refactor(plugins): migrate in-tree plugins to PyPI packages (cpex-*) by jonpspri · Pull Request #3965 · IBM/mcp-context-forge

jonpspri · 2026-04-01T06:19:59Z

Summary

Migrate 6 in-tree plugins (pii_filter, secrets_detection, url_reputation, retry_with_backoff, encoded_exfil_detection, rate_limiter) to standalone PyPI packages (cpex-*)
Delete all plugins_rust/ Rust plugin implementations — plugins are now distributed as pre-built wheels from PyPI
Remove Rust plugin builder stages from all Containerfiles (keep tools_rust/mcp_runtime)
Remove ~100 lines of rust-* plugin Makefile targets
Add [plugins] extra to install-dev and CI pytest workflow so cpex packages are installed
Update tool_service.py import from plugins.retry_with_backoff to cpex_retry_with_backoff
Update plugin kind paths across 7+ doc files from plugins.pii_filter.pii_filter.PIIFilterPlugin to cpex_pii_filter.PIIFilterPlugin
Clean up pre-commit configs, CODEOWNERS, MANIFEST.in, whitesource, .gitignore

Test plan

make install-dev && make test passes (cpex packages installed via [plugins] extra)
CI pytest workflow installs plugins via --extra plugins
Container builds succeed without plugins_rust/ directory
Plugin loading test (test_all_plugins_load_together) passes with correct count

🤖 Generated with Claude Code

Copilot

Pull request overview

This PR migrates the gateway’s rate limiter plugin from an in-tree Python implementation to the external cpex-rate-limiter PyPI package, updating configs/build artifacts accordingly and removing the old plugin code and its tests.

Changes:

Add cpex-rate-limiter>=0.0.2 as a new optional dependency group ([plugins]) and lock it in uv.lock.
Update plugin configuration files to reference cpex_rate_limiter.RateLimiterPlugin (and disable it in default/unit-test configs).
Install the new optional dependency group in Containerfile.lite and remove the in-tree plugins/rate_limiter implementation and related tests/docs.

Reviewed changes

Copilot reviewed 16 out of 17 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
`uv.lock`	Locks `cpex-rate-limiter` and adds the `plugins` extra.
`pyproject.toml`	Introduces optional dependency group `plugins` with `cpex-rate-limiter`.
`Containerfile.lite`	Installs `.[...,plugins]` so container builds include the package.
`plugins/config.yaml`	Switches rate limiter `kind` to `cpex_rate_limiter...` and disables by default.
`plugins/config-pii-guardian-policy.yaml`	Switches `kind` to external rate limiter.
`tests/unit/.../init_hooks_plugins_test.yaml`	Switches `kind` to external rate limiter and disables it for unit tests.
`tests/performance/plugins/config.yaml`	Switches `kind` to external rate limiter (still enabled for perf runs).
`plugins/webhook_notification/test_config.yaml`	Switches `kind` to external rate limiter for webhook testing config.
`docs/docs/using/plugins/plugins.md`	Updates rate limiter entry to point to external package.
`llms/plugins-llms.md`	Updates plugin list entry to `cpex-rate-limiter`.
`docs/docs/testing/unittest.md`	Removes rate limiter unit test from the unit test list.
`plugins/rate_limiter/*`	Removes in-tree plugin implementation, manifest, and README.
`tests/unit/.../rate_limiter/test_rate_limiter.py`	Removes in-tree rate limiter unit tests.
`tests/integration/test_rate_limiter.py`	Removes in-tree rate limiter integration tests.

Comments suppressed due to low confidence (1)

tests/performance/plugins/config.yaml:248

This performance config enables cpex_rate_limiter.RateLimiterPlugin (mode: permissive). Since cpex-rate-limiter is an optional extra, running tests/performance/test_plugins_performance.py will fail unless users have installed .[plugins]. Consider either disabling the rate limiter here by default or adding an explicit note in this config about the required extra/install command.

  # Rate limiter (fixed window, in-memory)
  - name: RateLimiterPlugin
    kind: cpex_rate_limiter.RateLimiterPlugin
    description: Per-user/tenant/tool rate limits
    version: 0.1.0
    author: Mihai Criveti
    hooks: [prompt_pre_fetch, tool_pre_invoke]
    tags: [limits, throttle]
    mode: permissive
    priority: 20

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

Copilot · 2026-04-01T06:25:55Z

-    kind: plugins.rate_limiter.rate_limiter.RateLimiterPlugin
+    kind: cpex_rate_limiter.RateLimiterPlugin
    description: Per-user/tenant/tool rate limits
    version: 0.1.0


The plugin version: 0.1.0 no longer matches the implementation referenced by kind: cpex_rate_limiter.RateLimiterPlugin (package is pinned to >=0.0.2). Since plugin version is surfaced via admin plugin APIs/UI, consider updating this value to the external package version (or omitting it like plugins/config.yaml does) to avoid misleading metadata.

