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 Phase Contract Audit

Risk addressed: the in-process contract between pipeline phases is argparse.Namespace. There are 426 `getattr(args, ...)` call sites across 16 modules. PipelineRequest.build_scratchpad_namespace (pipeline_request.py:332) hand-builds a Namespace with ~30 fields just to pass typed data between in-process stages. getattr(args, "field", None) swallows typos, so renames silently zero out values. New fields require parallel updates in PIPELINE_MCP_PROPERTIES, PipelineRequest, from_mapping, build_scratchpad_namespace, the CLI builder, and per-phase getattr calls. Tests pass argparse.Namespace() so phantom fields slip through.

PlanDraft and GenerationScratchpad dataclasses exist but _run_scratchpad operates on the dict payload and converts at the end via scratchpad_from_dict. The _filter_fields helper (pipeline_types.py:386) silently drops unknown keys.

What this skill produces

{
  "summary": {
    "total_getattr_sites": 426,
    "phantom_fields": 17,
    "phases_with_drift": 4,
    "fields_declared_but_unused": 8,
    "namespace_fields_total": 73,
    "typed_request_fields_total": 41
  },
  "phases": [
    {
      "phase": "route",
      "consumed_fields": ["route_brief", "material_type", "render_backend", ...],
      "declared_in": ["PipelineRequest"],
      "phantom_in_phase": [],
      "untyped_via_namespace_only": ["route_force"]
    },
    {
      "phase": "plan",
      "consumed_fields": [...],
      "declared_in": ["MaterialPlan", "PipelineRequest"],
      "phantom_in_phase": ["legacy_inspiration_seed"],
      "untyped_via_namespace_only": ["accept_inspiration_recommendations"]
    }
    /* ... per-phase ... */
  ],
  "phantom_fields_detail": [
    {
      "field": "legacy_inspiration_seed",
      "read_at": ["plan_builder.py:412"],
      "never_written_at": "no PipelineRequest entry, no Namespace builder line",
      "default_returned": "None",
      "blast_radius": "silent — production has never set this; renames break nothing but the field is dead"
    }
  ],
  "filter_fields_drift": [
    {
      "field": "branch_id",
      "declared_in": "MaterialPlan, GenerationScratchpad",
      "missing_from": "scratchpad_from_dict input shape",
      "behavior": "silently dropped at deserialization"
    }
  ],
  "duplicate_field_routing": [
    {
      "field": "design_variance",
      "paths": ["PipelineRequest.design_variance", "Namespace.design_variance", "scratchpad payload.design_variance"],
      "consistency": "ok"
    }
  ]
}

When to use

• Before adding a parameter to bgen orchestrate-material — to know how many sites need a parallel update.
• When a CLI flag works, an MCP call works, but the in-Python Namespace path silently no-ops.
• When _filter_fields drops a field you expected to survive.
• Before any refactor of pipeline_runner._scratchpad_args (pipeline_runner.py:806).
• To validate the typed-runtime promise that 45 MCP verbs hold end-to-end.

Inputs

• Repo path (default $PWD).
• Optional --phase <name> — restrict audit to one phase. Otherwise audits all six.

Procedure

1. Build the typed-field universe. Read brand_gen/pipeline_types.py and brand_gen/pipeline_request.py. Capture every dataclass field name from PipelineRequest, RoutingBrief, PlanDraft, MaterialPlan, GenerationScratchpad, PipelineResult, plus any nested types. Build a set keyed by (class, field_name).

1. Build the Namespace-field universe. Read PipelineRequest.build_scratchpad_namespace (pipeline_request.py:332) and pipeline_runner._scratchpad_args (pipeline_runner.py:806). Capture every attribute set on argparse.Namespace. These are "fields the pipeline says exist."

1. Build the consumed-field universe. Grep brand_gen/ for getattr(args, "([^"]+)". Group results by enclosing function/file. Tag each function with its phase using these mappings:

• _run_route / route_brief / classify_workflow_route_smart → route
• create_material_plan / _run_plan / plan_builder.py functions → plan
• validate_* / _run_critique → validate
• assemble_generation_scratchpad / execute_generation_scratchpad / _run_scratchpad / _run_generate → execute
• apply_generation_critique_policy / vlm_critique / auto_capture_* → review
• evolve_* / promote_* → evolve

1. Cross-reference to find phantoms. For each getattr(args, X), check whether X is in the typed-field universe AND in the Namespace-field universe. A phantom is a field consumed via getattr but absent from BOTH writers and ALL declared dataclasses. Mark these with their default-fallback so the user knows whether the silent-zero behavior matters.

1. Find dead declarations. For each typed dataclass field, check whether any getattr(args, field) references it. Fields declared but never read are dead.

1. Audit `_filter_fields` drift. Read pipeline_types.py:386 (_filter_fields) and every caller of from_dict (scratchpad_from_dict, plan_from_dict, etc.). For each typed dataclass, check whether the from_dict reconciliation walks the same field set as the dataclass declares. Catch nested-key fallbacks (pipeline_types.py:445-451).

1. Detect duplicate routing. Some fields traverse multiple paths (CLI arg → Namespace → request → scratchpad payload). For each, walk the chain and verify name consistency.

1. Rank blast radius. Phantom fields rank highest if read in the execute phase (live prompt content). Drift in _filter_fields ranks high for any persisted field. Dead declarations are low-priority.

Output schema

See "What this skill produces" above. Also emit a per-phase markdown table for human review:

| Phase | Consumed | Typed | Phantom | Drift |
|-------|----------|-------|---------|-------|
| plan  | 47       | 41    | 3       | 1     |

Reference files

• brand_gen/pipeline_request.py:332 — build_scratchpad_namespace (the Namespace contract).
• brand_gen/pipeline_runner.py:806 — _scratchpad_args (the second Namespace builder).
• brand_gen/pipeline_types.py:386, 445-451 — _filter_fields and nested fallback reconciliation.
• brand_gen/cli_builders.py — CLI args that flow into the Namespace.
• brand_gen/commands/generation.py — wires CLI to the runner.
• Project memory note: "New PipelineRunner params need 3 updates in brand_iterate_mcp.py: schema, Namespace, constructor." — verify the audit catches drift across these three sites.

Don't

• Don't fix anything. Output the matrix; let compound-engineering:ce-plan design the migration to typed phase IO.
• Don't grep blindly: getattr(obj, ...) is used elsewhere in the codebase. Filter to getattr(args, ...) and to functions tagged with a pipeline phase.
• Don't include test files in the consumed-field universe — tests pass throwaway Namespaces and would inflate the phantom count.