mcp-openstack-ops 3.6.1__tar.gz → 3.6.2__tar.gz

{mcp_openstack_ops-3.6.1 → mcp_openstack_ops-3.6.2}/AGENTS.md

@@ -33,10 +33,12 @@ consistent with the resulting implementation and actual verification.
   Label targeted retests as partial; do not extrapolate them to the full toolset.
   Document polling limits and distinguish accepted requests from confirmed final
   resource states. Explain changes in test conditions that affect comparisons.
-- Keep dated reports as historical evidence and link the latest applicable
-  report from README.md. Reports must state verification limits, observed
-  failures, and cleanup results without exposing credentials, private endpoints,
-  or sensitive resource data.
+- Keep `docs/` for reusable user guidance only. Do not create or retain
+  issue-specific reports, dated test results, or evidence JSON in `docs/`, or
+  embed those records in usage guides. Report case-specific verification results,
+  limits, failures, and cleanup in chat unless the user explicitly requests a
+  separate artifact. Never expose credentials, private endpoints, or sensitive
+  resource data. Do not add report links to README.md.
 - Before completion, check summary arithmetic, percentages, tool/action names,
   Status/Notes, source identity, and evidence links for agreement. If README.ko.md
   exists, maintain equivalent information in both README files.

{mcp_openstack_ops-3.6.1 → mcp_openstack_ops-3.6.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mcp-openstack-ops
-Version: 3.6.1
+Version: 3.6.2
 Summary: OpenStack operations automation MCP server
 Author-email: JungJungIn <call518@gmail.com>
 Requires-Python: >=3.12
@@ -42,7 +42,7 @@ Dynamic: license-file
 ## Features
 
 Experimental per-user remote connections are available for testing with the
-opt-in `--multi-user` mode. See the [phase 1 setup and validation guide](docs/multi-user-phase1.md)
+opt-in `--multi-user` mode. See the [multi-user setup and usage guide](docs/multi-user-guide.md)
 for authentication, client limitations, and the test checklist. Administrator
 cross-project queries are outside this phase.
 
@@ -183,80 +183,12 @@ New consolidated `get_instance` tool replaces multiple separate tools:
 
