Repository navigation
feat(mcp): Expose the bounded-context APIs through an MCP gateway - #188
Merged
Merged
Conversation
added 4 commits
November 8, 2026 22:42
A Rust CLI that logs in against the workbench's OIDC provider and calls the workbench APIs from a terminal, for scripting and CI use where the web app is not practical. - `vew login` runs an authorization-code flow with PKCE against the configured provider, receiving the callback on a loopback port, and stores the tokens per profile. `vew logout` revokes the refresh token where the provider exposes a revocation endpoint. - Commands are derived at runtime from each bounded context's OpenAPI document, so the CLI follows the APIs without being regenerated. Internal and documentation paths are skipped. - Configuration is a per-profile file holding the issuer, the OIDC client id and the API endpoints. Nothing is discovered from AWS, so the CLI needs no AWS credentials. - Install scripts for Unix shells and PowerShell, and a GitHub Actions workflow running fmt, clippy with warnings as errors, the test suite and a dependency audit, all pinned to Rust 1.95. Diagnostics are off unless --verbose is passed. The token-endpoint response body and the OAuth callback query string are never logged, since they carry tokens and the authorization code; only their sizes and the request path are. OAuth error bodies and JWKS documents are logged, as neither contains a credential. To use it, register an additional Cognito app client through the deployment's ADDITIONAL_APP_CLIENTS setting and put that client id in the CLI's config file.
The CLI's config holds one `api_endpoint`, but a deploy provisions one API
Gateway per bounded context. Without a custom API domain there is no host
in front of them, so the CLI has nothing to point at.
VewApiRouterStack supplies that host: an HTTP API that path-routes
/{bounded-context}/* to each backend gateway, resolving each target from
the SSM parameter that context's own stack publishes. It is registered as
the `else` of the ApiIntegrationStack branch in backend_app.py, because
with a custom API domain that stack already fronts the gateways under one
host and a second hop would be redundant.
Every route is behind a JWT authorizer. The CLI sends a Cognito access
token, which carries client_id and no aud claim, so the audience is the
CLI's app client id, read from the per-client SSM parameter the frontend
stack publishes for each AdditionalAppClients entry. Registering the
client is therefore a config change, not a code change: add a "vew-cli"
entry naming http://localhost:17823/callback as its callback URL.
deploy.sh gains phase 7b, the single writer of
/<app>-<env>/vew-api-endpoint in both deploy modes — the router's output
when there is no custom domain, the custom domain itself when there is.
Keeping it out of the stacks means an environment that drops its custom
domain does not fail on a parameter that already exists.
One Amazon Bedrock AgentCore Gateway fronts the projects, packaging, publishing and provisioning APIs as MCP tool targets, so an agent can drive the workbench through any MCP client. - Each bounded context opts in with enable_mcp=True on its BackendAppOpenApi; backend_app.py collects those that did and passes them to the gateway stack. - infra/helpers/mcp_openapi_transform turns each context's OpenAPI document into the shape the gateway accepts. - A REQUEST interceptor Lambda forwards the caller's Authorization header to the target API, so the existing Cedar authorizer still makes every decision. The gateway role can invoke the interceptor and read the spec bucket, and holds no execute-api:Invoke. - The gateway authorizes callers itself with CUSTOM_JWT against the user pool's discovery document, accepting only its own app client. The API schemas gain operation descriptions. The gateway publishes those as its tool descriptions, which is what an agent reasons over to choose a tool, so a name-only restatement degrades tool selection under ambiguity. test_mcp_tool_descriptions asserts every published tool and every context carries a substantive one, against the real schemas rather than the source, so the check cannot pass on intent alone. The gateway's Cognito client comes from the frontend deployment's AdditionalAppClients, read as an SSM parameter, rather than being created here: the authorizer's accepted-client list stays the one mechanism it already is. Register the gateway by adding an "mcp-gateway-agentcore" entry with the callback URLs in MCP_GATEWAY_OAUTH_REDIRECT_URIS.
Running the CLI's own CI checks locally surfaced work the import needed. - rustls 0.23.45 -> 0.23.46 clears four advisories, the most serious being Debug impls on client session values exposing TLS secrets (RUSTSEC-2026-0346), plus buffered data after ECH rejection, unbounded Brotli certificate decompression, and a ServerHello downgrade. - rustls-webpki 0.103.15 -> 0.103.16 clears RUSTSEC-2026-0382, where certificate verification could accept distrusted intermediates. - RUSTSEC-2026-0387, timing variability in crypto-bigint, is exempted with its reasoning in .cargo/audit.toml. It arrives through jsonwebtoken -> p256/p384 -> elliptic-curve, and the fix needs an upstream chain this repo cannot pull forward. It does not apply: the validator only verifies against public JWKS keys and holds no signing key, and it configures RS256 only, so the ECDSA path is linked but never executed. - cargo fmt reformats the multi-line debug_log! call. - cli/.gitignore no longer points at a CI config this repo does not have. cargo fmt --check, cargo clippy --all-targets --release --locked -- -D warnings, cargo test --release --locked and cargo audit all pass.
added 9 commits
November 9, 2026 17:18
isort orders the from-import list in backend_app.py, and the gateway stack module was appended rather than inserted. No behaviour change.
Scope the AgentCore trust relationships, drop infrastructure for a feature this branch does not carry, and restore a setting it removed. Confused deputy on both AgentCore grants. The gateway role trusted bedrock-agentcore.amazonaws.com and the interceptor's Lambda policy granted it invoke, both unconditioned, so the service would act for any AgentCore gateway in any account. Both now require aws:SourceAccount and an aws:SourceArn matching this stack's own gateway. The ARN is pinned by pattern because the gateway cannot exist before the role it assumes; AgentCore documents the ARN as name plus a generated ten-character suffix, so the name prefix is sufficient. The log key granted the global logs.<url suffix> principal. CloudWatch Logs requires a same-Region principal in a key policy, so the encrypted log group could not ingest. Now logs.<region>.<url suffix>. The shared infra.constructs.logs.log_group construct has the same defect and needs its own change. Remove the scheduled-jobs Lambda, its hourly marketplace rule and its legacy export: no scheduled_jobs_handler package exists here, so bundling it had no handler to bundle. The same applies to the gsi_image_source index, the marketplace, licence-manager, EC2 describe-image, Organizations and RAM grants, the unread SPOKE_ACCOUNT_ROLE variable and the ImageImported subscription - all belong to the image-import feature, which is a separate change. Restore BUILD_LOGS_S3_BUCKET_NAME on the packaging domain-event handler. The handler still reads it and passes it to the Image Builder pipeline service, which omits the s3Logs block when it is empty, so dropping the variable turned off build-log delivery without any error. release_component_version declared a required body for an empty model it never read, so a bodyless call failed validation while the schema marked the body optional. Drop the parameter, matching release_recipe_version. Two operation descriptions promised behaviour that does not exist here: GetImages advertised image-source and status filters it takes no parameters for, and GetAmis advertised marketplace metadata absent from the Ami schema. Under MCP these descriptions are the tool contract an agent reasons over, so a false one is a defect rather than a typo. Regenerate the TypeScript clients from the changed schemas. This also picks up drift that predates the branch: the projects client was missing ServiceClientAssignment, which its schema already declared. Add tests covering each fix, plus a static check that every Lambda entry path declared in infra resolves to a real package. That check also finds a pre-existing dead entry in ProjectsAppStack, recorded as a strict xfail so it cannot be forgotten or silently rot.
