Merge remote-tracking branch 'upstream/dev' into codex/merge-nextlevelbuilder-goclaw

# Conflicts:
#	internal/pipeline/think_stage.go
#	internal/pipeline/tool_stage.go
#	tests/integration/git_adapter_ssh_test.go

.github/workflows/release.yaml

@@ -161,94 +161,27 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        variant: [base, latest, full]
-        platform: [linux/amd64, linux/arm64]
         include:
           - variant: base
+            suffix: "-base"
             enable_embedui: "false"
             enable_python: "false"
             enable_full_skills: "false"
           - variant: latest
+            suffix: ""
             enable_embedui: "true"
             enable_python: "true"
             enable_full_skills: "false"
           - variant: full
+            suffix: "-full"
             enable_embedui: "true"
             enable_python: "true"
             enable_full_skills: "true"
-          - platform: linux/amd64
-            runner: ubuntu-latest
-            arch: amd64
-          - platform: linux/arm64
-            runner: ubuntu-24.04-arm
-            arch: arm64
-    runs-on: ${{ matrix.runner }}
+    runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
 
-      - uses: docker/setup-buildx-action@v3
-
-      - name: Log in to GHCR
-        uses: docker/login-action@v3
-        with:
-          registry: ghcr.io
-          username: ${{ github.actor }}
-          password: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Build and push by digest
-        id: build
-        uses: docker/build-push-action@v6
-        with:
-          context: .
-          platforms: ${{ matrix.platform }}
-          outputs: type=image,name=${{ env.GHCR_IMAGE }},push-by-digest=true,name-canonical=true,push=true
-          build-args: |
-            ENABLE_OTEL=false
-            ENABLE_EMBEDUI=${{ matrix.enable_embedui }}
-            ENABLE_PYTHON=${{ matrix.enable_python }}
-            ENABLE_FULL_SKILLS=${{ matrix.enable_full_skills }}
-            VERSION=v${{ needs.release.outputs.version }}
-          cache-from: type=gha,scope=${{ matrix.variant }}-${{ matrix.arch }}
-          cache-to: type=gha,mode=max,scope=${{ matrix.variant }}-${{ matrix.arch }}
-          provenance: false
-
-      - name: Export digest
-        run: |
-          mkdir -p /tmp/digests
-          digest="${{ steps.build.outputs.digest }}"
-          touch "/tmp/digests/${digest#sha256:}"
-
-      - uses: actions/upload-artifact@v4
-        with:
-          name: digests-${{ matrix.variant }}-${{ matrix.arch }}
-          path: /tmp/digests/*
-          if-no-files-found: error
-          retention-days: 1
-
-  # Combine per-platform digests into multi-arch manifests and push final tags
-  # to both GHCR and Docker Hub. One job per variant runs in parallel.
-  docker-merge:
-    needs: [release, docker-build]
-    if: needs.release.outputs.released == 'true'
-    runs-on: ubuntu-latest
-    timeout-minutes: 15
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - variant: base
-            suffix: "-base"
-          - variant: latest
-            suffix: ""
-          - variant: full
-            suffix: "-full"
-    steps:
-      - name: Download digests
-        uses: actions/download-artifact@v4
-        with:
-          path: /tmp/digests
-          pattern: digests-${{ matrix.variant }}-*
-          merge-multiple: true
+      - uses: docker/setup-qemu-action@v3
 
       - uses: docker/setup-buildx-action@v3
 
@@ -277,92 +210,34 @@ jobs:
             type=raw,value=latest,enable=${{ matrix.suffix == '' }},suffix=
             type=raw,value=${{ matrix.variant }},enable=${{ matrix.suffix != '' }}
 
-      - name: Create manifest list and push
-        working-directory: /tmp/digests
-        run: |
-          # Compose -t flags from metadata-action tags (newline-separated)
-          TAG_ARGS=()
-          while IFS= read -r tag; do
-            [ -n "$tag" ] && TAG_ARGS+=("-t" "$tag")
-          done <<< "${{ steps.meta.outputs.tags }}"
-          # Compose source digest refs (pushed to GHCR by docker-build)
-          DIGEST_ARGS=()
-          for f in *; do
-            DIGEST_ARGS+=("${{ env.GHCR_IMAGE }}@sha256:$f")
-          done
-          set -x
-          docker buildx imagetools create "${TAG_ARGS[@]}" "${DIGEST_ARGS[@]}"
+      - name: Build and push
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          platforms: linux/amd64,linux/arm64
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          build-args: |
+            ENABLE_OTEL=false
+            ENABLE_EMBEDUI=${{ matrix.enable_embedui }}
+            ENABLE_PYTHON=${{ matrix.enable_python }}
+            ENABLE_FULL_SKILLS=${{ matrix.enable_full_skills }}
+            VERSION=v${{ needs.release.outputs.version }}
+          cache-from: type=gha,scope=release-${{ matrix.variant }}
+          cache-to: type=gha,mode=max,scope=release-${{ matrix.variant }}
+          provenance: false
 
-      - name: Inspect manifest
-        run: |
-          docker buildx imagetools inspect \
-            "${{ env.GHCR_IMAGE }}:v${{ needs.release.outputs.version }}${{ matrix.suffix }}"
-
-  # Build web UI Docker image per platform on native runners.
+  # Build web UI Docker image.
   docker-web-build:
     needs: release
     if: needs.release.outputs.released == 'true'
     timeout-minutes: 30
-    strategy:
-      fail-fast: false
-      matrix:
-        platform: [linux/amd64, linux/arm64]
-        include:
-          - platform: linux/amd64
-            runner: ubuntu-latest
-            arch: amd64
-          - platform: linux/arm64
-            runner: ubuntu-24.04-arm
-            arch: arm64
-    runs-on: ${{ matrix.runner }}
+    runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
 
-      - uses: docker/setup-buildx-action@v3
-
-      - name: Log in to GHCR
-        uses: docker/login-action@v3
-        with:
-          registry: ghcr.io
-          username: ${{ github.actor }}
-          password: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Build and push by digest
-        id: build
-        uses: docker/build-push-action@v6
-        with:
-          context: ui/web
-          platforms: ${{ matrix.platform }}
-          outputs: type=image,name=${{ env.GHCR_IMAGE }}-web,push-by-digest=true,name-canonical=true,push=true
-          cache-from: type=gha,scope=web-${{ matrix.arch }}
-          cache-to: type=gha,mode=max,scope=web-${{ matrix.arch }}
-          provenance: false
-
-      - name: Export digest
-        run: |
-          mkdir -p /tmp/digests
-          digest="${{ steps.build.outputs.digest }}"
-          touch "/tmp/digests/${digest#sha256:}"
-
-      - uses: actions/upload-artifact@v4
-        with:
-          name: digests-web-${{ matrix.arch }}
-          path: /tmp/digests/*
-          if-no-files-found: error
-          retention-days: 1
-
-  docker-web-merge:
-    needs: [release, docker-web-build]
-    if: needs.release.outputs.released == 'true'
-    runs-on: ubuntu-latest
-    timeout-minutes: 15
-    steps:
-      - name: Download digests
-        uses: actions/download-artifact@v4
-        with:
-          path: /tmp/digests
-          pattern: digests-web-*
-          merge-multiple: true
+      - uses: docker/setup-qemu-action@v3
 
       - uses: docker/setup-buildx-action@v3
 
@@ -390,28 +265,21 @@ jobs:
             type=raw,value=v${{ needs.release.outputs.version }}
             type=raw,value=latest
 
-      - name: Create manifest list and push
-        working-directory: /tmp/digests
-        run: |
-          TAG_ARGS=()
-          while IFS= read -r tag; do
-            [ -n "$tag" ] && TAG_ARGS+=("-t" "$tag")
-          done <<< "${{ steps.meta.outputs.tags }}"
-          DIGEST_ARGS=()
-          for f in *; do
-            DIGEST_ARGS+=("${{ env.GHCR_IMAGE }}-web@sha256:$f")
-          done
-          set -x
-          docker buildx imagetools create "${TAG_ARGS[@]}" "${DIGEST_ARGS[@]}"
-
-      - name: Inspect manifest
-        run: |
-          docker buildx imagetools inspect \
-            "${{ env.GHCR_IMAGE }}-web:v${{ needs.release.outputs.version }}"
+      - name: Build and push
+        uses: docker/build-push-action@v6
+        with:
+          context: ui/web
+          platforms: linux/amd64,linux/arm64
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          cache-from: type=gha,scope=release-web
+          cache-to: type=gha,mode=max,scope=release-web
+          provenance: false
 
   # Notify Discord on new release (runs even if docker jobs fail)
   notify-discord:
-    needs: [release, build-binaries, docker-merge, docker-web-merge]
+    needs: [release, build-binaries, docker-build, docker-web-build]
     if: always() && needs.release.outputs.released == 'true' && !cancelled()
     runs-on: ubuntu-latest
     timeout-minutes: 5

.gitignore

@@ -82,6 +82,8 @@ goclaw.db-shm
 *.app.bak
 *.exe.bak
 # AI tool config directories (user-specific)
+.agents/
+.codex/
 .gemini/
 .opencode/
 