 ## 📊 OpenStack CLI vs MCP Tools Mapping
 
-**Latest targeted validation — 2026-11-02:** working tree based on `6a36b6e`.
-See the [fixes and final retest report](docs/expanded-validation-fixes-2026-11-02.md)
-and [sanitized evidence](docs/expanded-validation-fixes-2026-11-02.json).
-The final executable paths were tested with FastMCP **3.4.10** from `uv.lock`,
-two ordinary-user projects, **20 VMs and 20 one-GiB volumes**. Post-test tool
-wording/prompt corrections are identified separately in the report's source
-fingerprints; Claude replay remains unverified.
-
-| Verification group | Item | Result |
-| --- | --- | --- |
-| Live tool checks | Unique tools retested in limited scenarios | **14/14 (100%)**, 0 failed/unverified within this scope |
-| Live tool checks | Registered inventory not repeated in this final run | **91 of 105** tools; historical evidence below is not a current full sweep |
-| Named live harness cases | Fixture, isolation, lifecycle and grouped load checks | **101/102**, one Cinder storage-permission fixture failure, 0 skipped |
-| Local unit tests | Locked environment, `python -m pytest -q tests` | **199/199 (100%)**, 0 failed/skipped; 21 added regressions |
-| Bounded load requests | Shared-server, two-project bursts | **14/14**, actual overlap up to 8 |
-| Bounded load requests | Same shared server, 120 seconds, four workers | **209/209**, median 2.03 s, p95 2.70 s, max 3.24 s |
-| Backup recovery | After operator corrected one NFS directory permission | **6/6 fresh backups available**, both reader MCP state checks passed |
-| Claude client | Minimal actual model API preflight | Blocked before tools: credit/balance error; no final-source Claude replay |
-
-Network creation now uses the backend's default MTU when omitted; automatic
-1450 and explicit 1400 were verified. Both `set_volume` and `set_server_volume`
-attach/detach reached final states, with cross-project requests rejected. Both
-search wrappers now honor offset/case sensitivity and correct `has_more`.
-Default write restrictions remain in place. These successes do not certify all
-actions of those tools or all 105 registered tools.
-
-The Cinder fixture failure occurred through its API, independently of MCP. The
-operator corrected the failing directory's group-write permission; the original
-failure remains recorded alongside the successful targeted recovery. Actual
-Claude same-URL duplicate registration was not established as a product defect;
-single-entry/distinct-URL attempts also failed before tools. Restoring model API
-access is required for that controlled comparison.
-
-Earlier same-day results are consolidated in the final report's
-[sanitized evidence](docs/expanded-validation-fixes-2026-11-02.json), under
-`historical_evidence.earlier_runs`. The previous 33-tool classification and
-41 actual Claude template attempts belong to the pre-fix source, not this
-final source. The complete 105-tool snapshot below is also historical.
-
-**Validation snapshot — 2026-10-30** (main working tree based on `dd98993`,
-with the uncommitted fixes described in the report):
-
-<table border="1" rules="cols">
-  <thead>
-    <tr><th>Verification type</th><th>Item</th><th>Result</th></tr>
-  </thead>
-  <tbody>
-    <tr><td rowspan="5">Live tool checks</td><td>Registered tools invoked and classified</td><td><strong>105</strong></td></tr>
-    <tr><td>Passed within the exercised live scenarios</td><td><strong>74</strong></td></tr>
-    <tr><td>Implemented, live unverified: service/API unavailable</td><td><strong>28</strong></td></tr>
-    <tr><td>Intentional capability limitations</td><td><strong>3</strong></td></tr>
-    <tr><td>Reproduced code defects / state-check timeouts in the retest</td><td><strong>0 / 0</strong></td></tr>
-    <tr><td>Local unit tests</td><td>Test cases in <code>tests/test_*.py</code></td><td><strong>160/160 (100%)</strong></td></tr>
-  </tbody>
-</table>
-
-State polling allowed 180 seconds; soft reboot completed in 131.66 seconds
-(previously it exceeded the 90-second check). No reboot implementation change
-was needed.
-
-The 10 previously reproduced tool defects were corrected. The 28 unavailable
-paths are **not live-validated successes**: 24 Octavia tools, two Neutron QoS
-tools, one Aodh tool, and one Neutron port-forwarding tool. Each affected table
-row is marked **⚠️ Live unverified**. An absent tool/action is marked **❌ Not
-implemented** rather than presented as an available feature.
-
-✅ means an implemented API path; it does not certify every action or backend.
-The live-pass count covers only the scenarios in the
-[latest retest report](docs/main-live-retest-2026-10-30.md), with
-[sanitized evidence](docs/main-live-retest-2026-10-30.json). API policy, asynchronous
-completion, production HTTPS/client compatibility, load, and guest I/O limits
-are stated there. All test resources were cleaned up; the final audit found zero
-remaining test resources across 24 categories. See also
-[API prerequisites and result semantics](#api-prerequisites-and-result-semantics).
+✅ denotes an implemented API path, not certification of every action or
+OpenStack deployment. **⚠️ Live unverified** identifies paths without successful
+live validation; **❌ Not implemented** identifies unavailable capabilities.
+Service availability, permissions, and asynchronous completion affect results.
+See [API prerequisites and result semantics](#api-prerequisites-and-result-semantics)
+and the [multi-user setup guide](docs/multi-user-guide.md).
 
 **Detailed Mapping by Category**
 
@@ -302,7 +234,7 @@ remaining test resources across 24 categories. See also
 | `openstack server set` | `set_server_properties` (action="set") | ✅ | Server property setting |
 | `openstack server unset` | `set_server_properties` (action="unset") | ✅ | Server property unsetting |
 | `openstack server dump create` | `set_server_dump` | ❌ Unsupported | Returns an explicit unsupported result; no dump is created |
-| `openstack server event list` | `get_server_events` | ✅ | Server event tracking |
+| `openstack server event list` | `get_server_events` | ✅ | Pass the server name or UUID via `instance_name`, not `server_id`; event visibility follows Nova policy |
 | `openstack server group list` | `get_server_groups` | ✅ | Server group listing |
 | `openstack server group create/delete` | `set_server_group` | ✅ | Server group management |
 | `openstack flavor list` | `set_flavor` (action="list") | ✅ | Requires modification tools to be enabled for registration |
@@ -542,7 +474,7 @@ remaining test resources across 24 categories. See also
 
 ### API prerequisites and result semantics
 
-The feature table marks implemented API paths, not verification against every OpenStack deployment. The management and query paths have local mocked API tests. Selected paths also passed live tests on a single-node Kolla Ansible deployment; see [integration validation](docs/integration-validation.md) for coverage and skipped operations. A missing service, extension, required microversion, or permission is an operation failure, not an empty successful result.
+The feature table marks implemented API paths, not verification against every OpenStack deployment. Service and policy requirements must be checked in the target deployment. A missing service, extension, required microversion, or permission is an operation failure, not an empty successful result.
 
