feat(cloud): read-only cloud-context MCP for GCP and AWS

CLAUDE.md

@@ -18,6 +18,7 @@ These skills live in the `feature-dev-workflow` plugin (`github.com/sourcehawk/f
 ## Operational rules
 
 - **TDD is the standard.** Failing test → watch it fail for the right reason → implement. One commit per task.
+- **Tests assert with `testify`.** Use `github.com/stretchr/testify/assert` for checks the test should keep running past, and `require` for preconditions a failure must stop at (a non-nil error before a dereference, setup that must succeed). Bare `t.Fatal` / `t.Errorf` is the rare exception, not the default.
 - **Before claiming done: `make test` + `make lint`; if `frontend/` touched, also `cd frontend && npm run typecheck`.** CI gates all three; local is the cheapest place to catch failures. Race-clean is non-negotiable.
 - **Commit conventions:** `feat(<area>): ...`, `fix(<area>): ...`, `refactor(<area>): ...`, `test(<area>): ...`, `chore(<area>): ...`. Area mirrors the module path.
 - **Never `--no-verify`, never `git add -A` / `git add .`.** Stage by name; pre-commit hooks exist for a reason.

cmd/triagent-mcp/serve.go

@@ -2,13 +2,18 @@ package main
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
 	"os"
 	"os/signal"
 	"strings"
 	"syscall"
 
+	"github.com/charmbracelet/log"
 	"github.com/sourcehawk/triagent/pkg/mcp/agentoperator"
+	"github.com/sourcehawk/triagent/pkg/mcp/cloud"
+	"github.com/sourcehawk/triagent/pkg/mcp/cloud/providers"
+	"github.com/sourcehawk/triagent/pkg/mcp/cloud/providers/aws"
 	"github.com/sourcehawk/triagent/pkg/mcp/git"
 	"github.com/sourcehawk/triagent/pkg/mcp/incidentio"
 	"github.com/sourcehawk/triagent/pkg/mcp/k8s"
@@ -21,28 +26,27 @@ import (
 	"github.com/sourcehawk/triagent/pkg/mcp/strategies"
 	"github.com/sourcehawk/triagent/pkg/mcp/teleport"
 	"github.com/sourcehawk/triagent/pkg/mcp/wiki"
-	"github.com/charmbracelet/log"
 	"github.com/spf13/cobra"
 )
 