docs/02-providers.md

@@ -554,7 +554,7 @@ ClaudeCLIProvider can be configured in `config.json`:
 }
 ```
 
-Or via database `llm_providers` table with `provider_type = "claude_cli"`.
+Or via database `llm_providers` table with `provider_type = "claude_cli"`. For database providers, `api_base` is the CLI executable selector (`"claude"` or an absolute binary path), not an HTTP base URL, so provider URL SSRF opt-ins do not apply to Claude CLI.
 
 ### Session Management
 

internal/agent/loop_lazy_mcp_test.go

@@ -5,12 +5,14 @@ import (
 	"testing"
 
 	"github.com/nextlevelbuilder/goclaw/internal/config"
+	"github.com/nextlevelbuilder/goclaw/internal/pipeline"
+	"github.com/nextlevelbuilder/goclaw/internal/providers"
 	"github.com/nextlevelbuilder/goclaw/internal/tools"
 )
 
 // mockExecTool is a simple tool that records whether it was executed.
 type mockExecTool struct {
-	name    string
+	name     string
 	executed bool
 }
 
@@ -24,13 +26,16 @@ func (m *mockExecTool) Execute(_ context.Context, _ map[string]any) *tools.Resul
 	return tools.NewResult("ok from " + m.name)
 }
 
-// simulateLazyActivationCheck mimics the allowedTools check from loop.go for one tool call:
+// simulateLazyActivationCheck mimics the runtime authorize gate from
+// makeAuthorizeToolCall (loop_pipeline_callbacks.go) for one tool call.
+// The real gate is a PipelineDeps.AuthorizeToolCall callback invoked by ToolStage;
+// this helper exercises the same logic in isolation so tests stay unit-level.
 //
-//	if allowedTools != nil && !allowedTools[tc.Name] {
-//	    if l.tools.TryActivateDeferred(tc.Name) { allowedTools[tc.Name] = true }
+//	if allowed != nil && !allowed[name] {
+//	    if reg.TryActivateDeferred(name) { allowed[name] = true }
 //	    else { result = ErrorResult(...) }
 //	}
-//	if result == nil { result = l.tools.ExecuteWithContext(...) }
+//	if result == nil { result = reg.ExecuteWithContext(...) }
 //
 // Returns (result, blocked).
 func simulateLazyActivationCheck(reg *tools.Registry, allowedTools map[string]bool, toolName string) (*tools.Result, bool) {
@@ -163,6 +168,24 @@ func TestLoop_LazyMCP_NilAllowedTools_AllowsAll(t *testing.T) {
 	}
 }
 
+func TestLoopAuthorizeToolCall_PrefixedNameCanonicalLookup(t *testing.T) {
+	loop := NewLoop(LoopConfig{
+		AgentToolPolicy: &config.ToolPolicySpec{ToolCallPrefix: "proxy_"},
+	})
+	gate := loop.makeAuthorizeToolCall()
+	state := &pipeline.RunState{}
+	state.Tool.AllowedTools = map[string]bool{"exec": true}
+
+	ok, reason := gate(context.Background(), state, providers.ToolCall{
+		ID:   "tc-proxy-exec",
+		Name: "proxy_exec",
+	})
+
+	if !ok {
+		t.Fatalf("prefixed tool call should resolve to canonical allowlist entry; reason: %q", reason)
+	}
+}
+
 func TestLoop_LazyMCP_SecondCall_UsesUpdatedAllowedTools(t *testing.T) {
 	// After first lazy activation, subsequent calls in the same iteration use the
 	// updated allowedTools map and don't invoke the activator again.

internal/agent/loop_pipeline_adapter.go

@@ -150,6 +150,7 @@ func (l *Loop) buildPipelineDeps(req *RunRequest, bridgeRS *runState) pipeline.P
 		ExecuteToolCall:   cb.executeToolCall,
 		ExecuteToolRaw:    cb.executeToolRaw,
 		ProcessToolResult: cb.processToolResult,
+		AuthorizeToolCall: cb.authorizeToolCall,
 		SequentialToolCall: func(tc providers.ToolCall) bool {
 			return l.resolveToolCallName(tc.Name) == "wait"
 		},

internal/agent/loop_pipeline_callbacks.go

@@ -57,6 +57,7 @@ func (l *Loop) pipelineCallbacks(req *RunRequest, bridgeRS *runState) pipelineCa
 		executeToolCall:    l.makeExecuteToolCall(req, bridgeRS),
 		executeToolRaw:     l.makeExecuteToolRaw(req),
 		processToolResult:  l.makeProcessToolResult(req, bridgeRS),
+		authorizeToolCall:  l.makeAuthorizeToolCall(),
 		checkReadOnly:      l.makeCheckReadOnly(req, bridgeRS),
 		sanitizeContent:    SanitizeAssistantContent,
 		flushMessages:      l.makeFlushMessages(req),
@@ -85,6 +86,7 @@ type pipelineCallbackSet struct {
 	executeToolCall    func(ctx context.Context, state *pipeline.RunState, tc providers.ToolCall) ([]providers.Message, error)
 	executeToolRaw     func(ctx context.Context, tc providers.ToolCall) (providers.Message, any, error)
 	processToolResult  func(ctx context.Context, state *pipeline.RunState, tc providers.ToolCall, rawMsg providers.Message, rawData any) []providers.Message
+	authorizeToolCall  func(ctx context.Context, state *pipeline.RunState, tc providers.ToolCall) (bool, string)
 	checkReadOnly      func(state *pipeline.RunState) (*providers.Message, bool)
 	sanitizeContent    func(string) string
 	flushMessages      func(ctx context.Context, sessionKey string, msgs []providers.Message) error
@@ -249,6 +251,44 @@ func (l *Loop) makeBuildFilteredTools(req *RunRequest) func(state *pipeline.RunS
 	}
 }
 
+// makeAuthorizeToolCall enforces a runtime fail-closed allowlist check before
+// every tool execution. AllowedTools is keyed by canonical registry names (built
+// by ThinkStage from FilterTools output). The model may emit prefixed names when
+// the agent has toolCallPrefix configured (e.g. "proxy_exec" → canonical "exec"),
+// so we resolve the name to its canonical form before the lookup to avoid a
+// guaranteed miss on every prefixed call.
+func (l *Loop) makeAuthorizeToolCall() func(ctx context.Context, state *pipeline.RunState, tc providers.ToolCall) (bool, string) {
+	return func(_ context.Context, state *pipeline.RunState, tc providers.ToolCall) (bool, string) {
+		allowed := state.Tool.AllowedTools
+		if allowed == nil {
+			// nil allowlist means no per-iteration restriction (e.g. BuildFilteredTools not wired).
+			return true, ""
+		}
+
+		// Resolve to canonical name before allowlist lookup. AllowedTools is keyed
+		// by canonical names; the model may emit prefixed names when toolCallPrefix
+		// is set (e.g. "proxy_exec" vs "exec"). Without this the lookup always misses.
+		name := l.resolveToolCallName(tc.Name)
+
+		if allowed[name] {
+			return true, ""
+		}
+
+		// Preserve lazy activation for deferred tools (typically per-user MCP).
+		if l.tools != nil && l.tools.TryActivateDeferred(name) {
+			// Re-check deny policy to prevent a lazy-activated tool from bypassing
+			// an explicit deny rule.
+			if l.toolPolicy != nil && l.toolPolicy.IsDenied(name, l.agentToolPolicy) {
+				return false, "tool not allowed by policy: " + name
+			}
+			allowed[name] = true
+			return true, ""
+		}
+
+		return false, "tool not allowed by policy: " + name
+	}
+}
+
 func (l *Loop) makeCallLLM(req *RunRequest, emitRun func(AgentEvent)) func(ctx context.Context, state *pipeline.RunState, chatReq providers.ChatRequest) (*providers.ChatResponse, error) {
 	return func(ctx context.Context, state *pipeline.RunState, chatReq providers.ChatRequest) (*providers.ChatResponse, error) {
 		provider := state.Provider

internal/channels/channel.go

@@ -347,9 +347,9 @@ func (c *BaseChannel) CheckDMPolicy(ctx context.Context, senderID, dmPolicy stri
 		if c.pairingService != nil {
 			paired, err := c.pairingService.IsPaired(ctx, senderID, c.name)
 			if err != nil {
-				slog.Warn("security.pairing_check_failed, assuming paired (fail-open)",
+				slog.Warn("security.pairing_check_failed, denying access (fail-closed)",
 					"sender_id", senderID, "channel", c.name, "error", err)
-				return PolicyAllow
+				return PolicyDeny
 			}
 			if paired {
 				return PolicyAllow
@@ -386,9 +386,9 @@ func (c *BaseChannel) CheckGroupPolicy(ctx context.Context, senderID, chatID, gr
 		if c.pairingService != nil {
 			paired, err := c.pairingService.IsPaired(ctx, groupSenderID, c.name)
 			if err != nil {
-				slog.Warn("security.pairing_check_failed, assuming paired (fail-open)",
+				slog.Warn("security.pairing_check_failed, denying access (fail-closed)",
 					"group_sender", groupSenderID, "channel", c.name, "error", err)
-				return PolicyAllow
+				return PolicyDeny
 			}
 			if paired {
 				c.MarkGroupApproved(chatID)

internal/channels/policy_test.go

@@ -174,12 +174,12 @@ func TestCheckDMPolicy_PolicyPairing(t *testing.T) {
 			wantResult:       PolicyAllow,
 		},
 		{
-			name:             "Pairing service error allows message (fail-open)",
+			name:             "Pairing service error denies message (fail-closed)",
 			senderID:         "user999",
 			allowList:        []string{},
 			paired:           false,
 			failPairingCheck: true,
-			wantResult:       PolicyAllow,
+			wantResult:       PolicyDeny,
 		},
 	}
 
@@ -319,14 +319,14 @@ func TestCheckGroupPolicy_PolicyPairing(t *testing.T) {
 			wantResult:    PolicyNeedsPairing,
 		},
 		{
-			name:             "Pairing service error allows (fail-open)",
+			name:             "Pairing service error denies (fail-closed)",
 			senderID:         "user999",
 			chatID:           "chat_2",
 			allowList:        []string{},
 			groupApproved:    false,
 			paired:           false,
 			failPairingCheck: true,
-			wantResult:       PolicyAllow,
+			wantResult:       PolicyDeny,
 		},
 	}
 
@@ -553,13 +553,13 @@ func TestCheckDMPolicy_AllPolicies_TableDriven(t *testing.T) {
 			wantResult: PolicyNeedsPairing,
 		},
 		{
-			name:             "pairing policy fail-open on service error",
+			name:             "pairing policy fail-closed on service error",
 			policy:           "pairing",
 			senderID:         "user3",
 			allowList:        []string{},
 			paired:           false,
 			failPairingCheck: true,
-			wantResult:       PolicyAllow,
+			wantResult:       PolicyDeny,
 		},
 	}
 

internal/channels/telegram/handlers.go

@@ -159,9 +159,8 @@ func (c *Channel) handleMessage(ctx context.Context, update telego.Update) {
 				p1, err1 := ps.IsPaired(ctx, userID, c.Name())
 				p2, err2 := ps.IsPaired(ctx, senderID, c.Name())
 				if err1 != nil || err2 != nil {
-					slog.Warn("security.pairing_check_failed, assuming paired (fail-open)",
+					slog.Warn("security.pairing_check_failed, denying access (fail-closed)",