 - **Modification tools:** the MCP server registers `set_*` management tools only when `ALLOW_MODIFY_OPERATIONS=true`. Existing action guards also reject writes when disabled. Dedicated read-only `get_*` tools remain available. API authorization is enforced by the target cloud; agent, user, domain, type and QoS administration may require elevated OpenStack roles.
 - **Alarms:** `set_alarms` calls the Keystone catalog's `alarming` endpoint for the configured interface/region. It supports list/create/show/delete of project alarms and creates `gnocchi_resources_threshold` alarms. Aodh, Gnocchi, the specified resource and a collected metric are required. The default metric is `cpu_util`, which is not collected by every deployment. `metric`, `resource_type`, `aggregation_method`, `granularity`, `evaluation_periods`, and `alarm_actions` can be supplied. An empty action list creates an alarm without notifications. There is no simulation fallback.
@@ -744,7 +676,7 @@ unset OPENSTACK_MCP_TOKEN
 ```
 
 This Bearer configuration uses the server's configured OpenStack account. For
-experimental per-user authentication, see the [phase 1 guide](docs/multi-user-phase1.md).
+experimental per-user authentication, see the [multi-user guide](docs/multi-user-guide.md).
 See also the [Claude Code MCP documentation](https://code.claude.com/docs/en/mcp).
 
 ## Server Configuration

{mcp_openstack_ops-3.6.1 → mcp_openstack_ops-3.6.2}/README.md

@@ -24,7 +24,7 @@
 ## Features
 
 Experimental per-user remote connections are available for testing with the
-opt-in `--multi-user` mode. See the [phase 1 setup and validation guide](docs/multi-user-phase1.md)
+opt-in `--multi-user` mode. See the [multi-user setup and usage guide](docs/multi-user-guide.md)
 for authentication, client limitations, and the test checklist. Administrator
 cross-project queries are outside this phase.
 
@@ -165,80 +165,12 @@ New consolidated `get_instance` tool replaces multiple separate tools:
 
 ## 📊 OpenStack CLI vs MCP Tools Mapping
 
-**Latest targeted validation — 2026-11-02:** working tree based on `6a36b6e`.
-See the [fixes and final retest report](docs/expanded-validation-fixes-2026-11-02.md)
-and [sanitized evidence](docs/expanded-validation-fixes-2026-11-02.json).
-The final executable paths were tested with FastMCP **3.4.10** from `uv.lock`,
-two ordinary-user projects, **20 VMs and 20 one-GiB volumes**. Post-test tool
-wording/prompt corrections are identified separately in the report's source
-fingerprints; Claude replay remains unverified.
-
-| Verification group | Item | Result |
-| --- | --- | --- |
-| Live tool checks | Unique tools retested in limited scenarios | **14/14 (100%)**, 0 failed/unverified within this scope |
-| Live tool checks | Registered inventory not repeated in this final run | **91 of 105** tools; historical evidence below is not a current full sweep |
-| Named live harness cases | Fixture, isolation, lifecycle and grouped load checks | **101/102**, one Cinder storage-permission fixture failure, 0 skipped |
-| Local unit tests | Locked environment, `python -m pytest -q tests` | **199/199 (100%)**, 0 failed/skipped; 21 added regressions |
-| Bounded load requests | Shared-server, two-project bursts | **14/14**, actual overlap up to 8 |
-| Bounded load requests | Same shared server, 120 seconds, four workers | **209/209**, median 2.03 s, p95 2.70 s, max 3.24 s |
-| Backup recovery | After operator corrected one NFS directory permission | **6/6 fresh backups available**, both reader MCP state checks passed |
-| Claude client | Minimal actual model API preflight | Blocked before tools: credit/balance error; no final-source Claude replay |
-
-Network creation now uses the backend's default MTU when omitted; automatic
-1450 and explicit 1400 were verified. Both `set_volume` and `set_server_volume`
-attach/detach reached final states, with cross-project requests rejected. Both
-search wrappers now honor offset/case sensitivity and correct `has_more`.
-Default write restrictions remain in place. These successes do not certify all
-actions of those tools or all 105 registered tools.
-
-The Cinder fixture failure occurred through its API, independently of MCP. The
-operator corrected the failing directory's group-write permission; the original
-failure remains recorded alongside the successful targeted recovery. Actual
-Claude same-URL duplicate registration was not established as a product defect;
-single-entry/distinct-URL attempts also failed before tools. Restoring model API
-access is required for that controlled comparison.
-
-Earlier same-day results are consolidated in the final report's
-[sanitized evidence](docs/expanded-validation-fixes-2026-11-02.json), under
-`historical_evidence.earlier_runs`. The previous 33-tool classification and
-41 actual Claude template attempts belong to the pre-fix source, not this
-final source. The complete 105-tool snapshot below is also historical.
-
-**Validation snapshot — 2026-10-30** (main working tree based on `dd98993`,
-with the uncommitted fixes described in the report):
-
-<table border="1" rules="cols">
-  <thead>
-    <tr><th>Verification type</th><th>Item</th><th>Result</th></tr>
