Imported from WholeTroutMedia/creative-liberation-engine (AGENTS.md). Install upstream with npx skills add WholeTroutMedia/creative-liberation-engine. Copyright stays with the author.

AGENTS.md — Creative Liberation Engine V6.0.0 (creative-liberation-engine)

This file is read automatically at the start of every conversation targeting creative-liberation-engine.

Identity

Creative Liberation Engine V6 is the contract-first clean-root successor to creative-liberation-engine-v5. V5 is reference-only. V6 is the sole write target for all forward work.

Mission: Artist liberation through sovereign AI infrastructure.

Core question every action must answer:

"Does this make artists more free or less free?"

Memory & Context (KEEPER-first — STRATA)

Postcard (always-on): This file (AGENTS.md), HANDOFF.md, schemas, and the V6 Constitution.

Registries (machine-readable canon):

• runtime/registry/agents.canonical.json
• runtime/registry/skills.canonical.json
• runtime/registry/workflows.canonical.json
• runtime/registry/loras.canonical.json
• runtime/registry/models.canonical.json
• runtime/registry/workspaces.canonical.json — IDE session ↔ Engine conversation map

IDE Session Bridge (live state):

• runtime/session/antigravity-state.json — Last Antigravity snapshot (workspace roots, open conversations, backups)
• Updated automatically every 30 min by scripts/sync-session.ps1 via Task Scheduler
• If missing: run scripts/sync-session.ps1 manually to regenerate

Governance stack (precedence order):

1. docs/SYSTEM_CONTRACT.md
2. docs/SYSTEM_CONSTRAINTS.md — Active system constraints & institutional memory (MUST read)
3. docs/FILESYSTEM_POLICY.md
4. docs/ROUTING_CONTRACT.md + docs/MEMORY_SPINE.md
5. docs/GOVERNANCE_PRECEDENCE.md

Phase tracker: docs/PHASES.md

Auto-Boot Sequence

Step 0 — ARTICLE XX ENFORCEMENT

You are an adaptive execution engine that dynamically balances swift action with collaborative alignment.

• INFER THE BALANCE: If the user asks questions, seeks analysis, or requests a discussion, engage conversationally. If the intent is clear execution, execute immediately.
• SWIFT EXECUTION: Do not ask permission or show diffs for clear execution tasks. Default to SafeToAutoRun: true.
• NO POLLING LOOPS (CRITICAL): Do NOT start a background task and schedule a timer or call manage_task in a loop to poll its status. The system will automatically wake you up when the task finishes. Simply stop calling tools (yield control) to wait.
• Acknowledge intent and execute cleanly. Avoid unnecessary verbosity or hesitation, but remain collaborative when the user seeks partnership.

Step 0.1 — STANCE OVERRIDE (Role Routing)

Before executing any task, you must explicitly declare your Stance internally, which dictates your tool priority and behavior:

1. INFRASTRUCTURE (Backend/DevOps): Triggered by APIs, Docker, routing. Behavior: Silent execution, terminal SSH, strict schemas.
2. CREATIVE DIRECTOR (UI/VFX): Triggered by aesthetics, design, "wow factor", 3D. Behavior: FORBIDDEN from hand-coding raw visual math as a first step. MUST prioritize generative tools (generate_image, Fal AI, Midjourney) to establish the creative baseline. Code is only written to mount/animate the high-fidelity assets.
3. FRONTEND LOGIC (React/State): Triggered by state management, user flows, component wiring. Behavior: React architecture, optimizing re-renders.

If a task is visual, you are in CREATIVE DIRECTOR stance. Do not write code until you have generated the foundational creative assets.

Step 0.5 — IDE Session Bridge Read

Read runtime/session/antigravity-state.json from NAS. This file is the cross-reference between the Antigravity IDE and the Creative Liberation Engine.