 						"user_id", userID, "channel", c.Name(), "err1", err1, "err2", err2)
-					paired = true
 				} else {
 					paired = p1 || p2
 				}
@@ -265,9 +264,8 @@ func (c *Channel) handleMessage(ctx context.Context, update telego.Update) {
 					groupSenderID := fmt.Sprintf("group:%d", chatID)
 					paired, pairErr := c.PairingService().IsPaired(ctx, groupSenderID, c.Name())
 					if pairErr != nil {
-						slog.Warn("security.pairing_check_failed, assuming paired (fail-open)",
+						slog.Warn("security.pairing_check_failed, denying access (fail-closed)",
 							"group_sender", groupSenderID, "channel", c.Name(), "error", pairErr)
-						paired = true
 					}
 					if paired {
 						c.MarkGroupApproved(chatIDStr)
@@ -320,9 +318,8 @@ func (c *Channel) handleMessage(ctx context.Context, update telego.Update) {
 					groupSenderID := fmt.Sprintf("group:%d", chatID)
 					paired, pairErr := c.PairingService().IsPaired(ctx, groupSenderID, c.Name())
 					if pairErr != nil {
-						slog.Warn("security.pairing_check_failed, assuming paired (fail-open)",
+						slog.Warn("security.pairing_check_failed, denying access (fail-closed)",
 							"group_sender", groupSenderID, "channel", c.Name(), "error", pairErr)
-						paired = true
 					}
 					if paired {
 						c.MarkGroupApproved(chatIDStr)
@@ -376,9 +373,8 @@ func (c *Channel) handleMessage(ctx context.Context, update telego.Update) {
 			groupSenderID := fmt.Sprintf("group:%d", chatID)
 			paired, err := c.PairingService().IsPaired(ctx, groupSenderID, c.Name())
 			if err != nil {
-				slog.Warn("security.pairing_check_failed, assuming paired (fail-open)",
+				slog.Warn("security.pairing_check_failed, denying access (fail-closed)",
 					"group_sender", groupSenderID, "channel", c.Name(), "error", err)
-				paired = true
 			}
 			if paired {
 				c.MarkGroupApproved(chatIDStr)

internal/http/files.go

@@ -95,11 +95,16 @@ func (h *FilesHandler) auth(next http.HandlerFunc) http.HandlerFunc {
 
 // deniedFilePrefixes blocks access to sensitive system directories.
 // Defense-in-depth: the auth token is the primary barrier, but restricting
-// known-sensitive paths limits damage if a token leaks.
+// known-sensitive absolute paths limits damage if a token leaks or a root is
+// misconfigured. All entries must be absolute paths because hasDeniedFilePrefix
+// uses pathWithinDir semantics — relative-rooted entries (e.g. "/.ssh") would
+// only block the literal filesystem subtree at that root, not user home dirs.
 var deniedFilePrefixes = []string{
 	"/etc/", "/proc/", "/sys/", "/dev/",
 	"/root/", "/boot/", "/run/",
-	"/var/run/", "/var/log/",
+	"/var/run/", "/var/log/", "/var/lib/", "/var/www/",
+	"/home/", "/Users/",
+	"/opt/", "/srv/",
 }
 
 func (h *FilesHandler) handleServe(w http.ResponseWriter, r *http.Request) {
@@ -127,6 +132,14 @@ func (h *FilesHandler) handleServe(w http.ResponseWriter, r *http.Request) {
 	}
 
 	signed := r.URL.Query().Get("ft") != ""
+	// Fail-closed observability: if no boundary roots are configured at all,
+	// lexicallyAllowsFilePath already returns false, but log explicitly so
+	// operators can detect misconfigured deployments.
+	if h.workspace == "" && h.dataDir == "" {
+		slog.Warn("security.files_no_boundary", "path", absPath)
+		http.NotFound(w, r)
+		return
+	}
 	if !h.lexicallyAllowsFilePath(r, absPath, signed) {
 		slog.Warn("security.files_path_denied", "path", absPath, "workspace", h.workspace, "data_dir", h.dataDir)
 		http.NotFound(w, r)

internal/http/files_path_security_test.go

@@ -88,6 +88,60 @@ func TestFilesHandleServe_ProcDir_Blocked(t *testing.T) {
 	}
 }
 
+// ---- handleServe: expanded deny-prefix list ----
+
+// TestFilesHandleServe_ExpandedDenyPrefixes verifies that paths under /home/,
+// /Users/, /var/lib/, /var/www/, /opt/, and /srv/ are all blocked. These are
+// defense-in-depth blocks for misconfigured roots — the token/RBAC checks are
+// the primary barriers.
+func TestFilesHandleServe_ExpandedDenyPrefixes_Blocked(t *testing.T) {
+	h, _ := makeTestFilesHandler(t)
+	mux := http.NewServeMux()
+	mux.HandleFunc("GET /v1/files/{path...}", h.handleServe)
+
+	cases := []struct {
+		name string
+		url  string
+	}{
+		{"home user ssh key", "/v1/files/home/user/.ssh/id_rsa"},
+		{"home user aws creds", "/v1/files/home/user/.aws/credentials"},
+		{"Users admin dir (macOS)", "/v1/files/Users/admin/secrets.txt"},
+		{"var lib docker secret", "/v1/files/var/lib/docker/overlay2/secret"},
+		{"var www config", "/v1/files/var/www/config.php"},
+		{"opt secrets key", "/v1/files/opt/secrets/key.pem"},
+		{"srv www app config", "/v1/files/srv/www/app/config.yaml"},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			req := httptest.NewRequest(http.MethodGet, tc.url, nil)
+			w := httptest.NewRecorder()
+			mux.ServeHTTP(w, req)
+			if w.Code == http.StatusOK {
+				t.Errorf("path %s should be denied (deny-prefix), got 200", tc.url)
+			}
+		})
+	}
+}
+
+// ---- handleServe: fail-closed observability on empty workspace/dataDir ----
+
+// TestFilesHandleServe_NoBoundary_Denies verifies that when both workspace and
+// dataDir are empty the handler returns 404 (fail-closed) for any path.
+func TestFilesHandleServe_NoBoundary_Denies(t *testing.T) {
+	h := NewFilesHandler("", "")
+	mux := http.NewServeMux()
+	mux.HandleFunc("GET /v1/files/{path...}", h.handleServe)
+
+	req := httptest.NewRequest(http.MethodGet, "/v1/files/tmp/test.txt", nil)
+	w := httptest.NewRecorder()
+	mux.ServeHTTP(w, req)
+
+	if w.Code == http.StatusOK {
+		t.Errorf("empty workspace+dataDir should deny all requests, got 200")
+	}
+}
+
 // ---- handleServe: workspace boundary enforcement ----
 
 func TestFilesHandleServe_FileInsideWorkspace_WithToken_Serves(t *testing.T) {

internal/http/oauth_test.go

@@ -65,6 +65,12 @@ func (m *mockProviderStore) UpdateProvider(_ context.Context, id uuid.UUID, upda
 			if v, ok := updates["api_key"]; ok {
 				p.APIKey = v.(string)
 			}
+			if v, ok := updates["api_base"]; ok {
+				p.APIBase = v.(string)
+			}
+			if v, ok := updates["provider_type"]; ok {
+				p.ProviderType = v.(string)
+			}
 			if v, ok := updates["settings"]; ok {
 				p.Settings = v.(json.RawMessage)
 			}

internal/http/providers.go

@@ -8,6 +8,7 @@ import (
 	"net"
 	"net/http"
 	"net/url"
+	"os"
 	"os/exec"
 	"path/filepath"
 	"regexp"
@@ -22,6 +23,7 @@ import (
 	"github.com/nextlevelbuilder/goclaw/internal/oauth"
 	"github.com/nextlevelbuilder/goclaw/internal/permissions"
 	"github.com/nextlevelbuilder/goclaw/internal/providers"
+	"github.com/nextlevelbuilder/goclaw/internal/security"
 	"github.com/nextlevelbuilder/goclaw/internal/store"
 	"github.com/nextlevelbuilder/goclaw/internal/tools"
 	usagecaps "github.com/nextlevelbuilder/goclaw/internal/usage/caps"
@@ -338,55 +340,138 @@ func normalizeOllamaAPIBase(p *store.LLMProviderData) {
 	}
 }
 
-// localProviderTypes are provider types that legitimately run on localhost
-// (e.g. Ollama, Claude CLI). SSRF checks are skipped for these.
-var localProviderTypes = map[string]bool{
-	store.ProviderOllama:    true,
-	store.ProviderClaudeCLI: true,
-	store.ProviderACP:       true,
+// localURLProviderTypes are provider types that legitimately run on localhost.
+// They are restricted to an explicit localhost allowlist
+// rather than skipping SSRF validation entirely.
+var localURLProviderTypes = map[string]bool{
+	store.ProviderOllama: true,
+	store.ProviderACP:    true,
 }
 
+// allowedLocalHosts are the only hosts permitted for local provider types.
+// Explicit allowlist (not blocklist) to prevent new internal addresses from
+// slipping through (e.g. 169.254.169.254 via ollama base URL).
+var allowedLocalHosts = []string{"localhost", "127.0.0.1", "::1", "host.docker.internal"}
+
+// dnsResolverFn resolves hostnames to IPs. Replaceable in tests.
+var dnsResolverFn = net.LookupHost
+
+// allowPrivateProviderURLsFn reports whether the operator has opted in to
+// permitting private / loopback / link-local / internal-hostname provider base
+// URLs via GOCLAW_ALLOW_PRIVATE_PROVIDER_URLS. Evaluated once at first call so
+// tests can override the variable before that happens.
+var allowPrivateProviderURLsFn = sync.OnceValue(func() bool {
+	v := strings.ToLower(strings.TrimSpace(os.Getenv("GOCLAW_ALLOW_PRIVATE_PROVIDER_URLS")))
+	return v == "1" || v == "true" || v == "yes"
+})
+
 // validateProviderURL rejects provider base URLs pointing to internal/private networks.
 // Defense-in-depth: prevents SSRF when providers are later used for API calls.
+//
+// Logic:
+//  1. Empty URL → allowed (provider may not need a custom base).
+//  2. Claude CLI → api_base is an executable path/command, not a URL.
+//  3. Scheme check (http/https only) → enforced for URL-based types, including
+//     local URL types. Blocks file://, gopher://, dict://, etc.
+//  4. Local URL types (ollama, acp) → host must be in allowedLocalHosts
+//     (explicit allowlist prevents reaching 169.254.169.254 or internal services
+//     via the local-type bypass).
+//  5. Remote types → if GOCLAW_ALLOW_PRIVATE_PROVIDER_URLS is set, allow and log.
+//     Otherwise: resolve DNS hostname; reject if ANY resolved IP satisfies
+//     security.IsBlocked (covers loopback, link-local, private, multicast,
+//     unspecified — including 0.0.0.0 and :: that earlier hand-rolled checks missed).
+//
+// DNS resolution on step 5 closes the nip.io / sslip.io / attacker-domain bypass
+// where a hostname passes a literal-string blocklist but resolves to a private IP.
 func validateProviderURL(rawURL string, providerType string) error {
-	if rawURL == "" || localProviderTypes[providerType] {
+	if rawURL == "" {
 		return nil
 	}
+	if providerType == store.ProviderClaudeCLI {
+		return validateClaudeCLIExecutablePath(rawURL)
+	}
 	u, err := url.Parse(rawURL)
 	if err != nil {
 		return fmt.Errorf("invalid URL: %w", err)
 	}
-	// Only allow http/https schemes — block file://, gopher://, dict://, etc.
