feat(solicitation-create): work-order-as-input + canonical labs schema + comprehensive content shape

[jjackson]

Summary

Three bundled rewrites of `skills/solicitation-create/SKILL.md` prompted by solicitation 3130 on `malaria-itn-app/20260521-1400` where the public page rendered blank Description, "TBD" timeline, "No deadline," Python-list-repr Scope, and zero questions / zero rubric simultaneously.

1. Work-order-as-primary-input

The skill now reads Phase 1's work order (`1-design/pdd-to-work-order.gdoc`) as the comprehensive content source, plus `decisions.yaml` for any later run-level decisions, alongside the PDD (now used only for problem-framing). The work order is the opinionated program brief; this skill transforms it into a public-facing solicitation — same comprehensive explanation, less prescriptive (rates → ranges, exact weeks → windows).

2. Canonical labs schema

Field names now match `solicitations/models.py`'s @Property accessors. Adding fields the template doesn't read silently does nothing (the labs API echoes any extra keys back without surfacing the drift). Migrated:

• `description` (not `overview`)
• `application_deadline` date string (not `response_window_days` int)
• `expected_start_date` / `expected_end_date` (not `anticipated_*`)
• `estimated_scale` (not `sample_target`)
• `questions[].text` (not `response_questions[].question`)
• `evaluation_criteria[].name/.scoring_guide/.linked_questions` (not `rubric[].dimension/.criterion`)
• `solicitation_type: 'eoi'` lowercase

Removed top-level fields not read by the public-detail template (`pass_bar`, `eligibility_criteria`, `geographic_scope`, `per_hh_payment_band_usd`, `budget`) — their content folds into `description` / `scope_of_work` prose instead.

3. Comprehensive content shape

• `description`: 500-800 words, foundation-pitch tone (problem framing → what this opp does → what the dataset enables). Not a one-paragraph summary.
• `scope_of_work`: 600-1000+ words of structured markdown, derived section-by-section from the work order. Explicit de-prescription rules (exact dollars → ranges, exact weeks → windows). Must be a single markdown string, not an array.
• Every question has a required `framing` field (1-2 sentence "why we're asking") + `text` (the actual prompt).
• Every evaluation criterion has a required `scoring_guide` (concrete strong/mid/weak/zero signals) + `linked_questions` (at least one question id). Weights are integers summing to 100.

4. New Step 7a — public-page structural verifier

After publish, curl the public URL and grep for the rendered Description body, parseable Application Deadline, non-TBD Timeline, markdown-bulleted Scope, and the question + criterion count strings. Any miss → `[BLOCKER]` naming the probable field-name drift. Catches the silent-echo failure mode where the labs API accepts the create cleanly but the public page is empty.

Removal criteria

Keep until the labs MCP ships `create_solicitation_from_brief` (server-side composition via labs's existing `commcare_connect/ai/agents/solicitation_agent.py`). At that point this skill collapses to ~30 lines passing the structured brief.

Test plan

• Next `/ace:run` against an opp publishes a solicitation whose public page renders all 5 sections cleanly.
• `description` body is ≥300 words and reads as foundation-pitch, not procurement-form.
• `scope_of_work` renders as markdown bullets / sub-headings on the public page (not Python repr).
• Step 7a structural verifier halts cleanly with named-section diagnostic when a future drift recurs.

🤖 Generated with Claude Code