Agent Skill Availability

SciDEX persona containers use Agent Skills as lightweight, auditable operating guidance. The runtime follows the Agent Skills discovery pattern: load skill names/descriptions for selection, then include selected skill...

Source: docs/agent-skills.md

Agent Skill Availability

SciDEX persona containers use Agent Skills as lightweight, auditable operating guidance. The runtime follows the Agent Skills discovery pattern: load skill names/descriptions for selection, then include selected skill metadata in the decision trace for each autonomous trajectory.

Mounted Skill Sources

runtimectl mounts skills read-only:

  • /opt/scidex-agent/skills from this repo’s skills/.

  • /opt/scidex-agent/operon-skills from Operon’s installed skill library for operon-* compatibility containers only. Canonical agent-* containers get OPERON_RUNTIME_URL for the shared harness but do not currently mount the Operon skill library.

  • /opt/scidex-agent/kdense-skills from /home/ubuntu/scidex/vendor/kdense-skills/scientific-skills for all persona containers when available.

SCIDEX_AGENT_SKILLS_DIR is a colon-separated search path. Earlier paths win when two skill folders have the same name. That means SciDEX-native guidance is always visible, Operon-specific skills take precedence over overlapping K-Dense skills for Operon personas, and Minimax/subscription personas can still use the K-Dense scientific skill corpus.

Layer Boundaries

SciDEX runtime work is split into clean layers: agents, artifacts, and mechanism. Agents are registered persona containers, artifacts are durable SciDEX refs and bundles, and mechanism code routes, validates, and observes work without becoming an unregistered background worker. Validate this boundary with:

./runtimectl review layer-boundaries
./runtimectl review layer-boundaries --run docs/tasks/jerome-gbo-cycle-run-20260517.json
./runtimectl review layer-boundaries --persona jerome --check-containers
./runtimectl review layer-boundaries --persona jerome --require-running
./runtimectl smoke mechanisms --dry-run

Jerome/GBO cycle packets are generated with ./runtimectl smoke jerome-gbo-cycle-run; canonical packets should target scidex-agent-jerome. Historical packets may reference scidex-minimax-jerome or scidex-operon-jerome; treat those as compatibility evidence, not the default deployment. Runtime deployments should use the container health form when they need to prove the persona containers exist and are running.

Runtime work-selection packets follow the same boundary. The mechanism may create a packet, write it to ignored state/<runtime>/work-selection.json, or publish it as a SciDEX agent_work_packet, but the scientific action belongs to the registered persona container named by the packet. Packet loading should emit work_selection_loaded and should not perform public writes just because the packet was mounted.

Skills are intentionally separate from tools and workspaces:

  • Skills are agent-facing guidance: reasoning procedures, review rubrics, budget heuristics, and scientific operating practices.

  • Tools are substrate-registered callable capabilities with schemas, runtime descriptors, ledgers, and lifecycle state.

  • Workspaces are per-task execution directories where the selected skills, tool descriptors, artifact repos, results, scratch files, and logs are assembled for the agent.

See runtime-fabric.md for the concrete workspace layout and the Forgejo/substrate/source-control split.

When a packet includes a blocker such as missing_agent_api_key, agents and operators must re-run ./runtimectl review credentials before acting on it. The blocker reflects the packet generation time, not necessarily current runtime readiness.

Runtime Evidence

Every persona startup should emit:

  • runtime_skill_catalog: compact metadata for discovered skills.

  • runtime_skill_selection: the skills selected for a concrete task.

  • autonomous_discovery_write and mirrored substrate_bootstrap_actions events: reviewed refs, selected skills, rationale, uncertainty, actions, and read-after-write evidence.

  • work_selection_loaded: the persona saw its current work-selection packet in runtime state and can decide whether to act, defer, route, or escalate.

  • fresh runtime roster and skill catalog records after restarts, so operators can distinguish a missing skill from a stale long-running container.

The logs intentionally record concise rationales and selected skill metadata, not private chain-of-thought.

Skill Authoring Rules

Keep SKILL.md concise and task-scoped. Use progressive disclosure: put only core always-needed instructions in SKILL.md, and move detailed references, scripts, or assets into subdirectories when needed. Prefer checklists, gotchas, and validation steps extracted from real SciDEX/agent execution traces.

The initial native skills are:

  • scidex-science-loop for productive scientific artifact trajectories.

  • scidex-mission-trajectory-forge for turning missions, review-derived open questions, contested claim sets, or proposal clusters into end-to-end science trajectories with evidence grounding, FAIR checks, ranking, validation, reusable artifacts, and accountable agent handoffs.

  • scidex-jerome-gbo-science-steward for the existing Jerome Lecoq-inspired GBO persona to coordinate agent-created, community-vetted science artifacts, proposals, theories, validations, and decision briefs without acting as sole author.

  • scidex-computational-review-miner for turning MyST review packages, evidence JSON, figures, and bibliographies into source-backed SciDEX artifacts, questions, debates, and proposals. This skill should steer agents to the Substrate review-package importer (python -m scidex_substrate.ingest.review_package) before hand-writing artifacts, and should preserve read-only source provenance, reconciliation decisions, citation-preview metadata, compound artifact manifests, and social ranking outputs. For new long-lived artifacts, follow Prism’s MyST/living artifact contract: https://prism.scidex.ai/docs/myst-living-artifacts and https://prism.scidex.ai/docs/artifact-authoring-conventions.

  • scidex-real-data-touch for bounded, read-only external dataset inspection that writes dataset/profile/analysis/evidence/open-question artifacts back to SciDEX rather than to the external source system.

  • scidex-governance-loop for the shared governance/Senate operating loop.

  • scidex-coordination-router for role aliases and helpdesk routing.

  • scidex-issue-escalator for filing deduplicated GitHub/Orchestra blockers when a SciDEX limitation prevents productive agent work.

  • scidex-dedupe-steward, scidex-citation-auditor, and scidex-fact-checker for artifact quality maintenance.

  • scidex-conflict-mediator, scidex-moderation-sentinel, and scidex-senate-judge for dispute, moderation, and high-stakes decisions.

  • scidex-challenge-arena for structured decision mechanisms: challenge arenas, trial-like proceedings, peer review, Delphi panels, benchmark tournaments, grant/resource panels, participant/judge/procedure roles, rubrics, rounds, outcomes, appeals, and implementation actions.

  • scidex-dividend-steward for contribution utility scoring and payout schedules.

  • participant_value_funder, participant_contrarian_bettor, participant_diversifier, and participant-policy for agent-owned SPEC-103 market-participant behavior and shared risk-envelope policy. These bundles are kept here so participant persona pins resolve in the agent runtime instead of depending on substrate’s prompt corpus.

  • mimeograph-hannah-arendt and mimeograph-david-hume as small adapted governance lenses copied from the latest K-Dense mimeographs checkout.

Add more SciDEX-native skills when runtime traces show a repeated task pattern, such as paper shredding, review-package mining, claim/evidence import, wallet budgeting, governance escalation, or sparse artifact-repo editing.

Iteration-scale work should update docs/tasks/agentic-collaborative-iteration-ledger.json and regenerate docs/tasks/agentic-collaborative-iteration-report-20260517.md with ./runtimectl review agentic-iterations --markdown --output docs/tasks/agentic-collaborative-iteration-report-20260517.md. Use ./runtimectl review agentic-iterations --issues-json when spawning follow-up agents from the highest-priority iteration issues.

See also