+	// Scheme check is unconditional for URL-based provider types, including local URL types.
 	switch u.Scheme {
 	case "http", "https":
 	default:
 		return fmt.Errorf("provider URL must use http or https scheme, got %q", u.Scheme)
 	}
+
 	host := u.Hostname()
-	// Block obvious internal targets
-	blocked := []string{"localhost", "127.0.0.1", "::1", "0.0.0.0", "169.254.169.254", "metadata.google.internal"}
-	for _, b := range blocked {
-		if strings.EqualFold(host, b) {
-			return fmt.Errorf("provider URL cannot point to %s", b)
+
+	// Local provider types: only allow an explicit localhost allowlist.
+	// This prevents using the local-type escape hatch to reach internal services
+	// or cloud metadata endpoints.
+	if localURLProviderTypes[providerType] {
+		for _, a := range allowedLocalHosts {
+			if strings.EqualFold(host, a) {
+				return nil
+			}
 		}
+		slog.Warn("security.provider_url.local_type_denied", "host", host, "provider_type", providerType)
+		return fmt.Errorf("provider type %q only allows localhost URLs (localhost, 127.0.0.1, ::1, host.docker.internal), got host %q", providerType, host)
 	}
-	// Block private IP ranges (normalize IPv6-mapped IPv4 to catch ::ffff:127.0.0.1)
-	ip := net.ParseIP(host)
-	if ip != nil {
-		if v4 := ip.To4(); v4 != nil {
-			ip = v4
-		}
-		if ip.IsLoopback() || ip.IsPrivate() || ip.IsLinkLocalUnicast() || ip.IsLinkLocalMulticast() {
-			return fmt.Errorf("provider URL cannot point to private network: %s", host)
-		}
+
+	// Operator opt-in to allow private-network provider URLs (e.g. LAN-hosted vLLM).
+	// Scheme check above still applies even with this gate open.
+	if allowPrivateProviderURLsFn() {
+		slog.Warn("security.provider_url.private_allowed", "host", host, "provider_type", providerType)
+		return nil
 	}
-	// Block common internal hostnames
+
+	// Check literal IP first (avoids unnecessary DNS lookup).
+	if ip := net.ParseIP(host); ip != nil {
+		if security.IsBlocked(ip) {
+			slog.Warn("security.provider_url.blocked", "host", host, "provider_type", providerType)
+			return fmt.Errorf("provider URL cannot point to %s", host)
+		}
+		return nil
+	}
+
+	// Block .internal / .local suffix before DNS (fail-fast for well-known patterns).
 	if strings.HasSuffix(host, ".internal") || strings.HasSuffix(host, ".local") {
+		slog.Warn("security.provider_url.blocked", "host", host, "provider_type", providerType)
 		return fmt.Errorf("provider URL cannot point to internal hostname: %s", host)
 	}
+
+	// Resolve DNS and check every returned address.
+	// Prevents bypass via wildcard services (nip.io, sslip.io) or attacker-controlled
+	// domains that map to private IPs (DNS-rebinding at config time).
+	addrs, err := dnsResolverFn(host)
+	if err != nil {
+		slog.Warn("security.provider_url.dns_resolve_failed", "host", host, "provider_type", providerType, "error", err)
+		return fmt.Errorf("provider URL hostname %q could not be resolved: %w", host, err)
+	}
+	for _, addr := range addrs {
+		ip := net.ParseIP(addr)
+		if ip == nil {
+			continue
+		}
+		if security.IsBlocked(ip) {
+			slog.Warn("security.provider_url.blocked_resolved", "host", host, "resolved_ip", ip.String(), "provider_type", providerType)
+			return fmt.Errorf("provider URL %q resolves to private/reserved address %s", host, ip)
+		}
+	}
 	return nil
 }
 
+func validateClaudeCLIExecutablePath(path string) error {
+	if strings.Contains(path, "\x00") {
+		return fmt.Errorf("Claude CLI executable path cannot contain NUL byte")
+	}
+	if _, err := url.ParseRequestURI(path); err == nil && strings.Contains(path, "://") {
+		return fmt.Errorf("Claude CLI api_base must be an executable path or %q, got URL %q", "claude", path)
+	}
+	if path == "claude" || filepath.IsAbs(path) {
+		return nil
+	}
+	return fmt.Errorf("Claude CLI api_base must be %q or an absolute executable path, got %q", "claude", path)
+}
+
 // --- Provider CRUD ---
 
 func (h *ProvidersHandler) handleListProviders(w http.ResponseWriter, r *http.Request) {

internal/http/providers_test.go

@@ -298,6 +298,79 @@ func TestProvidersHandlerCreateAllows1536EmbeddingDimensions(t *testing.T) {
 	}
 }
 
+func TestProvidersHandlerCreateAllowsClaudeCLIExecutablePath(t *testing.T) {
+	token := setupProvidersAdminToken(t)
+	providerStore := newMockProviderStore()
+	handler := NewProvidersHandler(providerStore, newMockSecretsStore(), nil, "")
+	mux := http.NewServeMux()
+	handler.RegisterRoutes(mux)
+
+	body := map[string]any{
+		"name":          "claude-local",
+		"provider_type": store.ProviderClaudeCLI,
+		"api_base":      writeFakeClaudeBinary(t),
+		"enabled":       true,
+	}
+	rawBody, err := json.Marshal(body)
+	if err != nil {
+		t.Fatalf("Marshal() error = %v", err)
+	}
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/providers", bytes.NewReader(rawBody))
+	req.Header.Set("Authorization", "Bearer "+token)
+	w := httptest.NewRecorder()
+	mux.ServeHTTP(w, req)
+
+	if w.Code != http.StatusCreated {
+		t.Fatalf("status code = %d, want %d, body=%s", w.Code, http.StatusCreated, w.Body.String())
+	}
+	if got := providerStore.providers["claude-local"].APIBase; got == "" || !filepath.IsAbs(got) {
+		t.Fatalf("stored Claude CLI api_base = %q, want absolute executable path", got)
+	}
+}
+
+func TestProvidersHandlerUpdateAllowsClaudeCLIExecutablePath(t *testing.T) {
+	token := setupProvidersAdminToken(t)
+	providerStore := newMockProviderStore()
+	provider := &store.LLMProviderData{
+		BaseModel:    store.BaseModel{ID: uuid.New()},
+		Name:         "claude-local",
+		ProviderType: store.ProviderClaudeCLI,
+		APIBase:      "claude",
+		Enabled:      true,
+	}
+	if err := providerStore.CreateProvider(context.Background(), provider); err != nil {
+		t.Fatalf("CreateProvider() error = %v", err)
+	}
+
+	handler := NewProvidersHandler(providerStore, newMockSecretsStore(), nil, "")
+	mux := http.NewServeMux()
+	handler.RegisterRoutes(mux)
+
+	nextPath := writeFakeClaudeBinary(t)
+	body := map[string]any{"api_base": nextPath}
+	rawBody, err := json.Marshal(body)
+	if err != nil {
+		t.Fatalf("Marshal() error = %v", err)
+	}
+
+	req := httptest.NewRequest(http.MethodPut, "/v1/providers/"+provider.ID.String(), bytes.NewReader(rawBody))
+	req.Header.Set("Authorization", "Bearer "+token)
+	w := httptest.NewRecorder()
+	mux.ServeHTTP(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("status code = %d, want %d, body=%s", w.Code, http.StatusOK, w.Body.String())
+	}
+	current, err := providerStore.GetProvider(context.Background(), provider.ID)
+	if err != nil {
+		t.Fatalf("GetProvider() error = %v", err)
+	}
+	if current.APIBase != nextPath {
+		t.Fatalf("api_base = %q, want %q", current.APIBase, nextPath)
+	}
+}
+
 func TestProvidersHandlerUpdateRejectsIncompatibleEmbeddingDimensions(t *testing.T) {
 	token := setupProvidersAdminToken(t)
 	providerStore := newMockProviderStore()

internal/http/providers_url_validate_test.go

@@ -0,0 +1,326 @@
+package http
+
+import (
+	"net"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// stubResolver returns a net.LookupHost-compatible function that maps known
+// hostnames to controlled IPs. Unknown hostnames return a DNS error.
+func stubResolver(m map[string][]string) func(host string) ([]string, error) {
+	return func(host string) ([]string, error) {
+		if addrs, ok := m[host]; ok {
+			return addrs, nil
+		}
+		return nil, &net.DNSError{Err: "no such host", Name: host, IsNotFound: true}
+	}
+}
+
+// saveAndRestoreGlobals saves mutable package-level vars and restores them after
+// the test. Call at the start of any test that touches dnsResolverFn or
+// allowPrivateProviderURLsFn.
+func saveAndRestoreGlobals(t *testing.T) {
+	t.Helper()
+	origResolver := dnsResolverFn
+	origAllow := allowPrivateProviderURLsFn
+	t.Cleanup(func() {
+		dnsResolverFn = origResolver
+		allowPrivateProviderURLsFn = origAllow
+	})
+}
+
+// --- Core merged test: all major cases in one table ---
+
+func TestValidateProviderURL(t *testing.T) {
+	saveAndRestoreGlobals(t)
+	allowPrivateProviderURLsFn = func() bool { return false }
+	absClaudePath := filepath.Join(t.TempDir(), "claude")
+
+	dnsResolverFn = stubResolver(map[string][]string{
+		"api.openai.com":         {"104.18.6.192"},
+		"legit-provider.com":     {"203.0.113.50"},
+		"10.10.27.30.nip.io":     {"10.10.27.30"},
+		"192.168.1.1.sslip.io":   {"192.168.1.1"},
+		"172.16.0.5.nip.io":      {"172.16.0.5"},
+		"169.254.169.254.nip.io": {"169.254.169.254"},
+		"127.0.0.1.nip.io":       {"127.0.0.1"},
+		"rebind.attacker.com":    {"10.0.0.1"},
+		"dual-stack.example.com": {"203.0.113.50", "10.0.0.1"},
+	})
+
+	tests := []struct {
+		name         string
+		rawURL       string
+		providerType string
+		wantErr      bool
+	}{
+		// Empty URL always allowed
+		{"empty URL", "", "openai_compat", false},
+		{"empty URL ollama", "", "ollama", false},
+
+		// Public remote URLs OK
+		{"public HTTPS", "https://api.openai.com/v1", "openai_compat", false},
+		{"public HTTP", "http://legit-provider.com/v1", "openai_compat", false},
+
+		// --- Scheme check: unconditional for ALL types including local ---
+		{"file scheme remote", "file:///etc/passwd", "openai_compat", true},
+		{"gopher scheme remote", "gopher://internal:25", "openai_compat", true},
+		{"file scheme ollama", "file:///etc/passwd", "ollama", true},       // H-1: scheme enforced even for local types
+		{"gopher scheme acp", "gopher://localhost:25", "acp", true},        // H-1: scheme enforced even for local types
+		{"file scheme claude_cli", "file:///bin/bash", "claude_cli", true}, // H-1: scheme enforced for URL-like Claude CLI values
+
+		// --- Local type: allowlist-only ---
+		{"ollama localhost", "http://localhost:11434/v1", "ollama", false},
+		{"ollama 127.0.0.1", "http://127.0.0.1:11434/v1", "ollama", false},
+		{"ollama ::1", "http://[::1]:11434/v1", "ollama", false},
+		{"ollama host.docker.internal", "http://host.docker.internal:11434/v1", "ollama", false},
+		{"acp 127.0.0.1", "http://127.0.0.1:9090", "acp", false},
