new-adapter

name	new-adapter
description	Add a new vendor signal source — scaffold the adapter crate, then walk the schema/fixture/reachability/fetch checklist. Use when asked to support a new analytics, monitoring, or ticketing tool.

Add a vendor adapter

Run the scaffolder, then follow the checklist it prints:

scripts/new-adapter.sh <vendor>    # lowercase, e.g. "amplitude"

It generates crates/merge0-adapter-<vendor>/ (Cargo.toml, Adapter impl skeleton, golden test, fixture stub) and registers the crate in the workspace. The workspace then deliberately does not compile until you add the Source variant — a new source IS a schema change (invariant 2), so docs/signal-schema.md bumps its version in the same PR.

The parts that are decisions, not boilerplate:

1. Ground the payload shape in the vendor's real API (docs or an actual response), then invent example.com fixture data (invariant 4 — never real payloads, never credentials). Bless expected files with MERGE0_BLESS=1 cargo test -p merge0-adapter-<vendor> and review the blessed output — blessing without reading is how wrong normalization gets pinned as correct.
2. Document normalization decisions in the crate doc comment (affected_count semantics, severity mapping, join_keys) — see crates/merge0-adapter-openpanel/src/lib.rs for the expected depth. Populate join_keys wherever derivable (invariant 3); severity maps conservative (a quiet inbox that's right beats a busy one).
3. Make the source reachable by a shipped scout in config/scouts/ — the hygiene rule fails the build until every (source, kind) your goldens emit is selectable. This is the six-dead-sources bug (docs/decision-log.md #9); don't over-pin scout sources lists in tests, that's how a source ships unreachable.
4. Fetch layer: poller or webhook in crates/merge0-fetch (wiremock-tested; construction fails loudly when the source is enabled but unconfigured), then server ingest routing + config/sources.toml entry (secrets by env-var NAME only).
5. Malformed rows are skipped, never envelope failures — one corrupt row must not sink the page. And adapters never invent fields the schema doesn't have.

Finish with the verify skill (full, unscoped) and, if counts changed, revisit e2e assertions — new sources change triage arithmetic.