WFDASH-WEBAPP/docs/order-assistant.md

# AI order assistant runbook

Design: `docs/superpowers/specs/2026-10-24-ai-order-assistant-design.md`.
Plan: `docs/superpowers/plans/2026-10-24-ai-order-assistant.md`.

## What it does

An order that came from a quote should point at that quote. Today only an
order created *from* a quote carries the link; every other order — a
walk-in, an email intake, a counter sale — has to be linked by hand.

The order assistant works out which quote an order came from. On a
schedule it looks at orders that have no quote yet, asks the AI which of
the candidate quotes the order came from, and then either:

- **links** the order to the quote itself (when it is sure), or
- **suggests** the quote, and a person accepts it.

Once the order points at a quote, the assistant can also **fill the
order's lines** from that quote — the cost, the supplier, the SKU and the
part number — leaving them in **purple** until a person checks them.

The AI never changes a line's status, never creates a purchase order and
never sends anything. It decides which quote an order belongs to, and it
may pre-fill a line's prices from that quote.

## The auto-link rule, in plain words

The assistant links an order to a quote **only** when all of these are
true:

- The **"Auto-link when sure"** setting is on (it is on by default).
- The AI says it is **sure** about the match.
- **Every** real order line is paired to a quote line, and each of those
  pairs is a **sure** match.
- **Every** pair points at a line of the **one** chosen quote (not a
  mixture of two quotes).
- The chosen quote belongs to the **same company** as the order.

If any of those fail, the assistant only writes a **suggestion**. The
order page then shows "AI suggests Q…" with the reason, the doubts and the
line-by-line pairing, and the existing **Link** button is there to accept
it.

The assistant also refuses to link when PocketBase says no — for example
when a line on the order has already moved past `pending`, or when
another order already points at that quote. A refusal is **not** an error:
the assistant falls back to a suggestion.

## Filling the lines from the quote

With **"Fill lines from the quote"** on, the assistant prices the order's
lines from the quote they are paired with. This is a copy, not a new
price: no AI call is involved, the pairing is already on the lines.

A line may be filled only when **all** of these are true:

- It is still `pending`.
- It carries **no cost** of its own (`cost_price` is 0).
- Its `pricing_source` is not `manual` and not `quote` — a price a person
  set is never overwritten.
- It has no **reviewed** AI fill on it.
- It is paired to a quote line (`quote_item` is set) that actually has a
  cost.

What the fill writes:

| Field | Rule |
|---|---|
| `cost_price` | The quote line's cost |
| `selling_price` | Only when the order line has none of its own |
| `supplier`, `sku`, `matched_price_list_id` | Only when the order line has none of its own |
| `markup` + `markup_percent` | Written together, from cost and selling |
| `pricing_source` | `ai_quote` |
| `ai_fill` | The audit record of the fill (see below) |

Every line is **re-read immediately before** it is written, so a cost a
person typed in the meantime wins over the AI's.

If the customer's price and the quoted price both exist and differ by more
than **5%**, the fill records a doubt, e.g. `customer price R16.64, quoted
R18.20`. Doubts show on the line's marker.

### The purple review

A filled line is **pending review** while all three are true:

- `ai_fill.status` is `filled`,
- `ai_fill.reviewed_at` is empty, and
- `pricing_source` is still `ai_quote`.

While a line is pending it shows up **purple** on the order Process page,
with the bot marker next to it. Hover or focus the marker to see the quote
it was filled from, the cost, the selling price, the markup, the reason and
any doubts.

The purple ends the moment somebody touches the line:

- **Saving the line** records the review on `ai_fill` (`reviewed_at`,
  `reviewed_by`, and whether the prices were accepted unchanged or
  changed). The pricing source stays `ai_quote`.
- The one-click **"use this cost"** copy writes `pricing_source: "quote"`
  — the operator chose the quote's cost themselves — so the AI's
  provenance ends too.
- A line whose price the operator types by hand becomes `pricing_source:
  "manual"`.