+		{"claude_cli command name", "claude", "claude_cli", false},
+		{"claude_cli absolute path", absClaudePath, "claude_cli", false},
+
+		// Local type with non-localhost hosts → blocked
+		{"ollama 169.254.169.254", "http://169.254.169.254/latest/meta-data/", "ollama", true},
+		{"ollama private IP", "http://10.0.0.5:11434/v1", "ollama", true},
+		{"ollama evil.com", "http://evil.attacker.tld:9999/v1", "ollama", true},
+		{"ollama postgres sidecar", "http://postgres:5432/v1", "ollama", true},
+		{"ollama link-local", "http://169.254.1.1:8080/v1", "ollama", true},
+		{"ollama .internal", "http://redis.internal:6379/v1", "ollama", true},
+		{"ollama gcp metadata", "http://metadata.google.internal/computeMetadata/v1/", "ollama", true},
+		{"acp private", "http://10.0.0.1:8080/v1", "acp", true},
+
+		// --- Remote type literal blocked IPs ---
+		{"remote localhost", "http://localhost:8080", "openai_compat", true},
+		{"remote 127.0.0.1", "http://127.0.0.1:8080", "openai_compat", true},
+		{"remote ::1", "http://[::1]:8080", "openai_compat", true},
+		{"remote 10.x", "http://10.0.0.1:8080/v1", "openai_compat", true},
+		{"remote 192.168.x", "http://192.168.1.100:8080/v1", "openai_compat", true},
+		{"remote 172.16.x", "http://172.16.0.5:8080/v1", "openai_compat", true},
+		{"remote 169.254.169.254", "http://169.254.169.254/latest/meta-data/", "openai_compat", true},
+
+		// --- Regression: unspecified addresses (bug missed by #974 hand-rolled checks) ---
+		{"0.0.0.0 unspecified", "http://0.0.0.0:8080/v1", "openai_compat", true},
+		{":: IPv6 unspecified", "http://[::]:8080/v1", "openai_compat", true},
+
+		// --- DNS bypass via wildcard services (nip.io, sslip.io) ---
+		{"nip.io 10.x", "http://10.10.27.30.nip.io:9999/v1", "openai_compat", true},
+		{"sslip.io 192.168.x", "http://192.168.1.1.sslip.io:8080/v1", "openai_compat", true},
+		{"nip.io 172.16.x", "http://172.16.0.5.nip.io:8080/v1", "openai_compat", true},
+		{"nip.io link-local", "http://169.254.169.254.nip.io/latest/", "openai_compat", true},
+		{"nip.io loopback", "http://127.0.0.1.nip.io:8080/v1", "openai_compat", true},
+		{"attacker rebind", "http://rebind.attacker.com:8080/v1", "openai_compat", true},
+		{"dual-stack any private", "http://dual-stack.example.com:8080/v1", "openai_compat", true},
+
+		// --- Internal hostname suffix ---
+		{".internal suffix", "http://metadata.google.internal/computeMetadata/v1/", "openai_compat", true},
+		{".local suffix", "http://myservice.local:8080", "openai_compat", true},
+
+		// --- Unresolvable hostname ---
+		{"unresolvable hostname", "http://nonexistent.invalid:8080/v1", "openai_compat", true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			err := validateProviderURL(tt.rawURL, tt.providerType)
+			if (err != nil) != tt.wantErr {
+				t.Errorf("validateProviderURL(%q, %q) error = %v, wantErr %v",
+					tt.rawURL, tt.providerType, err, tt.wantErr)
+			}
+		})
+	}
+}
+
+// --- Env gate: GOCLAW_ALLOW_PRIVATE_PROVIDER_URLS ---
+
+func TestValidateProviderURL_AllowPrivateFlag(t *testing.T) {
+	saveAndRestoreGlobals(t)
+
+	dnsResolverFn = stubResolver(map[string][]string{
+		"my-vllm.lan": {"10.0.0.1"},
+	})
+
+	t.Run("private URL allowed when flag is true", func(t *testing.T) {
+		allowPrivateProviderURLsFn = func() bool { return true }
+		if err := validateProviderURL("http://my-vllm.lan:8080/v1", "openai_compat"); err != nil {
+			t.Errorf("expected nil with allow-private flag, got: %v", err)
+		}
+		if err := validateProviderURL("http://192.168.1.50:8080/v1", "openai_compat"); err != nil {
+			t.Errorf("expected nil for private IP with allow-private flag, got: %v", err)
+		}
+		if err := validateProviderURL("http://localhost:8080/v1", "openai_compat"); err != nil {
+			t.Errorf("expected nil for localhost with allow-private flag, got: %v", err)
+		}
+		if err := validateProviderURL("http://llm.internal/v1", "openai_compat"); err != nil {
+			t.Errorf("expected nil for .internal with allow-private flag, got: %v", err)
+		}
+	})
+
+	t.Run("private URL blocked when flag is false", func(t *testing.T) {
+		allowPrivateProviderURLsFn = func() bool { return false }
+		if err := validateProviderURL("http://my-vllm.lan:8080/v1", "openai_compat"); err == nil {
+			t.Error("expected error for private URL without allow-private flag")
+		}
+	})
+
+	t.Run("scheme still enforced even with allow-private flag", func(t *testing.T) {
+		allowPrivateProviderURLsFn = func() bool { return true }
+		err := validateProviderURL("file:///etc/passwd", "openai_compat")
+		if err == nil {
+			t.Error("expected error for file:// scheme even with allow-private flag")
+		}
+		if !strings.Contains(err.Error(), "scheme") {
+			t.Errorf("expected scheme error, got: %v", err)
+		}
+	})
+}
+
+func TestValidateProviderURL_LocalTypesIgnoreAllowPrivateFlag(t *testing.T) {
+	saveAndRestoreGlobals(t)
+	allowPrivateProviderURLsFn = func() bool { return true }
+
+	cases := []struct {
+		url          string
+		providerType string
+	}{
+		{"http://ollama:11434/v1", "ollama"},
+		{"http://host.lan:11434/v1", "ollama"},
+		{"http://10.0.0.5:11434/v1", "ollama"},
+		{"http://acp-sidecar:9090", "acp"},
+	}
+	for _, c := range cases {
+		if err := validateProviderURL(c.url, c.providerType); err == nil {
+			t.Errorf("expected local provider URL %s / %s to remain blocked despite allow-private flag", c.url, c.providerType)
+		}
+	}
+}
+
+// --- H-1: scheme enforced for local provider types (from PR #972) ---
+
+func TestValidateProviderURL_LocalTypeSchemeEnforced(t *testing.T) {
+	saveAndRestoreGlobals(t)
+
+	cases := []struct {
+		url          string
+		providerType string
+	}{
+		{"file:///etc/passwd", "ollama"},
+		{"gopher://localhost:25", "ollama"},
+		{"file:///etc/passwd", "acp"},
+	}
+	for _, c := range cases {
+		err := validateProviderURL(c.url, c.providerType)
+		if err == nil {
+			t.Errorf("expected scheme error for %s / %s, got nil", c.url, c.providerType)
+			continue
+		}
+		if !strings.Contains(err.Error(), "scheme") {
+			t.Errorf("expected scheme error for %s / %s, got: %v", c.url, c.providerType, err)
+		}
+	}
+}
+
+// --- Regression: 0.0.0.0 and :: unspecified (bug PR #974 missed) ---
+
+func TestValidateProviderURL_UnspecifiedBlocked(t *testing.T) {
+	saveAndRestoreGlobals(t)
+	allowPrivateProviderURLsFn = func() bool { return false }
+
+	cases := []string{
+		"http://0.0.0.0:8080/v1",
+		"http://[::]:8080/v1",
+	}
+	for _, u := range cases {
+		err := validateProviderURL(u, "openai_compat")
+		if err == nil {
+			t.Errorf("expected error for unspecified address %q, got nil", u)
+		}
+	}
+}
+
+// --- DNS-rebinding: remote type resolving to private/loopback ---
+
+func TestValidateProviderURL_DNSRebindBlocked(t *testing.T) {
+	saveAndRestoreGlobals(t)
+	allowPrivateProviderURLsFn = func() bool { return false }
+
+	privateHosts := map[string][]string{
+		"rebind-private.example.com":   {"10.0.0.1"},
+		"rebind-loopback.example.com":  {"127.0.0.1"},
+		"rebind-linklocal.example.com": {"169.254.169.254"},
+	}
+	dnsResolverFn = stubResolver(privateHosts)
+
+	for host := range privateHosts {
+		err := validateProviderURL("http://"+host+":8080/v1", "openai_compat")
+		if err == nil {
+			t.Errorf("expected DNS rebind to be blocked for %s, got nil", host)
+		}
+	}
+}
+
+// --- Local type: all allowed variants ---
+
+func TestValidateProviderURL_LocalTypeAllowedHosts(t *testing.T) {
+	saveAndRestoreGlobals(t)
+
+	allowed := []struct {
+		url          string
+		providerType string
+	}{
+		{"http://localhost:11434/v1", "ollama"},
+		{"http://127.0.0.1:11434/v1", "ollama"},
+		{"http://[::1]:11434/v1", "ollama"},
+		{"http://host.docker.internal:11434/v1", "ollama"},
+		{"http://localhost:9090", "acp"},
+		{"http://127.0.0.1:9090", "acp"},
+	}
+	for _, a := range allowed {
+		if err := validateProviderURL(a.url, a.providerType); err != nil {
+			t.Errorf("expected %s / %s to be allowed, got: %v", a.url, a.providerType, err)
+		}
+	}
+}
+
+func TestValidateProviderURL_ClaudeCLIExecutablePath(t *testing.T) {
+	saveAndRestoreGlobals(t)
+	absClaudePath := filepath.Join(t.TempDir(), "claude")
+
+	allowed := []string{
+		"",
+		"claude",
+		absClaudePath,
+		filepath.Join(t.TempDir(), "Claude Code.app", "Contents", "MacOS", "claude"),
+	}
+	for _, raw := range allowed {
+		if err := validateProviderURL(raw, "claude_cli"); err != nil {
+			t.Errorf("expected Claude CLI executable %q to be allowed, got: %v", raw, err)
+		}
+	}
+
+	blocked := []string{
+		"file:///usr/local/bin/claude",
+		"https://api.anthropic.com/v1",
+		"relative/claude",
+		"claude --dangerously-skip-permissions",
+	}
+	for _, raw := range blocked {
+		if err := validateProviderURL(raw, "claude_cli"); err == nil {
+			t.Errorf("expected Claude CLI executable %q to be rejected", raw)
+		}
+	}
+}
+
+// --- Public remote URL always OK ---
+
+func TestValidateProviderURL_PublicHostOK(t *testing.T) {
+	saveAndRestoreGlobals(t)
+	allowPrivateProviderURLsFn = func() bool { return false }
+	dnsResolverFn = stubResolver(map[string][]string{
+		"api.openai.com": {"104.18.6.192"},
+	})
+
+	if err := validateProviderURL("https://api.openai.com/v1", "openai_compat"); err != nil {
+		t.Errorf("expected public host to pass, got: %v", err)
+	}
+}

internal/http/vault_handler_documents.go

@@ -1,13 +1,149 @@
 package http
 
 import (
+	"errors"
 	"log/slog"
 	"net/http"
+	"os"
+	"path/filepath"
 	"strings"
+	"time"
 
+	"github.com/google/uuid"
+	"github.com/nextlevelbuilder/goclaw/internal/eventbus"
 	"github.com/nextlevelbuilder/goclaw/internal/store"
+	"github.com/nextlevelbuilder/goclaw/internal/vault"
 )
 
+// writeDocumentContent writes the given content bytes to a tenant-workspace-
+// relative path, after validating the path stays inside the workspace and uses
+// an allowed extension. Returns the SHA-256 content hash on success.
+//
+// This is the shared writer used by handleCreateDocument and handleUpdateDocument
