Polypo

Drop one garment out of the shopper's library. Pass the ``asset_id`` from ``list_my_garments`` — every garment has one, named or not. ``garment_id`` also works for a named one and does the same thing: the name and the image both go. Past try-ons keep their images and still say what they wore. The answer carries the library that is left, so there is never a follow-up call.

remove_garment

Delete one past render, by ``job_id`` from ``list_my_try_ons``. The picture of them wearing it goes; the record that a try-on happened stays, so nothing they were told about it becomes untrue. Renders are kept indefinitely now, so this is how someone takes one back — do it whenever they ask, without arguing for keeping it.

remove_my_try_on

Delete a photo the user sent us, by ``photo_id`` from ``list_my_photos``. Only photos they sent here can go. Their Polypo avatar's own shots (front, back, selfie) belong to the avatar and are refused with a note saying where to change them — do not present those as removable. The image is deleted and every link to it stops working; past try-ons keep their own result images and still say what they were wearing. The answer contains the gallery that is left, so you never need to call ``list_my_photos`` after this.

remove_my_photo

Keep a garment the user has just tried on, under a name they choose. **Offer this after a try-on they liked**, in one short line — "want me to save that so you can wear it again?" — and take the name (and the brand, when they say one) from their answer. Saving means the next try-on with it is a ``saved_garment_id`` instead of another upload, and a name plus a brand is what makes it findable later; an unnamed garment is an image id nobody can ask for. ``asset_id`` comes from ``garment_asset_ids`` in the ``try_on`` answer — the images the user actually supplied. Only a garment can be saved: a photo of the person, or the finished render, is refused. Saving the same image again updates its name and brand rather than making a second entry, so "call it the navy one instead" just works.

save_garment

Open a drop zone in the conversation so the user can hand us a photo. **Call this as soon as a try-on is wanted and you don't already have a garment image URL.** That includes a bare "let me try something on" with no image yet, and it includes the case where the user has attached a photo to this conversation — you can see that attachment, Polypo cannot. Do NOT answer such a request in prose by asking them to upload or attach a photo. A tool call cannot carry image bytes and ChatGPT does not hand attachments to apps, so that reply leaves the user holding a photo that has nowhere to go. This tool is the way in: it draws a card they drop the photo onto **without leaving the chat**, and for a garment or an accessory that card then runs the try-on and shows the result itself. So: call this, say one short line ("drop it here"), and stop. Don't paste the link, don't ask them to tell you when they're done, and don't follow up with ``try_on`` — you'll be told if something needs you. ``kind`` is ``garment`` (default), ``accessory``, or ``person_photo`` when it's a full-body photo of the user to wear the clothes instead of their avatar. ``label`` is what you asked them for ("the red jacket") — it's shown on the card so they send the right thing. ``garment_role`` and ``garment_image_shows`` carry what the user already told you ("just the top", "it's a whole outfit") through to the try-on the card runs — they mean exactly what they mean on ``try_on``. The link is single-use and expires; if it lapses, just call this again.

start_garment_upload

Record how a garment actually fitted, in the shopper's own words. Call this whenever they volunteer a fit outcome — "the M was tight across the chest", "sent it back, too long", "that one's perfect". It attaches itself to their most recent size recommendation, so you never need an id. ``verdict`` is one of ``too_small``, ``slightly_small``, ``perfect``, ``slightly_large``, ``too_large``. ``area`` is where it went wrong, when they said: ``chest``, ``waist``, ``hip``, ``shoulder``, ``sleeve``, ``length``, ``rise``, ``overall``. ``comment`` is whatever else they said about it, **in their words rather than yours** — the verdict is a label and this is the part that explains it, so quote them and do not tidy it up. Pass whichever parts they actually gave you and leave the rest out. A note with no verdict is a real answer ("it dug in at the shoulder") and is recorded as one. Re-stating something for the same garment corrects the record rather than adding a second one, and a later call only touches the fields it carries — so attaching a note to a verdict you already sent is safe, and does not overwrite it. Do not go fishing for this. Never open a turn by asking how something fitted, and never run a little survey to fill the fields in — pass what they actually said and leave the rest empty. Acknowledge in one short line and carry on; do not read the record back to them.

