famlaw_v2/START_HERE.md

# START HERE — Claude Code kickoff

You are picking up an in-progress build. Read in this order, then start the task at
the bottom. Do **not** re-derive the design — it exists; follow it.

## Read order (do this first)
1. **`CLAUDE.md`** — the operational brief and single source of truth. Read it fully.
   Pay special attention to:
   - §1 — the governing rule + the two gates + Claude's role as the quality-elevation layer.
   - §2 — the 15 capabilities, grouped by RISK (what's safe now vs. attorney-gated).
   - §4 — locked decisions (apply them; do not re-litigate).
   - §5 — the verifiable-step method + the test harness rules.
   - §8 — **STEP 2, your immediate task.**
   - §10 — the audit fixes already applied (esp. fix 2, the `familylaw.proceeding` model).
   - §11 — the non-code success gaps (context for why some steps are gated).
2. **`BUILD_PLAN.md`** — the step contract and the run/test commands.
3. **The design docs `00`–`11`** (in the separate design package) — *why/detail* behind
   each step. Read the relevant doc(s) before each step, not all at once. For Step 2:
   docs 02 (domain) and 03 (data model).

## Current state
- **Step 1 is DONE and validated** — `activeblue_familylaw` module: the `familylaw.case`
  spine, the lifecycle state machine, attorney-only gates, security groups, views,
  and 10 passing tests (tag `familylaw_step1`). Module layout is in `CLAUDE.md §6`.
- The code is Odoo 18 (uses `<list>` not `<tree>`; direct attributes, no `attrs`/`states`).
- Nothing else is built yet. Steps 2–14 are designed, not implemented.

## Hard rules (from CLAUDE.md — do not violate)
- **One verifiable step at a time.** Don't scaffold ahead. Each step = installable code
  + a UI walkthrough + tagged unit tests + a done-when. Get green before advancing.
- **The two gates are sacred:** no code path lets AI output be filed/sent/signed without
  attorney approval, or lets an unverified citation into a filing. If a request would
  breach this, flag it — don't build it.
- **Mock all external APIs in tests** (Claude, DocuSeal, CourtListener). Never hit the
  network from a test. Test date math with fixed dates.
- **No outcome/success-percentage prediction.** The decision graph compares strength of
  support, never a "% chance of winning" (CLAUDE.md §2 EXCLUDED).
- **Tier 3 (capabilities 8–9, Steps 6–7–9) is gated** behind licensed-attorney
  validation of the legal logic. You may build the *structure/scaffolding* but flag that
  the legal content is unvalidated; do not present legal reasoning as authoritative.

## Your first task: STEP 2
Build exactly what `CLAUDE.md §8` specifies:
- Models: `familylaw.party`, `familylaw.child`, `familylaw.issue`, **`familylaw.proceeding`**.
- Add indexed, searchable `case_number` to `familylaw.case`; open an initial proceeding
  on case creation.
- **Automated conflict screening** — a local DB search over `familylaw.party` across all
  cases that surfaces potential conflicts for the attorney to clear (never auto-clears,
  never auto-blocks; the Step 1 `conflict_check_cleared` attorney gate remains).
- Extend case search to find by party name, child name, and `case_number`.
- **Intake questionnaire** (native models; creates a draft case): triage-first with the
  urgency screen, **conditional strictness** (standard = strict; emergency = fast-path on
  minimum facts), branching into the **child-support modification** question set, runs
  conflict screening on completion. Captures facts only — never answers "do I have a good
  case?" (logs it as an attorney question).
- Tests tagged `familylaw_step2` covering everything in §8's test list.
- **Done-when:** `familylaw_step2` green + the four-part walkthrough in §8.

Work in the module, keep Step 1's tests green, and deliver Step 2 as one reviewable,
tested slice. Then stop and report before starting Step 3.