brand-gen-loop-shape-assessment

For Sage brand work in Pi, use the paste-ready prompt at docs/prompts/pi-sage-brand-gen-full-pipeline.md. Keep this link instead of copying the full prompt into skill bodies.

Brand-Gen Loop-Shape Assessment

Risk addressed: the route → plan → critique → scratchpad → generate flow is a sequential mutation pipeline, not a feedback graph. Each phase reads the previous JSON file and writes a new one. Critique is in-line and one-shot (pipeline_runner._run_critique); it cannot iterate the plan. Auto-feedback writes to iteration_memory which is read at next-run prompt assembly time. There is no in-flight "the critique fixed the plan" loop except for legacy VLM. max_iterations re-runs only _run_generate, not the planning phases. PipelineResult carries a single artifact set, not a candidate set.

The whole system models taste as a one-pass refinement — but the actual problem (aesthetic convergence) requires loops at the plan layer.

What this skill produces

{
  "summary": {
    "phases": ["route", "plan", "validate", "execute", "review", "evolve"],
    "phase_can_reenter_predecessor": false,
    "pipeline_result_carries_candidate_set": false,
    "scoring_feeds_into_route": false,
    "scoring_feeds_into_plan_in_same_run": false,
    "max_iterations_scope": "execute_only",
    "in_run_loops": [
      { "name": "max_iterations", "scope": "execute → review → execute", "controlled_by": "iteration count" }
    ],
    "across_run_loops": [
      { "name": "iteration_memory feedback", "scope": "next run's prompt assembly reads prior negatives" },
      { "name": "learnings promotion", "scope": "evolve writes to learnings.json; future runs prepare-phase reads it" }
    ]
  },
  "actual_state_machine": "(mermaid or graph dsl describing today's flow)",
  "desired_state_machine": "(graph showing plan↻ on critique block, candidate-set carried, scoring feedback into plan)",
  "delta": [
    {
      "transition": "validate.blocking → plan",
      "today": "blocking_findings → stop_reason; user must re-invoke orchestrate-material",
      "desired": "in-run plan revision driven by blocking findings",
      "minimum_change": "PipelineRunner._run_critique returns plan_revisions; runner re-enters _run_plan"
    },
    {
      "transition": "review.iterate → plan",
      "today": "iteration_memory mutated; loop exits; next run reads memory",
      "desired": "in-run plan revision driven by review.before_after_diffs",
      "minimum_change": "max_iterations covers plan re-entry; PipelineResult carries iteration_history"
    },
    {
      "transition": "execute → multiple_candidates",
      "today": "PipelineResult.artifacts is a single set; multi-candidate work is multiple runs",
      "desired": "PipelineResult.candidates: List[CandidateResult] when an Experiment is active",
      "minimum_change": "AestheticExperiment from brand-gen-experiment-modeling + List wrapper on artifacts"
    }
  ],
  "non_loops": [
    "scoring → route (would let the system reroute when a scored direction underperforms; doesn't exist)",
    "scoring → plan (in-run; would close the convergence loop; doesn't exist)"
  ],
  "blast_radius": {
    "introducing_plan_loop": "max_iterations semantics, PipelineResult shape, run_ledger schema",
    "introducing_candidate_set": "PipelineResult.artifacts, every consumer of artifacts (UI, MCP, brand-orchestrator)"
  }
}