fix(ai): Redact session input and gate state-changing tools on approval

internal/api/ai_session_handlers.go

@@ -14,6 +14,18 @@ import (
 	"github.com/whilesmartgo/agents"
 )
 
+// redactSessionInput strips known secret values from operator-supplied text
+// before it enters the transcript, so seeded file content or pasted output
+// never carries a credential to the model provider.
+func (s *Server) redactSessionInput(sess *ai.Session, text string) string {
+	secrets := s.systemSecretValues()
+	if sess.Deployment != "" {
+		secrets = s.deploymentSecretValues(sess.Deployment)
+	}
+	redacted, _ := ai.NewRedactor(secrets).Redact(text)
+	return redacted
+}
+
 // composeUserMessage merges a short message with optional bulky
 // context. The model sees both; the operator's transcript shows only
 // the message.
@@ -72,6 +84,8 @@ func (s *Server) advanceSession(c *gin.Context, sess *ai.Session) error {
 
 // aiRunner assembles the runner for one session turn. A session that does not
 // auto-run pauses before any tools run, surfacing them for per-call approval.
+// Even with auto-run on, a batch containing a state-changing tool pauses:
+// reads run free, writes always ask.
 func (s *Server) aiRunner(c *gin.Context, sess *ai.Session, engine agents.Engine) agents.Runner {
 	runner := agents.Runner{
 		Engine: engine,
@@ -82,8 +96,16 @@ func (s *Server) aiRunner(c *gin.Context, sess *ai.Session, engine agents.Engine
 			StepLimitMessage: aiStepLimitMessage,
 		},
 	}
-	if !sess.AutoRun {
-		runner.Approve = func(context.Context, []agents.ToolCall) (bool, error) { return false, nil }
+	runner.Approve = func(_ context.Context, calls []agents.ToolCall) (bool, error) {
+		if !sess.AutoRun {
+			return false, nil
+		}
+		for _, call := range calls {
+			if s.toolMutates(call.Name) {
+				return false, nil
+			}
+		}
+		return true, nil
 	}
 	return runner
 }
@@ -220,7 +242,7 @@ func (s *Server) createAISession(c *gin.Context) {
 	prompt := ai.BuildSessionPrompt(req.Scope, req.Deployment, s.config.AI.DocsURL)
 	sess := ai.NewSession(req.Scope, req.Deployment, req.AutoRun, sessionActorFrom(c), prompt)
 	content, display := composeUserMessage(req.Message, req.Context)
-	sess.AddUserMessage(content, display, req.Seed)
+	sess.AddUserMessage(s.redactSessionInput(sess, content), display, req.Seed)
 
 	if err := s.advanceSession(c, sess); err != nil {
 		c.JSON(http.StatusBadGateway, gin.H{"error": err.Error()})
@@ -303,7 +325,7 @@ func (s *Server) postAISessionMessage(c *gin.Context) {
 	}
 
 	content, display := composeUserMessage(req.Message, req.Context)
-	sess.AddUserMessage(content, display, false)
+	sess.AddUserMessage(s.redactSessionInput(sess, content), display, false)
 	if err := s.advanceSession(c, sess); err != nil {
 		c.JSON(http.StatusBadGateway, gin.H{"error": err.Error()})
 		return

internal/api/ai_session_test.go

@@ -5,6 +5,8 @@ import (
 	"encoding/json"
 	"net/http"
 	"net/http/httptest"
+	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 
@@ -305,6 +307,81 @@ func TestAISessionPerCallApproval(t *testing.T) {
 	}
 }
 
+func TestAutoRunStillPausesForMutatingTools(t *testing.T) {
+	s, tmpDir, ts := setupPlanTestServer(t)
+	createTestDeployment(t, tmpDir, "myapp", &models.ServiceMetadata{Name: "myapp"})
+
+	s.aiProvider = &scriptedProvider{responses: []*ai.Response{
+		{ToolCalls: []ai.ToolCall{{ID: "c1", Name: "write_deployment_file",
+			Arguments: `{"path":"conf/app.conf","content":"key = value\n"}`}}, Model: "scripted"},
+		{Content: "Written.", Model: "scripted"},
+	}}
+
+	resp, parsed := doJSON(t, http.MethodPost, ts.URL+"/api/ai/sessions", map[string]interface{}{
+		"scope": "deployment", "deployment": "myapp", "auto_run": true,
+		"message": "set key to value in conf/app.conf",
+	})
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("status = %d, body %v", resp.StatusCode, parsed)
+	}
+	// Auto-run must not silently execute a state-changing tool.
+	if parsed["status"] != "awaiting_approval" {
+		t.Fatalf("status = %v, want awaiting_approval", parsed["status"])
+	}
+	if _, err := os.Stat(filepath.Join(tmpDir, "myapp", "conf", "app.conf")); !os.IsNotExist(err) {
+		t.Fatal("the file must not be written before approval")
+	}
+
+	id := parsed["id"].(string)
+	resp, parsed = doJSON(t, http.MethodPost, ts.URL+"/api/ai/sessions/"+id+"/approve",
+		map[string]interface{}{"approved": map[string]bool{"c1": true}})
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("approve status = %d, body %v", resp.StatusCode, parsed)
+	}
+	if parsed["status"] != "ready" {
+		t.Errorf("status after approve = %v", parsed["status"])
+	}
+	data, err := os.ReadFile(filepath.Join(tmpDir, "myapp", "conf", "app.conf"))
+	if err != nil || string(data) != "key = value\n" {
+		t.Errorf("approved write not applied: %q err=%v", string(data), err)
+	}
+}
+
+func TestAISessionRedactsSeededSecrets(t *testing.T) {
+	s, tmpDir, ts := setupPlanTestServer(t)
+	createTestDeployment(t, tmpDir, "myapp", &models.ServiceMetadata{Name: "myapp"})
+	envContent := "DB_PASSWORD=hunter2secret\nAPP_NAME=myapp\n"
+	if err := os.WriteFile(filepath.Join(tmpDir, "myapp", ".env.flatrun"), []byte(envContent), 0600); err != nil {
+		t.Fatal(err)
+	}
+
+	stub := &scriptedProvider{responses: []*ai.Response{{Content: "Looks fine.", Model: "scripted"}}}
+	s.aiProvider = stub
+
+	resp, parsed := doJSON(t, http.MethodPost, ts.URL+"/api/ai/sessions", map[string]interface{}{
+		"scope":      "deployment",
+		"deployment": "myapp",
+		"auto_run":   true,
+		"message":    "Review this file.",
+		"context":    "```\nDB_PASSWORD=hunter2secret\nhost=db\n```",
+	})
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("status = %d, body %v", resp.StatusCode, parsed)
+	}
+
+	// The provider must never see the secret value from the seeded context.
+	var sent strings.Builder
+	for _, m := range stub.lastRequestMessages() {
+		sent.WriteString(m.Content)
+	}
+	if strings.Contains(sent.String(), "hunter2secret") {
+		t.Error("a deployment env value in seeded context reached the provider")
+	}
+	if !strings.Contains(sent.String(), "[REDACTED]") {
+		t.Error("expected the seeded secret to be replaced with a redaction marker")
+	}
+}
+
 func TestAISessionDisabledReturns503(t *testing.T) {
 	_, _, ts := setupPlanTestServer(t)
 	resp, parsed := doJSON(t, http.MethodPost, ts.URL+"/api/ai/sessions",

internal/api/ai_tools.go

@@ -20,11 +20,14 @@ import (
 
 const maxToolOutputChars = 8000
 
-// aiTool is one read-only investigation capability the model can call
-// to discover facts about this installation instead of guessing.
+// aiTool is one capability the model can call: read-only investigation
+// by default, or a state change when Mutates is set.
 type aiTool struct {
 	Spec ai.Tool
-	Run  func(s *Server, c *gin.Context, boundDeployment string, args map[string]interface{}) (string, error)
+	// Mutates marks a tool that changes state. A batch containing one always
+	// pauses for operator approval, even in an auto-run session.
+	Mutates bool
+	Run     func(s *Server, c *gin.Context, boundDeployment string, args map[string]interface{}) (string, error)
 }
 
 // destructiveCommand matches obviously state-changing shell tokens, so
@@ -304,6 +307,7 @@ func (s *Server) aiToolRegistry() map[string]aiTool {
 			},
 		},
 		"write_deployment_file": {
+			Mutates: true,
 			Spec: ai.Tool{
 				Name:        "write_deployment_file",
 				Description: "Create or overwrite a text file inside a deployment's directory, for example a config or compose file. Requires write access to the deployment; the path cannot escape the deployment directory.",
@@ -338,6 +342,7 @@ func (s *Server) aiToolRegistry() map[string]aiTool {
 			},
 		},
 		"run_quick_action": {
+			Mutates: true,
 			Spec: ai.Tool{
 				Name:        "run_quick_action",
 				Description: "Run one of a deployment's configured quick actions by its id. Requires write access to the deployment.",
@@ -370,6 +375,7 @@ func (s *Server) aiToolRegistry() map[string]aiTool {
 			},
 		},
 		"control_deployment": {
+			Mutates: true,
 			Spec: ai.Tool{
 				Name:        "control_deployment",
 				Description: "Start, stop, or restart a whole deployment. Requires write access to the deployment.",
@@ -608,6 +614,17 @@ func (s *Server) runAITool(c *gin.Context, boundDeployment string, call ai.ToolC
 	return result
 }
 
+// toolMutates reports whether a named tool changes state, covering both
+// built-in tools and plugin-contributed ones.
+func (s *Server) toolMutates(name string) bool {
+	if plugin, tool, ok := parsePluginToolName(name); ok {
+		spec, found := s.pluginToolSpec(plugin, tool)
+		return found && spec.Mutates
+	}
+	t, ok := s.aiToolRegistry()[name]
+	return ok && t.Mutates
+}
+
 // toolAllowedDeploymentWrite resolves the target deployment and verifies the actor may write
 // to it, gating a mutating plugin tool.
 func (s *Server) toolAllowedDeploymentWrite(c *gin.Context, boundDeployment string, args map[string]interface{}) (string, error) {