-  </thead>
-  <tbody>
-    <tr><td rowspan="5">Live tool checks</td><td>Registered tools invoked and classified</td><td><strong>105</strong></td></tr>
-    <tr><td>Passed within the exercised live scenarios</td><td><strong>74</strong></td></tr>
-    <tr><td>Implemented, live unverified: service/API unavailable</td><td><strong>28</strong></td></tr>
-    <tr><td>Intentional capability limitations</td><td><strong>3</strong></td></tr>
-    <tr><td>Reproduced code defects / state-check timeouts in the retest</td><td><strong>0 / 0</strong></td></tr>
-    <tr><td>Local unit tests</td><td>Test cases in <code>tests/test_*.py</code></td><td><strong>160/160 (100%)</strong></td></tr>
-  </tbody>
-</table>
-
-State polling allowed 180 seconds; soft reboot completed in 131.66 seconds
-(previously it exceeded the 90-second check). No reboot implementation change
-was needed.
-
-The 10 previously reproduced tool defects were corrected. The 28 unavailable
-paths are **not live-validated successes**: 24 Octavia tools, two Neutron QoS
-tools, one Aodh tool, and one Neutron port-forwarding tool. Each affected table
-row is marked **⚠️ Live unverified**. An absent tool/action is marked **❌ Not
-implemented** rather than presented as an available feature.
-
-✅ means an implemented API path; it does not certify every action or backend.
-The live-pass count covers only the scenarios in the
-[latest retest report](docs/main-live-retest-2026-10-30.md), with
-[sanitized evidence](docs/main-live-retest-2026-10-30.json). API policy, asynchronous
-completion, production HTTPS/client compatibility, load, and guest I/O limits
-are stated there. All test resources were cleaned up; the final audit found zero
-remaining test resources across 24 categories. See also
-[API prerequisites and result semantics](#api-prerequisites-and-result-semantics).
+✅ denotes an implemented API path, not certification of every action or
+OpenStack deployment. **⚠️ Live unverified** identifies paths without successful
+live validation; **❌ Not implemented** identifies unavailable capabilities.
+Service availability, permissions, and asynchronous completion affect results.
+See [API prerequisites and result semantics](#api-prerequisites-and-result-semantics)
+and the [multi-user setup guide](docs/multi-user-guide.md).
 
 **Detailed Mapping by Category**
 
@@ -284,7 +216,7 @@ remaining test resources across 24 categories. See also
 | `openstack server set` | `set_server_properties` (action="set") | ✅ | Server property setting |
 | `openstack server unset` | `set_server_properties` (action="unset") | ✅ | Server property unsetting |
 | `openstack server dump create` | `set_server_dump` | ❌ Unsupported | Returns an explicit unsupported result; no dump is created |
-| `openstack server event list` | `get_server_events` | ✅ | Server event tracking |
+| `openstack server event list` | `get_server_events` | ✅ | Pass the server name or UUID via `instance_name`, not `server_id`; event visibility follows Nova policy |
 | `openstack server group list` | `get_server_groups` | ✅ | Server group listing |
 | `openstack server group create/delete` | `set_server_group` | ✅ | Server group management |
 | `openstack flavor list` | `set_flavor` (action="list") | ✅ | Requires modification tools to be enabled for registration |
@@ -524,7 +456,7 @@ remaining test resources across 24 categories. See also
 
 ### API prerequisites and result semantics
 
-The feature table marks implemented API paths, not verification against every OpenStack deployment. The management and query paths have local mocked API tests. Selected paths also passed live tests on a single-node Kolla Ansible deployment; see [integration validation](docs/integration-validation.md) for coverage and skipped operations. A missing service, extension, required microversion, or permission is an operation failure, not an empty successful result.
+The feature table marks implemented API paths, not verification against every OpenStack deployment. Service and policy requirements must be checked in the target deployment. A missing service, extension, required microversion, or permission is an operation failure, not an empty successful result.
 
 - **Modification tools:** the MCP server registers `set_*` management tools only when `ALLOW_MODIFY_OPERATIONS=true`. Existing action guards also reject writes when disabled. Dedicated read-only `get_*` tools remain available. API authorization is enforced by the target cloud; agent, user, domain, type and QoS administration may require elevated OpenStack roles.
 - **Alarms:** `set_alarms` calls the Keystone catalog's `alarming` endpoint for the configured interface/region. It supports list/create/show/delete of project alarms and creates `gnocchi_resources_threshold` alarms. Aodh, Gnocchi, the specified resource and a collected metric are required. The default metric is `cpu_util`, which is not collected by every deployment. `metric`, `resource_type`, `aggregation_method`, `granularity`, `evaluation_periods`, and `alarm_actions` can be supplied. An empty action list creates an alarm without notifications. There is no simulation fallback.
@@ -726,7 +658,7 @@ unset OPENSTACK_MCP_TOKEN
 ```
 
