Guide for building a new LimaCharlie USP adapter in this repo — researching the source, implementing it by copying the closest existing adapter, writing thorough tests with a mock/fake of the source, optionally live-testing against real credentials, validating the mock's fidelity, and shipping a documented PR. Use when the user wants to add a new adapter / data source / integration (e.g. "add an adapter for <vendor>", "ingest <product> logs into LimaCharlie").
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Guide for building a new LimaCharlie USP adapter in this repo — researching the source, implementing it by copying the closest existing adapter, writing thorough tests with a mock/fake of the source, optionally live-testing against real credentials, validating the mock's fidelity, and shipping a documented PR. Use when the user wants to add a new adapter / data source / integration (e.g. "add an adapter for <vendor>", "ingest <product> logs into LimaCharlie").
Building a new USP adapter
A USP adapter reads events from some external source and ships them to
LimaCharlie as USP DataMessages. Sources vary widely — REST APIs you poll,
vendor SDKs (AWS/GCP/Azure), streaming sockets, mailboxes, local files, stdin —
so there is no single template. What every adapter shares is one contract
(see references/patterns.md); everything else you get by copying the existing
adapter closest to your source.
The strongest, best-tested examples to imitate are threatlocker/ and gmail/
(both have full mock-backed test suites). But pick by source type:
HTTP API you poll → threatlocker/, 1password/, gmail/
local files / OS sources → file/, evtx/, wel/, stdin/
references/patterns.md maps source types to examples and lists the shared
contract and the patterns worth copying regardless of source. Read it plus your
chosen example(s) before writing code.
Work through the phases in order. Don't skip research — getting the source's real
shape right up front is what makes the implementation and the mock correct.
Phase 0 — Scope it with the user
Before writing anything, confirm:
What source / vendor and which events they want ingested.
How you reach the source: a documented HTTP API, an official SDK, a
socket/stream, a file? This decides which existing adapter you copy.
Auth method: API key/token, OAuth refresh token, basic auth, cloud
credentials/service account, etc. This drives the config struct and the mock's
auth check.
Whether they have working credentials you can use to live-test (ask now;
you'll need it in Phase 5). If they do, agree how they'll provide them — env
vars are simplest. Never commit credentials. This repo already uses files
like *-do-not-commit.txt for local secrets; follow that and check
.gitignore.
Phase 1 — Research the source
Read the real vendor documentation / SDK reference — do not rely on memory. Use
WebFetch / WebSearch. Capture, before coding:
How you list/receive events — for an HTTP API: the endpoint(s), method,
request params, and how pagination works (page number, cursor/next,
has_more, offset/limit, time window) — pagination is the #1 source of adapter
bugs. For an SDK: which client method streams/lists the events and how it
signals "no more". For a stream/queue: how messages are received and
acknowledged.
The event shape — a real sample payload, the stable per-record id field
(for dedup) and the timestamp field (for event time), with exact
paths/formats. For HTTP, the response envelope (bare array vs. {"data":[...]}).
Rate limits / error model — status codes, retry-after, throttling.
Record the doc URLs and a couple of real sample payloads — sanitized
versions go into the mock fixtures and the README. Grab an OpenAPI/Swagger spec,
Postman collection, or SDK example if one exists; it's the most reliable source
of shapes.
Phase 2 — Implement by copying the closest adapter
Create package directory myadapter/ (lowercase; the Go package is
usp_myadapter). Most adapters are a single client.go; the well-tested ones
add *_test.go and sometimes split transport into api.go (as threatlocker
does for HTTP). Mirror the file layout of whichever adapter you're copying.
Every adapter, whatever the source, must honor the universal contract (detailed
in references/patterns.md):
Config struct embedding uspclient.ClientOptions as client_options
(json+yaml tags on every field) with a Validate() that fills defaults.
func NewMyAdapter(ctx, conf) (*MyAdapter, chan struct{}, error) and an
idempotent Close() error; the channel closes when the adapter stops itself.
Read from the source however it works, and ship each event verbatim as a
protocol.DataMessage{JsonPayload, EventType, TimestampMs} through the usp
client. Don't reshape payloads — LimaCharlie parses/maps in the cloud.