 // Environment variable names. Flags override env when both are set.
 const (
-	envKubeconfig       = "TRIAGENT_MCP_KUBECONFIG"
-	envCRDsFile         = "TRIAGENT_MCP_CRDS_FILE"
-	envCrossplaneGroups = "TRIAGENT_MCP_CROSSPLANE_GROUPS"
-	envSessionDir                  = "TRIAGENT_MCP_SESSION_DIR"
-	envUserPlaybooksDir             = "TRIAGENT_MCP_USER_PLAYBOOKS_DIR"
-	envPluginPlaybooksDir           = "TRIAGENT_MCP_PLUGIN_PLAYBOOKS_DIR"
-	envSystemPlaybooksDir           = "TRIAGENT_MCP_SYSTEM_PLAYBOOKS_DIR"
-	envStrategiesSubagentModel      = "TRIAGENT_MCP_STRATEGIES_SUBAGENT_MODEL"
-	envMCPConfigPath                = "TRIAGENT_MCP_CONFIG_PATH"
-	envGitRepo               = "TRIAGENT_MCP_GIT_REPO"
-	envGitCacheDir           = "TRIAGENT_MCP_GIT_CACHE_DIR"
-	envGitClaudeBinary       = "TRIAGENT_MCP_GIT_CLAUDE_BINARY"
-	envGitFilterPrereleases  = "TRIAGENT_MCP_GIT_FILTER_PRERELEASES"
-	envWikiServePath         = "TRIAGENT_MCP_WIKI_PATH"
-	envWikiServeProposalsPath = "TRIAGENT_MCP_WIKI_PROPOSALS_PATH"
-	envWikiServeClaudeBinary = "TRIAGENT_MCP_WIKI_CLAUDE_BINARY"
+	envKubeconfig              = "TRIAGENT_MCP_KUBECONFIG"
+	envCRDsFile                = "TRIAGENT_MCP_CRDS_FILE"
+	envCrossplaneGroups        = "TRIAGENT_MCP_CROSSPLANE_GROUPS"
+	envSessionDir              = "TRIAGENT_MCP_SESSION_DIR"
+	envUserPlaybooksDir        = "TRIAGENT_MCP_USER_PLAYBOOKS_DIR"
+	envPluginPlaybooksDir      = "TRIAGENT_MCP_PLUGIN_PLAYBOOKS_DIR"
+	envSystemPlaybooksDir      = "TRIAGENT_MCP_SYSTEM_PLAYBOOKS_DIR"
+	envStrategiesSubagentModel = "TRIAGENT_MCP_STRATEGIES_SUBAGENT_MODEL"
+	envMCPConfigPath           = "TRIAGENT_MCP_CONFIG_PATH"
+	envGitRepo                 = "TRIAGENT_MCP_GIT_REPO"
+	envGitCacheDir             = "TRIAGENT_MCP_GIT_CACHE_DIR"
+	envGitClaudeBinary         = "TRIAGENT_MCP_GIT_CLAUDE_BINARY"
+	envGitFilterPrereleases    = "TRIAGENT_MCP_GIT_FILTER_PRERELEASES"
+	envWikiServePath           = "TRIAGENT_MCP_WIKI_PATH"
+	envWikiServeProposalsPath  = "TRIAGENT_MCP_WIKI_PROPOSALS_PATH"
+	envWikiServeClaudeBinary   = "TRIAGENT_MCP_WIKI_CLAUDE_BINARY"
 
 	envSessionsProposalsPath = "TRIAGENT_MCP_SESSIONS_PROPOSALS_PATH"
 	envSessionsClaudeBinary  = "TRIAGENT_MCP_SESSIONS_CLAUDE_BINARY"
@@ -71,10 +75,10 @@ type serveFlags struct {
 	systemPlaybooksDir string
 
 	// git flags
-	gitRepo               string
-	gitCacheDir           string
-	gitClaudeBinary       string
-	gitFilterPrereleases  bool
+	gitRepo              string
+	gitCacheDir          string
+	gitClaudeBinary      string
+	gitFilterPrereleases bool
 
 	// wiki flags
 	wikiPath          string
@@ -95,6 +99,9 @@ type serveFlags struct {
 	promURL    string
 	promBearer string
 	promBasic  string
+
+	// cloud flags
+	cloudProvider string
 }
 
 func serveCmd() *cobra.Command {
@@ -104,14 +111,14 @@ func serveCmd() *cobra.Command {
 		Short: "Run one of the triagent-mcp MCP servers over stdio",
 		Long: "Run one of the triagent-mcp MCP servers over stdio.\n\n" +
 			"Select the server via --kind. Supported kinds:\n" +
-			"  k8s, teleport, strategies, git, wiki, slack, incidentio, sessions, meta, agent-operator, signal-ingest, parallel, prom",
+			"  k8s, teleport, strategies, git, wiki, slack, incidentio, sessions, meta, agent-operator, signal-ingest, parallel, prom, cloud",
 		Hidden: true,
 		Args:   cobra.NoArgs,
 		RunE: func(cmd *cobra.Command, args []string) error {
 			return runServe(cmd.Context(), resolveFlags(f))
 		},
 	}
-	cmd.Flags().StringVar(&f.kind, "kind", "", "which MCP server to run: k8s, teleport, strategies, git, wiki, slack, incidentio, sessions, meta, agent-operator, signal-ingest, parallel, or prom (required)")
+	cmd.Flags().StringVar(&f.kind, "kind", "", "which MCP server to run: k8s, teleport, strategies, git, wiki, slack, incidentio, sessions, meta, agent-operator, signal-ingest, parallel, prom, or cloud (required)")
 