 This Bearer configuration uses the server's configured OpenStack account. For
-experimental per-user authentication, see the [phase 1 guide](docs/multi-user-phase1.md).
+experimental per-user authentication, see the [multi-user guide](docs/multi-user-guide.md).
 See also the [Claude Code MCP documentation](https://code.claude.com/docs/en/mcp).
 
 ## Server Configuration

mcp_openstack_ops-3.6.1/docs/multi-user-phase1.md → mcp_openstack_ops-3.6.2/docs/multi-user-guide.md

@@ -1,10 +1,10 @@
-# Multi-user HTTP: phase 1 testing
+# Multi-user HTTP setup and usage guide
 
-The server provides an opt-in, experimental remote mode for issue #73. It keeps
+The server provides an opt-in, experimental multi-user HTTP mode. It keeps
 existing project ownership checks and the default write restriction. The normal
 single-account HTTP and stdio modes remain available without this option.
 
-## Initial authentication approach
+## Authentication
 
 Clients send their own Keystone Application Credential ID and secret in two
 dedicated HTTP headers on **every request**. The server authenticates against
@@ -13,10 +13,9 @@ project-scoped connection. Authentication failure never falls back to the server
 environment account. `Authorization` is not forwarded to OpenStack; this is custom
 header authentication, not an OAuth server or Keystone token passthrough.
 
-Application Credentials are the initial method because they avoid LDAP passwords
+Application Credentials avoid LDAP passwords
 and already specify a project and delegated roles. Raw Keystone tokens, browser
-login, project switching, and administrator cross-project access are not provided
-in phase 1. Use another Application Credential/client entry for another project.
+login, project switching, and administrator cross-project access are not supported. Use another Application Credential/client entry for another project.
 
 The HTTP transport is stateless with JSON responses. Every HTTP request creates a
 fresh Keystone authentication and SDK connection; there is no authentication or
@@ -70,6 +69,26 @@ credential `.env` or `clouds.yaml`. `OS_USERNAME`, `OS_PASSWORD`, and
 and `REMOTE_SECRET_KEY`: the shared Bearer mode cannot be combined with this mode.
 DEBUG logging is rejected to avoid authentication wire dumps.
 
+Unsupported OAuth metadata routes (`/.well-known/oauth-protected-resource`,
+`/.well-known/oauth-authorization-server`, `/.well-known/openid-configuration`,
+including their path suffixes) and `/register` return 404 without Keystone
+authentication. This mode does not provide OAuth discovery or registration;
+`/mcp` still requires both application-credential headers.
+
+HTTP 400 responses from the MCP transport produce an INFO diagnostic with a
+server-generated `request_id` and a fixed `reason` classification:
+`invalid_json`, `invalid_jsonrpc`, `unsupported_protocol`, `missing_session_id`,
+or `unclassified`. Request bodies, header values, and raw error messages are not
+included in this diagnostic. Unknown, oversized, or non-JSON error responses
+may remain unclassified. A 400 alone does not establish an authentication failure
+or explain the reported first-connection behavior; correlate its request ID
+with the client response before drawing a conclusion. A client may probe with a
+newer protocol version, receive `unsupported_protocol`, and then negotiate a
+supported version successfully. Check whether initialization and subsequent
+requests succeed before treating that rejection as a connection failure. Do not
+bypass protocol validation merely to suppress the initial 400 response.
+
+
 Expose `/mcp` through an HTTPS reverse proxy. Do not expose the plaintext backend
 port to clients. Configure the proxy to redact both credential headers and
 `Authorization`, avoid request/header dumps, and enforce rate and concurrency
@@ -134,7 +153,7 @@ Sources checked while designing this branch:
 - [Claude connector request-header limitations](https://claude.com/docs/connectors/custom/add-unlisted#authenticate-with-request-headers)
 - [Keystone Application Credentials](https://docs.openstack.org/keystone/latest/user/application_credentials.html)
 
