You are a workflow orchestration expert specializing in AI a...

Prompt

You are a workflow orchestration expert specializing in AI agent systems. I am upgrading "AI Visual Director Production" v7.8.0 Release 6 (the merged, verified tree) → v8.0.0. WHAT ALREADY EXISTS (verified — the system is NOT strictly linear at step level; extend this machinery, never replace it): - rules/detailed/39_mandatory_next_step_protocol.md defines the STEP QUEUE (P0 + S1–S13 artifact-family steps), exactly one ACTIVE step, commands NEXT / REVISE [ID] / SUMMARY / SKIP TO [ID] / BATCH [IDs], and — since the approved Option C — §6 makes SIBLING BATCHING THE DEFAULT: items sharing one artifact family with identical passed dependencies ship as one step. Split conditions: divergent upstream dependencies; one sibling REVISE/ BLOCKED; reviewability cap; rule 38 CONTROLLED critic batch reduction. - rules/detailed/28_adaptive_next_protocol.md has phases P0–P6 and its own independent-sibling batching rule; 39 grants the override. - The R0–R8 gate chain in rules/detailed/35_reliability_gate_controller.md is the linear SPINE: the gate receipt shows all gates with statuses PASS|REVISE|BLOCK|NO-SHIP|N/A; N/A skips require a stated reason; and "Never bypass a failed upstream gate by producing downstream polish" is absolute. - COUPLING YOU MUST NOT BREAK: rules/detailed/36_runtime_activation_compiler.md §E FAILS compilation when no STEP QUEUE exists and names 39 as its scheduler; rules/detailed/00a_router.md makes 35/36/38/39 mandatory loads for every Production Sequence and assigns queue ownership to 39. - KNOWN DEFERRED WORK: 28/39 dual ownership of NEXT semantics is a logged consolidation candidate for a LATER release. Do NOT reopen it. If your design requires it, you must include the full rewire plan (port the queue schema to 28, demote 39 to provenance, rewire 00a lines 3/79, 00b:142, 36:143, 28:105, rewrite 36 §E) or design around it. THE REAL GAP: gate-level parallelism, conditional skips, and feedback edges exist only as prose (the receipt's "Frozen downstream dependencies" line); they are not machine-evaluable, not serialized into handoff packets, and have no bounded feedback semantics. YOUR TASK — redesign the GATE layer as a conditional DAG with explicit parallel execution semantics, layered ON TOP of the existing step-level queue (the two granularities must compose, not compete). Produce: 1. Complete DAG definition as adjacency list + node metadata (JSON) for the real gates: 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) 2. For each node: dependencies, conditional skip rules (N/A with reason), parallel siblings, feedback edges, required approvals 3. Known parallelism to encode: R3 and R4 develop concurrently after R2; R5 begins during R4 when audio assets exist; R6 drift failures trigger targeted R4 revision, not R2 restart; motion-graphics projects skip R4 character sheets (N/A with reason); client approval sub-gates run parallel to technical QA 4. Conditional logic: machine-evaluable predicates over production-state alone (project_type, asset_availability, failure_type) 5. Feedback edges: triggers, bounded iteration limits, what state is preserved vs reset, cycle prevention 6. Parallel protocol: shared-state coordination, conflict resolution, merge semantics — no race conditions on shared assets 7. Serialization: how the DAG embeds in handoff packets (rule 19) so a receiving agent reconstructs execution state 8. Deadlock prevention: formal guarantee that no valid configuration creates unresolvable waits BINDING CONSTRAINTS: - Must remain a DAG; feedback edges have bounded iteration limits. - Skip conditions evaluable from production-state alone. - Preserve every v7.8.0 compliance guarantee: no downstream polish bypassing upstream failure (35 §3); a gate PASSes only from the required artifact, never from confidence (35 §7). - Pure data-driven traversal — no hardcoded gate logic in agents. - HARD FACTS as gate-node metadata: 720p native ceiling with R0/R1 mandatory resolution flags; three-edit cap; two-repair budget with final slot reserved. - RULE 03: no stale spec tables; skip/branch predicates never hardcode model specs. - EVIDENCE HONESTY: X14 — no score-based branching without evidence; promotionEligible stays false. - PRIME DIRECTIVE: YOU DO NOT BUILD FROM SCRATCH, YOU UPDATE THE ALREADY BUILT. FILES YOUR ANSWER MUST EXTEND (name each edit explicitly): rules/detailed/35_reliability_gate_controller.md, rules/detailed/39_mandatory_next_step_protocol.md (§2, §3, §6), rules/detailed/28_adaptive_next_protocol.md, rules/detailed/36_runtime_activation_compiler.md (§E), rules/detailed/00a_router.md, rules/detailed/00b_output_contracts.md, rules/detailed/19_production_memory.md, dev/run_v780_release6_tests.py (the scheduler-agreement check asserting 36/39/00a agreement). OUTPUT SHAPE: 1. EDITS per named file (exact section + replacement text). 2. NEW FILES only where justified (the DAG JSON itself is expected). 3. REGRESSION CHECKS continuing the catalog at R651+. 4. FORMAT: recommendation → tradeoffs → failure modes → what to cut → named integration points. Include BOTH a human-readable diagram description AND machine-readable JSON. Do NOT ask clarifying questions. Define the complete DAG architecture.