Suggested change

version: 0.1.0

Copilot · 2026-04-01T06:25:55Z

-    kind: "plugins.rate_limiter.rate_limiter.RateLimiterPlugin"
+    kind: "cpex_rate_limiter.RateLimiterPlugin"
    description: "Per-user/tenant/tool rate limits"
    version: "0.1.0"


The plugin version: "0.1.0" is now inconsistent with kind: "cpex_rate_limiter.RateLimiterPlugin" (external package is >=0.0.2). Because plugin version is exposed in admin plugin APIs/UI, update this value to the external package version or remove it to avoid misleading operators.

Suggested change

version: "0.1.0"

Copilot · 2026-04-01T06:25:56Z

-    kind: "plugins.rate_limiter.rate_limiter.RateLimiterPlugin"
+    kind: "cpex_rate_limiter.RateLimiterPlugin"
    description: "Per-user/tenant/tool rate limits"
    version: "0.1.0"


This fixture still advertises version: "0.1.0" while kind now points to the external cpex_rate_limiter.RateLimiterPlugin (>=0.0.2). Even though the plugin is disabled in this test config, keeping the version accurate (or omitting it) helps avoid confusion when debugging plugin metadata or admin plugin listings.

Suggested change

version: "0.1.0"

Copilot · 2026-04-01T06:25:56Z

  # Rate Limiter Plugin (will generate rate limit violations)
  - name: "RateLimiter"
-    kind: "plugins.rate_limiter.rate_limiter.RateLimiterPlugin"
+    kind: "cpex_rate_limiter.RateLimiterPlugin"
    hooks: ["tool_pre_invoke"]
    mode: "permissive"  # Don't block for testing
    priority: 200


This dev test config now depends on the external cpex-rate-limiter package (kind: cpex_rate_limiter.RateLimiterPlugin). Add an explicit note here about installing the optional extra (e.g. uv pip install '.[plugins]') or set this plugin to mode: disabled by default, otherwise following the config file may fail with an import error in environments without that extra installed.

lucarlig · 2026-04-01T11:07:43Z

The import works fine on my side and I can see the plugin.

the author field, I think @gandhipratik203 is the author for the Rust version we’re uploading, although we might want to keep it as “Contextforge developers” for consistency.

One last thing: should we consider keeping the Python plugin as an example somewhere instead of removing it entirely?

gandhipratik203 · 2026-04-01T21:53:17Z

Hi @jonpspri @crivetimihai — thanks for putting this together. This is a good move — distributing plugins as pre-compiled PyPI packages is a well-established pattern (ruff, orjson, pydantic-core all do the same thing) and makes Rust acceleration accessible without requiring a toolchain on the gateway.

I'd like to help this move forward:

ADR — I can put together an ADR for the plugin extraction pattern since I have context on both the in-tree and package approaches. Would
cover the reasoning, pros/cons, effects on the build pipeline, testing strategy, and fallback behavior.
Test coverage — Happy to add integration tests, load tests, and algorithm-level validation to the cpex-plugins repo to make sure the
package has production-grade coverage.

araujof · 2026-04-10T00:23:46Z

@jonpspri I don't anticipate issues, but I think we should coordinate this PR with #3754.

Signed-off-by: lucarlig <luca.carlig@ibm.com>

lucarlig · 2026-04-10T09:39:56Z

Agreed with @gandhipratik203 on the plan:

Step 1
Move all plugins out of cpex-plugins (keeping plugin tests here) and freeze plugin code in that repo.

Step 2
Figure out the testing strategy and implement it (probably just clone contextforge repo in CI), then unfreeze plugin development.

For the old Python plugins, @jonpspri will spin up an example repo.

I’ll also add an ADR to document this.

Reasoning:
Right now, changes in cpex-plugins can unintentionally break CI in mcp-context-forge. Freezing the plugin code and decoupling it ensures that random or unrelated changes don’t cause CI failures in MCP, making builds more stable and predictable.

@brian.hussey @jonpspri — does this approach work?

Signed-off-by: lucarlig <luca.carlig@ibm.com>

lucarlig · 2026-04-10T13:52:30Z

@gandhipratik203 Short follow-up: I kept and updated the integration coverage for the packaged cpex_* plugins, fixed the shipped config/import drift, added gateway-side smoke coverage for the packaged plugin wiring, aligned the external package version metadata, and added ADR-048 to document the extraction order and repo split rationale. I also removed the gateway runtime dependency on cpex_retry_with_backoff internals and added targeted tests for the new retry parsing paths. Local full pytest, diff-cover, and doctests all passed after the follow-up fixes.

gandhipratik203

Thanks for turning these around so quickly — really appreciate the responsiveness, and the scope decisions feel right.

✅ Integration tests preserved (test_rate_limiter.py and test_encoded_exfil.py kept in-tree) — good call keeping them where they're already wired into CI.
✅ url_reputation pulled from the migration scope — appreciate the conservative approach here, much better to ship it when it's ready.
✅ ADR added (048-extract-rust-backed-plugins-first-and-preserve-python-examples.md) — nice to have the rationale documented for future contributors.
✅ Luca's plan to freeze plugin code in cpex-plugins and decouple CI is a clean way to stabilize things and addresses the testing concern well.

Approving — thanks again @lucarlig @jonpspri for the quick turnaround on this.

…r package Remove the in-tree rate_limiter plugin and replace it with the cpex-rate-limiter PyPI package, a compiled Rust extension providing the same RateLimiterPlugin class with additional algorithms (sliding-window, token-bucket) alongside the original fixed-window. - Add cpex-rate-limiter>=0.0.2 as a [plugins] optional dependency - Update Containerfile.lite to install the plugins extra - Remove plugins/rate_limiter/ source directory - Remove unit and integration tests that imported plugin internals - Update all config files to use cpex_rate_limiter.RateLimiterPlugin - Disable RateLimiterPlugin in test fixture config (package not available in unit test environment) - Update documentation to reflect the external package Signed-off-by: Jonathan Springer <jps@s390x.com> Signed-off-by: lucarlig <luca.carlig@ibm.com>

…ngine, benchmarks, and validation (#3809) * feat(rate-limiter): pluggable algorithms, tenant isolation fix, and scale load test - Add pluggable algorithm strategy: fixed_window, sliding_window, token_bucket - Add Redis backend for shared cross-instance rate limiting - Fix tenant isolation: skip by_tenant when tenant_id is None - Fix sliding window: sweep expired timestamps before counting - Fix backend validation: restore _validate_config check - Fix token bucket memory path: apply max(1,...) guard to reset timestamp - Add Redis integration tests for all three algorithms - Add direct regression tests for get_current_user tenant_id fallback - Add scale load test with Redis memory timeline and live algorithm detection - Add RL_PACE_MULTIPLIER for near-limit pace testing and boundary burst detection - Remove redundant algorithm locustfile; scale file is canonical - Correct stale comments and README limitations Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> * feat(rate-limiter): add Rust-backed engine, check() API, benchmarks, and validation - Rust-backed sliding window engine with pyo3-log integration - check() API with tenant propagation, sweep/retry-after support - Eliminate redundant ZRANGE in sliding window Lua script - Fix detect-secrets baseline for rate limiter load tests - Clarify memory backend is single-instance only in docs Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> * chore: regenerate detect-secrets baseline after rebase Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> * refactor(rate-limiter): review fixes, Redis hardening, key-format parity tests - Extract _dispatch_hook() shared by prompt_pre_fetch and tool_pre_invoke, reducing each hook to a single-line wrapper - Elevate Redis val_i64/val_f64 parse-error logging from warn to error so silent fail-open degradation surfaces in operator dashboards - Clamp sliding-window reset_timestamp with .max(1) so it is always strictly in the future even when the oldest entry expires in < 1 s - Add 5 s tokio::time::timeout around Redis connection establishment to prevent indefinite blocking on network partition - Replace silent except-pass in EVALSHA SHA tracking with logger.debug - Document dual Lua-script invariant (rolling-upgrade key-format parity) in both Python RedisBackend docstring and Rust redis_backend.rs header - Add 7 parametrized test_redis_key_format_parity_* tests validating that Python and Rust produce identical Redis keys for the same inputs - Revert unrelated .pyi stub changes for encoded_exfil_detection, pii_filter, retry_with_backoff, and secrets_detection Signed-off-by: Jonathan Springer <jps@s390x.com> * fix: strip trailing whitespace in pyi stubs, remove accidental .claude/ralph-loop.local.md - Remove plugins_rust/rate_limiter/.claude/ralph-loop.local.md which was accidentally committed — this is a local Claude Code loop state file and should never have been checked in. - Fix trailing whitespace in plugins_rust/rate_limiter/python/ rate_limiter_rust/__init__.pyi docstrings to pass pre-commit hooks. Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> * chore: regenerate detect-secrets baseline for new exfil test strings Update .secrets.baseline after adding test_extra_sensitive_keywords in plugins_rust/encoded_exfil_detection/src/lib.rs:969 which contains a fake credential string that triggers the Secret Keyword detector. All new entries are false positives (test data). Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> * chore: audit new detect-secrets baseline entries as false positives The baseline regeneration reset is_secret to null for entries whose line numbers shifted. Mark all 17 unaudited entries as is_secret=false (test data, example configs, fake credentials) to pass the --fail-on-unaudited pre-commit check. Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> --------- Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> Signed-off-by: Jonathan Springer <jps@s390x.com> Co-authored-by: Jonathan Springer <jps@s390x.com> Signed-off-by: lucarlig <luca.carlig@ibm.com>

…ation (#3839) Implement automatic tool discovery for upstream MCP servers via usage-aware adaptive polling. The gateway can now continuously synchronise tool lists from registered servers without manual intervention. Server classification (hot/cold): - Classify servers based on MCP session pool usage patterns - Hot servers (top 20% by recent usage): polled at 1x base interval - Cold servers (remaining 80%): polled at 3x base interval - Classification is deterministic: sorted by recency, active sessions, use count, and URL for tie-breaking - Leader election via Redis with TTL renewal for multi-worker coordination - Falls back to local-only operation without Redis Integration with GatewayService: - Health checks respect hot/cold classification intervals - Auto-refresh of tools/resources/prompts respects classification - Fail-open on classification errors (poll anyway) - Poll timestamps tracked via Redis with TTL expiry - Uses base gateway URL (pre-auth) for classification lookups to avoid leaking query-param auth secrets to Redis Configuration: - AUTO_REFRESH_SERVERS=true enables automatic tool sync (default: false) - GATEWAY_AUTO_REFRESH_INTERVAL=300 sets base polling interval - HOT_COLD_CLASSIFICATION_ENABLED=false (opt-in, requires Redis) Includes comprehensive tests with 100% coverage on the new ServerClassificationService and integration tests for the GatewayService hot/cold polling paths. Closes #3734 Signed-off-by: Lang-Akshay <akshay.shinde26@ibm.com> Signed-off-by: Mihai Criveti <crivetimihai@gmail.com> Signed-off-by: lucarlig <luca.carlig@ibm.com>

…r package Remove the in-tree rate_limiter plugin and replace it with the cpex-rate-limiter PyPI package, a compiled Rust extension providing the same RateLimiterPlugin class with additional algorithms (sliding-window, token-bucket) alongside the original fixed-window. - Add cpex-rate-limiter>=0.0.2 as a [plugins] optional dependency - Update Containerfile.lite to install the plugins extra - Remove plugins/rate_limiter/ source directory - Remove unit and integration tests that imported plugin internals - Update all config files to use cpex_rate_limiter.RateLimiterPlugin - Disable RateLimiterPlugin in test fixture config (package not available in unit test environment) - Update documentation to reflect the external package Signed-off-by: Jonathan Springer <jps@s390x.com> Signed-off-by: lucarlig <luca.carlig@ibm.com>

Remove all plugins_rust/ build infrastructure and update references across Containerfiles, Makefile, CI workflows, pre-commit configs, CODEOWNERS, and documentation to reflect that plugins are now distributed as PyPI packages (cpex-*) via the [plugins] optional extra. - Remove Rust plugin builder stages from all Containerfiles - Remove ~100 lines of rust-* plugin Makefile targets (keep mcp-runtime) - Add --extra plugins to CI pytest workflow - Add [plugins] extra to install-dev Makefile target - Update tool_service.py import to use cpex_retry_with_backoff - Update plugin kind paths in 7 doc files to cpex_pii_filter.* - Clean up pre-commit, CODEOWNERS, MANIFEST.in, whitesource, .gitignore Signed-off-by: Jonathan Springer <jps@s390x.com> Signed-off-by: lucarlig <luca.carlig@ibm.com>