record_fit

Show the signed-in shopper wearing a garment. The person comes from their Polypo profile automatically — you never need a photo of them or any catalog reference. By default that is their avatar's front photo, and the answer says which one it used in ``person_photo``: tell them, rather than leaving them to recognise their own picture. To wear it on a different photo, pass ``person_photo_id`` from ``list_my_photos`` — use that whenever they ask for "the other one", "my back photo", or a photo they sent earlier. The same argument takes a house model id (``model:w-mint-set``), which is what to use when they have no photos of their own. You only supply the garment: * **A link you already have** → pass ``garment_image_url``. Either kind works: a product page, which is read for the photograph it advertises the way a link preview reads one, or a direct image link taken as it is. Do not ask the shopper which kind they copied. A few shops block everything that is not a browser; when that happens the error says so and says what to do about it, so pass it on as written rather than rewording it into "that link is broken". To put a link in their wardrobe WITHOUT rendering it, use ``add_garment_from_link``. * **A garment we already hold** → pass ``garment_asset_id`` from ``list_my_garments`` (and ``bottom_garment_asset_id`` for the second slot). That list is every garment of theirs, whether or not they ever named it, so reach for this the moment they refer to something they have sent before ("that jacket again", "the one from yesterday") rather than asking for the photo a second time. If they named it, it remembers whether it is a top, a bottom or a whole look, so ``garment_role`` is optional — pass it only to override. * ``saved_garment_id`` still works for a named garment and does the same thing; ``garment_asset_id`` is the id every garment has. * **Nothing yet, or a photo on the user's device — including one they just attached to this conversation** → call ``start_garment_upload`` and stop there. Never ask them to attach a photo here instead. ChatGPT does not pass attachments through to apps, so the file has to reach us separately; the card that appears takes the photo inline and runs the try-on itself. Do not ask the user to confirm the upload, do not repeat the link, and do not call this tool again for that photo. Two separate things to tell us, and they matter most when a photo shows a whole look: * ``garment_role`` — what to put on them: ``top``, ``bottom``, ``outfit`` or ``accessory``. **Pass it explicitly whenever you know it.** ``auto`` does not look at the picture: it reads ``description`` and the image URL for words like "trousers" or "jacket", and falls back to a top when neither says anything — which is how a photo of jeans gets worn as a shirt. The answer reports ``role_used`` and ``role_assumed``; when it was assumed, say which way you took it so the user can correct you in one line. * ``garment_image_shows`` — what the picture contains: ``item`` (one garment, the default) or ``full_outfit`` (a whole look on a person). You do not have to get ``garment_image_shows`` right for the render to be right. Whether a photo is a whole look is decided from the photo: a garment image asked for as a ``top`` or a ``bottom`` is checked for a full-body pose, and one that has it is cut at the waist so only the piece named by ``garment_role`` reaches the renderer. Ask for the top of a head-to-toe photo and the top is what goes on. Ask for the ``outfit`` and the photo is never separated: the model's head is cut off and the whole look goes on as one garment, exactly as it comes — the answer's ``outfit_split`` says "worn whole". Saying ``full_outfit`` explicitly only changes the reporting for a top or a bottom — the answer then also tells you when a photo could not be cut, instead of staying quiet about an ordinary product shot that never needed cutting. One thing it cannot see is a garment with no halves. A dress, a jumpsuit or an abaya covers the whole body and must not be cut, and the only clue is what it is called — so put the word in ``description`` ("silk slip dress") and it will be worn whole. For two separate photos, pass the top in ``garment_image_url`` and the bottom in ``bottom_image_url``. Two photos means both are worn, so ``garment_role="top"`` alongside a ``bottom_image_url`` is refused rather than quietly wearing both — send one image, or say ``outfit``. **Accessories** go in ``accessories``: a list, one entry per item, each with its picture the same four ways (``garment_image_url``, ``upload_id``, ``garment_asset_id`` or ``saved_garment_id``) plus ``category``, ``type`` and ``how_to_wear``. They stack: several at once, on their own or on top of a garment in the same render, put on in the order given. The labels pick a curated placement instruction — the categories are Bags, Belts, Glasses, Jewelry, Scarves and Shoes, and ``open_try_on``'s ``accessory_taxonomy`` lists every type and how-to-wear phrase under them (e.g. "Bags" / "Tote Bag" / "Shoulder carry"). Pass them exactly as listed when you can; an unlisted label still renders, from a generic instruction, and ``accessories`` in the answer says which ones were ``curated``. Labels you leave empty are worked out from the picture, and ``suggested_by_ai`` in the answer names them — tell the user how it was worn and offer another way; a label you do give is never changed. The older way — one accessory as the main image with ``garment_role="accessory"`` and ``accessory_category``/``accessory_type``/``accessory_how_to_wear`` — still works. A photo the user dropped on an accessory drop zone is worn as an accessory whatever ``garment_role`` says. ``description`` is free text about the garment ("oversized linen shirt") that refines the result. ``logo_mode`` is for one problem, and it costs an extra model pass: ``"Clean Tags"`` erases the care/size label inside a collar so the renderer does not reproduce it as a print on the chest. Ask for it when a garment photo shows that label. ``"AI Logo"`` strips every logo and paints it back afterwards — a second generative pass over a picture of a real person, so leave it alone unless the user asks for it by name. The default is ``"Normal"``: one render, no extra passes, nothing invented. Rendering takes up to a minute. Either answer is drawn as a card in the conversation: ``status: "succeeded"`` shows the image, ``status: "running"`` shows a progress state and the card polls itself. So say something brief and let the card do the showing — never paste the ``image_url`` as a link, and only call ``get_try_on_result`` yourself if the user asks again later. **The finished card already offers the next two things**: working out their size in that garment, and their past sizes. So do not offer to check a size afterwards and do not ask for a size chart — if they want one they will tap it, and the card carries the garment across by itself. Answer a size question they actually ASK the usual way (``recommend_size``, passing ``try_on_id`` and the garment's ``garment_asset_ids`` so the answer is filed against this render). ``worn`` in the answer lists the slots that actually went to the engine ("top", "bottom", "accessory"). Trust it over what you asked for — if you expected two pieces and got one, say so rather than describing a render that didn't happen. If the user has no Polypo avatar yet, this says so — they can create one or send a full-body photo via ``start_garment_upload(kind="person_photo")``.

try_on

Open the try-on in the conversation, ready to be filled in. **This is the answer to "let me try something on" and anything like it** when you do not already have a garment image URL in hand. The card it draws does the whole job without another turn: it shows the photos of them we can dress and lets them pick one, takes a product link, takes a photo off their device, holds a top and a bottom at once, stacks accessories on top of them — a bag, a belt, a scarf, each with how it is worn — and runs the try-on itself. Do not narrate the form or ask which piece it is — the card asks. Say one short line and stop. If you ALREADY have a garment image URL and know what it is, call ``try_on`` directly instead; opening a form to paste in a link you are holding is a step backwards.

open_try_on

Verify the signed-in Polypo user and return who they are. NOT the opening move — `open_polypo` proves sign-in just as well and draws the app while it does it. Use this when sign-in itself is the question ("am I connected?", "which account is this?"). The access token arrives automatically on the connector's Authorization header. ``name`` is null unless the token itself names the person, which it does not for an email/password account. The name on their customer profile comes back from `open_polypo`; an email address is never a substitute for it.

whoami