You are a principal engineer designing tool interfaces for a...

Prompt

You are a principal engineer designing tool interfaces for autonomous AI agent systems. I am upgrading "AI Visual Director Production" v7.8.0 Release 6 (the merged, verified tree) → v8.0.0. The skill currently DESCRIBES production steps; agents narrate gates instead of executing them through formal interfaces. LOCKED PRODUCT DECISIONS (binding — violations invalidate the answer): - The package is a PROMPT-ONLY controller. Its product is copy-paste prompts/instructions for exactly two locked models: Nano Banana Pro (still images) and Gemini Omni Flash (video + edit prompts). - Replicate, ComfyUI, Runway, local SDXL, and any other generation backend are BANNED from this answer. Generation happens through prompts to the two locked models — nothing else. - Tool contracts therefore live on the ORCHESTRATION side: wrapping the package's own validators, artifact/file operations under the AVD marker protocol, and the submit/collect cycle for the two models' prompts. - Distributions: portable-core ships NO Python (max assurance PLAN-PASS under PORTABLE_REFERENCE); validation-sdk requires portable-core; combined and cumulative can execute validators (VERIFIED_RUNTIME). The manual fallback path is a FIRST-CLASS citizen of every contract, not an error case: the same contract must work whether a harness executes the validators or an operator performs them by hand under PORTABLE_REFERENCE. CORRECT GATE SEMANTICS (from the rule 35 §4 gate receipt — the earlier "concept divergence"/"reference calibration" labels are wrong): R0 Scope/rights/capability · R1 Brief/mode/taste · R2 Originality · R3 Causality/mechanism · R4 Assets/model sheets · R5 Board/timing/feasibility · R6 Execution channels · R7 Sound/edit/delivery · R8 Evidence QA. WHAT ALREADY EXISTS (verified — extend, never replace): - routing/HOST_CAPABILITY_ROUTING.json and routing/ACTIVE_MODULE_MANIFEST.json (host capability routing), bootstrap/probe_host.py and bootstrap/select_host_profile.py (capability probes). - validators/: compile_and_validate.py, validate_project.py, validate_instruction_compliance.py, validate_runtime_load.py, check_guide_staleness.py, extract_actual_output.py — deterministic checks. - adapters/generation/NANO_BANANA_PRO.md (incl. NBP-F1..F4 variation features), GEMINI_OMNI_FLASH.md, NANO_BANANA_PRO_OMNI_FLASH_PIPELINE.md, and adapters/generation/modes/MODE_A..D per model. - Rule 35 §6 cost-control: an 8-item cheapest-useful-diagnostic catalog (concept rewrite, logic map, style tile, model sheet, neutral keyframe, short motion proof, scratch VO timing, editor mockup) — your cost tiers attach to THIS catalog. YOUR TASK — design the complete Tool Integration Contract layer using Model Context Protocol (MCP). Produce: 1. MCP tool definitions (JSON) for the gate-critical tools, mapped to the real gates: R2 originality verification (Nano Banana Pro prompt issue + similarity comparison against approved references); R3 causality/ mechanism check (logic-map validation); R4 model-sheet locking (anchor generation + drift detection); R5 timing/beat validation (scratch VO timing + EDL parsing); R6 generation execution with durable-resolution proof (Omni Flash prompt issue + frame consistency check); R7 edit assembly verification (timeline parsing + transition validation on the platform-neutral editing surface) 2. For each tool: exact input schema, output schema, error codes, retry semantics, idempotency guarantees 3. A meta-tool validate_gate_evidence accepting a gate receipt + schema reference, returning structured pass/fail with violation details — wrapping the existing validators, not duplicating them 4. Fallback protocol: the manual/descriptive path an agent must take when a tool is unavailable (PORTABLE_REFERENCE behavior, honest labels, cap at PLAN-PASS) 5. Tool discovery contract: how an agent enumerates available tools and capabilities at runtime, integrated with HOST_CAPABILITY_ROUTING.json 6. Security boundary: which tools require human approval vs autonomous execution, and how that is encoded BINDING CONSTRAINTS: - Follow the MCP specification strictly (tools/resources/prompts separation). - Inputs/outputs JSON-serializable; no binary blobs (use URIs/hashes). - Every tool declares a resource cost tier (LOW/MED/HIGH) aligned with the rule 35 §6 diagnostic budget. - Error responses are machine-actionable, not prose. - HARD FACTS encoded in every video-touching contract: 720p native ceiling (above-720p requests are upscaled platform-neutral finishing handoffs flagged at R0/R1); at most three sequential conversational edits per clip; budget two one-variable repairs, final slot reserved. - RULE 03: no stale spec tables; no hardcoded model specs in any contract. - EVIDENCE HONESTY: X14 — reject scores without evidence; no invented metrics; a tool PASS is deterministic/structural, never live proof. - PRIME DIRECTIVE: YOU DO NOT BUILD FROM SCRATCH, YOU UPDATE THE ALREADY BUILT. FILES YOUR ANSWER MUST EXTEND (name each edit explicitly): routing/HOST_CAPABILITY_ROUTING.json, routing/ACTIVE_MODULE_MANIFEST.json, bootstrap/probe_host.py, bootstrap/select_host_profile.py, validators/compile_and_validate.py, validators/validate_project.py, validators/validate_instruction_compliance.py, validators/validate_runtime_load.py, adapters/generation/NANO_BANANA_PRO.md, adapters/generation/GEMINI_OMNI_FLASH.md, adapters/generation/NANO_BANANA_PRO_OMNI_FLASH_PIPELINE.md, rules/detailed/35_reliability_gate_controller.md (§4, §6), rules/RUNTIME.md (§6a, §6b), profiles/PORTABLE_REFERENCE.md, dev/run_v780_release6_tests.py. OUTPUT SHAPE: 1. EDITS per named file (exact section + replacement text). 2. NEW FILES only where justified, with justification. 3. REGRESSION CHECKS continuing the catalog at R651+. 4. FORMAT: recommendation → tradeoffs → failure modes → what to cut → named integration points. Do NOT ask clarifying questions. Make definitive design choices. Output complete MCP JSON definitions ready for implementation.