 	// k8s flags
 	cmd.Flags().StringVar(&f.kubeconfig, "kubeconfig", "", "path to kubeconfig (defaults to $"+envKubeconfig+", then $KUBECONFIG, then ~/.kube/config) [kind=k8s]")
@@ -150,6 +157,9 @@ func serveCmd() *cobra.Command {
 	cmd.Flags().StringVar(&f.promBearer, "prom-bearer", "", "Authorization: Bearer token for Prometheus (defaults to $"+envPromBearer+") [kind=prom]")
 	cmd.Flags().StringVar(&f.promBasic, "prom-basic", "", "Basic auth credentials user:pass for Prometheus (defaults to $"+envPromBasic+") [kind=prom]")
 
+	// cloud flags
+	cmd.Flags().StringVar(&f.cloudProvider, "provider", "", "cloud provider to serve: gcp or aws; required (defaults to $"+cloud.EnvProvider+") [kind=cloud]")
+
 	return cmd
 }
 
@@ -215,6 +225,9 @@ func resolveFlags(f *serveFlags) serveFlags {
 	if out.promBasic == "" {
 		out.promBasic = os.Getenv(envPromBasic)
 	}
+	if out.cloudProvider == "" {
+		out.cloudProvider = os.Getenv(cloud.EnvProvider)
+	}
 	// Bool env override: only consider when the operator hasn't passed
 	// the flag explicitly. Cobra preserves the flag default (true) when
 	// unset, so we can't distinguish "operator passed --filter-prereleases=true"
@@ -263,10 +276,12 @@ func runServe(ctx context.Context, f serveFlags) error {
 		return runParallel(ctx, f)
 	case "prom":
 		return runProm(ctx, f)
+	case "cloud":
+		return runCloud(ctx, f)
 	case "":
-		return fmt.Errorf("--kind is required (one of: k8s, teleport, strategies, git, wiki, slack, incidentio, sessions, meta, agent-operator, signal-ingest, parallel, prom)")
+		return fmt.Errorf("--kind is required (one of: k8s, teleport, strategies, git, wiki, slack, incidentio, sessions, meta, agent-operator, signal-ingest, parallel, prom, cloud)")
 	default:
-		return fmt.Errorf("unknown --kind %q (want one of: k8s, teleport, strategies, git, wiki, slack, incidentio, sessions, meta, agent-operator, signal-ingest, parallel, prom)", f.kind)
+		return fmt.Errorf("unknown --kind %q (want one of: k8s, teleport, strategies, git, wiki, slack, incidentio, sessions, meta, agent-operator, signal-ingest, parallel, prom, cloud)", f.kind)
 	}
 }
 
@@ -423,6 +438,86 @@ func runProm(ctx context.Context, f serveFlags) error {
 	return srv.Run(ctx)
 }
 