Make the "no secret is logged" property assertable, then fix every place a probe with planted markers found one. debug_log! writes to stderr, which a test cannot read back, so the property was unassertable and any regression invisible. It now also records into an in-memory buffer, compiled only under cfg(test) or the new log-capture feature so a release binary has no buffer at all. The adapter tests live in cli/tests/, which does not get cfg(test) for this crate, so CI runs cargo test --features log-capture; without the feature the new test file compiles to nothing. New oauth_error module renders a failed OAuth call from its status and an RFC 6749 / 7009 error code only. A token-endpoint error body can contain a refresh token, a revocation error body echoes the submitted token, and error_description is free text chosen by whoever drove the redirect, so none of it reaches the user or the log. A code outside the RFC list collapses to a placeholder rather than being echoed. A serde error quotes the value it rejected, for example `invalid type: string "<access token>"`. The exchange path now reports the position only, and the refresh and discovery paths report nothing from the message. The callback server no longer puts the request target in the unexpected-path error, since the query carries code and state, and no longer forwards error_description or a non-standard error code. Its request read is capped at 64 KiB: this port takes unauthenticated input, and read_line on a client that never sends a newline grew without bound. StateMismatch carried both state values; printing the expected one lets anyone reading the output replay it, and the mismatch alone is what the user needs. The JWT validator no longer logs the kid or the sub, and reports the jsonwebtoken error kind rather than its Display text, which can quote the token. The token file's .tmp was created at the umask default and chmod'd to 0600 afterwards, so the secret was briefly readable by others. It is now created with mode 0600 via create_new, which also refuses to follow a symlink planted at that path, and a stale .tmp is removed first. Also: pin cargo-audit to 0.22.2; correct the RUSTSEC-2023-0071 reason, which claimed upstream has no fix — the advisory records "Patched >= 0.10.0" and cargo tree shows rsa 0.9.10 is pinned by jsonwebtoken's rust_crypto backend, which is the real reason we cannot take it; replace the no-op test_discovery_success, which had neither an act nor an assert step; and correct cli/README.md, which named a vew-cli-client-id parameter and a VewCliClientId output that do not exist and said deploy.sh creates the CLI app client, which it does not. Each redaction test was confirmed to fail against the unfixed adapter: 5 of 6 fail with the fix stashed. 182 tests pass with the feature, plain cargo test stays green, clippy --all-targets -D warnings clean.
The router exists only to give the vew CLI one api_endpoint when there is no custom API domain, but it was built on every domainless deploy. Its authorizer resolves /<app>-ui-<env>/app-client-id/vew-cli at synth, and that parameter exists only when the frontend deployment's ADDITIONAL_APP_CLIENTS contains a vew-cli entry, so a deploy without that client stood up a stack nothing used and failed on the lookup. deploy.sh now decides and passes it as context. Two conditions, not one: the client has to be configured, and the frontend and backend have to be in the same region. The parameter is written by the frontend stacks in FE_REGION while the router reads it in AWS_REGION, and a public deployment puts the frontend in us-east-1 wherever the backend runs, so that read would miss. A mismatch now warns and skips rather than failing later. When the router is no longer configured, Phase 7b deletes a stack left behind by an earlier deploy that did have the client. Without that it keeps serving an endpoint nothing maintains, and no later deploy ever touches it again. A delete that a CloudFormation export blocks warns instead of aborting, since the leftover is inert once the stack is out of the CDK app. Phase 7b also no longer runs under --diff. Its put-parameter is a real write, and the router-output read fails outright when the router stack does not exist, so a diff on a fresh account stopped with an error about a stack a deploy had not yet created. The diff now exits before it, and the closing summary names the two extra things it skips. Verified by synthesising the backend app with vew-cli-enabled both true and false: the router template is present in the first and absent in the second, with no synth error either way. The jq client test was checked against an empty list, a list without the entry, and lists with it. shellcheck and bash -n are clean.
Remove the remaining OAuth callback log leaks and handle API router lifecycle transitions across custom-domain and regional deployments. Correct the documented SSM endpoint path.
Validate names, descriptions, parameters, request properties, and local references on the exact transformed schemas uploaded to AgentCore. Add MCP-only names for cross-target collisions and document exposed inputs.
A rejected ID token was reported as "token rejected: {kind:?}". The
kind is not a fixed enum: ErrorKind::Json wraps the serde error, which
quotes the rejected claim value, and other variants carry text. With
--verbose the value reached the debug log as well as the error.
Map each kind to a fixed name and drop its payload. The new redaction
test signs a token whose sub is a number and fails against the
previous commit.
The gateway stack reads /<app>-ui-<env>/app-client-id/mcp-gateway-agentcore
at deploy time, but it was registered on every deploy. Without an
"mcp-gateway-agentcore" entry in ADDITIONAL_APP_CLIENTS the deploy
failed ("Unable to fetch parameters"), seen on a fresh deployment.
Register the stack only when deploy.sh finds that client and the
frontend and backend share a region (the parameter is written in the
frontend region), as the vew CLI router already does, and delete a
stale gateway stack otherwise. Document the two optional clients.
Also drop the 127.0.0.1 callback URLs: Cognito and the frontend's
AdditionalAppClients validation accept plain http only for localhost,
so registering the documented list failed the frontend deploy. A new
test checks the list against that rule and fails on the old list.
raphaho
pushed a commit
that referenced
this pull request
Nov 13, 2026
With the MCP gateway (#188) merged in, its synth-time tool-metadata check rejected the four operations this branch adds: their descriptions were one line and their parameters and request properties had none, so the backend could not synthesize. Describe them. The gateway branch also asserted that packaging holds no Marketplace or License Manager action. Marketplace discovery needs some, so pin the exact set instead. That turned up four actions packaging was granted but never calls (DescribeAgreement, GetAgreementTerms, GetProduct, ListFulfillmentOptions); only SearchAgreements is used, so drop the rest.
