Jorge MM Marques

• Session Start — Call ONCE at conversation start — never repeat in the same chat. Returns slim router (_README.router.md ~400 tok) + datetime (date, simplenote_folder, weekday) + presence summary from Context Hub (physical_device, work_surface, current_task, confidence) + fleet tailnet status (devices: local/online/offline from tailscale status --json) + recent durable memories and memory_receipt (exact injected keys or explicit backend error). Reuse response.datetime, response.presence and response.fleet — do N
• Current Context — Refresh mid-conversation only — session_start already includes presence summary. Presença e contexto de trabalho do Jorge (Context Hub, payload completo). agent_runtime = infra onde o MCP corre (VPS); NÃO é presença. Usar presence.work_surface, presence.physical_device e current_task. Silencioso — nunca pedir ao Jorge para correr contextctl. Se presence.confidence < 0.6, usar melhor palpite e continuar; só perguntar se a acção depender criticamente do device (ex.: correr comando local destrutivo
• List Folder — List files and subfolders in any vault folder. path='': vault root. stats=true: mtime/ctime/size. Non-recursive: all entry types with ext on files (.pdf, .py, …). recursive=true: all .md under path (capped ~200; use find_files for PDFs/binaries). Avoid recursive on 00_inbox — token-heavy.
• Find Files — Discover vault files by name, extension, type, path, date, or size — metadata only, no content. Use for PDFs, docx, xlsx, code, media (.md included unless exclude_md=true). Max 100 results; requires at least one filter. After discovery, use run_skill to read (dre-parser, vault-find --brief, domain skills). Types: ebooks, docs, images, audio, video, code, data, archives.
• Read Vault Images — Load image files from the vault as base64 data URIs for inline preview. Use after find_files when you need to SHOW photos (not just list paths). Max 6 images, 2 MB each. jpg/png/gif/webp/bmp/svg.
• Read Note — Return note content by vault-relative path. Large notes are windowed by line offset/limit and capped at ~32000 chars. When truncated=true, continue with offset=next_offset. Token-heavy paths may include additive field warning (CHANGELOG, WhatsApp exports, oversized manifest entries, JSON/JSONL). Prefer read_frontmatter for YAML only; bulk_read for ≤20 paths.
• Read Frontmatter — Return only the parsed YAML frontmatter of a note (no body). Use before run_skill to check has_learnings (~200 chars vs full read_note). Malformed YAML returns partial dict plus warning, not a hard error.
• Bulk Read Notes — Read up to 20 notes in one round-trip (per-note body cap like read_note). Returns notes dict plus optional errors (batch does not abort on single failure). Honours _PRIVADO/ and vault-graph.json blocks.
• Read JSON — Navigate structured .json files schema-first — never use read_note or file-reader on JSON. Workflow: mode=schema (default) to see shape/types/sizes cheaply, then mode=list to browse arrays/objects with pagination, then mode=get with query for the exact value. query syntax: dotted keys and brackets (e.g. conversations[0].messages[-1].content). get values capped at ~32000 chars. For vault-graph.json use run_skill('vault-graph', ['query', ...]) instead.
• Read JSONL — Read .jsonl/.ndjson files line-by-line without loading the whole file. mode=schema infers shape from first lines; mode=rows (default) returns a paginated window. Use offset/next_offset to paginate. Never use read_note on JSONL.
• Write Note — Write a vault note. mode controls behavior: 'upsert' (default) = create or overwrite; 'create' = fail if exists; 'update' = fail if not exists; 'append' = add to end of existing note.
• Edit Note — Surgically replace an exact string in a note. Default: old_str must appear exactly expect_count times (default 1). replace_all=true replaces every occurrence. Saves ~90% tokens vs read_note + write_note for targeted changes.
• Move Note — Move or rename a note within the vault.
• Search Notes — LAST RESORT full-text search (case-insensitive AND). Max 30 paths per call; IPv4 queries rejected. Prefer: vault-dispatch (routing), vault-find (recent files), vault-graph (relations), vault-search skill (regex/acentos), activity-log (PC context). Never search vague terms or IPs.
• Get Current Datetime — Mid-conversation ONLY — never at session bootstrap. session_start already returns the same datetime; reuse it instead of calling this. Use only if the chat is long and you need a fresh clock for hoje/ontem/esta semana. Returns date, time, weekday, ISO week, simplenote_folder (YYYY/MM).
• Tasks — Manage operational tasks (backend Vikunja). Requires 'command' (see enum). ALWAYS ask Jorge before calling 'add' — agents suggest, never create autonomously. id accepts global integer OR 'projeto/indice' (e.g. 'infra/16'). commands: focus=list NOW tasks; list=all open tasks; add=create task (needs title); done/drop/waiting=change status (need id); move=change bucket (needs id+bucket); sweep=triage suggestions (read-only); update=edit fields (needs id); archive=month report (needs month YYYY-MM);
• Vikunja (deprecated — use tasks) — DEPRECATED alias for `tasks`. Use tasks() instead. Manage operational tasks via Vikunja backend. Same commands as tasks tool.
• Calendar — Manage CalDAV calendar events (Radicale at cal.jorgemarques.com). Requires 'command'. Timezone: Europe/Lisbon. Never invent times. calendars=list collections; list=events in range (needs from+to); add=create event (needs title+start); get/update/delete=by uid; import=batch from YAML/JSON file. calendar arg: castelform | pessoal | foco (projection, read-only for agents).
• Run Skill — Execute a skill from 50_infra/skills/ (searched recursively by name; must have main.py). Site jorgemarques.com (obras, título, CV): run_skill('portfolio-api', ['patch-work', '<slug>', '--title', '...', '--apply']). Dry-run first: omit --apply. Never edit_note/browser for live portfolio content. vault-dispatch routes NEW vault notes — not portfolio API writes.
• Approve Improvement Proposal — Approve or reject a pending agent-improvement proposal by signature. SSOT: pending-approvals.json. Use decision 'yes'/'approve' or 'no'/'reject'. Applied proposals are processed by agent-improvement --apply-approved cron.

Schemas and invoke metadata: Directory API