lib/perplexity.py etc)
- fix-strict-date-filtering.md - date filtering is enforced in scripts/lib/dates.py
New planning goes in docs/plans/ following the ce:plan convention.
Root vendor/package/ was an accidentally committed extracted npm tarball
(steipete-bird-0.8.0). Zero importers: the real vendored X client lives
at scripts/lib/vendor/bird-search/, referenced by scripts/lib/bird_x.py
and tests/test_bird_x.py.
Removes 215 files + 1 .tgz, dropping repo from 406 to 191 files and
clearing the claude.ai skill-upload 200-file cap.
Adds /vendor/ to .gitignore (leading slash so scripts/lib/vendor/ is unaffected).
Add extract_transcript_highlights() that scores sentences by specificity
(numbers, proper nouns, topic relevance) and filters YouTube filler
(subscribe, welcome back, etc). Top 5 highlights shown as structured
bullets in compact output. Full transcript moved to collapsible <details>
block so the LLM reads highlights first, full text on demand.
SKILL.md updated to instruct the judge agent to quote highlights
directly in synthesis, same as Reddit top comments.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
TRANSCRIPT_MAX_WORDS raised from 500 to 5000 so the LLM gets the full
content of most videos (up to ~25 minutes). Removed the second 200-char
truncation in render.py that was reducing transcripts to a single sentence
before the judge agent ever saw them.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When Cloudflare blocks requests to bsky.social or public.api.bsky.app
with a 403, the error was swallowed by a generic except clause and
reported as "Bluesky auth failed" - misleading users into thinking
their credentials were wrong.
Now _create_session() preserves the specific error in _session_error,
and search_bluesky() surfaces it. Cloudflare 403s get a clear message
about network-level blocks. Actual 401s say "Invalid credentials".
Closes#69
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Updated frontmatter description to be more search-friendly for ClawHub.
Added 12 new tags: deep-research, twitter, bluesky, recency, news,
citations, multi-source, social-media, analysis, web-search, ai-skill,
clawhub. Also added 11 GitHub repo topics.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The short alias /last30 only works on some platforms. Claude Code requires
the full /last30days name, so the follow-up suggestions after comparison
research were producing "Unknown skill: last30" errors.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>