Round 1 (blockers + high): - Restore exclude-newer = "10 days" in pyproject.toml; replace stale langchain/requests pins with cpex-* per-package overrides anchored to 2026-04-09 so the plugins resolve newer than the global window - Guard cpex_retry_with_backoff import in tool_service.py with try/except ImportError; falls back to (None, True) for the Python pipeline when the optional [plugins] extra is not installed - Delete orphaned .github/workflows/rust-plugins.yml and the associated test cases in tests/unit/test_rust_plugins_workflow.py; drop the workflow card from docs/docs/architecture/explorer.html - Delete orphaned docs/docs/using/plugins/rust-plugins.md and remove it from docs/docs/using/plugins/.pages mkdocs nav - Harden docker-entrypoint.sh install_plugin_requirements: canonicalize /app and the resolved requirements path with readlink -f and require the path to live under /app/, log non-comment lines from the requirements file before pip runs, and skip cleanly on validation failure - Delete PLUGIN-MIGRATION-PLAN.md (one-time planning doc) - Add COPY plugins/requirements.txt to Containerfile.scratch (the layered Containerfile.lite already had it; the broad COPY . in Containerfile already includes it) Round 2 (medium + low): - Bump cpex-* version pin floors in pyproject.toml [plugins] to match resolved versions in uv.lock (cpex-rate-limiter>=0.0.3, cpex-encoded-exfil-detection>=0.2.0, cpex-pii-filter>=0.2.0, cpex-url-reputation>=0.1.1) - Add Prerequisites section to tests/performance/PLUGIN_PROFILING.md documenting the [plugins] extra requirement - Add Status: Partially superseded note to ADR-041 explaining that plugins_rust/ was removed when in-tree Rust plugins migrated to PyPI packages - Document upgrade semantics in plugins/requirements.txt header (pip without --upgrade skips already-satisfied constraints) - Add importlib.util.find_spec() precheck to tests/performance/test_plugins_performance.py main(); the script now skips cleanly with an actionable message if any of the five cpex packages referenced by the perf config are missing - Rename tests/unit/test_rust_plugins_workflow.py to test_go_toolchain_pinning.py to match its remaining contents (Go workflow pin and Makefile toolchain assertion) Follow-ups tracked in #4116 and IBM/cpex-plugins#21 for the longer-term tool_service.py refactor that will eliminate the cross-package import entirely. Signed-off-by: Jonathan Springer <jps@s390x.com> Signed-off-by: lucarlig <luca.carlig@ibm.com>

Signed-off-by: lucarlig <luca.carlig@ibm.com>

brian-hussey

Approval as @lucarlig and @gandhipratik203 have previously approved.

…3965) * refactor(plugins): replace in-tree rate_limiter with cpex-rate-limiter package Remove the in-tree rate_limiter plugin and replace it with the cpex-rate-limiter PyPI package, a compiled Rust extension providing the same RateLimiterPlugin class with additional algorithms (sliding-window, token-bucket) alongside the original fixed-window. - Add cpex-rate-limiter>=0.0.2 as a [plugins] optional dependency - Update Containerfile.lite to install the plugins extra - Remove plugins/rate_limiter/ source directory - Remove unit and integration tests that imported plugin internals - Update all config files to use cpex_rate_limiter.RateLimiterPlugin - Disable RateLimiterPlugin in test fixture config (package not available in unit test environment) - Update documentation to reflect the external package Signed-off-by: Jonathan Springer <jps@s390x.com> Signed-off-by: lucarlig <luca.carlig@ibm.com> * feat(rate-limiter): pluggable algorithms with Rust-backed execution engine, benchmarks, and validation (#3809) * feat(rate-limiter): pluggable algorithms, tenant isolation fix, and scale load test - Add pluggable algorithm strategy: fixed_window, sliding_window, token_bucket - Add Redis backend for shared cross-instance rate limiting - Fix tenant isolation: skip by_tenant when tenant_id is None - Fix sliding window: sweep expired timestamps before counting - Fix backend validation: restore _validate_config check - Fix token bucket memory path: apply max(1,...) guard to reset timestamp - Add Redis integration tests for all three algorithms - Add direct regression tests for get_current_user tenant_id fallback - Add scale load test with Redis memory timeline and live algorithm detection - Add RL_PACE_MULTIPLIER for near-limit pace testing and boundary burst detection - Remove redundant algorithm locustfile; scale file is canonical - Correct stale comments and README limitations Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> * feat(rate-limiter): add Rust-backed engine, check() API, benchmarks, and validation - Rust-backed sliding window engine with pyo3-log integration - check() API with tenant propagation, sweep/retry-after support - Eliminate redundant ZRANGE in sliding window Lua script - Fix detect-secrets baseline for rate limiter load tests - Clarify memory backend is single-instance only in docs Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> * chore: regenerate detect-secrets baseline after rebase Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> * refactor(rate-limiter): review fixes, Redis hardening, key-format parity tests - Extract _dispatch_hook() shared by prompt_pre_fetch and tool_pre_invoke, reducing each hook to a single-line wrapper - Elevate Redis val_i64/val_f64 parse-error logging from warn to error so silent fail-open degradation surfaces in operator dashboards - Clamp sliding-window reset_timestamp with .max(1) so it is always strictly in the future even when the oldest entry expires in < 1 s - Add 5 s tokio::time::timeout around Redis connection establishment to prevent indefinite blocking on network partition - Replace silent except-pass in EVALSHA SHA tracking with logger.debug - Document dual Lua-script invariant (rolling-upgrade key-format parity) in both Python RedisBackend docstring and Rust redis_backend.rs header - Add 7 parametrized test_redis_key_format_parity_* tests validating that Python and Rust produce identical Redis keys for the same inputs - Revert unrelated .pyi stub changes for encoded_exfil_detection, pii_filter, retry_with_backoff, and secrets_detection Signed-off-by: Jonathan Springer <jps@s390x.com> * fix: strip trailing whitespace in pyi stubs, remove accidental .claude/ralph-loop.local.md - Remove plugins_rust/rate_limiter/.claude/ralph-loop.local.md which was accidentally committed — this is a local Claude Code loop state file and should never have been checked in. - Fix trailing whitespace in plugins_rust/rate_limiter/python/ rate_limiter_rust/__init__.pyi docstrings to pass pre-commit hooks. Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> * chore: regenerate detect-secrets baseline for new exfil test strings Update .secrets.baseline after adding test_extra_sensitive_keywords in plugins_rust/encoded_exfil_detection/src/lib.rs:969 which contains a fake credential string that triggers the Secret Keyword detector. All new entries are false positives (test data). Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> * chore: audit new detect-secrets baseline entries as false positives The baseline regeneration reset is_secret to null for entries whose line numbers shifted. Mark all 17 unaudited entries as is_secret=false (test data, example configs, fake credentials) to pass the --fail-on-unaudited pre-commit check. Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> --------- Signed-off-by: Pratik Gandhi <gandhipratik203@gmail.com> Signed-off-by: Jonathan Springer <jps@s390x.com> Co-authored-by: Jonathan Springer <jps@s390x.com> Signed-off-by: lucarlig <luca.carlig@ibm.com> * feat(discovery): add automatic tool discovery with hot/cold classification (#3839) Implement automatic tool discovery for upstream MCP servers via usage-aware adaptive polling. The gateway can now continuously synchronise tool lists from registered servers without manual intervention. Server classification (hot/cold): - Classify servers based on MCP session pool usage patterns - Hot servers (top 20% by recent usage): polled at 1x base interval - Cold servers (remaining 80%): polled at 3x base interval - Classification is deterministic: sorted by recency, active sessions, use count, and URL for tie-breaking - Leader election via Redis with TTL renewal for multi-worker coordination - Falls back to local-only operation without Redis Integration with GatewayService: - Health checks respect hot/cold classification intervals - Auto-refresh of tools/resources/prompts respects classification - Fail-open on classification errors (poll anyway) - Poll timestamps tracked via Redis with TTL expiry - Uses base gateway URL (pre-auth) for classification lookups to avoid leaking query-param auth secrets to Redis Configuration: - AUTO_REFRESH_SERVERS=true enables automatic tool sync (default: false) - GATEWAY_AUTO_REFRESH_INTERVAL=300 sets base polling interval - HOT_COLD_CLASSIFICATION_ENABLED=false (opt-in, requires Redis) Includes comprehensive tests with 100% coverage on the new ServerClassificationService and integration tests for the GatewayService hot/cold polling paths. Closes #3734 Signed-off-by: Lang-Akshay <akshay.shinde26@ibm.com> Signed-off-by: Mihai Criveti <crivetimihai@gmail.com> Signed-off-by: lucarlig <luca.carlig@ibm.com> * refactor(plugins): replace in-tree rate_limiter with cpex-rate-limiter package Remove the in-tree rate_limiter plugin and replace it with the cpex-rate-limiter PyPI package, a compiled Rust extension providing the same RateLimiterPlugin class with additional algorithms (sliding-window, token-bucket) alongside the original fixed-window. - Add cpex-rate-limiter>=0.0.2 as a [plugins] optional dependency - Update Containerfile.lite to install the plugins extra - Remove plugins/rate_limiter/ source directory - Remove unit and integration tests that imported plugin internals - Update all config files to use cpex_rate_limiter.RateLimiterPlugin - Disable RateLimiterPlugin in test fixture config (package not available in unit test environment) - Update documentation to reflect the external package Signed-off-by: Jonathan Springer <jps@s390x.com> Signed-off-by: lucarlig <luca.carlig@ibm.com> * refactor(plugins): update build, CI, and docs for PyPI plugin migration Remove all plugins_rust/ build infrastructure and update references across Containerfiles, Makefile, CI workflows, pre-commit configs, CODEOWNERS, and documentation to reflect that plugins are now distributed as PyPI packages (cpex-*) via the [plugins] optional extra. - Remove Rust plugin builder stages from all Containerfiles - Remove ~100 lines of rust-* plugin Makefile targets (keep mcp-runtime) - Add --extra plugins to CI pytest workflow - Add [plugins] extra to install-dev Makefile target - Update tool_service.py import to use cpex_retry_with_backoff - Update plugin kind paths in 7 doc files to cpex_pii_filter.* - Clean up pre-commit, CODEOWNERS, MANIFEST.in, whitesource, .gitignore Signed-off-by: Jonathan Springer <jps@s390x.com> Signed-off-by: lucarlig <luca.carlig@ibm.com> * fix(plugins): address PR review findings on PyPI plugin migration Round 1 (blockers + high): - Restore exclude-newer = "10 days" in pyproject.toml; replace stale langchain/requests pins with cpex-* per-package overrides anchored to 2026-04-09 so the plugins resolve newer than the global window - Guard cpex_retry_with_backoff import in tool_service.py with try/except ImportError; falls back to (None, True) for the Python pipeline when the optional [plugins] extra is not installed - Delete orphaned .github/workflows/rust-plugins.yml and the associated test cases in tests/unit/test_rust_plugins_workflow.py; drop the workflow card from docs/docs/architecture/explorer.html - Delete orphaned docs/docs/using/plugins/rust-plugins.md and remove it from docs/docs/using/plugins/.pages mkdocs nav - Harden docker-entrypoint.sh install_plugin_requirements: canonicalize /app and the resolved requirements path with readlink -f and require the path to live under /app/, log non-comment lines from the requirements file before pip runs, and skip cleanly on validation failure - Delete PLUGIN-MIGRATION-PLAN.md (one-time planning doc) - Add COPY plugins/requirements.txt to Containerfile.scratch (the layered Containerfile.lite already had it; the broad COPY . in Containerfile already includes it) Round 2 (medium + low): - Bump cpex-* version pin floors in pyproject.toml [plugins] to match resolved versions in uv.lock (cpex-rate-limiter>=0.0.3, cpex-encoded-exfil-detection>=0.2.0, cpex-pii-filter>=0.2.0, cpex-url-reputation>=0.1.1) - Add Prerequisites section to tests/performance/PLUGIN_PROFILING.md documenting the [plugins] extra requirement - Add Status: Partially superseded note to ADR-041 explaining that plugins_rust/ was removed when in-tree Rust plugins migrated to PyPI packages - Document upgrade semantics in plugins/requirements.txt header (pip without --upgrade skips already-satisfied constraints) - Add importlib.util.find_spec() precheck to tests/performance/test_plugins_performance.py main(); the script now skips cleanly with an actionable message if any of the five cpex packages referenced by the perf config are missing - Rename tests/unit/test_rust_plugins_workflow.py to test_go_toolchain_pinning.py to match its remaining contents (Go workflow pin and Makefile toolchain assertion) Follow-ups tracked in #4116 and IBM/cpex-plugins#21 for the longer-term tool_service.py refactor that will eliminate the cross-package import entirely. Signed-off-by: Jonathan Springer <jps@s390x.com> Signed-off-by: lucarlig <luca.carlig@ibm.com> * revert: restore tests changes from PR #3965 Signed-off-by: lucarlig <luca.carlig@ibm.com> * fix(ci): align plugin tests with PyPI migration Signed-off-by: lucarlig <luca.carlig@ibm.com> * test: remove legacy plugin test skip infrastructure Signed-off-by: lucarlig <luca.carlig@ibm.com> * test: align packaged plugin tests with rust shims Signed-off-by: lucarlig <luca.carlig@ibm.com> * test: cover retry policy import path in tool service Signed-off-by: lucarlig <luca.carlig@ibm.com> * fix: harden cpex plugin migration paths Signed-off-by: lucarlig <luca.carlig@ibm.com> * test: cover retry policy parser branches Signed-off-by: lucarlig <luca.carlig@ibm.com> * test: cover plugin requirements entrypoint path Signed-off-by: lucarlig <luca.carlig@ibm.com> --------- Signed-off-by: lucarlig <luca.carlig@ibm.com> Signed-off-by: Jonathan Springer <jps@s390x.com> Co-authored-by: Pratik Gandhi <gandhipratik203@gmail.com> Co-authored-by: Lang-Akshay <akshay.shinde26@ibm.com> Co-authored-by: lucarlig <luca.carlig@ibm.com>

jonpspri requested review from araujof, crivetimihai, kevalmahajan, madhav165 and terylt as code owners April 1, 2026 06:20

Copilot AI review requested due to automatic review settings April 1, 2026 06:20

Copilot started reviewing on behalf of jonpspri April 1, 2026 06:21 View session

Copilot AI reviewed Apr 1, 2026

View reviewed changes

jonpspri force-pushed the jps-plugin-pypi branch 2 times, most recently from a0061cb to bed1dc6 Compare April 1, 2026 10:14

jonpspri mentioned this pull request Apr 1, 2026

feat(rate-limiter): pluggable algorithms with Rust-backed execution engine, benchmarks, and validation #3809

Merged

jonpspri force-pushed the jps-plugin-pypi branch from bed1dc6 to f101a18 Compare April 1, 2026 10:41

lucarlig previously approved these changes Apr 1, 2026

View reviewed changes

jonpspri marked this pull request as draft April 9, 2026 22:13

jonpspri dismissed lucarlig’s stale review via 6831640 April 10, 2026 00:06

jonpspri changed the title ~~refactor(plugins): replace in-tree rate_limiter with cpex-rate-limiter package~~ refactor(plugins): migrate in-tree plugins to PyPI packages (cpex-*) Apr 10, 2026

This was referenced Apr 10, 2026

[CHORE]: Refactor tool_service.py to remove direct cpex_retry_with_backoff import #4116

Open

Add to_rust_native_policy() method to RetryWithBackoffPlugin IBM/cpex-plugins#21

Open

jonpspri force-pushed the jps-plugin-pypi branch from 6831640 to 1ee8234 Compare April 10, 2026 07:05

jonpspri marked this pull request as ready for review April 10, 2026 07:10

jonpspri requested a review from dima-zakharov as a code owner April 10, 2026 07:10

jonpspri assigned lucarlig and gandhipratik203 Apr 10, 2026

lucarlig added a commit that referenced this pull request Apr 10, 2026

revert: restore tests changes from PR #3965

171c80e

Signed-off-by: lucarlig <luca.carlig@ibm.com>

lucarlig added a commit that referenced this pull request Apr 10, 2026

revert: restore tests changes from PR #3965

aafb714

Signed-off-by: lucarlig <luca.carlig@ibm.com>

lucarlig force-pushed the jps-plugin-pypi branch from 171c80e to e3ab892 Compare April 10, 2026 09:52

lucarlig force-pushed the jps-plugin-pypi branch 2 times, most recently from e7163a0 to 751be5b Compare April 10, 2026 10:33

lucarlig requested a review from gandhipratik203 April 10, 2026 13:47

gandhipratik203 previously approved these changes Apr 10, 2026

View reviewed changes

brian-hussey self-assigned this Apr 10, 2026

lucarlig dismissed gandhipratik203’s stale review via e63b14f April 10, 2026 14:05

jonpspri and others added 14 commits April 10, 2026 15:05

revert: restore tests changes from PR #3965

06fdce4

Signed-off-by: lucarlig <luca.carlig@ibm.com>

fix(ci): align plugin tests with PyPI migration

863e3b1

Signed-off-by: lucarlig <luca.carlig@ibm.com>

test: remove legacy plugin test skip infrastructure

479b921

Signed-off-by: lucarlig <luca.carlig@ibm.com>

test: align packaged plugin tests with rust shims

8f10d0d

Signed-off-by: lucarlig <luca.carlig@ibm.com>

test: cover retry policy import path in tool service

284829f

Signed-off-by: lucarlig <luca.carlig@ibm.com>

fix: harden cpex plugin migration paths

a89c599

Signed-off-by: lucarlig <luca.carlig@ibm.com>

test: cover retry policy parser branches

3f6559a

Signed-off-by: lucarlig <luca.carlig@ibm.com>

test: cover plugin requirements entrypoint path

fc5a959

Signed-off-by: lucarlig <luca.carlig@ibm.com>

lucarlig force-pushed the jps-plugin-pypi branch from e63b14f to fc5a959 Compare April 10, 2026 14:05

brian-hussey approved these changes Apr 10, 2026

View reviewed changes

brian-hussey merged commit ca4957a into main Apr 10, 2026
27 checks passed

brian-hussey deleted the jps-plugin-pypi branch April 10, 2026 14:58

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

refactor(plugins): migrate in-tree plugins to PyPI packages (cpex-*)#3965

refactor(plugins): migrate in-tree plugins to PyPI packages (cpex-*)#3965
brian-hussey merged 14 commits intomainfrom
jps-plugin-pypi

jonpspri commented Apr 1, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 1, 2026

Uh oh!

Copilot AI Apr 1, 2026

Uh oh!

Copilot AI Apr 1, 2026

Uh oh!

Copilot AI Apr 1, 2026

Uh oh!

lucarlig commented Apr 1, 2026

Uh oh!

gandhipratik203 commented Apr 1, 2026 •

edited

Loading

Uh oh!

araujof commented Apr 10, 2026

Uh oh!

lucarlig commented Apr 10, 2026

Uh oh!

lucarlig commented Apr 10, 2026 •

edited

Loading

Uh oh!

gandhipratik203 left a comment

Uh oh!

brian-hussey left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

Conversation

jonpspri commented Apr 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

lucarlig commented Apr 1, 2026

Uh oh!

gandhipratik203 commented Apr 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

araujof commented Apr 10, 2026

Uh oh!

lucarlig commented Apr 10, 2026

Uh oh!

lucarlig commented Apr 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

gandhipratik203 left a comment

Choose a reason for hiding this comment

Uh oh!

brian-hussey left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

jonpspri commented Apr 1, 2026 •

edited

Loading

gandhipratik203 commented Apr 1, 2026 •

edited

Loading

lucarlig commented Apr 10, 2026 •

edited

Loading