fix(api): keep provider refresh waiters single-flight

[WH-2099]

Summary

This is a follow-up to #37388 for the provider-list refresh path tracked under #34264.

The previous single-flight lock prevented ordinary cache-miss stampedes, but waiters could still fall through to the plugin daemon after the lock TTL deadline.
That made the Redis lock TTL act as both a crashed-owner lease and a waiter deadline.

This PR makes the coordination contract explicit:

• Redis lock TTL is only a lock lease.
• Waiters keep reading the current provider generation and cache until they either read cache or acquire the refresh lock.
• Redis cache or lock failures still fall back to direct daemon discovery to preserve availability.
• Generation changes naturally restart the cache/lock path on the new generation.

CI Stability

This PR also fixes #38277.

test_inspect_skill_maps_invalid_archives_to_service_errors generated ZIP bytes during pytest collection.
Because zipfile.writestr(name, payload) stores the current timestamp in ZIP entry metadata, xdist workers could collect different parametrized node ids and fail before running tests.

The ZIP fixture now uses a fixed entry timestamp, and the parametrized cases have stable ids.

Tests

• env DEBUG=false HTTP_PROXY= HTTPS_PROXY= ALL_PROXY= NO_PROXY= http_proxy= https_proxy= all_proxy= no_proxy= uv run --project api pytest -o addopts='' api/tests/unit_tests/services/plugin/test_plugin_service.py api/tests/unit_tests/core/plugin/test_model_runtime_adapter.py -q
• uv run --project api ruff check api/core/plugin/plugin_service.py api/tests/unit_tests/services/plugin/test_plugin_service.py api/tests/unit_tests/core/plugin/test_model_runtime_adapter.py
• env -u HTTP_PROXY -u HTTPS_PROXY -u ALL_PROXY -u NO_PROXY -u http_proxy -u https_proxy -u all_proxy -u no_proxy DEBUG=false uv run --project api pytest -p no:benchmark -n 2 api/tests/unit_tests/services/test_agent_config_service.py::test_inspect_skill_maps_invalid_archives_to_service_errors
• git diff --check

Refs #34264
Fixes #38277

[github-actions]

Pyrefly Type Coverage

Metric	Base	PR	Delta
Type coverage	51.87%	51.87%	+0.00%
Strict coverage	51.39%	51.39%	+0.00%
Typed symbols	31,536	31,542	+6
Untyped symbols	29,539	29,539	0
Modules	2952	2952	0

[codecov]

Codecov Report

❌ Patch coverage is 91.80328% with 5 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.27%. Comparing base (61650d3) to head (b57853c).

Files with missing lines	Patch %	Lines
api/core/plugin/plugin_service.py	91.80%	4 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #38226      +/-   ##
==========================================
- Coverage   85.27%   85.27%   -0.01%     
==========================================
  Files        4982     4982              
  Lines      263256   263230      -26     
  Branches    50022    50014       -8     
==========================================
- Hits       224504   224480      -24     
- Misses      34366    34367       +1     
+ Partials     4386     4383       -3     

Flag	Coverage Δ
api	85.50% <91.80%> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[cqjjjzr]

I feel that we should not make this an infinite wait. A good idea would be introduce a hard timeout that once exceeded, fail the whole process.

[Copilot]

Pull request overview

This PR refines the plugin provider-list refresh coordination in PluginService so concurrent cache-miss waiters remain single-flight across Redis lock TTL boundaries (treating lock TTL as a lease rather than a waiter deadline), while still preserving the “fallback to direct daemon discovery” behavior when Redis cache/lock coordination fails.

Changes:

• Replaces the prior “wait with deadline” approach with a loop that keeps re-checking the current generation cache and retrying lock acquisition until cache is filled or the caller becomes refresh owner.
• Tightens typing around Redis locks via a small internal Protocol and updates the lock acquisition/release helpers accordingly.
• Updates and extends unit tests to cover (a) continuing to wait past lock TTL and (b) restarting the lock/cache path after generation changes.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
api/core/plugin/plugin_service.py	Reworks provider-cache coordination to keep waiters single-flight across lock TTL and generation changes.
api/tests/unit_tests/services/plugin/test_plugin_service.py	Updates and adds tests validating the new wait/lock/generation behavior.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

Pyrefly Diff

base → PR

--- /tmp/pyrefly_base.txt	2026-07-02 04:15:00.682688556 +0000
+++ /tmp/pyrefly_pr.txt	2026-07-02 04:14:52.046710658 +0000
@@ -4619,8 +4619,6 @@
    --> tests/unit_tests/core/plugin/test_backwards_invocation_app.py:417:74
 ERROR Argument `SimpleNamespace` is not assignable to parameter `tenant` with type `Tenant` in function `core.plugin.backwards_invocation.model.PluginModelBackwardsInvocation.invoke_summary` [bad-argument-type]
   --> tests/unit_tests/core/plugin/test_backwards_invocation_model.py:54:72
-ERROR Generator function should return `Generator` [bad-return]
-  --> tests/unit_tests/core/plugin/test_model_runtime_adapter.py:72:51
 ERROR Class member `_BadString.__str__` overrides a member in a parent class but is missing an `@override` decorator [missing-override-decorator]
    --> tests/unit_tests/core/plugin/test_plugin_entities.py:238:17
 ERROR Argument `TestPluginDaemonEntities.test_credential_type_helpers._FakeCredential` is not assignable to parameter `self` with type `CredentialType` in function `core.plugin.entities.plugin_daemon.CredentialType.get_name` [bad-argument-type]
@@ -6827,8 +6825,6 @@
   --> tests/unit_tests/services/plugin/test_oauth_service.py:68:34
 ERROR Object of class `FunctionType` has no attribute `assert_called_once_with` [missing-attribute]
   --> tests/unit_tests/services/plugin/test_oauth_service.py:95:9
-ERROR Generator function should return `Generator` [bad-return]
-  --> tests/unit_tests/services/plugin/test_plugin_service.py:18:51
 ERROR Class member `DummyRetrieval.get_pipeline_templates` overrides parent class `PipelineTemplateRetrievalBase` in an inconsistent manner [bad-override]
  --> tests/unit_tests/services/rag_pipeline/pipeline_template/test_pipeline_template_base.py:5:9
 ERROR Class member `DummyRetrieval.get_pipeline_templates` overrides a member in a parent class but is missing an `@override` decorator [missing-override-decorator]

[Copilot]

Pull request overview

Copilot reviewed 3 out of 4 changed files in this pull request and generated 1 comment.

[cqjjjzr]

LGTM