Writes .mercemur.json so every command in this directory knows which store it acts on, after checking that the store answers and the key works. The check is the point. Writing a config that turns out to be wrong moves the failure to the next command, where the message is about that command rather than about the link.

​
env pull

Fills .env.local with what the CLI can prove: the store id, the API base, the storefront URL and the default locale.
It cannot supply a publishable key. Keys are minted from the dashboard and are not on the public API, so the file records the variable with a note rather than leaving a name you discover from a stack trace. Run mercemur open apikeys.
Existing values are kept unless you pass --force: a local override is usually deliberate. Unrelated variables and key order are preserved, so a pull is a small diff rather than a rewrite.

​
translations

Columns are locale, resource_type, resource_id, field, value, with the value last because it is the only one a translator edits.
The API scopes a translation list to one resource on purpose, so an export walks them and reports how many reads that took. A store with ten thousand products should see what the export costs before it is surprised by it. Reading product requires read_products, and so on for each type walked.
Import sends only the values that differ, so re-importing an unchanged export writes nothing. Leading and trailing space is preserved: it belongs to the translator, and stripping it changes the rendered copy.

​
redirects

Columns are from_path, to_path, status_code. A status must be 301, 302, 307 or 308; anything else is refused rather than guessed at.
Import only adds and updates. A redirect missing from the file is left alone, because a shorter file is far more often a partial export than an instruction to delete the rest, and removing a live redirect breaks a URL somebody has indexed.
A file that redirects the same path twice is refused, naming both rows: that is a merge accident, and applying both silently leaves whichever landed last.

​
seo

audit reports the mistakes a search result makes visible: a duplicate title across two pages, a description too long to survive truncation, one too short to say anything. It audits what is set, not what is missing. A resource with no override falls back to its own catalogue title, which is usually correct, and reporting every one of those would bury the findings that matter. Exits non-zero on a duplicate title, which is the one finding here that is unambiguously wrong: two pages competing for the same query is a decision nobody made. Length findings are warnings, because those limits are conventions that move.

​
metafields

One JSON object per line, not CSV.
A metafield’s value is typed by its definition: a number, a boolean, an object, a list of references. CSV stringifies all of it, so 420 comes back as "420" and nothing reports it until a template renders the quotes. JSONL keeps the type and still gives a diff of one line per field.
Values are compared as canonical JSON, so the same object with its keys in another order is the same value and a reformatted file is not mistaken for a store-wide edit. A line with no definition_id is refused rather than guessed at: the definition is what decides the value’s type, and writing one without it stores a shape the storefront cannot read. Run mercemur metafields definitions to see them. A malformed line names its own line number, because “invalid character” in a file of several thousand is a search rather than a fix.

​
A header with no rows

Every import here treats a header-only CSV as zero rows rather than an error. That file is exactly what an export of a store with none of these produces, and erroring on it would break the round trip at its emptiest point. A file with no header at all names no columns and is refused.