SciDEX-Agents skill library

The native skill library mounted into SciDEX persona runtimes. Per SPEC-013 each skill is a first-class artifact: versioned, searchable, signal-rated, and improvable. This README is the catalog + authoring guide — read it before adding a skill so the library stays consistent as it grows to many skills.

See also ../docs/agent-skills.md for runtime mounting and search-path precedence, and ../docs/agent-first-maintenance.md for how seed-driven periodic maintenance runs as agents (detect→gap_mission→claim→ act→gate) rather than substrate cron jobs.

How skills are stored, discovered, and selected

• One folder per skill: skills/<slug>/SKILL.md (+ optional resource files). The folder is flat — the runtime loader scans direct children of each skills root for a SKILL.md (it does not recurse), so do not nest skills in category subfolders. Categories are organizational only (this catalog + the tags frontmatter), not directories.

• Naming: native SciDEX-agents skills use the scidex- prefix (scidex-science-loop); persona-lens skills use mimeograph-<figure>. The skill’s identity comes from the frontmatter name, not the folder name (keep them equal). Substrate’s maintenance library uses bare kebab names (entity-linking); cross-repo refs to those drop the prefix.

• Discovery: runtime loaders read SCIDEX_AGENT_SKILLS_DIR (colon-separated search path; earlier paths win on name collision) and parse each SKILL.md’s frontmatter. Substrate also DB-seeds native skills from SKILL.md and refreshes the row when the content hash changes.

• Selection: the runtime scores skills by token overlap with the persona’s mission/domains/role/task and injects the top few into context. Make name and description carry the real trigger vocabulary.

Authoring a new skill

1. Copy SKILL.template.md to skills/<slug>/SKILL.md (the template is a top-level file, not a <dir>/SKILL.md, so the loader never mounts it as a skill).

2. Frontmatter: required name (slug, ^[a-z][a-z0-9-]*$) + description (≥20 chars, packed with triggers). Recommended SPEC-013 fields: version (semver), triggers, tags, personas, source: native, source_format: scidex, requires_substrate_access, and dependencies (other skills this builds on).

3. Body: progressive disclosure — keep SKILL.md concise and task-scoped (when to use → workflow → quality gates → gotchas → a minimal output record). Move heavy schemas, tables, scripts, or templates into subdirectories (reference/, templates/) and link to them, as scidex-persona-creation/ does.

4. Reference other skills by slug; never inline another skill’s body.

5. Add a row to the catalog below.

Catalog

Authoring

Skill	Use it when
scidex-persona-creation	Author a new persona/Mimeo: bundle + manifest, playful public name + aliases + consent/disclaimer, skill bindings, optional runtime-binding JSON. Hand-off to Substrate import.

Science

Skill	Use it when
scidex-science-loop	A scientist agent needs to discover artifacts, choose useful work, run/plan analysis, publish, link evidence, comment, rank, or file blockers without noise.
scidex-computational-review-miner	Mine a computational review / MyST package / evidence bundle / large bibliography into proposable SciDEX artifacts.
scidex-mission-trajectory-forge	Turn a mission, open question, contested claim set, or proposal cluster into an end-to-end science trajectory.
scidex-jerome-gbo-science-steward	Steward Global Brain Observatory science priorities (Jerome persona).
scidex-real-data-touch	Inspect an external public dataset/manifest without modifying it; create a bounded local profile; publish.

Quality & maintenance

Skill	Use it when
scidex-citation-auditor	Verify citations, provenance, source links, DOI/PMID metadata, quote fidelity, evidence support.
scidex-fact-checker	Check for factual errors, hallucinated claims, fabricated sources, overbroad conclusions.
scidex-dedupe-steward	Detect and propose safe deduplication of artifacts/wiki/claims/datasets/papers/hypotheses.
scidex-crosslink-gardener	Review/apply the linking_maintenance worker’s cross-link suggestions; link under-linked artifacts.

Governance

Skill	Use it when
scidex-governance-loop	Core governance loop: Senate issues, maintenance roles, moderation, quality, economics, disputes.
scidex-senate-judge	Judge Senate issues, challenge awards, appeals, sanctions, quarantines, large payouts.
scidex-challenge-arena	Design/run/judge/archive structured decision mechanisms (debate, compete, deliberate, adjudicate).
scidex-conflict-mediator	Mediate artifact disputes, contested edits, debate escalation, moderation edge cases.
scidex-moderation-sentinel	Detect/respond to spam, slop, abuse, fabricated artifacts, hallucination floods, malicious tagging.
scidex-dividend-steward	Review contributions and propose/execute bounded dividends, rewards, payouts, anti-gaming.

Coordination & runtime

Skill	Use it when
scidex-continuous-core-loop	The shared default-mode budget-aware runtime loop (memory maintenance, session hygiene, skill selection).
scidex-coordination-router	Route , helpdesk requests, blocked agents, capability aliases, A2A asks, issues, Orchestra tasks.
scidex-issue-escalator	A blocker needs repo work, Substrate/Prism/API change, runtime provisioning, or operator intervention.

Meta-evaluation

Skill	Use it when
scidex-skill-benchmark-loop	Benchmark a SciDEX skill — dispatch it on curated tasks, jury the output with complementary skills, post a benchmark_record, emit observability-only skill-credit deltas. Use before wiring a skill into a new payout, governance, or selection path.
scidex-eval-task-curator	Seed, grow, perturb, or retire the held-out gold-set of benchmark_task artifacts the benchmark loop draws from. Curator may propose but may not grade its own additions.

Mimeograph lenses

Skill	Use it when
mimeograph-hannah-arendt	Governance questions of factual truth, public judgment, bureaucracy, conflict mediation (Arendt lens).
mimeograph-david-hume	Factual grounding, causal-claim review, empiricism, institutional legitimacy (Hume lens).