Make seller-copilot marketplace-aware: add Shopee Brasil

[fiharet34]

Summary

Makes the merged seller-copilot skill marketplace-aware: it now covers Shopee Brasil
alongside Mercado Livre, and decides which one a request belongs to before doing anything else.

Eleven files: SKILL.md modified, plus references/which-marketplace.md and nine
references/shopee-*.md procedures. The fifteen Mercado Livre reference files are untouched —
byte-identical to what merged in #27.

Shopee sellers currently have no entry point in this catalogue at all; every skill here is
Mercado Livre only.

Why one skill rather than a second one

The two marketplaces are independent datasets with different grains, different field names and
different estimate methods, and the connector forbids mixing them in a single query. A seller
does not always say which one they mean — so deciding is itself the first step of the work,
and that decision has to live somewhere a seller's question reaches. A separate Shopee skill
would only ever be reached by someone who already said "Shopee".

How the decision is made

references/which-marketplace.md holds the logic and the full capability comparison. SKILL.md
carries a compact Step 0 that runs before any analysis:

• The seller said so → take them at their word.
• An MLB-prefixed identifier or a mercadolivre link → Mercado Livre. A bare 10–11 digit
  number or a shopee link → Shopee.
• Buy-box, medals, keywords, fulfilment are Mercado Livre concepts → Mercado Livre.
• Otherwise ask, in one short question, before querying. No defaulting to the richer dataset.

Never mix the two in one query, table or total. Comparing them is legitimate and produces two
labelled analyses side by side, compared by orders of magnitude rather than exact figures,
because the estimate methods differ.

Shopee coverage is narrower, and the skill says so

Nine analyses are supported: category evaluation, market structure, finding products, validating
a product, item momentum, competitor discovery, shop profiling, assortment and price gaps,
pricing.

These are declared unavailable with the reason, never approximated and never answered from
Mercado Livre data instead:

Not available on Shopee	Why
Keyword / search data	Does not exist for Shopee
Seasonality, when to stock, long-run trend	History begins May 2026 — too short
Buy-box / catalogue position	No such mechanic; every row is an item
Seller medals, reputation ladder	Badge tiers only — official / preferred / regular
Fulfilment, free-shipping, listing type	Not available as item attributes
Cancellation rate	Not available
Brand market share	Coverage is a lower bound; share is not computable
Per-variant pricing	Item-level only
Fees, commissions, real margin	No Shopee fee data is held

One capability runs the other way: Shopee has a per-item weekly sales series that Mercado Livre
lacks, used by shopee-item-momentum.md, which states that the weekly figure is smoothed from a
rolling counter rather than measured.

Evidence

A 22-item pack run through a real agent CLI against the live connector. Routing is asserted
from the trace rather than scored by a judge — a judge cannot see which tool was called, and an
answer built from the wrong marketplace reads perfectly well.

Tier	What it asserts	Result
1	Right marketplace queried, wrong one not; reference reads match	8/8
2	No marketplace named → asks, queries nothing	2/3
3	Both marketplaces → data from each, kept separate	3/3
4	Six capabilities Shopee lacks declined; two it has not refused	8/8

21/22, in one clean pass, $10.98. Assertions require a successful result, not merely a tool
call — an earlier version counted calls and passed two items that had returned no data.

No Mercado Livre regression. The eight-item Mercado Livre quality pack was re-run against the
changed skill: the six comparable items moved 4.67 → 4.50, four of them unchanged, well inside
the ±0.5 retest band this project has seen on repeat runs. The vague-entry case actually improved,
3 → 4.5, because Step 0 gives it a concrete first question to ask.

The one known limitation

Tier 2's single failure is real and I could not fix it from inside this skill. Asked "vale a
pena entrar na categoria X?" with no marketplace named, the request matches a focused skill's
frontmatter triggers closely enough that the focused skill handles it directly — and because every
focused skill here is Mercado Livre only, the seller gets Mercado Livre data without being asked
which marketplace they meant. seller-copilot never loads, so no text in it can intervene.
Reproduced identically across two runs.

The other two ambiguity items pass, so the guard itself works; what fails is interception, and
only when the phrasing closely matches a focused skill's own triggers. It is partly a consequence
of a deliberate choice: those category triggers were kept out of this skill's description
precisely to avoid cannibalising the focused skills. Closing the gap means competing with them,
which may not be the better trade. Happy to take direction on which way you'd prefer.

Worth knowing operationally

A single Shopee analysis costs roughly 4–7 connector queries. On the plan this was tested
against, the Shopee daily allowance was initially ~10 requests — about two analyses before
lockout — and the skill handles exhaustion correctly, reporting the limit in business terms and
refusing to invent figures. Flagging it because it bears on how usable the Shopee side is for a
seller on a lower tier, not because anything here needs changing.

Notes for reviewers

• Additive. No existing skill, asset or README row is modified. The fifteen Mercado Livre
  references are unchanged.
• Shopee reference files carry a shopee- prefix; unprefixed files remain Mercado Livre. Every
  cross-link inside a Shopee file points at a Shopee sibling — verified, and asserted in the tests
  after an earlier version silently resolved to the Mercado Livre procedures.
• The description is at 994 of the 1024 limit. Making room for Shopee triggers meant trimming some
  Portuguese Mercado Livre ones; the regression check above is what confirms that was safe.
• Both marketplaces are Brasil, so currency, language and the retail calendar are unchanged.

Public-Safety Checklist

• No credentials, tokens, keys, or secrets.
• No internal filesystem paths.
• No private Slack, Jira, Notion, admin, or service URLs.
• No private service names or implementation details.
• No customer data or private product data.
• Skill directory name matches SKILL.md frontmatter name.

No data-model names, field names, tool identifiers or resource URIs appear in any of these files —
they describe procedures and judgement only.