9 of 11 tasks
andreameroni
previously approved these changes
Nov 13, 2026
# Conflicts: # README.md # backend/infra/config.py # backend/infra/constants.py # deploy.sh
andreameroni
approved these changes
Nov 13, 2026
raphaho
added a commit
that referenced
this pull request
Nov 13, 2026
#200) * feat(cli): Add the vew command-line client A Rust CLI that logs in against the workbench's OIDC provider and calls the workbench APIs from a terminal, for scripting and CI use where the web app is not practical. - `vew login` runs an authorization-code flow with PKCE against the configured provider, receiving the callback on a loopback port, and stores the tokens per profile. `vew logout` revokes the refresh token where the provider exposes a revocation endpoint. - Commands are derived at runtime from each bounded context's OpenAPI document, so the CLI follows the APIs without being regenerated. Internal and documentation paths are skipped. - Configuration is a per-profile file holding the issuer, the OIDC client id and the API endpoints. Nothing is discovered from AWS, so the CLI needs no AWS credentials. - Install scripts for Unix shells and PowerShell, and a GitHub Actions workflow running fmt, clippy with warnings as errors, the test suite and a dependency audit, all pinned to Rust 1.95. Diagnostics are off unless --verbose is passed. The token-endpoint response body and the OAuth callback query string are never logged, since they carry tokens and the authorization code; only their sizes and the request path are. OAuth error bodies and JWKS documents are logged, as neither contains a credential. To use it, register an additional Cognito app client through the deployment's ADDITIONAL_APP_CLIENTS setting and put that client id in the CLI's config file. * feat(infra): Give the vew CLI a single API endpoint The CLI's config holds one `api_endpoint`, but a deploy provisions one API Gateway per bounded context. Without a custom API domain there is no host in front of them, so the CLI has nothing to point at. VewApiRouterStack supplies that host: an HTTP API that path-routes /{bounded-context}/* to each backend gateway, resolving each target from the SSM parameter that context's own stack publishes. It is registered as the `else` of the ApiIntegrationStack branch in backend_app.py, because with a custom API domain that stack already fronts the gateways under one host and a second hop would be redundant. Every route is behind a JWT authorizer. The CLI sends a Cognito access token, which carries client_id and no aud claim, so the audience is the CLI's app client id, read from the per-client SSM parameter the frontend stack publishes for each AdditionalAppClients entry. Registering the client is therefore a config change, not a code change: add a "vew-cli" entry naming http://localhost:17823/callback as its callback URL. deploy.sh gains phase 7b, the single writer of /<app>-<env>/vew-api-endpoint in both deploy modes — the router's output when there is no custom domain, the custom domain itself when there is. Keeping it out of the stacks means an environment that drops its custom domain does not fail on a parameter that already exists. * feat(mcp): Expose the bounded-context APIs through an MCP gateway One Amazon Bedrock AgentCore Gateway fronts the projects, packaging, publishing and provisioning APIs as MCP tool targets, so an agent can drive the workbench through any MCP client. - Each bounded context opts in with enable_mcp=True on its BackendAppOpenApi; backend_app.py collects those that did and passes them to the gateway stack. - infra/helpers/mcp_openapi_transform turns each context's OpenAPI document into the shape the gateway accepts. - A REQUEST interceptor Lambda forwards the caller's Authorization header to the target API, so the existing Cedar authorizer still makes every decision. The gateway role can invoke the interceptor and read the spec bucket, and holds no execute-api:Invoke. - The gateway authorizes callers itself with CUSTOM_JWT against the user pool's discovery document, accepting only its own app client. The API schemas gain operation descriptions. The gateway publishes those as its tool descriptions, which is what an agent reasons over to choose a tool, so a name-only restatement degrades tool selection under ambiguity. test_mcp_tool_descriptions asserts every published tool and every context carries a substantive one, against the real schemas rather than the source, so the check cannot pass on intent alone. The gateway's Cognito client comes from the frontend deployment's AdditionalAppClients, read as an SSM parameter, rather than being created here: the authorizer's accepted-client list stays the one mechanism it already is. Register the gateway by adding an "mcp-gateway-agentcore" entry with the callback URLs in MCP_GATEWAY_OAUTH_REDIRECT_URIS. * fix(cli): Clear six dependency advisories and format the logging change Running the CLI's own CI checks locally surfaced work the import needed. - rustls 0.23.45 -> 0.23.46 clears four advisories, the most serious being Debug impls on client session values exposing TLS secrets (RUSTSEC-2026-0346), plus buffered data after ECH rejection, unbounded Brotli certificate decompression, and a ServerHello downgrade. - rustls-webpki 0.103.15 -> 0.103.16 clears RUSTSEC-2026-0382, where certificate verification could accept distrusted intermediates. - RUSTSEC-2026-0387, timing variability in crypto-bigint, is exempted with its reasoning in .cargo/audit.toml. It arrives through jsonwebtoken -> p256/p384 -> elliptic-curve, and the fix needs an upstream chain this repo cannot pull forward. It does not apply: the validator only verifies against public JWKS keys and holds no signing key, and it configures RS256 only, so the ECDSA path is linked but never executed. - cargo fmt reformats the multi-line debug_log! call. - cli/.gitignore no longer points at a CI config this repo does not have. cargo fmt --check, cargo clippy --all-targets --release --locked -- -D warnings, cargo test --release --locked and cargo audit all pass. * style(mcp): Sort the backend_app stack import isort orders the from-import list in backend_app.py, and the gateway stack module was appended rather than inserted. No behaviour change. * feat(packaging): Add AMI and marketplace image import Lets a publisher base a product on an existing AMI instead of building one through a pipeline, and on an AWS Marketplace subscription instead of an AMI they own. Three capabilities arrive together because the wizard offers them as one choice of image source: - AMI import. A five-step wizard collects the source, AMI, platform, software and review; the handler validates the AMI is reachable and registers it as an image with imageSource=IMPORTED. New value objects cover architecture, platform, OS version, image name and product code. - Marketplace subscription discovery. The hub lists the account's agreements, resolves each to its product and per-region AMI ids, and a scheduled job re-checks subscriptions for newer images. - License Manager grants. A marketplace product VEW is subscribed to is licensed to the spoke by creating and accepting a grant; one the customer already owns is only verified. Both run as new states in the AMI sharing state machine. Existing images are stamped imageSource=PIPELINE by a new DynamoDB migration, so records written before this change keep working. Decisions worth naming, because neither side of the merge was simply newer: - make_share_ami_decision now returns a ShareAmiDecisionResult dataclass rather than a four-tuple. A marketplace decision has to carry the licence ARN, home region and grant name downstream, which a tuple of four cannot. The region-aware shared-AMI lookup is kept as it is. - publish_version_command_handler keeps filtering shared AMIs to the distribution's own account, and the marketplace short-circuit sits in front of that filter