github mvanhorn/last30days-skill v3.0.9
v3.0.9 - The Self-Debug Release
on GitHub

latest releases: v3.18.0, v3.17.0, v3.16.0...
3 months ago

v3.0.9 - The Self-Debug Release

Highlights

v3.0.9 is live. New user-facing capabilities, broader cross-platform support, and a skill that now runs reliably on Claude Code, Codex, Hermes, Gemini, claude.ai, and OpenClaw. The headline fix: the engine refuses "birthday gift for 40 year old" style queries with a clarifying question instead of 5 minutes of junk output. The headline feature: TikTok and YouTube top comments now render alongside Reddit's, so the most-engaged voice from every source makes it into the synthesis.

The label - "The Self-Debug Release": I handed 5 separate Opus 4.7 instances their own failed outputs and asked them to debug themselves. Three converged on "SKILL.md is too big and the LAWs are too deep." Two converged on "the engine should refuse demographic-shopping queries outright" and "the WebSearch Sources reminder is overriding LAW 1." I copy-pasted their diagnoses into code. Validation: 5/5 canonical compliance on the topics that had failed.

New capabilities

  • TikTok and YouTube top comments render alongside Reddit's. PR #260 made the top-engagement comment from each TikTok video and YouTube video first-class in the output - same prominent 💬 Top comment treatment Reddit's top comment already got. This is the biggest user-facing output change since 3.0.0 and it was never announced. The community inspiration trace: @uppinote20's original push for richer Reddit comments (PR #143) seeded the pattern; this PR generalized it across TikTok and YouTube. PR #265 followed up by fixing the ScrapeCreators url= param + new response shape for YouTube comments/transcripts so the enrichment actually works.

  • last30days runs on Hermes AI Agent now. @stephenmcconnachie's PR (#228) added Hermes as a first-class deploy target. scripts/sync.sh detects ~/.hermes/skills/research and deploys the full skill (SKILL.md, scripts, lib modules, fixtures) to Hermes's skills directory alongside Claude Code and Codex. This is one of the biggest surface-area expansions in v3 - last30days is now usable inside the Hermes agent's research workflows without any manual wiring.

  • Multi-key SCRAPECREATORS_API_KEY rotation. @zaydiscold's PR (#268) added automatic key rotation. Set SCRAPECREATORS_API_KEY_1, SCRAPECREATORS_API_KEY_2, etc. and the engine rotates when a key hits rate limits instead of failing the whole run. For power users running daily queries, this is the difference between rate-limit 429s and zero-touch reliability.

  • The skill works on Windows now. @Chelebii's PR (#227) stabilized the vendored Bird X search client on Windows. Previously the bundled X backend had subtle runtime issues on Windows terminals; now it runs clean. Pair this with @Gujiassh's UTF-8 encoding fix (#225) for saved output and Windows users get the full v3 experience without workarounds.

  • Linux permission checks stopped false-warning. @george231224's PR (#216) fixed check_perms on Linux by preferring GNU stat's syntax over the BSD stat that the skill was calling. Linux users were getting spurious permission warnings on .env files that were already correctly 600-chmod'd. Now the check matches reality.

  • Gemini CLI got a first-class install path. @hnshah's docs PR (#224) added the Gemini CLI install note and workaround for a rough edge in the Gemini skill loader. Gemini users now have a one-paragraph install flow in the README instead of having to reverse-engineer the plugin layout.

  • Offline quality evaluation. @j-sperling's PR (#233) added eval_topics.json as a fixture. Contributors and I can now run quality-regression checks on synthesis output without burning live API credits. This is the scaffolding that made the plan 015 validation gate affordable - without eval fixtures, testing 5/5 canonical compliance on every release would cost real money every time. Ships as contributor infrastructure but shows up as stability for end users.

  • Reddit client got a cleaner HTTP layer. @iliaal shipped three architecture PRs back-to-back (#207, #208, #209) that consolidated Reddit's HTTP handling into http.get(params=...), rejected garbage input in _parse_date, and unified _sc_headers into http.scrapecreators_headers. End-user benefit: fewer flaky timeouts, fewer "weird parse error" crashes, a codebase that's easier for future contributors to touch without breaking Reddit. These aren't sexy PRs; they're the kind of refactor that prevents six future bug reports.

  • The --days=N flag keeps working. @BryanTegomoh's PR (#230) restored backcompat for the legacy --days alias so anyone who'd scripted against it in 2.x doesn't break on v3. Small PR, meaningful reliability gain for existing users.

  • INCLUDE_SOURCES has a sane default. @hnshah's PR (#223) defaulted the env var to empty string instead of unset. Missing env no longer breaks source inclusion on fresh installs.

  • Version metadata stays in sync. @Gujiassh's PR (#217) aligned the SKILL.md version header with the sync target version, and @shalomma's PR (#229) closed the remaining drift between the SKILL.md header and plugin.json. "Which version am I actually on" is no longer an adventure.

  • Bird X engagement handling got hardened. @j-sperling's PR (#234) made bird_x skip all-None engagement dicts instead of crashing on them. Rare condition, but the kind of thing that silently kills a run on a specific topic.

  • Dev workflow hygiene. @j-sperling's gitignore PR (#232) dropped .venv, .coverage, htmlcov, and .memsearch from the tracked tree. Contributor quality-of-life; keeps PR diffs clean.

  • The skill installs to claude.ai. PRs #242 and #244 shipped scripts/build-skill.sh plus the .gitattributes + export-ignore plumbing that packages last30days into a claude.ai-upload-ready .skill file under the 200-file cap. The skill is no longer Claude-Code-only - it installs directly on claude.ai, too. README has the upload workflow.

  • OpenAI Codex CLI discovers the skill natively. PR #219 added .agents/skills/last30days/SKILL.md as a real file (not symlinked - Codex's loader skips symlinks) plus .codex-plugin/plugin.json as the namespace marker. The skill now shows up as last30days:last30days when Codex runs in a checkout. Inspired by @Jah-yee (#153) and @dannyshmueli on X.

  • /last30days as a slash command. PR #267 added commands/last30days.md so plugin users can type /last30days <topic> and Claude Code autocomplete prefix-matches it to the canonical /last30days:last30days form. No more typing the double-namespace.

The self-debug technique, for anyone rebuilding this elsewhere

The breakthrough wasn't the individual fixes. It was the realization that instead of guessing why the model was ignoring the rules, I should ask the model. Five separate Opus 4.7 sessions debugged their own outputs:

  • "Did you read SKILL.md?" → "I tried Read, hit the 25K token cap, and bailed instead of chunked-reading."
  • "Why the trailing Sources block?" → "The WebSearch tool's own reminder said MANDATORY. Precedence was unclear."
  • "Why the section headers?" → "I had strong priors on Peter Steinberger and wrote my thesis instead of passing through."
  • "Why the wrong file?" → "I read .agents/skills/last30days/SKILL.md first because it appeared in the path glob."

Three of the five said "move the LAWs to the top." Two said "make the engine enforce it so the model can't not comply." I shipped both. That's the whole technique: when the LLM-under-orchestration keeps breaking the contract, don't argue with it - ask it to debug itself, and build structural enforcement around whatever it names as the root cause.

Thank you

Community PR authors since v3.0.0:

v3 roadmap contributors (issues and PRs that shaped the v3 feature set):

The five Opus 4.7 instances that debugged their own failures on v3.0.7 and v3.0.8 and converged on the fixes. The convergence was the breakthrough; this release is their diagnosis in code.

Install / Update

/plugin marketplace add mvanhorn/last30days-skill
/plugin install last30days@last30days-skill

Or if already installed: 

/plugin update last30days
/reload-plugins

Verify

cat ~/.claude/plugins/cache/last30days-skill/last30days/*/.claude-plugin/plugin.json | grep version

Should print "version": "3.0.9".

Smoke test

/last30days birthday gift for 40 year old

Should ask a clarifying question before running. If it runs the engine anyway, the cache is stale - repeat the plugin update.

Full Changelog: v3.0.5...v3.0.9

What's Changed

  • feat: multi-key rotation for SCRAPECREATORS_API_KEY by @zaydiscold in #268
  • feat: make default fun level actually surface comedy by @mvanhorn in #272
  • revert: default fun surfacing (#272) by @mvanhorn in #273
  • feat: wire compare.sh and CLAUDE.md for /last30days-beta channel by @mvanhorn in #276
  • feat: release v3.0.6 - promote plans 003-009 from private beta to public by @mvanhorn in #277
  • fix: v3.0.7 - restore mandatory first-line badge + pin SKILL_ROOT + skill-specificity anchor by @mvanhorn in #278
  • fix: v3.0.8 - SKILL.md was too big and LAWs too deep (model-self-debug grounded) by @mvanhorn in #279
  • fix: v3.0.9 - engine refuses Class 1 keyword traps, delete stale SKILL.md files, reinforce LAW 1 over WebSearch by @mvanhorn in #280
  • docs: v3.0.9 release notes - The Self-Debug Release by @mvanhorn in #281

New Contributors

Full Changelog: v3.0.5...v3.0.9

Check out latest releases or
releases around mvanhorn/last30days-skill v3.0.9

Don't miss a new last30days-skill release

NewReleases is sending notifications on new releases.

Get notifications