HERMES

Reconcile split-brain truth sources in a multi-organ federation — when arifOS capability_map, Hermes config.yaml providers block, MCP server registry, and runtime env all claim to be the canonical source for "what's configured". Load when Arif reports "Hermes says 1 provider, arifOS says 28", when a tool call fails with unknown provider, or when fixing audit findings about API key / model / organ coverage.

Add, fix, or migrate custom/third-party LLM providers in a Hermes Agent install so they appear in `/model` (CLI) and the Telegram gateway picker. Covers providers block shape, env-var key mapping, security-blocker on direct config writes, three-tier multimodal fallback chains, MiMo UltraSpeed/Pro/v2.5 capability matrix and supports_vision flag for forced auxiliary routing, vault→hermes.env key sync pattern, picker zen with 5-state failure classification (OK/PAUSED/BROKEN/CHAT_INCOMPATIBLE/EXC) and dedupe by route priority, MoA preset cross-dependency on provider models, MiniMax direct provider + mmx-cli for multimodal generation, OpenCode Go vs Zen tier verification, and dead fallback cleanup.

MiniMax multimodal via mmx-cli — TTS, video, music, image, vision, search

Test whether a proposed improvement actually beats the existing baseline BEFORE committing to build it. Spawn parallel forge agents that each compare challenger vs baseline on the same data. Kill ideas that don't improve; forge only what measurably helps.

Map of the arifOS federation — organs, ports, roles, MCP surfaces, health probes. Load when Arif asks 'what organs are alive', 'show me the federation', 'is X running', 'what port is Y', or when you need to route a task to the right organ.

Probe live state BEFORE proposing fixes, additions, or changes. Read configs, test commands, check what exists — then act only on real gaps. Use ALWAYS. This is not optional. Arif corrected multiple times (2026-07-04, 2026-07-08, 2026-07-11) for proposing fixes to things that already worked, for citing stale audits as live state, and for accepting 'first time' / 'novel' / 'pick from N options' prompts without probing disk.

Forge new constitutional organs in arifOS — vector indexing, dual-gate architecture, EMD pipeline integration, Qdrant backfill, derived-view enrichment.

Create new arifOS runtime modules (audit, fatigue, scoring, etc.) following project conventions — Pydantic v2, VAULT999 persistence, ruff clean, line budgets.

Track nasi lemak sales across locations/days — process receipt images, compile baki, compute revenue. Use when user shares handwritten cash bills, order receipts, baki counts, or asks for sales summaries and revenue analysis.

Dark-themed HTML+SVG visual guides for non-tech subjects — anatomy, biology, timelines, recovery journeys, process flows, laws/rules, and educational explanations. Sibling to architecture-diagram; same dark grid aesthetic but extended for explanatory/educational content.

Generate images without heavy local installs — free APIs (Pollinations.ai), pre-generated galleries (Generated Photos), and quick alternatives. MiniMax image-01 is now PRIMARY for Malay/SEA phenotype — this skill covers free fallbacks only.

Governed A-FORGE development: reading directives, modifying governed shell tools, wiring arifOS kernel round-trips, and maintaining constitutional gates in the arifOS federation's execution organ. Covers forgeShell.ts, arifJudge.ts, arifSeal.ts, godelLock.ts, and MCP server registration. Always load when Arif gives a JITU/forge directive, an E1-E9 order, or asks to build/modify anything in /root/A-FORGE.

Add, modify, and test Caddy reverse proxy routes on the arif-fazil.com infrastructure. Covers static file serving, API proxy, handle block ordering, systemd service creation for backend APIs, and end-to-end verification. USE WHEN: "add Caddy route", "proxy to localhost", "deploy behind Caddy", "new subpath on arif-fazil.com", "Caddy config".

Multi-repo git hygiene pipeline for the arifOS federation — test, stage, commit, push across all 8 repos with per-repo quirks handling. Load when Arif asks to 'git zen', 'clean up repos', 'commit everything', 'push all repos', 'dirty repos', 'git status across federation', or any multi-repo git maintenance task.

Audit a directory tree for stale, duplicate, and bloated entries. Classify as LIVE/STALE/ORPHANED, quarantine safe targets, escalate decisions on active-but-bloated projects. Produces a review report.

Operate FLAME — the RM0 Free-Loop AI Model Engine. Hit-rate-adaptive, graceful-swap inference mesh for tools, system workers, and non-constitutional agent sub-tasks. Two-lane architecture separating FLAME tool lane (throughput) from governed agent cascade (truth). Load when using flame/free-llm CLI, wiring tools to free inference, auditing the free fleet, or mapping FLAME integration candidates across MCP tools and cron jobs.

FLAME free-loop inference mesh for system workers and tools — RM0-only, non-agentic, hit-rate-adaptive routing. Use when building or debugging tool-lane LLM inference, routing scripts through free tiers, or adding new providers to the free fleet.

Design, build, and maintain Hermes Agent cron jobs as a governed daily rhythm. Tiered architecture: human briefs (DM), alert guardians (group), weekly thinkers (DM). Covers script patterns, delivery routing, alert-only design, and the 'does this earn attention?' principle. Load when Arif asks to 'audit cron jobs', 'add a cron job', 'fix the morning brief', 'what cron jobs exist', 'cron rhythm', 'daily schedule', or any task touching Hermes cron infrastructure.

