feat(plugin): allow disabling the tenant plugin model providers cache

[GareArc]

Summary

Plugins assigned to a workspace from an enterprise admin console never appear in that workspace's model provider list (ESQ1-190, Zendesk #3547).

PluginService.fetch_plugin_model_providers caches the per-tenant provider list in Redis for 24h. Enterprise installs plugins by calling the plugin daemon directly, bypassing this API, so nothing here ever learns the cache is stale and the newly assigned plugin stays invisible until the TTL expires.

A cache is only correct if the system that owns the writes can invalidate it. This adds PLUGIN_MODEL_PROVIDERS_CACHE_ENABLED so deployments where something else installs plugins can turn it off and read through to the daemon, which does own the writes. The daemon-side replacement cache is langgenius/dify-plugin-daemon#780.

Defaults to true, so nothing changes for anyone installing plugins through this API — including all OSS deployments, where the existing invalidate_plugin_model_providers_cache call on plugin lifecycle keeps the cache correct.

The early return sits before the generation read and the refresh lock, so the disabled path touches no Redis at all rather than degrading to a slower cached path.

Second commit

Reverts b6099d0 (#39468), the inner-API endpoint that let an external system invalidate this cache. That approach was abandoned: it required wiring the dify endpoint and inner API key into every enterprise deployment, coupled enterprise to dify's cache internals, and did nothing for any other installer. Turning the cache off where it cannot be invalidated is the simpler contract.

What this deliberately does not do

An earlier revision also added PLUGIN_MODEL_SCHEMA_CACHE_ENABLED. That has been dropped.

The model schema cache has a related but much narrower problem: its key is {tenant_id}:{provider}:{model_type}:{model}:{user_id}, and provider carries no version, so a plugin upgrade leaves a stale schema behind. But it self-limits — PLUGIN_MODEL_SCHEMA_CACHE_TTL is 1h, versus 24h for the provider list — and it only bites if the affected model is exercised inside that window. Most upgrades are additive, so the visible symptom is "the new parameter has not shown up yet".

More to the point, there is no daemon-side replacement for it, so a switch there would only let operators delete a cache and get nothing back. Making the key version-aware is the right fix, but the plugin build is discarded in _to_provider_entity before anything is cached, and every way of threading it back through was worse than the 1h staleness it would prevent. Left for separate work.

Testing

• test_fetch_plugin_model_providers_bypasses_redis_when_cache_disabled asserts the daemon is the only source when disabled, and that get/mget/setex/lock are all untouched.
• uv run pytest tests/unit_tests/services/plugin/ tests/unit_tests/core/plugin/ — 572 passed.
• ruff clean.

[github-actions]

Pyrefly Type Coverage

Metric	Base	PR	Delta
Type coverage	55.10%	55.11%	+0.01%
Strict coverage	54.59%	54.60%	+0.01%
Typed symbols	35,801	35,799	-2
Untyped symbols	29,447	29,433	-14
Modules	3078	3076	-2

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.09%. Comparing base (9fd1fea) to head (35ad555).
⚠️ Report is 11 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #39632      +/-   ##
==========================================
- Coverage   86.10%   86.09%   -0.02%     
==========================================
  Files        5076     5073       -3     
  Lines      287044   286890     -154     
  Branches    57241    57200      -41     
==========================================
- Hits       247168   247003     -165     
- Misses      35124    35136      +12     
+ Partials     4752     4751       -1     

Flag	Coverage Δ
api	86.31% <100.00%> (-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:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[Copilot]

✅ Ready to approve

The functional change is well-scoped and covered by a unit test; only a small docstring clarification is recommended.

Note: this review does not count toward required approvals for merging.

Pull request overview

This PR adds an operator-controlled switch to disable the per-tenant plugin model providers Redis cache so enterprise deployments (where plugins may be installed out-of-band via the daemon) don’t serve stale provider lists for up to 24 hours.

Changes:

• Add PLUGIN_MODEL_PROVIDERS_CACHE_ENABLED (default true) and document it in env examples.
• Add an uncached fetch path in PluginService.fetch_plugin_model_providers that bypasses Redis entirely when disabled.
• Remove the previously-added inner API endpoint (and its unit tests) used to invalidate this cache externally.

File summaries

File	Description
docker/envs/core-services/shared.env.example	Adds the new env flag to the docker deployment example.
api/.env.example	Adds the new env flag to the backend env example.
api/configs/feature/init.py	Introduces PLUGIN_MODEL_PROVIDERS_CACHE_ENABLED in PluginConfig.
api/core/plugin/plugin_service.py	Implements the Redis-bypass path and refactors uncached fetch into a helper.
api/tests/unit_tests/services/plugin/test_plugin_service.py	Adds a unit test asserting no Redis calls occur when the cache is disabled.
api/controllers/inner_api/workspace/plugin_model_providers.py	Removes the inner API invalidation endpoint.
api/controllers/inner_api/init.py	Removes the namespace import/export for the deleted endpoint.
api/tests/unit_tests/controllers/inner_api/workspace/test_plugin_model_providers.py	Removes unit tests for the deleted endpoint.

Review details

• Files reviewed: 8/8 changed files
• Comments generated: 1
• Review effort level: Low

Note

Your feedback helps us improve the quality of this feature.
Please use 👍 or 👎 to tell us whether this assessment is correct.

---

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