geo-pipeline — Codex orchestrator for recomby-geo

On Claude Code each stage is a bare slash command (/01-intake … /07-reaudit). Codex CLI has no bare custom commands, so this skill is the Codex entry point: it carries the orchestration rules and routes to the per-stage specification files, which are the single source of truth shared with the Claude Code side. Do not duplicate stage logic here — read the stage file and follow it.

Per-stage specs (read the one you're running): commands/01-intake.md, commands/02-audit.md, commands/03-gap.md, commands/04-content-brief.md, commands/05-production.md, commands/06-distribution.md, commands/07-reaudit.md (relative to this plugin's root). Full directory convention + dependency graph: orchestrator/run.md.

How to run a stage on Codex

1. Identify the client folder clients/<slug>/ and the stage the user wants.

2. Read the matching commands/0X-*.md and execute its Procedure verbatim.

3. Validate every JSON artifact the stage writes against its schema before moving on (Codex has no built-in schema validation — run it explicitly):

   python3 - <<'PY'
   import json, jsonschema
   pairs = {
     "brand_context.json": "brand_context.schema.json",
     "visibility_baseline.json": "visibility_baseline.schema.json",
     "content_priorities.json": "content_priorities.schema.json",
   }  # see schemas/ for the full set incl. attribution_diff + review_feedback
   # jsonschema.Draft202012Validator(json.load(open("plugins/recomby-geo/schemas/<file>"))).validate(json.load(open("clients/<slug>/<artifact>")))
   print("validate each artifact against plugins/recomby-geo/schemas/*.schema.json")
   PY
   

4. Stop at the human-in-loop points (see rules below).

Stage order (never run out of order)

inputs/ → 01-intake → brand_context.json
        → 02-audit  → visibility_baseline.json
        → 03-gap    → content_priorities.json
        → 04-content-brief → briefs/<id>.md (+ .html, REQUIRED-FILL slots)
            [EXPERT FILLS THE SLOTS — not the AI]
          04 Step 9 verifies fills → status: ready-for-production
        → 05-production → drafts/<id>.md (+ review .html)
        → 06-distribution → distribution/<id>.json + publish-bundle.md
            [PUBLISH + WAIT 7+ days]
        → 07-reaudit (monthly) → reaudit/round-N.json → feeds next 03-gap

02 needs 01; 03 needs 01+02; 04 needs 01+03; 05 needs 04 (filled); 06 needs 05; 07 needs a prior 02.

Hard rules (orchestration level)

1. The expert fills briefs — not the AI. 05-production refuses to run unless briefs/<id>.meta.json status is ready-for-production. Never auto-fill REQUIRED-FILL slots; pause the pipeline if the expert is unavailable. This human-in-loop checkpoint is the entire moat.
2. Schema validation is enforced. If an artifact fails its schema, fix before moving on. Schemas: schemas/*.schema.json.
3. One client per folder. Don't share JSON across clients/<slug>/ folders or factor out "common" context.
4. Never run a stage out of order (see graph above).

Client review HTML

Stages 04 and 05 produce interactive HTML for the client via the geo-review-html skill (also in this plugin). On Codex this works the same as on Claude Code — the stage spec already calls render_html.py.

Requirements

The capability skills' scripts (e.g. seo-geo-optimizer, geo-review-html) need python3 on PATH; the schema validation step needs jsonschema (pip install jsonschema). These are the same dependencies as the Claude Code side.