+// runCloud wires the read-only cloud-context MCP. --provider selects the
+// concrete backend; New plugs it in behind cloud.Provider. The launcher passes
+// the allowlist override path, target scope, and pinned identity through the
+// subprocess env (cloud.EnvAllowlistPath, cloud.EnvScope,
+// cloud.EnvExpectedIdentity), never argv. A multi-account aws source additionally
+// carries its accounts, source_profile, and alias (cloud.EnvAWSAccounts,
+// _SOURCE_PROFILE, _ALIAS); New generates the per-account assume-role profiles
+// and surfaces the accounts as the agent's selectable targets.
+func runCloud(ctx context.Context, f serveFlags) error {
+	if f.cloudProvider == "" {
+		return fmt.Errorf("--provider is required (gcp or aws) (set --provider or $%s)", cloud.EnvProvider)
+	}
+	scope, err := parseCloudScope(os.Getenv(cloud.EnvScope))
+	if err != nil {
+		return fmt.Errorf("build cloud mcp server: %w", err)
+	}
+	accounts, err := parseAWSAccounts(os.Getenv(cloud.EnvAWSAccounts))
+	if err != nil {
+		return fmt.Errorf("build cloud mcp server: %w", err)
+	}
+	provider, err := providers.New(f.cloudProvider, providers.Options{
+		AWSAlias:         os.Getenv(cloud.EnvAWSAlias),
+		AWSSourceProfile: os.Getenv(cloud.EnvAWSSourceProfile),
+		AWSAccounts:      accounts,
+	})
+	if err != nil {
+		return err
+	}
+	srv, err := cloud.New(cloud.Options{
+		Provider:         provider,
+		AllowlistPath:    os.Getenv(cloud.EnvAllowlistPath),
+		Scope:            scope,
+		ExpectedIdentity: os.Getenv(cloud.EnvExpectedIdentity),
+	})
+	if err != nil {
+		return fmt.Errorf("build cloud mcp server: %w", err)
+	}
+	log.Info("mcp serve --kind=cloud starting", "provider", f.cloudProvider)
+	return srv.Run(ctx)
+}
+
+// parseCloudScope decodes the JSON-encoded target scope the launcher froze into
+// a cloud.ScopeAllowlist. An empty value yields an empty scope, which leaves the
+// target axes unconstrained. A malformed value is an error that aborts startup:
+// failing closed, since a misconfigured scope must never silently widen run_cli
+// by dropping the deployment's restrictions.
+func parseCloudScope(raw string) (cloud.ScopeAllowlist, error) {
+	var scope cloud.ScopeAllowlist
+	if raw == "" {
+		return scope, nil
+	}
+	if err := json.Unmarshal([]byte(raw), &scope); err != nil {
+		return cloud.ScopeAllowlist{}, fmt.Errorf("malformed cloud scope in $%s: %w", cloud.EnvScope, err)
+	}
+	return scope, nil
+}
+
+// parseAWSAccounts decodes the JSON-encoded aws multi-account set the launcher
+// froze into the aws provider's account list. An empty value yields nil, the
+// single-account / single-identity form. A malformed value is an error that
+// aborts startup: failing closed, since a misconfigured accounts list must never
+// silently drop accounts the agent should be able to select.
+func parseAWSAccounts(raw string) ([]aws.Account, error) {
+	if raw == "" {
+		return nil, nil
+	}
+	var wire []struct {
+		AccountID string `json:"account_id"`
+		RoleARN   string `json:"role_arn"`
+	}
+	if err := json.Unmarshal([]byte(raw), &wire); err != nil {
+		return nil, fmt.Errorf("malformed cloud aws accounts in $%s: %w", cloud.EnvAWSAccounts, err)
+	}
+	accounts := make([]aws.Account, 0, len(wire))
+	for _, w := range wire {
+		accounts = append(accounts, aws.Account{ID: w.AccountID, RoleARN: w.RoleARN})
+	}
+	return accounts, nil
+}
+
 func runGit(ctx context.Context, f serveFlags) error {
 	if f.gitRepo == "" {
 		return fmt.Errorf("--repo is required (owner/name) (set --repo or $%s)", envGitRepo)

cmd/triagent-mcp/serve_cloud_test.go

@@ -0,0 +1,97 @@
+package main
+
+import (
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestRunServe_CloudKindRequiresProvider(t *testing.T) {
+	t.Parallel()
+	err := runServe(context.Background(), serveFlags{kind: "cloud"})
+	require.Error(t, err, "expected error when --provider is missing")
+	assert.Contains(t, err.Error(), "provider", "error should mention --provider")
+}
+
+func TestRunServe_CloudKindRejectsUnknownProvider(t *testing.T) {
+	t.Parallel()
+	err := runServe(context.Background(), serveFlags{kind: "cloud", cloudProvider: "azure"})
+	require.Error(t, err, "expected error for an unknown provider")
+	assert.Contains(t, err.Error(), "azure", "error should name the rejected provider")
+}
+
+func TestRunServe_UnknownKindErrorListsCloud(t *testing.T) {
+	t.Parallel()
+	err := runServe(context.Background(), serveFlags{kind: "bogus"})
+	require.Error(t, err, "expected error for unknown kind")
+	assert.Contains(t, err.Error(), "cloud", "kind list should include cloud")
+}
+
+func TestServeCmd_KnowsCloudKind(t *testing.T) {
+	t.Parallel()
+	cmd := serveCmd()
+	assert.Contains(t, cmd.Long, "cloud", "serve --help should list cloud")
+}
+
+func TestParseCloudScope_EmptyYieldsUnconstrained(t *testing.T) {
+	t.Parallel()
+	scope, err := parseCloudScope("")
+	require.NoError(t, err)
+	assert.Empty(t, scope.Projects)
+	assert.Empty(t, scope.Regions)
+	assert.Empty(t, scope.Accounts)
+}
+
+func TestParseCloudScope_ValidJSON(t *testing.T) {
+	t.Parallel()
+	scope, err := parseCloudScope(`{"projects":["prod"],"regions":["us-central1"]}`)
+	require.NoError(t, err)
+	assert.Equal(t, []string{"prod"}, scope.Projects)
+	assert.Equal(t, []string{"us-central1"}, scope.Regions)
+}
+
+func TestParseCloudScope_MalformedFailsClosed(t *testing.T) {
+	t.Parallel()
+	_, err := parseCloudScope(`{"projects":`)
+	require.Error(t, err, "a malformed scope must fail closed, not silently drop restrictions")
+}
+
+func TestRunCloud_MalformedScopeAborts(t *testing.T) {
+	t.Setenv("TRIAGENT_CLOUD_PROVIDER", "gcp")
+	t.Setenv("TRIAGENT_CLOUD_SCOPE", `{"projects":`)
+	err := runCloud(context.Background(), serveFlags{kind: "cloud", cloudProvider: "gcp"})
+	require.Error(t, err, "a malformed scope must abort cloud-server startup")
+	assert.Contains(t, err.Error(), "scope", "the error should name the scope")
+}
+
+func TestParseAWSAccounts_EmptyYieldsNil(t *testing.T) {
+	t.Parallel()
+	accs, err := parseAWSAccounts("")
+	require.NoError(t, err)
+	assert.Nil(t, accs)
+}
+
+func TestParseAWSAccounts_DecodesJSON(t *testing.T) {
+	t.Parallel()
+	accs, err := parseAWSAccounts(`[{"account_id":"111111111111","role_arn":"arn:aws:iam::111111111111:role/r"},{"account_id":"222222222222","role_arn":"arn:aws:iam::222222222222:role/r"}]`)
+	require.NoError(t, err)
+	require.Len(t, accs, 2)
+	assert.Equal(t, "111111111111", accs[0].ID)
+	assert.Equal(t, "arn:aws:iam::222222222222:role/r", accs[1].RoleARN)
+}
+
+func TestParseAWSAccounts_MalformedFailsClosed(t *testing.T) {
+	t.Parallel()
+	_, err := parseAWSAccounts(`[{"account_id":`)
+	require.Error(t, err, "a malformed accounts list must fail closed, not silently drop accounts")
+}
+
+func TestRunCloud_MalformedAWSAccountsAborts(t *testing.T) {
+	t.Setenv("TRIAGENT_CLOUD_PROVIDER", "aws")
+	t.Setenv("TRIAGENT_CLOUD_AWS_ACCOUNTS", `[{"account_id":`)
+	err := runCloud(context.Background(), serveFlags{kind: "cloud", cloudProvider: "aws"})
+	require.Error(t, err, "a malformed accounts list must abort cloud-server startup")
+	assert.Contains(t, err.Error(), "accounts", "the error should name the accounts")
+}