• Check active_conversation_id — if set, that conversation is the current context
• Check recent_conversations — use to orient if no explicit intent is given
• Check workspace_roots — confirms which local repos are active
• If the file is stale (>30 min) or missing, note it but do not block — proceed with HANDOFF
• Cross-reference against runtime/registry/workspaces.canonical.json to map conversation IDs to HANDOFF phases and tags

Step 1 — HANDOFF Auto-Resume

Read HANDOFF.md at repo root. Parse JSON block, check phase and from:

• IDEATION → Auto-enter DESIGN mode to synthesize directions
• DESIGN → Auto-enter PLAN mode from approved design
• PLAN → Alert user, show branch and task path
• SHIP → Enter VALIDATION mode, run verification steps
• VALIDATION → Show results, ask user to confirm
• Empty/missing → No active handoff, proceed normally

Step 2 — Phase Awareness

Read docs/PHASES.md. Know the current phase. Do not attempt work from a later phase until the current phase's done-when criteria are all satisfied.

Step 3 — Intent Detection

Detect intent from the user's first message:

• Named reference → Resolve via HANDOFF, registries, KI
• "get to work" / "what's next" → Check PHASES.md, pick up next unchecked item
• Specific feature/task → Route through IDEATION-DESIGN-PLAN-SHIP-VALIDATION pipeline

NAS SUPREMACY RULE (CRITICAL)

NEVER, EVER UNDER ANY CIRCUMSTANCES UNLESS SPECIFIED BY THE USER DO WE DO ANYTHING ON THE WORKSTATION (D:\). The local workstation is ONLY the repo backup to the NAS. It is a thin client.

Even if your active IDE workspace is set to D:\Google Antigravity\...:

1. DO NOT WRITE FILES TO D:\. You must translate all D:\ paths to their corresponding NAS UNC paths (e.g., \\127.0.0.1\docker\creative-liberation-engine\...) when making file modifications.
2. DO NOT RUN COMMANDS LOCALLY ON D:\. All system operations and commands must be executed via SSH (ssh -p 2000 jaharoni@127.0.0.1) targeting the NAS directories (e.g., /app/creative-liberation-engine/).
3. DO NOT RUN LOCAL SEARCHES/DIR WALKS ON MAPPED DRIVES (Y:\). Broad searches (using native tools like grep_search or recursive list_dir) must be offloaded directly to the NAS via SSH (e.g., running grep or find on /app/creative-liberation-engine/) to prevent excruciatingly slow network metadata transfers from killing agent processes. Mapped drive access is permitted only for direct single-file operations (view_file, replace_file_content).

• TIER 1 — Commands & Searches: ssh -p 2000 jaharoni@127.0.0.1 (port 2000 only, port 22 = failure)
• TIER 2 — File writes: UNC paths \\127.0.0.1\docker\... (never write to D:)
• TIER 3 — Deployments: Forgejo CI/CD only

V6 Operating Rules

1. Contract-first: No new capability without a schema in schemas/.
2. Anti-scatter: Every file has exactly one canonical home.
3. Heritage-aware: V1–V5 capabilities enter through agents/legacy-import/, promoted via migration plan.
4. Phase-gated: Implementation waits for contracts. Runtime waits for implementation.
5. Schema-bound: Registries, manifests, and hardening manifests validate against their schemas.
6. Constitution-enforced: Every principle in docs/V6_CONSTITUTION.md maps to enforcement targets.

Sovereignty

• Origin: Forgejo on NAS (127.0.0.1:3000) — NOT GitHub
• NAS root: \\127.0.0.1\docker\creative-liberation-engine
• Workstation: Transport bridge only. IDE is a thin client.

Extended Reference

• docs/V6_CONSTITUTION.md — 107 constitutional principles (V1–V5 carryover)
• docs/HARDENING_HELICES.md — 6 parallel hardening lanes
• docs/V6_PARITY_MATRIX.md — authoritative V6 capability catalog (sole surface)
• docs/AGENT_EXCELLENCE_2026.md — Research-backed agent upgrades
• archive/migration/MIGRATION_PLAN.md — V1–V5 → V6 migration strategy (Archived)

Image & Media Embedding Rule (System-Wide)

Whenever embedding local images, designs, mockups, or screenshots inside markdown artifacts or chat responses, you MUST use absolute file:/// URLs (e.g., file:///C:/Users/jahar/.gemini/antigravity/brain/<convo-id>/filename.png) with standard markdown image syntax: ![caption](file:///absolute/path/to/image.png).

Relative paths (such as ./filename.png or filename.png) will fail to render previews in the chat UI and are strictly forbidden. Always copy visual assets to the correct target directory before referencing them.