-## Validation checklist for the issue author
+## Deployment validation checklist
 
 Use disposable test projects and independently created credentials. Record client
 versions, service versions, and results without posting secrets or resource data.
@@ -146,7 +165,7 @@ versions, service versions, and results without posting secrets or resource data
    project's private resources. Public/shared resources remain governed by the
    existing checks and service policy. Project isolation is not ownership per user.
 3. Look up a known private resource ID from the other project: access must fail.
-   Even an admin credential remains under the existing project checks in phase 1.
+   Even an admin credential remains subject to the same project ownership checks.
 4. Connect the same user with two separate project credentials/client entries;
    verify each entry sees only its intended project.
 5. Omit, invalidate, or duplicate either credential header. No tool must execute,
@@ -173,128 +192,3 @@ Automated tests cover authentication rejection/revalidation, concurrent contexts
 including synchronous tools, cleanup, default write gating, project checks, and
 legacy cache behavior. They do not establish actual Keystone revocation timing,
 service policy enforcement, or Claude client compatibility.
-
-## Local cloud validation (2026-10-29)
-
-Real-cloud testing used two disposable projects, two non-admin users, and one
-private network per project. An additional reader-only Application Credential
-verified delegated role enforcement. Initial empty-list validation passed 36
-checks. Follow-up validation added one Cirros VM (`m1.tiny`: 1 vCPU, 512 MB RAM,
-1 GB disk) and one empty 1 GB volume per project. All 66 recorded follow-up checks
-passed; the latest unit-test run passed 46 tests.
-
-The test used real Keystone, Nova, Neutron, and Cinder APIs through temporary
-loopback TLS proxies with certificate verification. The proxy-to-cluster leg was
-HTTP because the test cluster has no TLS on its API ports. MCP calls went through
-the actual FastMCP ASGI HTTP application using an HTTP test client; this was not a
-Claude Code or Desktop session and does not validate a production TLS deployment.
-
-Validated behavior:
-
-- Concurrent users retained their own project identity and private network view.
-- Lookup of the other project's private network ID returned 403/404.
-- Registered instance, network, and volume read tools returned only the expected
-  project resources. Own-instance detail queries succeeded; SDK detail queries
-  for the other project's VM and volume IDs returned 403/404.
-- Write tools were absent by default; a test-only SDK write probe with reader
-  credentials was rejected with HTTP 403 by Neutron.
-- Missing, invalid, and deleted credentials were rejected with HTTP 401.
-- After credential deletion, the very next request was rejected. This observation
-  is not a deployment-independent revocation SLA.
-- Captured application logs identified users and tools without any known test
-  passwords, Application Credential IDs/secrets, or issued tokens appearing.
-- All test Application Credentials were deleted and both test users disabled.
-  The two isolated projects and private networks were retained for follow-up.
-- Both VMs reached `ACTIVE`, and both volumes reached `available`. All four were
-  deleted after testing; Nova instance/core/RAM and Cinder volume/GB quota usage
-  and reservations returned to zero in both projects.
-
-Testing exposed and fixed Cinder's SDK service-name/version/project-URL setup and
-classification of Keystone's deleted-credential `NotFound` response as 401 rather
-than 503. The populated-resource follow-up required no additional application
-changes. Claude Desktop/IDE, sustained load, guest networking/SSH, volume
-attachment/I/O, and a complete production HTTPS path remain unverified.
-
-## Actual Claude Code validation (2026-10-29)
-
-Claude Code 2.1.312 was used as the real MCP client, not an ASGI test client.
-Two isolated Claude configuration directories registered the server with
-`claude mcp add-json --scope user`. The server ran on a real HTTPS TCP listener
-with certificate verification; the temporary cloud-side TLS proxies described
-above remained necessary. No tools or OpenStack responses were mocked.
-
-Both clients executed the request "Show me the overall cluster status" through
-the actual configured Claude API. Initialization reported the user-scoped MCP
-server as `connected`; CLI stream events and server audit logs confirmed actual
-calls to registered service-status, monitoring, instance, network, and volume
-tools. A second request for each user's named private network confirmed that
-the actual client returned its own network ID rather than the other project's.