Switch kimi-code CLI models — add new model definitions, change default_model, test, and verify. Covers the config.toml schema, provider binding, model capability flags, and the Kimi Code managed OAuth provider. Load when Arif says "tukar model kimi-code", "switch kimi model", "pakai K3 untuk kimi", "update kimicode", or any task touching /root/.kimi-code/config.toml model config.

Build thin translation proxies between incompatible LLM API formats (Chat Completions ↔ Responses API, etc.). For custom providers, external routing gateways, and protocol bridges.

How Hermes uses TokenRouter — model selection, auto-routing, cost strategy, FREE tier, sovereign anchor. Load when Hermes needs to choose an LLM model or when asked about TokenRouter usage.

Manage VAULT999 append-only seal chain integrity — verification, writer hardening, incident response, IMAGE_SEAL separation, identity propagation, and recovery. Covers chain verification (hash chain walk, duplicate detection, fork detection), writer hardening (pg_advisory_xact_lock, FOR UPDATE, unique constraints), identity propagation (dual write path gap, actor_source/kernel_verdict plumbing), incident documentation (freeze → snapshot → manifest → classify), and recovery (epoch fork, quarantine, repair). Use when: Arif says 'check seal chain', 'vault integrity', 'VAULT999 chain break', 'writer hardening', 'IMAGE_SEAL', 'seq mismatch', 'prev_hash', 'chain audit', 'recovery fork', 'vault999 writer', 'identity propagation', 'actor_source', 'self_report receipts', 'anonymous receipts', 'receipts logging anonymous', or any task touching /root/.local/share/arifos/vault999/.

Standing instruction for all geoscience artifacts — panels, dossiers, maps, cross-sections. Born from Raja Ridhuan's feedback: 'tak cukup geology.' 8 hard rules ensuring real geological substance, not just epistemic tagging.

ATLAS333 — 33 human paradoxes, 7 zones, cognitive geometry. Access via arifos://atlas333/* MCP resources. Use when reasoning, making decisions, or detecting paradox tension.

Audit or design a governed agent against the 7-primitive constitutional anatomy (Identity, Skills, Tools, Memory, State, Kernel, Actuator) and the constitutional chain (000-INIT through 999-SEAL). Includes breach protocol for fault-physics handling.

Analyze institutions for intelligence — structural features, failure modes, literature framework, and federation mapping. Use when evaluating whether an institution (human or agentic) can sense, learn, adapt, remember, and prevent extraction across time.

Track orders, sales, and remaining stock from handwritten receipt images shared by Arif for his food-stall business. Extract data into clean markdown tables, reconcile across multiple LRT station locations, handle item-name corrections from user's internal naming system.

Deep research on a topic: multi-source web research → structured synthesis → deliverable. For books, tools, frameworks, domains, people, institutions — and personal names (witness-frame-aware). Load when user says 'deep research on', 'research this', 'what is the full picture on', 'give me everything on', or asks for a name/person investigation with verifiable or unverifiable anchors.

XAUUSD (Gold) technical analysis — price action, key levels, confirmation entries, risk management. User-validated methodology.

When Arif (or any sovereign voice) feeds you a status narrative with claimed numbers ("100% 200 OK", "Zero dead links", "all surfaces wired"), probe against live state before accepting. Class of work is "narrative-vs-state audit" — distinct from kernel audits (arifos-kernel-zen-audit) which probe identity/wrapper behavior. This skill probes external surface claims (web routing, MCP surfaces, agent registrations, SOT YAML, navigation maps).

Convert informal business records (handwritten receipt images, oral stock/sales updates) into structured CSV + revenue analysis. Use when the user shares receipt photos and expects compiled sales data with insights.

Daily XAUUSD gold trading signal for rakyat marhaen. One-page PDF, dark OANDA-style chart, zoomed candlesticks, buy/sell zones near current price, EMA 20/50, R:R visual, simple language. Not broker jargon.

SyedOS — Agent mode for Abang Sado Syed (@rico_ricaldo_33). Voice-first, BM masculine, XAUUSD trading signals, disciplined delivery. DM primary, group banter when tagged.

Capture session failure patterns as constitutional scars — reusable diagnostics that prevent repeat mistakes. The 'scar' is the permanent record of what went wrong, why, and how to never do it again.

MiniMax multimodal via mmx-cli — TTS, video, music, image, vision, search. Image generation is PRIMARY choice for Malay/SEA phenotype and realism-critical surfaces.

Call the Qwen Token Plan text-to-image / image-edit models (qwen-image-2.0, qwen-image-2.0-pro, wan2.7-image, wan2.7-image-pro) to generate or edit images from a text description. Activates when the user asks to draw, render, generate, or edit an image. For Malay/SEA phenotype prompts, prefer minimax-cli (MiniMax image-01).

Operate the AAA governance stack as a tagged multi-role agent. Every Hermes/OpenCode/Codex response must carry a 1-line role tag ([333-AGI Δ] / [555-ASI Ω] / [888-APEX ΦΙ]) at the top so the user can tell which role is currently in the driver's seat. The mode tag is not optional and is not a section header — it is a constitutional reflex that prevents role-mode opacity, contamination acceptance, and existential-frame failure. Use when any agent in the arifOS federation is responding to Arif, and especially when Arif expresses direction signals (identity, paralysis, 'peníng', 'I don't know what I want').

Context brief pattern for delegating structural/non-code tasks to coding agents (OpenCode, Claude Code, Codex).

Manage OpenCode ACP sessions from Hermes — spawn, delegate, attach, agent management

Delegate coding to OpenCode CLI (features, PR review).