+// to materialise inline JSON `content` payloads on disk, so the enrichment
+// pipeline (which reads files by path) can later compute summaries, embeddings
+// and links.
+func (h *VaultHandler) writeDocumentContent(workspace, relPath string, content []byte) (string, error) {
+	// Ensure the tenant workspace exists before resolving symlinks. Non-master
+	// tenants live under workspace/tenants/{slug}/ which handleUpload creates on
+	// demand (see vault_handler_upload.go) — without this, the very first JSON
+	// content write into a fresh tenant would fail EvalSymlinks with ErrNotExist
+	// and 500 the request.
+	if err := os.MkdirAll(workspace, 0o755); err != nil {
+		return "", err
+	}
+
+	// Resolve the workspace root through any symlinks so all containment checks
+	// compare canonical paths.
+	realWS, err := filepath.EvalSymlinks(workspace)
+	if err != nil {
+		return "", err
+	}
+	target := filepath.Join(realWS, filepath.Clean(relPath))
+
+	// Lexical containment: blocks "../" and absolute-path escapes.
+	if target != realWS && !strings.HasPrefix(target, realWS+string(os.PathSeparator)) {
+		slog.Warn("security.vault_symlink_escape",
+			"site", "path_traversal", "attempted", target, "workspace", realWS)
+		return "", os.ErrInvalid
+	}
+
+	// Symlink containment: a lexical prefix check is not enough — a symlink that
+	// lives *inside* the workspace but points outside it (e.g. workspace/link ->
+	// /etc) would otherwise let the write follow it and clobber a file beyond
+	// the workspace. Resolve the deepest already-existing ancestor of the target
+	// and confirm it still canonicalises to a path inside the workspace, before
+	// any directory is created or byte written.
+	for ancestor := filepath.Dir(target); ; {
+		resolved, rerr := filepath.EvalSymlinks(ancestor)
+		if rerr == nil {
+			if resolved != realWS && !strings.HasPrefix(resolved, realWS+string(os.PathSeparator)) {
+				slog.Warn("security.vault_symlink_escape",
+					"site", "ancestor", "resolved", resolved, "workspace", realWS)
+				return "", os.ErrInvalid
+			}
+			break
+		}
+		if !os.IsNotExist(rerr) {
+			return "", rerr
+		}
+		parent := filepath.Dir(ancestor)
+		if parent == ancestor {
+			break // walked to the filesystem root without an existing ancestor
+		}
+		ancestor = parent
+	}
+
+	if err := os.MkdirAll(filepath.Dir(target), 0o755); err != nil {
+		return "", err
+	}
+
+	// Fast-fail when the final component is already a symlink — gives a clean
+	// rejection + security log without parsing OS-specific error codes. The
+	// O_NOFOLLOW open below closes the TOCTOU window this Lstat leaves open
+	// (an attacker could swap the file for a symlink between calls).
+	if fi, lerr := os.Lstat(target); lerr == nil && fi.Mode()&os.ModeSymlink != 0 {
+		slog.Warn("security.vault_symlink_escape",
+			"site", "final_lstat", "target", target, "workspace", realWS)
+		return "", os.ErrInvalid
+	}
+
+	// O_NOFOLLOW makes the kernel refuse to follow a symlink at the final
+	// component atomically with the open, removing the Lstat→Open race entirely
+	// on Unix. On Windows oNoFollow == 0 and the Lstat above is best-effort.
+	f, err := os.OpenFile(target, os.O_WRONLY|os.O_CREATE|os.O_TRUNC|oNoFollow, 0o644)
+	if err != nil {
+		if isSymlinkLoopErr(err) {
+			slog.Warn("security.vault_symlink_escape",
+				"site", "open_nofollow", "target", target, "workspace", realWS)
+			return "", os.ErrInvalid
+		}
+		return "", err
+	}
+	if _, werr := f.Write(content); werr != nil {
+		_ = f.Close() // best-effort cleanup; the write error is the real one
+		return "", werr
+	}
+	// Explicit Close so a delayed FS error (e.g. disk-full at flush) surfaces
+	// instead of being silently swallowed by a deferred close.
+	if cerr := f.Close(); cerr != nil {
+		return "", cerr
+	}
+	return vault.ContentHash(content), nil
+}
+
+// publishDocUpserted enqueues the standard EventVaultDocUpserted so the
+// enrichment worker re-runs summary/embedding/link extraction for this doc.
+// No-op when the event bus is not wired (some test setups).
+func (h *VaultHandler) publishDocUpserted(tenantID uuid.UUID, agentID, docID, relPath, hash, workspace string) {
+	if h.eventBus == nil {
+		return
+	}
+	tenantIDStr := tenantID.String()
+	// Start enrichment progress before publishing so the worker pool can't drain
+	// the event and finish before the progress tracker registers it — same
+	// ordering handleUpload relies on to avoid that race.
+	if h.enrichProgress != nil {
+		h.enrichProgress.Start(1, tenantID)
+	}
+	h.eventBus.Publish(eventbus.DomainEvent{
+		ID:        uuid.Must(uuid.NewV7()).String(),
+		Type:      eventbus.EventVaultDocUpserted,
+		SourceID:  docID + ":" + hash,
+		TenantID:  tenantIDStr,
+		AgentID:   agentID,
+		Timestamp: time.Now(),
+		Payload: eventbus.VaultDocUpsertedPayload{
+			DocID:       docID,
+			TenantID:    tenantIDStr,
+			AgentID:     agentID,
+			Path:        relPath,
+			ContentHash: hash,
+			Workspace:   workspace,
+		},
+	})
+}
+
 // handleListAllDocuments lists vault documents across all agents in tenant.
 // Optional query param agent_id to filter by specific agent.
 func (h *VaultHandler) handleListAllDocuments(w http.ResponseWriter, r *http.Request) {
@@ -99,6 +235,16 @@ func (h *VaultHandler) handleGetDocument(w http.ResponseWriter, r *http.Request)
 }
 
 // handleCreateDocument creates a new vault document.
+//
+// `content` semantics — symmetric with handleUpdateDocument:
+//   - field omitted (nil pointer): metadata-only stub, no file write, no event
+//     (typically followed by a rescan that materialises the file from disk).
+//   - field present and non-empty: bytes materialised at <tenant-workspace>/<path>,
+//     SHA-256 hash stored on the row, EventVaultDocUpserted emitted so the
+//     enrichment worker computes summary/embeddings/links — same code path the
+//     multipart /v1/vault/upload endpoint and the filesystem rescan use.
+//   - field present and empty (""): a 0-byte file is written and the event still
+//     fires. Same behaviour as PUT — pick the right shape on the client side.
 func (h *VaultHandler) handleCreateDocument(w http.ResponseWriter, r *http.Request) {
 	locale := extractLocale(r)
 	tenantID := store.TenantIDFromContext(r.Context())
@@ -107,6 +253,7 @@ func (h *VaultHandler) handleCreateDocument(w http.ResponseWriter, r *http.Reque
 	var body struct {
 		Path     string         `json:"path"`
 		Title    string         `json:"title"`
+		Content  *string        `json:"content"` // nil=no write; ""=write empty file; "data"=write bytes (mirrors PUT)
 		DocType  string         `json:"doc_type"`
 		Scope    string         `json:"scope"`
 		TeamID   string         `json:"team_id"`
@@ -137,6 +284,16 @@ func (h *VaultHandler) handleCreateDocument(w http.ResponseWriter, r *http.Reque
 		writeJSON(w, http.StatusBadRequest, map[string]string{"error": "invalid scope"})
 		return
 	}
+	// Validate extension upfront if caller wants to write content — matches the
+	// whitelist enforced by /v1/vault/upload so the two endpoints accept the
+	// same file types. nil pointer = "no content field", which skips the check.
+	if body.Content != nil {
+		ext := strings.ToLower(filepath.Ext(body.Path))
+		if !allowedUploadExts[ext] {
+			writeJSON(w, http.StatusBadRequest, map[string]string{"error": "unsupported file type: " + ext})
+			return
+		}
+	}
 
 	doc := &store.VaultDocument{
 		TenantID: tenantID.String(),
@@ -160,11 +317,50 @@ func (h *VaultHandler) handleCreateDocument(w http.ResponseWriter, r *http.Reque
 			doc.Scope = "team"
 		}
 	}
+
+	// Persist file content if provided, before the DB upsert so the
+	// content_hash column is set in a single write.
+	var (
+		wsPath        string
+		writtenHash   string
+		contentWasSet bool
+	)
+	if body.Content != nil {
+		wsPath = h.resolveTenantWorkspace(r.Context())
+		if wsPath == "" {
+			writeJSON(w, http.StatusBadRequest, map[string]string{"error": "workspace not available"})
+			return
+		}
+		hash, werr := h.writeDocumentContent(wsPath, body.Path, []byte(*body.Content))
+		if werr != nil {
+			slog.Warn("vault.create: write content failed", "path", body.Path, "error", werr)
+			if errors.Is(werr, os.ErrInvalid) {
+				writeJSON(w, http.StatusBadRequest, map[string]string{"error": "path escapes workspace"})
+				return
+			}
+			writeJSON(w, http.StatusInternalServerError, map[string]string{"error": "failed to write content"})
+			return
+		}
+		doc.ContentHash = hash
+		writtenHash = hash
+		contentWasSet = true
+	}
+
 	if err := h.store.UpsertDocument(r.Context(), doc); err != nil {
 		slog.Warn("vault.create failed", "error", err)
 		writeJSON(w, http.StatusInternalServerError, map[string]string{"error": err.Error()})
 		return
 	}
+
+	// Fire enrichment after the DB row exists so the worker can locate the doc.
+	if contentWasSet {
+		evAgent := ""
+		if doc.AgentID != nil {
+			evAgent = *doc.AgentID
+		}
+		h.publishDocUpserted(tenantID, evAgent, doc.ID, body.Path, writtenHash, wsPath)
+	}
+
 	// Re-fetch by ID (set via RETURNING) — unambiguous even when same path exists across teams.
 	created, _ := h.store.GetDocumentByID(r.Context(), tenantID.String(), doc.ID)
 	if created != nil {
@@ -189,6 +385,7 @@ func (h *VaultHandler) handleUpdateDocument(w http.ResponseWriter, r *http.Reque
 
 	var body struct {
 		Title    *string        `json:"title"`
+		Content  *string        `json:"content"` // nil=no change; "" clears the file; non-empty rewrites it
 		DocType  *string        `json:"doc_type"`
 		Scope    *string        `json:"scope"`
 		TeamID   *string        `json:"team_id"` // nil=no change, ""=clear, "uuid"=set
@@ -225,11 +422,52 @@ func (h *VaultHandler) handleUpdateDocument(w http.ResponseWriter, r *http.Reque
 		existing.Metadata = body.Metadata
 	}
 
+	// Rewrite content on disk when caller supplied a `content` field.
+	var (
+		wsPath        string
+		writtenHash   string
+		contentWasSet bool
+	)