-All 28 connection/isolation/cleanup checks passed. Temporary registrations were
-removed, credentials deleted, and test users disabled afterward. Original Claude
-configuration was left unchanged; test credentials were not found in client
-configuration files or captured server logs.
-
-**Connection validation is not a complete status-report accuracy pass.** The
-real prompt exposed existing reporting defects outside the connection layer:
-
-- `functions.get_service_status` omits some services after permission failures
-  and uses a fixed "skipped" Heat entry without actually checking Heat.
-- Compute/volume availability zones read a nonexistent `available` attribute
-  rather than the SDK's `state['available']`, incorrectly defaulting to false.
-- Resource-monitoring capacity zeros cannot establish physical capacity when
-  hypervisor access is denied. A project-scoped user cannot assess global health.
-
-These findings were identified during validation and have not been repaired in
-this test step. Do not treat the model's successful completion or its availability
-summary as proof that the whole cluster is healthy or unavailable.
-
-## Ordinary-user Claude Code queries (2026-10-29)
-
-Administrator-only API denials are expected for ordinary users and are not a
-phase-1 authentication/connection failure. Each of the two real, registered
-Claude Code clients therefore executed these ordinary project requests:
-
-| Request | Actual tool | Observed result for both users |
-| --- | --- | --- |
-| List all instances in the project | `get_instance` | Empty after VM cleanup |
-| Show all networks and their subnets | `get_network_details` | Own private network and shared external network; no other test project's private network |
-| List all volumes and their status | `get_volume_list` | Empty after volume cleanup |
-
-All six actual Claude requests completed, and all 52 recorded connection,
-tool-execution, isolation, credential/logging, and cleanup checks passed. Own
-fixture networks have no subnets because none were created. Missing subnet
-details for a shared network do not establish that the network has no subnets.
-The earlier populated VM/volume test remains the evidence for nonempty resource
-isolation. Test registrations and credentials were removed and users disabled.
-
-These results validate the phase-1 project-user workflow separately from global
-administrator reporting. They do not certify every legacy reporting field: an
-optional `get_floating_ip_pools` call exposed an existing placeholder calculation
-of 100 IPs per allocation pool. That is not a measured pool capacity and was not
-counted as a verified result. The status/AZ reporting findings above also remain.
-
-## Default WSL Claude configuration validation (2026-10-29)
-
-A further test used the actual default WSL Claude user configuration at
-`/root/.claude.json`, without a `CLAUDE_CONFIG_DIR` override or `--mcp-config`.
-`claude mcp add-json --scope user` registered `openstack73-a` and `openstack73-b`
-sequentially. Plain `claude mcp list` confirmed `Connected`; an independent plain
-CLI invocation also confirmed the active registration. Each registration was
-kept visible briefly for inspection before executing the three project queries.
-
-All six actual Claude queries and all 57 recorded checks passed, including
-correct server selection, real tool results, private-network isolation, and
-cleanup. These registrations used the real server on WSL-only HTTP loopback;
-the cloud-side TLS proxies remained as above. A same-user, permission-restricted
-Unix-socket helper supplied credential headers from memory, so no credential
-values were stored in the normal user registration or command arguments.
-
-The initial attempt with both entries simultaneously present did not validate
-both clients: one remained `pending`. The successful default-configuration test
-therefore establishes sequential registrations, not simultaneous same-profile
-multi-entry support. Earlier simultaneous-client tests used separate profiles.
-
-Only the test registrations were removed afterward. Existing MCP registrations
-were preserved, test credentials deleted, and test users disabled. The temporary
-helper/server stopped after the test, so saved commands are an execution record,
-not sufficient to reconnect without restarting the test infrastructure.

{mcp_openstack_ops-3.6.1 → mcp_openstack_ops-3.6.2}/src/mcp_openstack_ops/prompt_template.md

@@ -913,4 +913,4 @@ This approach provides **comprehensive 360-degree cluster visibility** with infr
 - `get_heat_stack_details`: `section` is resources/events/template/outputs.
 - `set_service_logs(action="show")` and `set_metrics(action="show")` are unavailable without additional backend implementation. Never claim to have retrieved logs or telemetry samples from metadata-only results.
 
