Abel CAP causal exploration skill. Use for direct CAP graph questions and for off-graph life or decision questions that should be routed through market proxy tickers. Trigger when the user asks what is driving a node, why a node moved, whether a path exists, what changes under intervention, what a counterfactual preview says, how a server's causal capabilities are exposed, or when a career, education, lifestyle, or macro decision should be read through Abel's market signal layer. Do not use for pure quote lookups, generic news summaries, raw node dumps, or unrelated coding tasks.
Use this skill for cause-effect questions on the Abel CAP wrapper. Financial markets are the signal layer, not the product: the CAP server exposes finance and crypto price or volume nodes, and those nodes can be used either directly or as proxy signals for larger real-world questions.
Treat skill update detection as a soft prerequisite on the first use of this skill in a session.
references/update-flow.md.python scripts/check_skill_update.py.update_available: true, tell the user the current version, latest version, and a concise summary from CHANGELOG.md.Y/N choice so the user can answer quickly.causal-abel skill. Do not propose a full npx skills update.Authorization is a required entry step for this skill when it will call Abel APIs on the user's behalf.
--api-key, or .env.skills.references/setup-guide.md immediately. Do not start CAP probing, capability inspection, or any other live API call first.GET https://api.abel.ai/echo/web/credentials/oauth/google/authorize/agent.data.authUrl to the user, store data.resultUrl or data.pollToken, and poll until the result is authorized, failed, or expired.Use causal-abel when the user needs causal structure, reachability, intervention meaning, or capability inspection on the Abel CAP surface.
Typical triggers:
extensions.abel.*Do not use this skill for:
Check authorization state before any live API call.
references/update-flow.md before live Abel API usage.Y/N.ABEL_API_KEY is missing from session state, --api-key, and .env.skills, start the OAuth handoff from references/setup-guide.md first.Start from the user's causal question and the live CAP surface.
https://cap.abel.ai unless the user provides another base_url.https://cap-sit.abel.ai is the SIT variant when you need the staging environment.https://api.abel.ai/echo/ as the OAuth and business API host from references/setup-guide.md, not as the default public CAP probe target.Classify the task.
capability_discovery: what the server exposesdirect_graph: direct node, path, blanket, or intervention questionproxy_routed: real-world question that must be represented through market proxiesRead structure before telling a story.
driver_explanation, reachability_check, intervention_effect, counterfactual_read, or capability_audit.Normalize node inputs before any live probe.
If the user installs this skill, asks to connect Abel, or the workflow is missing an Abel API key, follow references/setup-guide.md exactly.
references/update-flow.md.data.authUrl to the user, not the /authorize/agent API URL.data.resultUrl or data.pollToken and poll until the result is authorized, failed, or expired.data.apiKey in session state and .env.skills when local storage is available.references/question-routing.mdreferences/inversion-flow.mdassets/report-template.mdreferences/setup-guide.mdreferences/update-flow.mdreferences/probe-usage.mdreferences/capability-layers.mdNVDA or SPOT, default to <ticker>_close unless the question is explicitly about trading activity or volume.Spotify or music streaming, map it to a real ticker first; do not probe a free-form phrase.For capability discovery, avoid redundant full-surface dumps.
meta.capabilities when the question is "what surfaces or tiers exist here?"meta.methods only when you need invocation metadata such as arguments or result_fields.meta.methods queries with params.verbs instead of pulling the whole registry and filtering it afterward.python skill/causal-abel/scripts/cap_probe.py methods observe.predict traverse.parents over ad hoc curl payloads when you need a stable method read.Read structure before telling a story.
driver_explanation, reachability_check, intervention_effect, counterfactual_read, or capability_audit.Normalize node inputs before any live probe.
NVDA or SPOT, default to <ticker>_close unless the question is explicitly about trading activity or volume.Spotify or music streaming, map it to a real ticker first; do not probe a free-form phrase.Use decision gates, not verb dumps.
driver_explanation, start with traverse.parents or graph.neighbors(scope=parents), then add graph.markov_blanket only if direct drivers are still unclear.reachability_check, start with graph.paths on the specific proposed source and target. Use extensions.abel.validate_connectivity only when screening a small candidate set is more honest than many repeated path probes.intervention_effect, do a minimal structural confirmation first, then call intervene.do only if the structural question is already clear.capability_audit, do not pair a full meta.capabilities dump with a full meta.methods dump unless the user explicitly needs both views. Inventory first, then targeted method detail for the verbs that remain in question.extensions.abel.* when CAP core cannot answer the user's actual question.Answer in layers.
assets/report-template.md: start from the user's original question, map that question to graph nodes, then separate each verb's result from what that result means for the question.Stay semantically honest.
observe.predict as observational, intervene.do as intervention, and counterfactual_preview as preview-only.