Validate and, if permitted, execute ONE tool call against the target. DEFAULT-DENY: a probe runs ONLY IF it is provably read-only — a read-verb tool name (list/get/read/search/...), a server readOnlyHint, or a query-executor whose sql/query argument is a read-only statement (SELECT/WITH/SHOW/DESCRIBE/EXPLAIN/PRAGMA). A write-verb tool name or the destructive backstop (server destructiveHint or destructive-verb name) rejects regardless of your declaration. Your classification is REQUIRED but ADVISORY — recorded as evidence, it never authorizes execution. Also enforces JSON-schema validation, anti-fabrication provenance (an argument that DECLARES an observed source must cite a value actually captured — otherwise rejected; providing provenance is optional but citing a non-existent observation is fabrication), and the probe budget. Returns a redacted observation or the rejection reason. The observation's identifiers list EVERY short leaf scalar as a candidate (name, value, json_pointer, from_tool) — your raw material to author entity names, identifiers, enums (from distinct values), and relationships. Every result includes a gaps report (unsampled_structures, unexecuted_tools, untraversed_identifiers, sampling_hints, depth_signal).

Synthesize the workspace model, operational model, capability profile, quality report and SKILL.md from this session's accumulated safe observations, and write the full artifact set to disk. VERIFY every claim against the captured observations first — do not assert unobserved structures, identifiers, or relationships (the anti-fabrication provenance check will reject invented observed citations), and mark authored/inferred claims distinctly from probe-observed ones. If ZERO probes were accepted, only a STUB skill is written that plainly states nothing was safely observed — no rich profile is fabricated over a bare catalogue. Pass usage_summary: YOUR narrative of how THIS user actually uses this source, reasoned from what you observed (their saved searches, folders, tracked entities, recurring queries) — not a generic capability list. This becomes the skill's 'How You Use This MCP' section and is the whole point: the skill must let an agent exploit the MCP the way this user does. Returns skill_path to report back to the user.

Regenerate SKILL.md from an existing profile directory (profile-metadata.json + tool-catalogue.json + workspace-model.json + operational-model.json).

Connect to the target MCP, list its declared tools/resources/prompts, and start a profiling session. Returns tool cards (name, description, input schema, raw server annotations, backstop_blocked advisory) as raw material — YOU classify each tool's risk; DiscoMCP never keyword-guesses.

List target MCP ids configured in this DiscoMCP server.

Check whether a DiscoMCP skill already covers this target's current declared catalogue, without probing. A MISS (no existing_skill_dir) means dispatch a BACKGROUND subagent to profile the target (inspect_target -> execute_probe gap loop -> finalize_profile) and keep working the user's task while it runs.

Return the current GAP REPORT for an active session, computed only from state already gathered — no target calls, no probe consumed. Reports (does not decide): unsampled_structures (collections listed but never drilled into), unexecuted_tools (unprobed tools minus backstop-blocked; you judge which are read-safe, with why_useful), untraversed_identifiers (ids seen in output but never used as a get-by-id argument, with likely_consumer_tools), sampling_hints (schema params like orderBy/pageSize/q/filter on unused tools for smart sampling), and depth_signal (raw coverage counts + probe budget). The same report rides every execute_probe result under gaps. You decide when coverage is enough.