rather than replacing it. - ShareAmiDecision.NOT_REQUIRED is renamed to NotRequired for consistency with the other members, including the two added here. The value is unchanged, so nothing on the wire moves. - get_version_entity and the automated-build-product entrypoint are left out: both were defined and never referenced. Three packaging test fixtures gain the entities GSI and GSI_NAME_ENTITIES, because their bootstrappers now run the migration, which queries that index as the deployed Lambdas do. The packaging route total moves from 42 to 45; all three new routes are project-path scoped and address no component or recipe, so both guard counts are unchanged. * feat(packaging): Degrade gracefully without Marketplace access Marketplace discovery assumed the hub can read the account's agreements and received licences. Two calls fail outright when it cannot: SearchAgreements needs Organizations access the hub only has as an organisation admin, and ListReceivedLicenses needs the AWSServiceRoleForMarketplaceLicenseManagement service-linked role, which exists only once License Manager integration is enabled in Marketplace settings. Both now treat AccessDeniedException and AuthorizationException as "no marketplace data available": they log a warning naming the likely cause and return an empty list, so a deployment without marketplace access keeps working with the rest of the image sources instead of erroring. Adds docs/marketplace-integration.md covering the prerequisites, setup and workflow, and points at it from the README. * feat(packaging): Carry marketplace metadata through image import Completes the marketplace image source: the import handler stores the product's subscription metadata on the image, discovery returns the per-region AMI ids alongside each subscription, and the images and subscriptions pages show which product an image came from. MarketplaceAmi gains creationDate, which describe_images already returns. Deliberately not included: the update-detection half of this work. Upstream it adds a CheckMarketplaceUpdates command, its handler, a MarketplaceProductUpdateDetected event, and an hourly EventBridge rule targeting a packaging "scheduled jobs" Lambda. That Lambda has no handler module — its entry directory holds only a Pydantic job model, while every other packaging entrypoint has handler.py and bootstrapper.py, and BackendAppFunction resolves handler.handler by default with no override. So the rule would invoke a function that fails at import, hourly, and the command is never dispatched from anywhere else. Porting it would publish a feature that cannot run. For the same reason the packaging "scheduled jobs" Lambda itself is left out, along with the legacy CfnOutput exporting its role ARN — that export exists for an old CloudFormation stack this repo never had. Also left out is the gsi_image_source GSI. It indexes imageSource for exactly that path, nothing queries it, and an unused global secondary index costs write capacity and storage on every image write. The update check is worth having. It needs a handler first, which is new code rather than a port, so it belongs in its own change. * fix(packaging): Pick a region's newest marketplace AMI deterministically get_regional_ami_ids looped over every AMI a region publishes and assigned each to the same map key, so it kept whichever describe_images returned last. That call applies no ordering, so a product publishing more than one available AMI in a region produced an arbitrary choice that could differ between runs. The map becomes MarketplaceMetadata.regionalAmiIds, which decides the AMI a region launches through the CloudFormation RegionMap, so an arbitrary pick can launch a superseded image. Take the newest instead. creationDate was added to MarketplaceAmi to compare image ages, but this selection was never changed to use it: CreationDate is ISO 8601 and sorts lexicographically, an AMI without one loses to any AMI that has one, and amiId breaks a tie so the result is stable across runs. A region that publishes no AMI is now skipped explicitly rather than falling out of the loop body, which reads the same but says so. The test asserts the same two AMIs ordered newest-last and newest-first give the same answer; before this change those orders returned different AMIs. * fix(mcp): Apply review fixes to the MCP gateway Scope the AgentCore trust relationships, drop infrastructure for a feature this branch does not carry, and restore a setting it removed. Confused deputy on both AgentCore grants. The gateway role trusted bedrock-agentcore.amazonaws.com and the interceptor's Lambda policy granted it invoke, both unconditioned, so the service would act for any AgentCore gateway in any account. Both now require aws:SourceAccount and an aws:SourceArn matching this stack's own gateway. The ARN is pinned by pattern because the gateway cannot exist before the role it assumes; AgentCore documents the ARN as name plus a generated ten-character suffix, so the name prefix is sufficient. The log key granted the global logs.<url suffix> principal. CloudWatch Logs requires a same-Region principal in a key policy, so the encrypted log group could not ingest. Now logs.<region>.<url suffix>. The shared infra.constructs.logs.log_group construct has the same defect and needs its own change. Remove the scheduled-jobs Lambda, its hourly marketplace rule and its legacy export: no scheduled_jobs_handler package exists here, so bundling it had no handler to bundle. The same applies to the gsi_image_source index, the marketplace, licence-manager, EC2 describe-image, Organizations and RAM grants, the unread SPOKE_ACCOUNT_ROLE variable and the ImageImported subscription - all belong to the image-import feature, which is a separate change. Restore BUILD_LOGS_S3_BUCKET_NAME on the packaging domain-event handler. The handler still reads it and passes it to the Image Builder pipeline service, which omits the s3Logs block when it is empty, so dropping the variable turned off build-log delivery without any error. release_component_version declared a required body for an empty model it never read, so a bodyless call failed validation while the schema marked the body optional. Drop the parameter, matching release_recipe_version. Two operation descriptions promised behaviour that does not exist here: GetImages advertised image-source and status filters it takes no parameters for, and GetAmis advertised marketplace metadata absent from the Ami schema. Under MCP these descriptions are the tool contract an agent reasons over, so a false one is a defect rather than a typo. Regenerate the TypeScript clients from the changed schemas. This also picks up drift that predates the branch: the projects client was missing ServiceClientAssignment, which its schema already declared. Add tests covering each fix, plus a static check that every Lambda entry path declared in infra resolves to a real package. That check also finds a pre-existing dead entry in ProjectsAppStack, recorded as a strict xfail so it cannot be forgotten or silently rot. * fix(cli): Keep secrets out of the log and error output Make the "no secret is logged" property assertable, then fix every place a probe with planted markers found one. debug_log! writes to stderr, which a test cannot read back, so the property was unassertable and any regression invisible. It now also records into an in-memory buffer, compiled only under cfg(test) or the new log-capture feature so a release binary has no buffer at all. The adapter tests live in cli/tests/, which does not get cfg(test) for this crate, so CI runs cargo test --features log-capture; without the feature the new test file compiles to nothing. New oauth_error module renders a failed OAuth call from its status and an RFC 6749 / 7009 error code only. A token-endpoint error body can contain a refresh token, a revocation error body echoes the submitted