Beatless

Master meta-router for the 7-stage research pipeline. Use when user discusses research at the workflow level — 'how do I start research', 'what's the research pipeline', 'I have a research idea, what's next', 'guide me through a research sprint', 'what stage am I at', 'is my sprint working', '我有个研究想法', '下一步该做什么', '研究流程'. Routes to the right /research-* command based on sprint state + user intent.

Drive the new 6-stage research pipeline (init→survey→axiom→propose→dispatch→loop→review) via Hermes cron. Replaces deprecated auto-research and research-host cron drivers. Reads ledgers/<tag>/TODO.md, picks next stage, invokes claude -p with the appropriate /research-* command. Halts on SOTA/budget/BLOCK/human-flag.

Delegate ACADEMIC tasks to OpenAI Codex CLI. Distinct from codex-router (engineering/debugging). Three modes — novelty audit, citation verify, claim-vs-code-vs-numbers — all conservative + citation-anchored. Used by /research-* commands and Hermes research-cycle cron.

Delegate ACADEMIC tasks to Gemini CLI. Distinct from gemini-router (engineering/long-context). Three modes — cross-domain mining (Flash), challenger feasibility audit (Pro), arxiv probe (Pro). All challenger-framed; never agreement-seeking. Always paired with codex-academic citation-verify.

Audit the Astro blog at ~/claw/blog/ for missing zh translations, broken links, stale posts, and one-off content fixes. v2 reads from canonical path ~/claw/blog/src/content/blogs/ and treats draft:true entries as 'translated, awaiting human flip' so the audit no longer false-positives on translated drafts (closes Finding F2).

Daily strict-regulation audit of Hermes Agent + Beatless. Three-model parallel audit (Opus 4.7 + Codex gpt-5.3-codex + Gemini 3.1-pro) all run the SAME audit task simultaneously; final synthesis pass collapses three independent reports into one expert report. Output is committed as a blog post mdx and recorded into Hermes memory so the next cron tick is aware of prior findings.

Crawl Tier 1/2 lab tech-report feeds (RSS/Atom) into ~/obsidian-vault/feeds/<YYYY-MM-DD>/<lab>-<slug>.md. Output is feed-style notes (NOT papers — papers go to Zotero via paper-harvest). 404s logged and skipped. Rate-limit-friendly with 2s sleep between feeds. Dedup by destination path so re-runs are idempotent.

Convergence-driven GitHub PR pipeline. Picks targets from Tier 0 (curated repos.tier0.yaml) and Tier 1 (network-derived from followed-users events), filters via Tier 2 (BIG_ORGS exclusion) and Tier 3 (per-issue gates), then delegates the actual PR work to claude-code-router. Triple-pass review (codex correctness + gemini architecture + sonnet adversarial) before submit.

Triage open PRs we own and claimed issues. Read reviewer comments, fix CI breakage, address review points, and reply humbly. Never opens new PRs (that's github-pr's job), never claims new issues. Heavy lifting delegated to claude-code-router with a tightly-scoped permission boundary.

Harvest new papers from arXiv (lab+keyword filter), OpenReview (CCF-A venues), CVF, ACL Anthology, and HuggingFace Papers. Dedup against Zotero (arxiv_id → DOI → URL → title-similarity 0.85). Push fresh items into Zotero collections, capped per tick. Hermes memory tracks per-source last-fetched markers so we don't re-poll empty windows.

Mirror a Zotero collection into the Obsidian vault as `@<citekey>.md` notes, one per item. Skips items already mirrored (by Zotero key); writes new notes with extracted metadata + abstract. Records sync receipts so blog-draft can detect newly available paper notes.

Delegate a task to OpenAI Codex CLI (`codex exec`). Preset for terminal debugging, running test suites, code-correctness review, and shell-driven fix loops. Returns structured pass/fail with evidence.

Delegate a task to Gemini CLI (`gemini -p`). Preset for long-context reading, research, translation review, and adversarial second-opinion passes. The 1M+ context window is its key affordance.

Delegate a coding task to the Claude Code CLI (`claude -p`). Preset for heavy file edits, multi-file refactors, GitHub PR creation, and codebase exploration. The caller passes a working directory + task; this skill spawns a subprocess, captures stdout/stderr, and returns a structured result.

Decides whether an issue is worth pursuing as a PR. Use AFTER the Python preflight has fetched the issue body, labels, comments, and CONTRIBUTING.md excerpt — this skill does the JUDGMENT (dispute detection, AI-policy interpretation, direction alignment) that regex cannot do reliably. Returns a structured verdict: proceed / block / yield / ambiguous. Invoke from /github-pr Phase 2.5 before cloning any repo.