### The board badge

The Orders board shows an **"AI x/y"** badge on an order when the AI
filled some of its lines: `x` is how many lines still need a review, `y`
is how many the AI filled. The badge is hidden when nothing is pending.

### `order_items.ai_fill`

| Field | Meaning |
|---|---|
| `status` | `filled`, `skipped` or `error` |
| `source` | `quote` |
| `quote_id`, `quote_number`, `quote_item_id` | Where the numbers came from |
| `cost_price`, `selling_price`, `markup` | What was written |
| `supplier_id`, `sku` | What was written |
| `reason` | Why the line was paired with that quote line |
| `doubts` | Anything a person should check (e.g. the 5% price difference) |
| `model`, `attempted_at` | Which model, and when |
| `reviewed_at`, `reviewed_by` | Set when a person checked the line |
| `accepted_unchanged` | True when the person kept both prices |
| `supplier_changed` | True when the person changed the supplier |

## Going live (manual steps, in this order)

1. **Migrate PocketBase.** Run against the live PB:
   `npm run pb:migrate:order-assistant`
   (adds `orders.ai_quote_match`, `order_items.ai_fill` and the
   `order_assistant_*` fields on `ai_settings`). If the app is deployed
   before this, PocketBase drops the unknown fields silently and the
   assistant's result lands nowhere.
2. **Publish the PB hook change.** `ai_quote_match` and `ai_fill` were
   added to `IGNORED_FIELDS` in `pocketbase/pb_hooks/audit_log_utils.js`,
   so the AI's own writes do not flood the audit log. Publish it the same
   way the other hook changes are published
   (`npm run pb:hooks:publish`).
3. **Deploy** the app. Pushing to main does not auto-deploy; it is a
   manual deploy that takes 5 to 10 minutes.
4. **Clear the old rule-based suggestions.** Quote matching is AI-only:
   the rule-based matcher no longer runs (not at intake, not on the order
   page, not in the sweep). The suggestions it left behind are wrong far
   too often — on 2026-10-24 there were 2,888 on unlinked orders, one
   quote suggested to 309 of them. Clear them — **dry run first**:
   ```bash
   npx tsx scripts/2026-10-24-clear-rule-quote-suggestions.ts
   npx tsx scripts/2026-10-24-clear-rule-quote-suggestions.ts --apply
   ```
   It only clears `potential_quote` on orders with no hard link whose
   `quote_match_method` is not an AI one; links and AI suggestions stay.
5. **Activate the n8n clock** (below).
6. **Turn it on** in Settings, AI: the global "AI enabled" switch must be
   on, then switch on **"Order assistant enabled"** in the Order assistant
   section and click Save. Defaults: every 10 minutes, auto-link on, fill
   lines on.

The tick does nothing until both switches are on, so steps 5 and 6 are
safe in either order.

## n8n workflow: "WFDASH Order assistant tick"

n8n only acts as a clock. The app decides whether a run is due (AI and
the assistant on, and at least "Run every" minutes since the last run), so
the setting in the app controls the real frequency.

Spec:

| Node | Setting |
|---|---|
| Name | `WFDASH Order assistant tick` |
| Trigger | Schedule Trigger, every 5 minutes |
| Action | HTTP Request |
| Method / URL | `POST` `https://wfdash.apphub.co.za/api/v1/orders/assistant/tick` |
| Authentication | Generic Credential Type, Header Auth, **the same credential** the quote assistant tick workflow uses (an `Authorization: Bearer <AGENT_API_KEY>` header). Do not create a new secret. |
| Body | none |
| Settings tab | Retry On Fail: **off**; Timeout: **120000 ms** (120 s) |
| Status | **inactive** until the app is deployed and migrated |

Click-by-click:

1. n8n, Workflows, Create workflow. Name it `WFDASH Order assistant tick`.
2. Add node, Schedule Trigger. Trigger Interval: Minutes, Minutes Between
   Triggers: `5`.
3. Add node, HTTP Request, connect it after the trigger.
   - Method: `POST`.
   - URL: `https://wfdash.apphub.co.za/api/v1/orders/assistant/tick`.
   - Authentication: Generic Credential Type, Generic Auth Type: Header
     Auth. Pick the credential used by the quote assistant tick workflow
     (open that workflow and read its HTTP Request node to see which one).
   - Options, Add option, Timeout: `120000`.
   - Settings tab on the node: turn **Retry On Fail** off.
4. Save. Leave the workflow **inactive** until the steps above are done.
5. After deploy, click Execute workflow once. Expected response while the
   assistant is off: `{"ran":false,"reason":"disabled"}`. When on and not
   yet due: `{"ran":false,"reason":"not_due"}`. While an earlier run is
   still going: `{"ran":false,"reason":"running"}`. When it ran:
   `{"ran":true,"summary":{...}}`.
6. Toggle the workflow Active.

A 401 means the credential is wrong (it must send the app's
`AGENT_API_KEY` as a Bearer token). A tick that takes long is normal when
there is work. Each run is bounded: it matches **at most 10 orders** and
fills **at most 10 orders**, and it stops starting new orders once 90
seconds have passed, so a run normally answers inside the 120 s timeout;
leftovers carry to the next run.

An order the tick has nothing to do with does **not** use up that budget:
the run looks through up to 200 unlinked orders to find 10 it can work
on, and an order whose lines the fill could not write is skipped rather
than counted.

Runs never overlap. A run stamps its start time before it does any work and
writes its summary when it ends. While a start is newer than the last
summary and less than 10 minutes old, other ticks answer
`{"ran":false,"reason":"running"}`. If a run dies without writing its
summary, the next tick after those 10 minutes takes over.

## Running it by hand

- **Order page** — the **"Match with AI"** button on the quote match card
  runs the assistant for that one order and reloads the order. It also
  fills the lines when the order ends up linked.
- **MCP tool** — `orders.assistant_match` with `{"orderId": "..."}`.

Both force a run past the "already looked at these lines" rule.

## Reading `ai_quote_match`

`orders.ai_quote_match` (JSON) is the record of what the assistant decided
about an order. Read it in the PB admin.

| Field | Meaning |
|---|---|
| `status` | `matched` (the order was linked), `suggested`, `no_match` or `error` |
| `quote_id`, `quote_number` | The quote the assistant chose |
| `confidence` | `high`, `medium` or `low` |
| `reason`, `doubts` | The AI's plain-language reasoning and anything to check |
| `lines` | The pairing: order line id → quote line id, with `same_item` (`sure` / `likely` / `unsure`) |
| `auto_linked` | True only when the assistant itself created the link |
| `lines_hash` | Fingerprint of the order lines the run looked at — the same lines are not looked at twice |
| `model` | The model that answered |
| `run_at` | When the run happened |
| `error` | Set when `status` is `error` |

## Settings (Settings → AI → Order assistant)

| Setting | Default | Meaning |
|---|---|---|
| Order assistant enabled | off | The tick does nothing while this is off |
| Run every (minutes) | 10 | How often the tick may run |
| Auto-link when sure | on | May the assistant link the order itself? |
| Fill lines from the quote | on | May the assistant pre-fill the order's lines from the linked quote? |
| Prompt | empty | Overrides the built-in prompt when set |

`orders.ai_quote_match` and `order_items.ai_fill` are in `IGNORED_FIELDS`
of the audit hook, so the assistant's own writes do not flood the audit
log.

## Not in this phase

- The assistant does not price a line from scratch — it only copies the
  cost, supplier, SKU and part number from the quote line the order line
  is already paired with.
- It never creates a purchase order, never sends an email and never
  changes a line's status.
- There is no POS fallback: the assistant fills only from a quote.