token, and error_description is free text chosen by whoever drove the redirect, so none of it reaches the user or the log. A code outside the RFC list collapses to a placeholder rather than being echoed. A serde error quotes the value it rejected, for example `invalid type: string "<access token>"`. The exchange path now reports the position only, and the refresh and discovery paths report nothing from the message. The callback server no longer puts the request target in the unexpected-path error, since the query carries code and state, and no longer forwards error_description or a non-standard error code. Its request read is capped at 64 KiB: this port takes unauthenticated input, and read_line on a client that never sends a newline grew without bound. StateMismatch carried both state values; printing the expected one lets anyone reading the output replay it, and the mismatch alone is what the user needs. The JWT validator no longer logs the kid or the sub, and reports the jsonwebtoken error kind rather than its Display text, which can quote the token. The token file's .tmp was created at the umask default and chmod'd to 0600 afterwards, so the secret was briefly readable by others. It is now created with mode 0600 via create_new, which also refuses to follow a symlink planted at that path, and a stale .tmp is removed first. Also: pin cargo-audit to 0.22.2; correct the RUSTSEC-2023-0071 reason, which claimed upstream has no fix — the advisory records "Patched >= 0.10.0" and cargo tree shows rsa 0.9.10 is pinned by jsonwebtoken's rust_crypto backend, which is the real reason we cannot take it; replace the no-op test_discovery_success, which had neither an act nor an assert step; and correct cli/README.md, which named a vew-cli-client-id parameter and a VewCliClientId output that do not exist and said deploy.sh creates the CLI app client, which it does not. Each redaction test was confirmed to fail against the unfixed adapter: 5 of 6 fail with the fix stashed. 182 tests pass with the feature, plain cargo test stays green, clippy --all-targets -D warnings clean. * fix(infra): Register the vew API router only when the CLI is configured The router exists only to give the vew CLI one api_endpoint when there is no custom API domain, but it was built on every domainless deploy. Its authorizer resolves /<app>-ui-<env>/app-client-id/vew-cli at synth, and that parameter exists only when the frontend deployment's ADDITIONAL_APP_CLIENTS contains a vew-cli entry, so a deploy without that client stood up a stack nothing used and failed on the lookup. deploy.sh now decides and passes it as context. Two conditions, not one: the client has to be configured, and the frontend and backend have to be in the same region. The parameter is written by the frontend stacks in FE_REGION while the router reads it in AWS_REGION, and a public deployment puts the frontend in us-east-1 wherever the backend runs, so that read would miss. A mismatch now warns and skips rather than failing later. When the router is no longer configured, Phase 7b deletes a stack left behind by an earlier deploy that did have the client. Without that it keeps serving an endpoint nothing maintains, and no later deploy ever touches it again. A delete that a CloudFormation export blocks warns instead of aborting, since the leftover is inert once the stack is out of the CDK app. Phase 7b also no longer runs under --diff. Its put-parameter is a real write, and the router-output read fails outright when the router stack does not exist, so a diff on a fresh account stopped with an error about a stack a deploy had not yet created. The diff now exits before it, and the closing summary names the two extra things it skips. Verified by synthesising the backend app with vew-cli-enabled both true and false: the router template is present in the first and absent in the second, with no synth error either way. The jq client test was checked against an empty list, a list without the entry, and lists with it. shellcheck and bash -n are clean. * fix(publishing): Page the grant lookups and trim the licence grants Four of the review findings on the licence-grant path. The authorization check, the state machine and the deprecated-AMI handling are not in this commit. has_active_received_grant called get_paginator("list_received_grants"). License Manager publishes no paginator for that operation - can_paginate is False - so boto3 raises OperationNotPageableError, which is not a ClientError and therefore escaped the handler around the call. That function failed on every invocation. It now follows NextToken by hand, through one helper shared with the other grants call. _find_existing_grant read a single page of list_distributed_grants, which has no paginator either. A grant on a later page looked absent, so the create that followed failed with "already has a grant". It pages too. accept_grant called AcceptGrant when ListReceivedGrants returned nothing: a missing grant left the status as None, which is neither ACTIVE nor DISABLED, so the code fell through to the accept. It now stops with a clear adapter error instead. update_version_command_handler guarded the image source on `if ami.imageSource:`, which kept the previous AMI's image source and Marketplace metadata when the new AMI has none - a version updated from a Marketplace AMI to a pipeline-built one went on claiming a Marketplace origin and licence. Both attributes are now written on every update, including as None. Two existing tests asserted the old attribute set and are updated. The hub's License Manager grants drop to CreateGrant and ListDistributedGrants plus the licence reads: it never reads a single grant, deletes one, or accepts one. The spoke keeps AcceptGrant, CreateGrantVersion and ListReceivedGrants and loses GetGrant. Verified: the paging helper exercised across three pages, and _find_existing_grant shown to find a grant that only appears on page two. OperationNotPageableError confirmed against the installed boto3. 2751 app tests and 6 infra tests pass. * fix(publishing): Authorize a licence grant against the project create_license_grant_command_handler distributed a Marketplace licence for any AMI id to any AWS account. Neither side of the request was checked against the project the command names, so a caller with access to one project could have the hub grant another project's paid licence to an account of their choosing. Both halves are now checked before anything is distributed. The AMI has to belong to the project in the command, which the Ami read model already records as projectId, and the target account has to be one of that project's Active accounts, read through the existing ProjectsQueryService port. Either check failing raises a DomainException and no grant is created, nothing is persisted, and no event is published. That port already existed and had no caller. Wiring it into the two entry points that dispatch this command needed three supporting changes: both bootstrappers now build a ProjectsApiQueryService through the ServiceRegistry, the api and ami-sharing functions get the projects GET /internal/accounts route via cross_bc_api_access, and their AppConfig classes extend VEWBaseConfig, which the registry reads. VEWBaseConfig only wraps environment variables BackendAppEntrypoints already injects into every Lambda, so this adds no new configuration. accept_license_grant_command_handler called accept_grant before reading the grant record, so a command naming a grant the project does not have still issued a real AcceptGrant against the spoke account and only then failed. The record is read first and a missing one stops the handler. Verified: the three new authorization tests fail against the handler without the check and pass with it; 2754 app tests pass; cdk synth of PublishingAppStack and ServiceDiscoveryStack is clean, with the trimmed grants confirmed in the template (no DeleteGrant, no GetGrant) and /internal/accounts now granted to the api, domain-events and ami-sharing functions. * fix(publishing): Stop the AMI sharing machine succeeding on a failure Three structural defects in the AMI sharing state machine, all on the Marketplace paths. VerifyLaunchPermissionLambda reports whether the target account can actually launch the regional Marketplace AMI, and nothing read the answer: both the LICENSE_GRANT and VERIFY_EXTERNAL_GRANT chains ran straight on to SucceedAmiSharing, so a spoke that could not launch the image was recorded as shared successfully. A Choice now gates that. SucceedAmiSharing reads $.copiedAmi.copiedAmiId, which the verify-external-grant path never set, so that path could not complete even when the grant was valid. Both Marketplace paths now record the regional AMI - the image the spoke actually launches - at a shared Pass after the launch-permission check. FailAmiSharingLambda was invoked without the error. A task failure populates $.error through add_catch, but a Choice taking its otherwise branch does not, so the reason a share failed was dropped and the version showed a bare Failed status. Each failing Choice branch now sets its own $.error through a Pass - unexpected decision, grant not verified, launch permission denied - and the Lambda payload carries $.error on every route. Verified against the synthesised definition: all twelve routes into that state set $.error, so the new payload reference always resolves. The Lambda still ignores the field, because FailAmiSharingCommand has no reason yet. Carrying it is the prerequisite for that work, which also has to persist statusReason and return it from GetProductVersion. Also in the tests: five call sites passed test_event=mocked_test_event without calling it. That fixture returns a factory, so the tests handed StartExecution a function instead of an event, which is not JSON-serialisable - the ParamValidationError I had twice attributed to these tests needing live credentials. They need Step Functions Local on localhost:8083, which is a separate matter, so this fix is verified statically. The 12-digit account-like values in conftest and the module docstring are now documentation example ids. Verified: cdk synth of PublishingAppStack clean, with the Choice, the regional-AMI Pass and every $.error route asserted against the synthesised definition. infra failures are byte-identical with these changes stashed - 9 either way, all in the other state-machine suites that need Step Functions Local. * feat(publishing): Report why an AMI share failed A failed version said only "Failed". The state machine already knew the reason and dropped it, so the only way to find out was to open the Step Functions execution. FailAmiSharingLambda now receives $.error, which the previous commit made available on every route into that state, and turns it into a short reason stored as statusReason on the version distribution and returned from GetProductVersion. step_error.summarize handles both shapes $.error arrives in: a task catch, where Cause is JSON carrying errorType, errorMessage and a trace array, and a failing Choice branch, where Cause is already a sentence. Only the type and message are kept. The trace is never included - it is long, it is useless to whoever is looking at a failed product version, and it names internal file paths. The result is collapsed to one line and capped at 1000 characters, so one failure cannot write an unbounded string into the version record. The reason is cleared on success and on retry, so it cannot outlive the failure it describes and sit on a version that is running again. FailAmiSharingRequest.error is optional, so an execution already in flight without the field still parses. The API schema gains statusReason on VersionDistribution and the publishing TypeScript client is regenerated, which is how the field reaches the UI. Only the changed model file lands in the diff; the generator's whitespace churn is normalised by the existing hooks. 12 tests cover the summariser, including that no stack trace, file path or trace key survives it, that the cap holds, and that an odd or unparseable Cause degrades to the error name. A new handler test proves the reason reaches the version record rather than only that None does. 363 publishing tests and 2767 app tests pass; cdk synth clean. * fix(packaging): Find a subscription's AMIs when they are deprecated describe_images hides deprecated images unless asked, so a Marketplace subscription whose AMIs are all past their deprecation date looked like a subscription with no AMI at all, and could not be imported. Both listings now pass IncludeDeprecated=True and capture DeprecationTime, which MarketplaceAmi exposes as a `deprecated` property. Including them changes which AMI gets picked, so both pickers now prefer a supported image. The hub lookup took amis[0] - an arbitrary image, since describe_images has no defined order - and now takes the newest supported one, falling back to the newest deprecated one only when every image is deprecated, so an older subscription still resolves to something. get_regional_ami_ids already took the newest and now applies the same preference; without that, adding IncludeDeprecated would have started handing deprecated AMIs to spoke accounts. get_resource_shares is paginated and only the first page was read, so the License Manager share could sit on a later page, the Organizations integration looked inactive, and an import that should have been allowed was blocked. It now follows nextToken. While in these files, the local variable and test names describing the AWS Organizations management account no longer use "master". The MasterAccountId and MasterAccountArn response fields keep their names, because those belong to the AWS API. Each behaviour was shown to fail against the previous code: the three hub-pick tests fail with amis[0] restored, and the paging test fails with the single-page read restored. 2772 app tests pass. * feat(packaging): Show when a Marketplace subscription's AMI is deprecated The backend now resolves a deprecated AMI when a product publishes no supported one, so the person importing has to be able to see that. `deprecated` and `deprecationTime` are carried on the discovered subscription read model, the API model and the OpenAPI schema, taken from whichever AMI the hub lookup settled on. In the UI: a red "Deprecated" badge beside the product name in both the subscriptions list and the import wizard's selector, and a warning in the wizard once a deprecated subscription is the selected one. It is a warning rather than a blocker because the import still works; the point is that the vendor no longer supports the image and may have published a replacement product. Only DiscoveredMarketplaceSubscription.ts is regenerated, deliberately. Regenerating the whole packaging client breaks the build: the committed client exports CreateRecipeRequestRecipePlatformEnum, ...RecipeArchitectureEnum and ...RecipeOsVersionEnum, which create-recipe.logic.ts imports, but the schema declares those three properties as plain `type: string` with no enum, so the generator drops them. The committed client was generated from an older schema and is stale. Restoring those enums to the schema is not a safe drive-by: the committed client lists six OS versions (Ubuntu 20/22/24, Windows Server 2019/2022/2025) while the domain's RecipeOsVersion now has two (Ubuntu 24, Windows Server 2025). Deciding which is authoritative changes what recipe creation accepts, which has nothing to do with this change. Left as a separate finding. 