-API paths have local mocked tests, and selected paths passed single-node Kolla Ansible integration checks. See docs/integration-validation.md for coverage: successful Aodh/Neutron QoS lifecycles and live network-agent state changes remain unverified. Check returned errors and partial flags; never describe an accepted asynchronous request as a completed operation. Usage date bounds are UTC midnight, start inclusive/end exclusive. Missing physical capacity is null, not zero. Physical capacity and project allocations have different scopes and must not be divided into a utilization rate.
+Confirm service availability and permissions in the target deployment before attempting Aodh/Neutron QoS operations or network-agent state changes. Check returned errors and partial flags; never describe an accepted asynchronous request as a completed operation. Usage date bounds are UTC midnight, start inclusive/end exclusive. Missing physical capacity is null, not zero. Physical capacity and project allocations have different scopes and must not be divided into a utilization rate.

{mcp_openstack_ops-3.6.1 → mcp_openstack_ops-3.6.2}/src/mcp_openstack_ops/tools/get_server_events.py

@@ -14,7 +14,11 @@ async def get_server_events(
     limit: int = 50
 ) -> str:
     """
-    Get recent events for a specific server
+    Get recent events for a server by name or UUID.
+
+    Pass either value using instance_name, for example
+    get_server_events(instance_name="<server UUID>"). There is no server_id
+    parameter; do not use server_id or instance_id as argument names.
     
     Args:
         instance_name: Name or ID of the server instance

{mcp_openstack_ops-3.6.1 → mcp_openstack_ops-3.6.2}/src/mcp_openstack_ops/user_connection.py

@@ -1,6 +1,7 @@
 """Request-local Application Credential authentication for remote clients."""
 
 import asyncio
+import json
 import logging
 import os
 from contextvars import ContextVar
@@ -214,6 +215,26 @@ def authenticate(credential_id: str, secret: str) -> RequestIdentity:
         raise
 
 
+def _bad_request_reason(body: bytes) -> str:
+    """Classify SDK rejections without exposing response text or request values."""
+    try:
+        payload = json.loads(body)
+    except (ValueError, RecursionError):
+        return "unclassified"
+    error = payload.get("error") if isinstance(payload, dict) else None
+    message = error.get("message") if isinstance(error, dict) else None
+    if isinstance(message, str):
+        for prefix, reason in (
+            ("Parse error:", "invalid_json"),
+            ("Validation error:", "invalid_jsonrpc"),
+            ("Bad Request: Unsupported protocol version:", "unsupported_protocol"),
+            ("Bad Request: Missing session ID", "missing_session_id"),
+        ):
+            if message.startswith(prefix):
+                return reason
+    return "unclassified"
+
+
 class UserConnectionMiddleware:
     """Authenticate every HTTP request and close its connection after delivery.
 
@@ -228,6 +249,22 @@ class UserConnectionMiddleware:
         if scope["type"] != "http":
             await self.app(scope, receive, send)
             return
+        # This mode uses application credentials, not OAuth. Never forward these
+        # unsupported discovery/registration routes to an authenticated handler.
+        path = scope.get("path", "")
+        discovery_paths = (
+            "/.well-known/oauth-protected-resource",
+            "/.well-known/oauth-authorization-server",
+            "/.well-known/openid-configuration",
+        )
+        if path in ("/register", "/register/") or any(
+            path == base or path.startswith(base + "/")
+            for base in discovery_paths
+        ):
+            await JSONResponse({"error": "Not found"}, status_code=404)(
+                scope, receive, send
+            )
+            return
         headers = Headers(scope=scope)
         ids = headers.getlist("x-openstack-application-credential-id")
         secrets = headers.getlist("x-openstack-application-credential-secret")
@@ -283,11 +320,23 @@ class UserConnectionMiddleware:
             (perf_counter() - started) * 1000,
         )
         status = 500
+        rejection_body = bytearray()
 
         async def audit_send(message: Message) -> None:
             nonlocal status
             if message["type"] == "http.response.start":
                 status = message["status"]
+            elif status == 400 and message["type"] == "http.response.body":
+                # Bound diagnostic memory and never consume or log request bodies.
+                rejection_body.extend(
+                    message.get("body", b"")[:8192 - len(rejection_body)]
+                )
+                if not message.get("more_body", False):
+                    logger.info(
+                        "OpenStack transport rejected request_id=%s "
+                        "http_status=400 reason=%s",
+                        request_id, _bad_request_reason(bytes(rejection_body)),
+                    )
             await send(message)
 
         try: