# fusion_accounting_followup — Cursor / Claude Context

## Purpose

AI-augmented customer follow-ups (dunning) — a Fusion-native replacement
for (and coexisting with) Odoo Enterprise's `account_followup` module.
Ships in Phase 4 of the fusion_accounting roadmap.

## Architecture

Hybrid: the engine (`fusion.followup.engine`, AbstractModel) is the
SINGLE write surface for the follow-up lifecycle. Everything else
(controllers, OWL components, AI tools, wizards, cron) routes through
the engine's 7-method public API:

- `get_overdue_for_partner(partner)`
- `compute_followup_level(partner)`
- `send_followup_email(partner, level=None, force=False)`
- `escalate_to_next_level(partner)`
- `pause_followup(partner, until_date=None)`
- `reset_followup(partner)`
- `snapshot_followup_history(partner, limit=50)`

Pure-Python services live in `services/`:

- `overdue_aging` — 6 buckets (current, 1-30, 31-60, 61-90, 91-120, 120+)
- `level_resolver` — match aging to a `fusion.followup.level`
- `risk_scorer` — 0-100 payment-risk score plus structured drivers
- `tone_selector` — gentle / firm / legal based on level + risk
- `followup_text_generator` + `followup_text_prompt` — LLM-generated
  follow-up text with a templated fallback that keeps the feature
  usable offline

Persisted models in `models/`:

- `fusion.followup.level` — level definition (delay_days, tone,
  mail_template_id, requires_manual_review, sequence)
- `fusion.followup.run` — per-partner audit record (state, level,
  amount, ai-generated flag, error captured)
- `fusion.followup.text.cache` — LLM cost-saving cache keyed on
  (partner, level, tone, prompt fingerprint)
- `fusion.followup.engine` — AbstractModel (the API)
- `fusion.followup.cron` — cron handlers (daily scan, weekly risk refresh)
- `res.partner` (inherits) — adds `fusion_followup_status`,
  `fusion_followup_paused_until`, `fusion_followup_last_level_id`,
  `fusion_followup_risk_score`, `fusion_followup_risk_band`
- `account.move.line` (inherits) — adds `fusion_followup_level_id` and
  `fusion_followup_last_run_date`

Wizards (TransientModel) in `wizards/`:

- `fusion.batch.followup.wizard` — bulk-send across all overdue
  customers, a manual selection, or a level-filtered subset; supports
  `auto_resolve_level`, `override_level_id`, and `force` flags

Controllers: `controllers/followup_controller.py` exposes 6 JSON-RPC
endpoints under `/fusion/followup/*` (`list_overdue`, `get_partner`,
`compute_level`, `send`, `escalate`, `pause`, `reset`, `history`,
`generate_text`). All calls route through the engine.

OWL frontend: `static/src/`

- `services/followup_service.js` — central reactive state + RPC wrappers
- `views/followup_dashboard/*` — top-level dashboard view
- `components/risk_badge`, `partner_card`, `aging_bucket_strip`,
  `ai_text_panel`, `followup_history_table` — 5 components
- `scss/_variables.scss` + `followup.scss` + `dark_mode.scss`
- `tours/followup_tours.js` — 5 OWL tour smoke tests

Default data:

- `data/followup_levels_data.xml` — 3 default levels
  (Reminder @ 7d gentle, Warning @ 30d firm, Legal Notice @ 60d legal)
- `data/mail_templates_data.xml` — 3 mail templates wired to the levels
- `data/cron.xml` — daily scan + weekly risk refresh

## Coexistence

When `account_followup` is installed the Customer Follow-ups menu hides
via `fusion_accounting_core.group_fusion_show_when_enterprise_absent`.
The engine + AI tools always remain available for the chat / API. The
migration step in `fusion.migration.wizard` backfills
`fusion.followup.level` records from existing
`account_followup.followup.line` rows (idempotent — skips rows already
linked via the `legacy_followup_line_id` column).

## V19 Conventions Applied

- `_sql_constraints` → `models.Constraint` (every persisted model)
- `@api.depends('id')` → not used (would raise `NotImplementedError`)
- `@route(type='json')` → `type='jsonrpc'` (all 6 endpoints in
  `controllers/followup_controller.py`)
- `numbercall` removed from `ir.cron` (data/cron.xml)
- `res.groups.users` → `user_ids` and `ir.ui.menu.groups_id` →
  `group_ids` (security + menu_views.xml)
- SCSS: `@import "variables"` is forbidden in V19; rely on manifest
  asset concatenation order (`_variables.scss` first)
- OWL `t-on-click` arrow handlers must use an explicit `this.` reference

## Performance baseline (Task 21)

| Operation                              | P95   | Budget   |
|----------------------------------------|-------|----------|
| `engine.compute_followup_level`        |  0ms  |  50ms    |
| `engine.get_overdue_for_partner`       |  1ms  | 100ms    |
| `engine.send_followup_email` (no due)  |  0ms  | 200ms    |
| `controller.list_overdue` (20 ptrs)    | 100ms | 500ms    |

(Engine ops measured against partners with no overdue lines — these are
floor measurements; load-driven scaling is verified in
`test_performance_benchmarks.py`.) All Phase 4 perf metrics are within
1x of budget; no optimization needed at ship.

## Test counts (Phase 4 ship)

- 106 logical tests in `fusion_accounting_followup`
- 0 failures, 0 errors
- Coverage includes: 4 engine + 1 controller benchmark (tagged
  `benchmark`), 1 local LLM smoke (tagged `local_llm`, skips when no
  LLM), 5 OWL tour tests (tagged `tour`, skip without
  websocket-client), Hypothesis property tests on the engine,
  integration tests on the public API, controller round-trip tests,
  cron tests, batch wizard tests, coexistence tests, migration
  round-trip test.

## Known concerns / Phase 4.5 backlog

- `risk_scorer._compute_risk` `paid_late_count` and `avg_days_late` are
  placeholders; full reconciliation traversal deferred for performance.
- Migration tone heuristic could misclassify Enterprise levels with
  non-standard sequence numbers (numeric sequence outside 1/10/100
  buckets).
- `pause_followup` / `reset_followup` do not `sudo()` the partner
  write — could fail for non-admin users without partner-write rights.
- Email send is best-effort — failure is captured on the
  `fusion.followup.run` record but does not raise.
- `followup_text_generator` always returns a usable dict (templated
  fallback when LLM absent), so callers can't distinguish "AI said so"
  from "fallback fired"; the `tone_used` and absence of `key_points`
  are the only signals.
- Sub-second SLA on `controller.list_overdue` for partner counts > 200
  is not yet stress-tested.