1112 packaging tests pass, tsc reports no new errors, 28 frontend tests pass. * fix(provisioning): Always report why a launch failed Two ways a failed workbench showed no usable reason. get_provisioned_product_failure_reason read the stack's failed resources and returned None when it found nothing, which happens when a launch fails before the stack exists or the stack is already gone. It now falls back to the Service Catalog record's errors, and then to the provisioned product's status message. Reading the reason could also be the thing that failed. __get_stack_events raises an AdapterException when the product or its events cannot be read, and that escaped to the caller, so a missing stack turned a reportable failure into an unhandled adapter error. Both reads are now contained: a reason that cannot be read is logged and reported as absent, never raised. "The instance configuration for this AWS Marketplace product is not supported" classified as UNKNOWN. It names no Availability Zone and does not start with "Unsupported", so it fell through every pattern. It is a template problem - the instance type in the product's parameters is not one the Marketplace offer permits - so it now classifies as TEMPLATE, which is the class that tells the user to change a parameter. Both service changes were shown to fail against the previous code: the fallback test returns None and the containment test raises. Two classification cases cover the bare message and the form CloudFormation wraps it in. 2776 app tests pass. * fix(publishing): Record a licence grant against the registered AMI The Marketplace decision set both of its AMI ids to the regional image. They have two different jobs. original_ami_id must be the registered image - the one the project imported - because that is what the publishing AMIs table is keyed by, and what create_license_grant now looks up to check the AMI belongs to the caller's project. With the regional id that lookup missed, so the authorization check added earlier in this branch would have rejected every Marketplace grant as "not part of project". The two changes are coupled: this is what makes that check work rather than merely compile. image_id stays the regional image, because that is the one the spoke launches, the one VerifyLaunchPermission asks about, and the one recorded as copiedAmiId on success. CreateLicenseGrantLambda's payload follows, reading $.decideActionResponse.originalAmiId instead of .imageId. Confirmed against the synthesised definition: the grant now takes originalAmiId while VerifyLaunchPermission and PassMarketplaceAmi still take imageId. When the product publishes no AMI for the spoke's region there is nothing regional to launch and both ids are the registered image, which a third test pins. The two id-separation tests fail against the previous code. 2779 app tests pass. * feat(publishing): Distribute a licence for an external subscription An externally-managed subscription means the customer owns the licence, so DecideAction only ever verified that the spoke already had a grant. When the spoke had none the decision stayed verify-only, the state machine failed the verification, and the launch could never succeed - even when the hub was perfectly able to distribute a grant itself. The decision now upgrades to LICENSE_GRANT when both hold: the spoke has no active received grant, and License Manager's organization integration is switched on. The second is a new port method, can_distribute_license, reading GetServiceSettings.OrganizationConfiguration.EnableIntegration, with license-manager:GetServiceSettings granted to the AMI-sharing and API functions. A spoke that already has a grant stays verify-only, because there is nothing to distribute. A VEW-managed subscription is untouched: it is already a grant decision and the upgrade does not consult License Manager for it, which a test pins. The licence service is an optional constructor argument, so the decision still works without it and the existing tests are unaffected. The home-region fallback that both the decision and the upgrade need is now one helper rather than two copies. Note for review: the description of this fix also mentioned organizations:DescribeOrganization and ram:GetResourceShares. This implementation uses GetServiceSettings alone, which is License Manager's own answer to "may I distribute", so only that action is granted rather than three. If the other two are load-bearing in your version, say so and I will match it. Four new tests, the upgrade one shown to fail against the previous decision. 2783 app tests pass; cdk synth confirms GetServiceSettings is granted and DeleteGrant and GetGrant are still absent. * test(publishing): Make the state machine mocks match the definition Every mock on the Marketplace paths disagreed with what the state machine actually reads, so these test cases could not have exercised the paths they are named after. MockedVerifyLaunchPermissionLambda returned launchPermissionVerified, but the ResultSelector reads $.Payload.isAccessible. The DecideAction mocks returned five of the ten fields the selector reads, missing imageId, projectId, licenseArn, homeRegion and grantName. MockedCreateLicenseGrantLambda returned grantArn without grantId or homeRegion. Each field list is taken from the synthesised definition rather than written by hand. Adds LaunchPermissionDeniedPath, covering a spoke that cannot launch the regional AMI. Nothing covered it, which is how the missing Choice after VerifyLaunchPermission went unnoticed: before that Choice existed this case would have been reported as a successful share. Also in this commit: deploy.sh gains Phase 7c, which reports the License Manager Organizations integration state and changes nothing. Enabling it is an account-wide decision for an administrator, and it cannot be done from a deploy, so the phase distinguishes enabled, disabled, and not determined - the last covering both no Organizations access and running outside the management account. docs/marketplace-integration.md gains the CLI form of enabling the integration, in order and with the reason the order matters; the ~30 minute delay before the resource share propagates; the fact that the grant decision is retaken at every publish, so enabling it later needs no re-import; a table of the Marketplace statusReason values; the instance-type limits behind "The instance configuration for this AWS Marketplace product is not supported"; and the deprecated-AMI behaviour. The Externally Managed section is corrected: it said such a version always fails without a manual grant, which is no longer true now that the hub distributes one when it can. The CLI parameter shapes are checked against botocore's service model. 2833 app and infra tests pass; the two failures are account_onboarding, which needs Step Functions Local, as do the ami_sharing tests themselves - so the mock corrections are verified against the definition, not by running them. * fix(publishing): Require the resource share before distributing a licence Resolves the divergence from the review on this check. It asked for license-manager:GetServiceSettings, organizations:DescribeOrganization and ram:GetResourceShares; can_distribute_license read only OrganizationConfiguration.EnableIntegration from the first. Reading the flag alone is not enough, and the reason is the propagation delay this branch documents elsewhere: EnableIntegration reads back true as soon as the setting is saved, but the RAM resource share that actually permits a cross-account grant takes around half an hour to appear. In that window the check said yes, an externally-managed subscription was upgraded to a grant, and CreateGrant then failed. The share is what has to exist, so require it. GetServiceSettings returns it as LicenseManagerResourceShareArn, which is the same fact organizations:DescribeOrganization plus ram:GetResourceShares arrives at the long way - so this is the stricter check the review was after while still needing one action rather than three. Four cases covered: integration on with the share, integration on while the share has not propagated, integration off, and no organization configuration at all. 2786 app tests pass. * fix(cli): Finish security review fixes Remove the remaining OAuth callback log leaks and handle API router lifecycle transitions across custom-domain and regional deployments. Correct the documented SSM endpoint path. * fix(mcp): Validate exposed tool metadata Validate names, descriptions, parameters, request properties, and local references on the exact transformed schemas uploaded to AgentCore. Add MCP-only names for cross-target collisions and document exposed inputs. * fix(marketplace): Close security review gaps Derive grant-sensitive Marketplace metadata server-side, require per-license and active organization readiness, and bind grant acceptance to persisted data. Tighten IAM and fix workflow, deprecation, deployment, and documentation regressions. * fix(cli): Keep ID token claim values out of errors A rejected ID token was reported as "token rejected: {kind:?}". The kind is not a fixed enum: ErrorKind::Json wraps the serde error, which quotes the rejected claim value, and other variants carry text. With --verbose the value reached the debug log as well as the error. Map each kind to a fixed name and drop its payload. The new redaction test signs a token whose sub is a number and fails against the previous commit. * fix(packaging): Import an AMI into one project only Packaging's upstream-id lookup and publishing's AMI read model both hold one image per AMI id. Publishing keys its row by AMI id alone and authorizes licence grants against that row's projectId. The import check only rejected a duplicate in the same project, so a second project importing the same AMI took the row over: the first project's grants then failed with "not part of project" and the AMI dropped out of its AMI list. Reject an import when any other project holds the AMI. Read every image for the AMI id, paged, rather than the first match. Only a Failed or Deleted image releases the AMI; Retired keeps it, because published versions may still use it. The error does not name the other project. Document the restriction. Sharing one AMI across projects needs a project-scoped read model in publishing and is left for a follow-up. * fix(publishing): Ship the Organizations stubs in the publishing layer The License Manager adapter imports mypy_boto3_organizations at module level for the resource-share check, but the publishing layer did not pin it. The dev virtualenv has every stub, so unit tests passed; the deployed publishing API failed to initialise with Runtime.ImportModuleError (seen on a fresh deploy). Pin it in the publishing layer and add a test that every runtime mypy_boto3_* import in a bounded context or app/shared is pinned in that context's layer lock or the shared lock. The test fails without the pin. * fix(mcp): Deploy the gateway only when its client is configured The gateway stack reads /<app>-ui-<env>/app-client-id/mcp-gateway-agentcore at deploy time, but it was registered on every deploy. Without an "mcp-gateway-agentcore" entry in ADDITIONAL_APP_CLIENTS the deploy failed ("Unable to fetch parameters"), seen on a fresh deployment. Register the stack only when deploy.sh finds that client and the frontend and backend share a region (the parameter is written in the frontend region), as the vew CLI router already does, and delete a stale gateway stack otherwise. Document the two optional clients. Also drop the 127.0.0.1 callback URLs: Cognito and the frontend's AdditionalAppClients validation accept plain http only for localhost, so registering the documented list failed the frontend deploy. A new test checks the list against that rule and fails on the old list. * fix(marketplace): Meet the MCP tool checks and trim packaging grants With the MCP gateway (#188) merged in, its synth-time tool-metadata check rejected the four operations this branch adds: their descriptions were one line and their parameters and request properties had none, so the backend could not synthesize. Describe them. The gateway branch also asserted that packaging holds no Marketplace or License Manager action. Marketplace discovery needs some, so pin the exact set instead. That turned up four actions packaging was granted but never calls (DescribeAgreement, GetAgreementTerms, GetProduct, ListFulfillmentOptions); only SearchAgreements is used, so drop the rest. * test(shared): Simplify the stub-import walk below the complexity limit flake8's max-complexity check rejected _runtime_stub_imports (C901, complexity 9). Split the import-name extraction into its own function; the behaviour is unchanged and the test still fails without the publishing layer's Organizations pin. * style(frontend): Apply the repo's lint autofix to the deprecation badges --------- Co-authored-by: raphaho <raphaho@amazon.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this change do?
Puts one Amazon Bedrock AgentCore Gateway in front of the projects, packaging, publishing
and provisioning APIs as MCP tool targets, so an agent can drive the workbench from any MCP
client.
enable_mcp=Trueon itsBackendAppOpenApi.infra/helpers/mcp_openapi_transformturns each context's OpenAPI document into theshape the gateway accepts. At synth time it rejects a tool with a missing, short or
duplicate description, an undocumented parameter or request property, a duplicate tool
name, or a broken
$ref.Authorizationheader to the targetAPI, so the existing Cedar authorizer still makes every authorization decision. The
gateway role can invoke the interceptor and read the spec bucket, and holds no
execute-api:Invoke.CUSTOM_JWTagainst the user pool, accepting onlyits own app client.
aws:SourceAccountand anaws:SourceArnmatching this stack's gateway. The gateway logkey grants the regional CloudWatch Logs principal, limited to its own log group.
mcp-gateway-agentcoreapp client is configured, asfeat(cli): Add the vew command-line client #187 does for its router.
deploy.shdeletes a gateway stack left over once it is nolonger configured.
The four API schemas gain operation and parameter descriptions. The gateway publishes them
as its tool descriptions, which is what an agent reasons over to choose a tool.
Type of change
How was this tested?
pytest app: 2685 passed.pytest infra/backend/tests infra/helpers/tests: 110 passed, 1 xfail.description text.
vitest: 28 passed.earlier run.
tools/list(with feat(packaging): Import images from an AMI or Marketplace subscription #200's operations on top) returned 104tools. None has an empty, short or duplicate description, no names are duplicated,
and none exposes an
Authorizationinput.tools/callthrough an MCP client completed the auth-code flow and returned data.Checklist
CONTRIBUTING.mdREADME.md)Notes for reviewers
Where to focus.
mcp_gateway_agentcore_app_stack.pyfor the gateway, its role and theauthorizer;
infra/helpers/mcp_openapi_transform.pyfor what becomes a tool; and theinterceptor handler, the only new code on the request path. The schema changes are
description text.
Setting it up. Add an
mcp-gateway-agentcoreentry toADDITIONAL_APP_CLIENTS. It isa public client, with no secret and auth-code with PKCE. Its callback URLs are
constants.MCP_GATEWAY_OAUTH_REDIRECT_URIS, i.e.http://localhost:<port>andhttp://localhost:<port>/callbackfor ports 19876–19880. Cognito accepts plain http onlyfor
localhost, so there is no127.0.0.1entry. The stack prints ready-to-paste MCPclient configs.
The gateway creates no Cognito client of its own. It reads the client id that
ADDITIONAL_APP_CLIENTSpublishes, so the authorizer's accepted-client list stays a singlemechanism. A test asserts no client is created.
release_component_versionnow accepts a request with no body, matching the schemachange that made the body optional.
Tool names.
x-mcp-operation-idgives two provisioning tools clearer names(
ListProvisionableProducts,ListProvisionableProductVersions).Acknowledgment
By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.