Encompass: Support Guzzle 8

[mlahargou]

🤖

iFixit is moving to Guzzle 8 in iFixit/ifixit#65442, and this client breaks on it. Some requests fail without a full response, like a garbled reply or a cert error. The client asks Guzzle for the response anyway, and on Guzzle 8 that's a fatal error instead of an ApiException. A request that times out halfway through the body also comes back as an ApiException with status 200, so our retry code thinks Encompass answered.

Changes

• Helper functions: \GuzzleHttp\json_encode() is gone in Guzzle 8 and \GuzzleHttp\Psr7\build_query() has been gone since psr7 2 (iFixit monkeypatches it back). Now json_encode(..., JSON_THROW_ON_ERROR) and Psr7\Query::build(), which work on both.
• Failed transfers: a real HTTP error response keeps its status, headers and body. Any other failure is now ApiException code 0 with the Guzzle exception as getPrevious(). Connection errors and timeouts before a response still aren't wrapped, same as before.

The code is generated by swagger-codegen and the templates aren't in this repo, so rerunning build.sh would undo this.

What each failure becomes, sync call (async matches)

From a local server, timeout => 1:

Case	Before, Guzzle 7.15.5	Before, Guzzle 8.2.0	After, both
200	returns	returns	returns
400 with JSON body	ApiException(400) + body	same	same
500	ApiException(500) + body	same	same
Connection refused	ConnectException	ConnectException	same, unwrapped
Timeout before headers	ConnectException	NetworkTimeoutException	same, unwrapped
Timeout mid-body	ConnectException	ApiException(200), empty body	7: ConnectException; 8: ApiException(0), previous ResponseTimeoutException
Connection dropped mid-body	ApiException(200), empty body	ApiException(200), empty body	ApiException(0), previous RequestException (7) / ResponseTransferException (8)
Garbled reply (curl error 1)	ApiException(0)	fatal Error: undefined method getResponse()	ApiException(0), previous RequestException

QA

• vendor/bin/phpcs --standard=ruleset.xml EncompassAPI/ passes.
• Encompass tests pass in iFixit/ifixit#65442 with the submodule on this branch.

🤖 Generated with Claude Code

[mlahargou]

🤖

QA by Claude

Ran on cominor (michaellahar-2) against e20a3c4, from iFixit/ifixit in three setups: master with the old SDK (1108083, Guzzle 7.15.5), master with the submodule moved to e20a3c4 (Guzzle 7.15.5), and branch guzzle-8 (Guzzle 8.2.0), whose submodule already points at e20a3c4. iFixit/ifixit#65442 bumps the submodule to this commit; its CI integration job (GitSubmodulesLineageTest) waits on this PR merging.

Tested

• Tests touching Encompass, same counts old SDK vs new SDK on master: all OK both times. EncompassServiceTest 16/77, EncompassDropshipmentUpdaterTest 1/6, DropshipmentTest 1/4, ShopifyInventoryServiceTest 4/32, WarehouseInventoryServiceTest 1/4, GMCSupplementalFeedTest 12/36, ShopifyProductServiceTest 17/121, ShopifyWebhookProductBodyFactoryTest 43/560, ProductDataAPITest 26/212 (tests/assertions). On guzzle-8 the same, except EncompassServiceTest 19/92 (the branch adds 3 tests).

• Psalm: psalm --no-cache per file on all 19 files under iFixit/Service/Encompass/: "No errors found" on master + new SDK and on guzzle-8.

• No Guzzle function calls left: grep of EncompassAPI/Encompass for build_query, GuzzleHttp\json_encode, ->getResponse() finds nothing outside ApiException.php. Master's EncompassMonkeypatch.php still loads with the new SDK (build_query(["a"=>"b c"]) returns a=b%20c, defined only there), so no redeclare conflict.

• Live reads against test.encompass.com (cominor's config; not production), through EncompassService with a request-logging Guzzle middleware. Only search, partsInformation and orderStatus were called. Request URL and body (credentials redacted) were byte-identical in all three setups:

  Call	Request	Result, all 3 setups
  search('iPhone 6')	POST /restfulservice/search	200, errorCode 100, 200 parts, data sha1 ff6483c7… (old and new SDK, same minute)
  searchPartInfo sku 1110741, basePN 12538335	POST /restfulservice/partsInformation	200, errorCode 100, 2393 bytes, sha1 f319cd8f…
  orderStatus dropshipment 14309, US3510511	POST /restfulservice/orderStatus	real 400, errorCode -208 "invalid referenceNumber"

  No query strings on any request (all bodies are JSON), so Query::build([]) adds no ?. My first old-SDK search run gave a different sha1 (78330 vs 78355 bytes); rerunning the old SDK minutes later matched the new one, so that was the test server's data changing.

• Real 4xx error path unchanged: orderStatus with ENCQA-DOES-NOT-EXIST gave EncompassApiException code 400 wrapping Encompass\SDK\ApiException code 400, 10 headers, body {"status":{"httpCode":"400","errorCode":"-208",...}} (120 bytes), previous null, in all three setups. A POST to a nonexistent path on the same host gave ApiException 404, 11 headers, 235032-byte HTML body, previous null, in all three. The only difference: Guzzle 8's message has no trailing newline.

• Connection refused (127.0.0.1:9): GuzzleHttp\Exception\ConnectException, not wrapped, in all three setups. On Guzzle 8 it extends NetworkException, which #65442's retry catches.

• Timeout mid-body (local socket sends a 200 header plus 10 of 1000 bytes, client timeout 1s): Guzzle 7, old and new SDK both throw an unwrapped ConnectException (cURL 28), so master's behavior is unchanged. Guzzle 8 + new SDK throws ApiException code 0, null headers and body, previous = ResponseTimeoutException (a ResponseTransferException), which is what #65442's retry checks for.

Not tested

• Any write endpoint (createOrder, cancel, brand/exploded-view, RoundTrip). Their request builders got the same json_encode/Query::build swap, but I only ran the three reads above.
• Async (...Async) methods; nothing in iFixit/ifixit calls them, so the new errorResponseFrom() branch in the rejection handler ran only through code reading.
• A 429/5xx from the real gateway (retry paths are covered by EncompassServiceTest with mocks only).
• Production Encompass credentials or data.

Manual changes applied

• Moved 3P/EncompassAPIphpSDK to e20a3c4 on the master checkout and back to 1108083 afterwards. Scratch probe scripts ran from stdin; no repo files were edited. The checkout is back on its workspace branch with matching vendor/.

[mlahargou]

LGTM. CR 📱 QA 📱