my-hermes-skills

Changelog

23 entries across 17 skills, newest first, from each skill's CHANGELOG.md. Every push that bumps a version is also published as a GitHub release.

November 2026

  1. writing-reportsv2.0.0

    Changed

    • Reports are now written as short research papers: structured Abstract (Background, Objective, Methods, Results, Conclusions), then numbered Introduction, Methods, Results, Discussion, Limitations and Conclusion sections. The Verdict / Findings / Caveats layout is retired.
    • Every number carries an evidence type (measured, third-party benchmark, vendor-reported, community report, derived), and claims are hedged to match.
    • Report types (benchmark study, evidence review, technical investigation, technical guide) adjust the Methods and Results sections.
    • Search goes through Camofox instead of SearXNG.

    Added

    • Rules for rewriting an older report into the new format without changing its facts.
  2. impeccablev4.9.0

    Changed

    • Synced to upstream Impeccable 4.9.0 (engine 0.1.23) via npx impeccable install -y --providers=hermes --global into a scratch HOME, then copied in; files remain byte-identical to that install.
  3. writing-reportsv1.1.0

    Changed

    • References are numbered in order of first citation ([n] Title: <URL>), and every sourced claim carries an inline [n] marker. The docs site renders these as linked superscripts with a hover preview and back-links.
    • The template drops the --- separators between sections.
  4. slack-mention-watchv2.7.1

    Fixed

    • A queued review update is dropped when SLACK_REVIEW_AGENT changed since it was queued (review: dropped queued update for <id> (...: agent is now <agent>)), instead of being sent to the old agent's session.
    • Queued targets that no longer need a review (merged, approved, ...) are no longer pulled into a follow-up run. A failed resume is not retried in the same tick; its update stays queued.
    • Each target's PR/ticket status is checked once per tick, not once per use.
    • A started review run is saved to SLACK_REVIEW_STATE right away, so a later error in the same tick no longer loses it and the target is not reviewed twice.
    • PR review asks match whole words only: "looks", "checked" or "testing" alone no longer count as a review request.
    • The read-only guard also allows gh run/gh workflow view/list, gh search, gh status, gh issue status, git branch listing flags and git remote -v/show/get-url.
  5. slack-mention-watchv2.7.0

    Fixed

    • Review agent: only explicit review asks in the Mention: text start a review (PTAL, "can you review", "please check this", ...), not any PR link or ticket key in the block. Targets come from the mention, else the thread root only. Other pings get review: skipped (not a review request); asks without a target review: skipped (no target in mention or thread root).
    • Review dedupe is per target: one new target no longer re-reviews the others, and nothing is reviewed again just because 24 h passed (PRs: until the head commit changes; tickets: SLACK_REVIEW_DAYS, 7).
    • Review runs are read-only for every agent: git/gh guard wrappers first on the agent's PATH run only an allowlist of read-only git/gh commands (gh api: GET only) and block the rest, aliases included; claude no longer allows gh api (it allowed POSTs, e.g. PR comments). Argv checks, not a sandbox (see "Write guarantee"). SLACK_REVIEW_GUARD=0 turns them off.

    Added

    • Follow-ups resume the target's existing omp/claude session (review: resumed <id> (..., old..new)) with a brief holding the new Slack block(s) and, for PRs, only the old..new diff. A session still running gets the update queued (review: busy <id> (update queued)), delivered on a later tick.
    • review_spawn.py --sessions prints the session table (read only).
    • Opt-in Jev second opinion for review asks (SLACK_REVIEW_JEV=1 + TYPESAFE_API_KEY, SLACK_REVIEW_JEV_MIN 0.8, log in SLACK_REVIEW_JEV_LOG); the Jev shadow log asks the same review question.

    Changed

    • SLACK_REVIEW_STATE holds a session table per target (session, agent, runs, reviewed head commit, pid, queued updates); entries are pruned after 30 days without a run. Pre-2.7 files are migrated on the next save; migrated entries are never resumed. The review: already running line is gone.
  6. auditing-worktreesv1.0.0

    • Initial tracked version.
  7. camofox-browserv1.0.0

    • Initial tracked version.
  8. covidence-full-text-retrievalv1.2.0

    • Initial tracked version.
  9. covidence-full-text-reviewv2.0.0

    • Initial tracked version.
  10. covidence-screeningv1.0.0

    • Initial tracked version.
  11. demo-runbookv0.1.0

    • Initial tracked version.
  12. hermes-deepseek-cachev1.1.0

    • Initial tracked version.
  13. ilearning-autoadvancev5.0.0

    • Initial tracked version.
  14. impeccablev4.8.0

    • Initial tracked version.
  15. multi-repo-pr-review-merge-roundv1.0.0

    • Initial tracked version.
  16. oh-my-ghosttyv1.0.0

    • Initial tracked version.
  17. sci-hub-accessv1.1.0

    • Initial tracked version.
  18. slack-mention-watchv2.6.0

    Added

    • Opt-in Jev shadow mode (SLACK_MENTION_JEV_SHADOW=1 + TYPESAFE_API_KEY): logs TypeSafe Jev's emoji verdict and already-handled probability for each new mention to ~/.hermes/scripts/.slack-mention-jev.jsonl beside the regex result. Sends mention and thread text to api.typesafe.ai; off by default; changes no output or reaction.

    Changed

    • Digest cron example uses --reasoning-effort high, not max (max was ~88% of the job's spend).
  19. slack-mention-watchv2.5.0

    • Initial tracked version.
  20. slack-scanv1.0.0

    • Initial tracked version.
  21. superpowers-using-superpowersv1.0.0

    • Initial tracked version.
  22. tldrv1.1.0

    • Initial tracked version.
  23. writing-reportsv1.0.0

    • Initial tracked version.