Patterns worth copying from the well-tested adapters (all in
references/patterns.md): a test seam that puts the external dependency
behind a small interface so a fake/mock can stand in (threatlocker's uspSink
unexported newThreatLockerAdapter(ctx,conf,sink)); lifecycle via
utils.NewEvent() + WaitGroup + sync.OnceClose; exactly-once shipping
via utils.NewLocalDeduper keyed on a stable id when the source can re-deliver;
robust transport (HTTPError + isTransientError, backoff, stop on 401/403)
for HTTP sources; JSON fidelity via utils.UnmarshalCleanJSON/utils.Dict; and
backpressure handling on uspclient.ErrorBufferFull.
Then register the adapter (see references/registration.md — easiest is to
grep an existing adapter like threatlocker across these files and replicate):
containers/conf/all.go: import the package, add a field to GeneralConfigs.
containers/general/tool.go: import it, add the else if method == "myadapter"
dispatch branch.
root README.md: add a short section with a run example.
Build with go build ./containers/general and go fmt ./myadapter/....
Phase 3 — Write a mock/fake of the source
The whole point is end-to-end tests that run without live credentials, by
substituting a stand-in for the external source through the test seam. Model on
threatlocker/mock_test.go and gmail/mock_test.go:
For an HTTP source: an httptest.NewServer whose handler reproduces the
real API — checks auth the way the real API does, paginates over an in-memory
dataset with the same mechanism, returns the same envelope shape.
For an SDK / non-HTTP source: a fake implementing the small interface your
adapter calls, returning realistic results and exercising the same "more
data / no more" and error behaviors.
Realistic fixtures matching the Phase-1 sample payloads — nested objects,
arrays, mixed scalar types, the real id and timestamp field names/formats.
End-to-end tests to write:
All events ship; event type, parsed timestamp, and verbatim payload match.
Re-reading/re-polling does not re-ship (dedup works, when applicable).
An event appearing mid-run ships exactly once.
A multi-page / multi-batch dataset is fully consumed, each event once.
Multiple event types (if applicable) are tagged correctly.
Bad auth → adapter stops, ships nothing.
Phase 4 — Run the tests
Standard go test:
go test ./myadapter/...
All tests must pass and be fully enabled — never skip a test or weaken an
assertion to get green. If a run_local_test.sh ever exists at the repo root,
prefer it (it handles local system dependencies). There isn't one today, so
go test is correct for a self-contained adapter.
Phase 5 — Live-test with real credentials (if available)
If the user provided credentials in Phase 0, run the real adapter against the
live source to prove it works end-to-end. Easiest is via the built binary
(matches production), pointed at a test OID:
Confirm a new sensor appears in the user's LimaCharlie org and events flow.
Capture the real responses/messages while it runs (log the raw payloads, or
use a proxy) — these are gold for Phase 5b. If you write a throwaway integration
test to do the capture, remove it before the PR; do not leave a skipped test
behind.
Phase 5b — Validate the mock against reality
With real data in hand, diff it against your mock:
Does the real pagination / batching match what the mock assumes (sizes,
the end-of-data signal, cursor/has_more)?
Is the envelope / result shape exactly what the mock returns?
Do the id and timestamp field paths/formats match? Are timestamps in a
format your parser handles?
Do the fixtures carry the field types the real data has (big integers,
nullable fields, nested arrays)?
Does auth rejection behave as the mock simulates?
Update the mock and fixtures to mirror captured reality, so the tests would have
caught any shape mistake and future maintainers can evolve the adapter
confidently without live credentials. Re-run Phase 4.
If the user has no credentials, say so plainly, make the mock as faithful as
possible from the documented shapes (Phase 1), and note in the PR that it hasn't
been live-verified.
Phase 6 — Document and open a draft PR
Write myadapter/README.md (model on threatlocker/README.md):
One-line description + links to the vendor and the API/SDK docs (Phase 1).
Authentication: how to obtain the credential and how it's supplied.
Configuration table: every config key, required?, description, defaults.
What the data looks like: which events are collected, the event types, a
sample payload.
How it works (polling/streaming, pagination/batching, dedup) and
operational caveats (rate limits, required filters, region/instance, etc.).
CLI and YAML run examples.
Add the short root-README.md entry too and make sure the example matches.
PR body: what it ingests, API/SDK docs link, auth method, event types/sample
data, how it was tested (mock coverage + whether live-verified), follow-ups. Be
careful to stage only your adapter's files — don't sweep in unrelated
working-tree changes.
Checklist
Source researched from real docs; auth, how-events-arrive, pagination/
batching, id+timestamp fields, sample payload recorded.
