{
  "version": 3,
  "created_at": "2026-07-10T22:17:25.544Z",
  "task_summary": "Build a disease-agnostic drug-program orchestration skill + specialist agent",
  "agents": [],
  "phases": [
    {
      "name": "Plan",
      "delegations": [
        {
          "steps": [
            {
              "title": "Extract the EoE arc into a reusable stage spec",
              "description": "Parse the Project Tolera plan JSONs and key deliverables (business_plan.md, scientific_plan.md, pmhc_landscape_report, PFDD/patient_voice, omics/target files) already in this project to derive the canonical 10-stage sequence, the inputs/outputs of each stage, and every point where a human decision was actually required. Produce `stage_spec.md` (the generalized pipeline blueprint) and `checkpoint_map.csv` (stage → gate question → what human input resolves it → go/no-go criteria). This grounds the skill in what actually worked, not a hypothetical flow."
            },
            {
              "title": "Design the checkpoint + program-dossier model",
              "description": "Define the human-in-the-loop protocol: which decisions Claude Science makes autonomously vs. must surface via ask_user (target selection, modality choice, go/no-go gates, spend/financing assumptions, any claim needing clinical validation), and how confidence/uncertainty triggers a checkpoint. Define the `program_dossier.json` state object that threads through all stages (disease, decisions made, artifacts produced per stage, open questions, gate status). Write this up as `checkpoint_protocol.md` and a `program_dossier.schema.json`."
            },
            {
              "title": "Author the drug-program-orchestration skill (SKILL.md)",
              "description": "Write the conductor SKILL.md: purpose, when-to-trigger, the 10 stages each with (goal, which existing skill/connector to load, inputs, outputs, gate criteria, human-checkpoint), disease-agnostic guidance, and the dossier-threading protocol. Explicitly reference sub-skills (indication-dossier, patient-centered-market-and-survey, systematic-review-orchestration, omics-target-mining, the protein-design skills, figure-composer, etc.) so the conductor composes rather than reimplements. Include the two gap-stages (target/modality selection rubric; scientific-plan assembly template)."
            },
            {
              "title": "Ship kernel.py helpers + templates",
              "description": "Author `kernel.py` at the skill root with helper functions: dossier init/load/save/update, a stage-status tracker, a checkpoint-formatter (turns a gate into a structured ask_user call), and a deliverable-manifest builder. Add `assets/` templates: scientific_plan_template.md, program_summary_template.md, and the target/modality scoring rubric CSV. Functions only (no top-level classes/decorators) so the sidecar auto-loads."
            },
            {
              "title": "Dry-run validate the stage map on a fresh disease",
              "description": "Sanity-check the pipeline by running its planning logic (not full compute) against a different indication than EoE to confirm it is genuinely disease-agnostic: resolve stage 1–5 inputs, confirm each references the right skill/connector, and confirm checkpoints fire at the right places. Produce `dryrun_<disease>.md` showing the stage map and checkpoint schedule the orchestrator would present. Catches EoE-specific assumptions leaking into the generalized flow."
            },
            {
              "title": "Publish the skill and create the specialist agent",
              "description": "Publish drug-program-orchestration via host.skills. Then (after confirming full-access vs restricted skill set with you) create a specialist agent profile via host.agents.create that preloads the orchestration skill and names the relevant data/regulatory connectors, with instructions embedding the human-checkpoint philosophy. Provide the switch instructions. Deliver a short README (`orchestration_readme.md`) explaining how to invoke it (‘run the program pipeline for <disease>’) and how the checkpoints work."
            }
          ]
        }
      ],
      "id": "phase-0"
    }
  ],
  "feasibility": {
    "rationale": "Nearly every pipeline stage already exists as a catalog skill and is proven by the EoE/Project Tolera work in this project. The build is primarily an orchestration layer (conductor skill) plus a specialist agent profile that preloads it — both are well-supported by the customize SDK. The genuinely new logic is stage-gating, human-checkpoint definition, program-dossier state, and two gap stages (target/modality selection, scientific-plan assembly), which are authored as prose+templates rather than heavy compute.",
    "confidence": "high"
  }
}