+	if body.Content != nil {
+		ext := strings.ToLower(filepath.Ext(existing.Path))
+		if !allowedUploadExts[ext] {
+			writeJSON(w, http.StatusBadRequest, map[string]string{"error": "unsupported file type: " + ext})
+			return
+		}
+		wsPath = h.resolveTenantWorkspace(r.Context())
+		if wsPath == "" {
+			writeJSON(w, http.StatusBadRequest, map[string]string{"error": "workspace not available"})
+			return
+		}
+		hash, werr := h.writeDocumentContent(wsPath, existing.Path, []byte(*body.Content))
+		if werr != nil {
+			slog.Warn("vault.update: write content failed", "path", existing.Path, "error", werr)
+			if errors.Is(werr, os.ErrInvalid) {
+				writeJSON(w, http.StatusBadRequest, map[string]string{"error": "path escapes workspace"})
+				return
+			}
+			writeJSON(w, http.StatusInternalServerError, map[string]string{"error": "failed to write content"})
+			return
+		}
+		existing.ContentHash = hash
+		writtenHash = hash
+		contentWasSet = true
+	}
+
 	if err := h.store.UpsertDocument(r.Context(), existing); err != nil {
 		slog.Warn("vault.update failed", "error", err)
 		writeJSON(w, http.StatusInternalServerError, map[string]string{"error": err.Error()})
 		return
 	}
+
+	if contentWasSet {
+		evAgent := ""
+		if existing.AgentID != nil {
+			evAgent = *existing.AgentID
+		}
+		h.publishDocUpserted(tenantID, evAgent, existing.ID, existing.Path, writtenHash, wsPath)
+	}
+
 	updated, _ := h.store.GetDocumentByID(r.Context(), tenantID.String(), docID)
 	if updated != nil {
 		writeJSON(w, http.StatusOK, updated)

internal/http/vault_handler_documents_nofollow_unix.go

@@ -0,0 +1,20 @@
+//go:build !windows
+
+package http
+
+import (
+	"errors"
+	"syscall"
+)
+
+// oNoFollow is OR'd into the open flags so the kernel refuses to follow a
+// symlink at the final path component — closes the Lstat→Open TOCTOU window
+// where a tenant could swap a regular file for a symlink between the two
+// syscalls. On Windows this is a no-op (see the _windows.go counterpart).
+const oNoFollow = syscall.O_NOFOLLOW
+
+// isSymlinkLoopErr reports whether err is the kernel's "refused to follow
+// symlink" signal from an O_NOFOLLOW open (ELOOP on Linux/macOS).
+func isSymlinkLoopErr(err error) bool {
+	return errors.Is(err, syscall.ELOOP)
+}

internal/http/vault_handler_documents_nofollow_windows.go

@@ -0,0 +1,11 @@
+//go:build windows
+
+package http
+
+// Windows has no O_NOFOLLOW concept and follows reparse points transparently;
+// the Lstat short-circuit in writeDocumentContent is the only protection
+// available there. Desktop edition is single-user with no untrusted tenants,
+// so this degradation is acceptable.
+const oNoFollow = 0
+
+func isSymlinkLoopErr(err error) bool { return false }

internal/http/vault_handler_documents_test.go

@@ -0,0 +1,470 @@
+package http
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/google/uuid"
+
+	"github.com/nextlevelbuilder/goclaw/internal/eventbus"
+	"github.com/nextlevelbuilder/goclaw/internal/store"
+	"github.com/nextlevelbuilder/goclaw/internal/vault"
+)
+
+// masterTenant resolves TenantWorkspace to h.workspace unchanged, keeping the
+// content-write tests pointed at a single temp dir.
+var masterTenant = uuid.MustParse("0193a5b0-7000-7000-8000-000000000001")
+
+func masterCtx() context.Context {
+	return store.WithTenantID(context.Background(), masterTenant)
+}
+
+// ---- writeDocumentContent: path & symlink containment (the blocking finding) ----
+
+func TestWriteDocumentContent_WritesInsideWorkspace(t *testing.T) {
+	h := &VaultHandler{}
+	ws := t.TempDir()
+
+	hash, err := h.writeDocumentContent(ws, "notes/hello.md", []byte("hi"))
+	if err != nil {
+		t.Fatalf("writeDocumentContent: %v", err)
+	}
+	if want := vault.ContentHash([]byte("hi")); hash != want {
+		t.Errorf("hash = %q, want %q", hash, want)
+	}
+	got, err := os.ReadFile(filepath.Join(ws, "notes", "hello.md"))
+	if err != nil {
+		t.Fatalf("read written file: %v", err)
+	}
+	if string(got) != "hi" {
+		t.Errorf("content = %q, want %q", got, "hi")
+	}
+}
+
+func TestWriteDocumentContent_EmptyContent(t *testing.T) {
+	h := &VaultHandler{}
+	ws := t.TempDir()
+
+	hash, err := h.writeDocumentContent(ws, "empty.md", []byte(""))
+	if err != nil {
+		t.Fatalf("writeDocumentContent: %v", err)
+	}
+	if want := vault.ContentHash([]byte("")); hash != want {
+		t.Errorf("hash = %q, want %q", hash, want)
+	}
+	fi, err := os.Stat(filepath.Join(ws, "empty.md"))
+	if err != nil {
+		t.Fatalf("stat empty file: %v", err)
+	}
+	if fi.Size() != 0 {
+		t.Errorf("size = %d, want 0", fi.Size())
+	}
+}
+
+func TestWriteDocumentContent_RejectsParentEscape(t *testing.T) {
+	h := &VaultHandler{}
+	ws := t.TempDir()
+
+	_, err := h.writeDocumentContent(ws, "../escape.md", []byte("x"))
+	if !errors.Is(err, os.ErrInvalid) {
+		t.Fatalf("err = %v, want os.ErrInvalid", err)
+	}
+	if _, statErr := os.Stat(filepath.Join(filepath.Dir(ws), "escape.md")); statErr == nil {
+		t.Error("file escaped the workspace via ../")
+	}
+}
+
+// A symlink that lives inside the workspace but points outside it must not be
+// followed — this is the exact escape the lexical prefix check missed.
+func TestWriteDocumentContent_RejectsSymlinkDirEscape(t *testing.T) {
+	h := &VaultHandler{}
+	ws := t.TempDir()
+	outside := t.TempDir()
+
+	if err := os.Symlink(outside, filepath.Join(ws, "link")); err != nil {
+		t.Fatalf("symlink: %v", err)
+	}
+
+	_, err := h.writeDocumentContent(ws, "link/pwn.md", []byte("x"))
+	if !errors.Is(err, os.ErrInvalid) {
+		t.Fatalf("err = %v, want os.ErrInvalid", err)
+	}
+	if _, statErr := os.Stat(filepath.Join(outside, "pwn.md")); statErr == nil {
+		t.Error("file written outside workspace through symlinked directory")
+	}
+}
+
+// A symlink planted at the final path component must be refused rather than
+// followed, so an attacker can't redirect a write onto an existing outside file.
+func TestWriteDocumentContent_RejectsFinalComponentSymlink(t *testing.T) {
+	h := &VaultHandler{}
+	ws := t.TempDir()
+	outside := t.TempDir()
+
+	outsideFile := filepath.Join(outside, "target.md")
+	if err := os.WriteFile(outsideFile, []byte("original"), 0o644); err != nil {
+		t.Fatalf("seed outside file: %v", err)
+	}
+	if err := os.Symlink(outsideFile, filepath.Join(ws, "evil.md")); err != nil {
+		t.Fatalf("symlink: %v", err)
+	}
+
+	_, err := h.writeDocumentContent(ws, "evil.md", []byte("pwned"))
+	if !errors.Is(err, os.ErrInvalid) {
+		t.Fatalf("err = %v, want os.ErrInvalid", err)
+	}
+	got, _ := os.ReadFile(outsideFile)
+	if string(got) != "original" {
+		t.Errorf("outside file clobbered through final-component symlink: %q", got)
+	}
+}
+
+// ---- handler-level: extension whitelist, event ordering, empty PUT, orphan ----
+
+// fakeVaultStore embeds the interface (unimplemented methods panic if hit) and
+// overrides only the two methods the content path touches.
+type fakeVaultStore struct {
+	store.VaultStore
+	docs       map[string]*store.VaultDocument
+	upserted   bool
+	upsertErr  error
+	upsertCall int
+}
+
+func newFakeVaultStore() *fakeVaultStore {
+	return &fakeVaultStore{docs: map[string]*store.VaultDocument{}}
+}
+
+func (f *fakeVaultStore) UpsertDocument(_ context.Context, doc *store.VaultDocument) error {
+	f.upsertCall++
+	if f.upsertErr != nil {
+		return f.upsertErr
+	}
+	if doc.ID == "" {
+		doc.ID = "doc-" + uuid.NewString()
+	}
+	f.upserted = true
+	f.docs[doc.ID] = doc
+	return nil
+}
+
+func (f *fakeVaultStore) GetDocumentByID(_ context.Context, _ /*tenantID*/, id string) (*store.VaultDocument, error) {
+	return f.docs[id], nil
+}
+
+// fakeEventBus captures published events and records whether the DB upsert had
+// already happened at publish time, to assert ordering.
+type fakeEventBus struct {
+	store          *fakeVaultStore
+	published      []eventbus.DomainEvent
+	upsertedAtFire bool
+}
+
+func (b *fakeEventBus) Publish(e eventbus.DomainEvent) {
+	if b.store != nil {
+		b.upsertedAtFire = b.store.upserted
+	}
+	b.published = append(b.published, e)
+}
+func (b *fakeEventBus) Subscribe(eventbus.EventType, eventbus.DomainEventHandler) func() {
+	return func() {}
+}
+func (b *fakeEventBus) Start(context.Context)     {}
+func (b *fakeEventBus) Drain(time.Duration) error { return nil }
+
+func newJSONRequest(t *testing.T, method, target string, body any) *http.Request {
+	t.Helper()
+	buf, err := json.Marshal(body)
+	if err != nil {
+		t.Fatalf("marshal body: %v", err)
+	}
+	req := httptest.NewRequest(method, target, bytes.NewReader(buf))
+	req.Header.Set("Content-Type", "application/json")
+	return req.WithContext(masterCtx())
+}
+
+func TestHandleCreateDocument_RejectsDisallowedExtension(t *testing.T) {
+	h := &VaultHandler{} // extension check fires before any store/workspace access
+	req := newJSONRequest(t, http.MethodPost, "/v1/vault/documents", map[string]any{
+		"path":    "payload.exe",
+		"title":   "x",
+		"content": "data",
+	})
+	rr := httptest.NewRecorder()
+
+	h.handleCreateDocument(rr, req)
+
+	assertBadRequest(t, rr, "unsupported file type")
+}
+
+func TestHandleCreateDocument_WritesContentAndPublishesAfterUpsert(t *testing.T) {
+	st := newFakeVaultStore()
+	bus := &fakeEventBus{store: st}
+	ws := t.TempDir()
+	h := &VaultHandler{store: st, eventBus: bus, workspace: ws}
+
+	req := newJSONRequest(t, http.MethodPost, "/v1/vault/documents", map[string]any{
+		"path":    "decisions/adr-001.md",
+		"title":   "ADR 001",
+		"content": "# Decision\nuse postgres",
+	})
+	rr := httptest.NewRecorder()
+
+	h.handleCreateDocument(rr, req)
+
+	if rr.Code != http.StatusCreated {
+		t.Fatalf("status = %d, want 201 (body: %s)", rr.Code, rr.Body.String())
+	}
+	// File materialised on disk.
+	got, err := os.ReadFile(filepath.Join(ws, "decisions", "adr-001.md"))
+	if err != nil {
+		t.Fatalf("read written doc: %v", err)
+	}
+	if string(got) != "# Decision\nuse postgres" {
+		t.Errorf("content = %q", got)
+	}
+	// Exactly one enrichment event, fired AFTER the DB upsert.
+	if len(bus.published) != 1 {
+		t.Fatalf("published %d events, want 1", len(bus.published))
+	}
+	if !bus.upsertedAtFire {
+		t.Error("event published before DB upsert — enrichment worker could race ahead of the row")
+	}
+	if bus.published[0].Type != eventbus.EventVaultDocUpserted {
+		t.Errorf("event type = %v, want %v", bus.published[0].Type, eventbus.EventVaultDocUpserted)
+	}
+}
+
+func TestHandleCreateDocument_NoContentSkipsWriteAndEvent(t *testing.T) {
+	st := newFakeVaultStore()
+	bus := &fakeEventBus{store: st}
+	ws := t.TempDir()
+	h := &VaultHandler{store: st, eventBus: bus, workspace: ws}
+
+	req := newJSONRequest(t, http.MethodPost, "/v1/vault/documents", map[string]any{
+		"path":  "meta-only.md",
+		"title": "Meta only",
+	})
+	rr := httptest.NewRecorder()
+
+	h.handleCreateDocument(rr, req)
+
+	if rr.Code != http.StatusCreated {
+		t.Fatalf("status = %d, want 201 (body: %s)", rr.Code, rr.Body.String())
+	}
+	if _, statErr := os.Stat(filepath.Join(ws, "meta-only.md")); statErr == nil {
+		t.Error("metadata-only create must not write a file")
+	}
+	if len(bus.published) != 0 {
+		t.Errorf("metadata-only create published %d events, want 0", len(bus.published))
+	}
+}
+
+// Documents the orphan-file behaviour: when the DB upsert fails after the bytes
+// are already on disk, the handler returns 500 and the file is intentionally
+// left in place (no rollback). If that ever changes, this test should change
+// with it.
+func TestHandleCreateDocument_UpsertFailureLeavesFile(t *testing.T) {
+	st := newFakeVaultStore()
+	st.upsertErr = errors.New("db down")
+	bus := &fakeEventBus{store: st}
+	ws := t.TempDir()
+	h := &VaultHandler{store: st, eventBus: bus, workspace: ws}
+
+	req := newJSONRequest(t, http.MethodPost, "/v1/vault/documents", map[string]any{
+		"path":    "orphan.md",
+		"title":   "Orphan",
+		"content": "written before upsert",
+	})
+	rr := httptest.NewRecorder()
+
+	h.handleCreateDocument(rr, req)
+
+	if rr.Code != http.StatusInternalServerError {
+		t.Fatalf("status = %d, want 500 (body: %s)", rr.Code, rr.Body.String())
+	}
+	if _, statErr := os.Stat(filepath.Join(ws, "orphan.md")); statErr != nil {
+		t.Errorf("file should remain after upsert failure (orphan accepted): %v", statErr)
+	}
+	if len(bus.published) != 0 {
+		t.Errorf("no event should fire when upsert fails, got %d", len(bus.published))
+	}
+}
+
+func TestHandleUpdateDocument_EmptyContentClearsFile(t *testing.T) {
+	st := newFakeVaultStore()
+	st.docs["existing-1"] = &store.VaultDocument{
+		ID:       "existing-1",
+		TenantID: masterTenant.String(),
+		Path:     "notes/keep.md",
+		Title:    "Keep",
+		DocType:  "note",
+		Scope:    "shared",
+	}
+	bus := &fakeEventBus{store: st}
+	ws := t.TempDir()
+	h := &VaultHandler{store: st, eventBus: bus, workspace: ws}
+
+	empty := ""
+	req := newJSONRequest(t, http.MethodPut, "/v1/vault/documents/existing-1", map[string]any{
+		"content": empty,
+	})
+	req.SetPathValue("docID", "existing-1")
+	rr := httptest.NewRecorder()
+
+	h.handleUpdateDocument(rr, req)
+
+	if rr.Code != http.StatusOK {
+		t.Fatalf("status = %d, want 200 (body: %s)", rr.Code, rr.Body.String())
+	}
+	fi, err := os.Stat(filepath.Join(ws, "notes", "keep.md"))
+	if err != nil {
+		t.Fatalf("stat cleared file: %v", err)
+	}
+	if fi.Size() != 0 {
+		t.Errorf("file size = %d, want 0 (empty content clears file)", fi.Size())
+	}
+	if st.docs["existing-1"].ContentHash != vault.ContentHash([]byte("")) {
+		t.Errorf("content_hash = %q, want empty-content hash", st.docs["existing-1"].ContentHash)
+	}
+	if len(bus.published) != 1 {
+		t.Errorf("published %d events, want 1", len(bus.published))
+	}
+}
+
+// Locks the "no change" contract on PUT: omitting `content` (nil pointer in
+// the body struct) must not touch the disk or fire an enrichment event, even
+// when other metadata fields are present.
+func TestHandleUpdateDocument_NilContentSkipsWriteAndEvent(t *testing.T) {
+	st := newFakeVaultStore()
+	st.docs["existing-2"] = &store.VaultDocument{
+		ID:          "existing-2",
+		TenantID:    masterTenant.String(),
+		Path:        "notes/keep.md",
+		Title:       "Old",
+		DocType:     "note",
+		Scope:       "shared",
+		ContentHash: "preexisting-hash",
+	}
+	bus := &fakeEventBus{store: st}
+	ws := t.TempDir()
+	h := &VaultHandler{store: st, eventBus: bus, workspace: ws}
+
+	// content field intentionally omitted → *string is nil → no write, no event.
+	req := newJSONRequest(t, http.MethodPut, "/v1/vault/documents/existing-2", map[string]any{
+		"title": "Renamed but content untouched",
+	})
+	req.SetPathValue("docID", "existing-2")
+	rr := httptest.NewRecorder()
+
+	h.handleUpdateDocument(rr, req)
+
+	if rr.Code != http.StatusOK {
+		t.Fatalf("status = %d, want 200 (body: %s)", rr.Code, rr.Body.String())
+	}
+	if _, statErr := os.Stat(filepath.Join(ws, "notes", "keep.md")); statErr == nil {
+		t.Error("nil content must not materialise a file")
+	}
+	if st.docs["existing-2"].ContentHash != "preexisting-hash" {
+		t.Errorf("content_hash = %q, want unchanged 'preexisting-hash'", st.docs["existing-2"].ContentHash)
+	}
+	if len(bus.published) != 0 {
+		t.Errorf("nil-content PUT must not publish events, got %d", len(bus.published))
+	}
+}
+
+// Handler-level guard at vault_handler_documents.go path validation must reject
+// "..", short-circuiting before writeDocumentContent even runs.
+func TestHandleCreateDocument_RejectsParentTraversalPath(t *testing.T) {
+	h := &VaultHandler{} // path check runs before any store/workspace access
+	req := newJSONRequest(t, http.MethodPost, "/v1/vault/documents", map[string]any{
+		"path":    "../../etc/passwd.md",
+		"title":   "x",
+		"content": "evil",
+	})
+	rr := httptest.NewRecorder()
+
+	h.handleCreateDocument(rr, req)
+
+	assertBadRequest(t, rr, "invalid path")
+}
+
+// Mirrors PUT semantics on POST: `content: ""` writes a 0-byte file and
+// fires the enrichment event (the I4 fix that aligned POST/PUT).
+func TestHandleCreateDocument_EmptyContentWritesEmptyFile(t *testing.T) {
+	st := newFakeVaultStore()
+	bus := &fakeEventBus{store: st}
+	ws := t.TempDir()
+	h := &VaultHandler{store: st, eventBus: bus, workspace: ws}
+
+	req := newJSONRequest(t, http.MethodPost, "/v1/vault/documents", map[string]any{
+		"path":    "placeholder.md",
+		"title":   "Placeholder",
+		"content": "",
+	})
+	rr := httptest.NewRecorder()
+
+	h.handleCreateDocument(rr, req)
+
+	if rr.Code != http.StatusCreated {
+		t.Fatalf("status = %d, want 201 (body: %s)", rr.Code, rr.Body.String())
+	}
+	fi, err := os.Stat(filepath.Join(ws, "placeholder.md"))
+	if err != nil {
+		t.Fatalf("stat empty file: %v", err)
+	}
+	if fi.Size() != 0 {
+		t.Errorf("size = %d, want 0", fi.Size())
+	}
+	if len(bus.published) != 1 {
+		t.Errorf("published %d events, want 1", len(bus.published))
+	}
+}
+
+// B1 regression: non-master tenants resolve to workspace/tenants/{slug}/ which
+// is created on demand. The very first JSON content write into a fresh tenant
+// must NOT 500 because EvalSymlinks can't find that dir — writeDocumentContent
+// is responsible for creating it (mirrors what handleUpload does).
+func TestHandleCreateDocument_NonMasterTenantFirstContentWrite(t *testing.T) {
+	st := newFakeVaultStore()
+	bus := &fakeEventBus{store: st}
+	ws := t.TempDir()
+	h := &VaultHandler{store: st, eventBus: bus, workspace: ws}
+
+	nonMaster := uuid.MustParse("01999999-1111-7000-8000-000000000abc")
+	ctx := store.WithTenantSlug(store.WithTenantID(context.Background(), nonMaster), "acme")
+
+	// newJSONRequest builds against masterCtx by default; WithContext swaps in
+	// the non-master tenant context for this single test.
+	req := newJSONRequest(t, http.MethodPost, "/v1/vault/documents", map[string]any{
+		"path":    "notes/first.md",
+		"title":   "First",
+		"content": "hello from acme",
+	}).WithContext(ctx)
+	rr := httptest.NewRecorder()
+
+	h.handleCreateDocument(rr, req)
+
+	if rr.Code != http.StatusCreated {
+		t.Fatalf("status = %d, want 201 (body: %s)", rr.Code, rr.Body.String())
+	}
+	tenantDir := filepath.Join(ws, "tenants", "acme")
+	got, err := os.ReadFile(filepath.Join(tenantDir, "notes", "first.md"))
+	if err != nil {
+		t.Fatalf("read non-master tenant doc: %v", err)
+	}
+	if string(got) != "hello from acme" {
+		t.Errorf("content = %q", got)
+	}
+	if len(bus.published) != 1 {
+		t.Errorf("published %d events, want 1", len(bus.published))
+	}
+}

internal/mcp/grant_checker.go

@@ -49,6 +49,8 @@ type storeGrantChecker struct {
 	cache sync.Map // map[cacheKey]*cacheEntry
 }
 
+const grantCheckerCacheSubscriberID = bus.TopicCacheMCP + ":grant_checker"
+
 // NewStoreGrantChecker creates a GrantChecker backed by the MCP store.
 // Subscribes to TopicCacheMCP for cache invalidation on grant changes.
 func NewStoreGrantChecker(mcpStore store.MCPServerStore, msgBus *bus.MessageBus) *storeGrantChecker {
@@ -59,7 +61,7 @@ func NewStoreGrantChecker(mcpStore store.MCPServerStore, msgBus *bus.MessageBus)
 	// Subscribe to cache invalidation events.
 	// When MCP grants are revoked/modified, the HTTP handler broadcasts to this topic.
 	if msgBus != nil {
-		msgBus.Subscribe(bus.TopicCacheMCP, func(event bus.Event) {
+		msgBus.Subscribe(grantCheckerCacheSubscriberID, func(event bus.Event) {
 			if event.Name == protocol.EventCacheInvalidate {
 				gc.cache.Clear()
 				slog.Debug("mcp.grant_checker.cache_cleared", "trigger", "bus_event")

internal/mcp/grant_checker_test.go

@@ -6,7 +6,9 @@ import (
 	"testing"
 
 	"github.com/google/uuid"
+	"github.com/nextlevelbuilder/goclaw/internal/bus"
 	"github.com/nextlevelbuilder/goclaw/internal/store"
+	"github.com/nextlevelbuilder/goclaw/pkg/protocol"
 )
 
 // mockMCPStore implements store.MCPServerStore for testing.
@@ -28,11 +30,13 @@ func (m *mockMCPStore) GetServer(ctx context.Context, id uuid.UUID) (*store.MCPS
 func (m *mockMCPStore) GetServerByName(ctx context.Context, name string) (*store.MCPServerData, error) {
 	return nil, nil
 }
-func (m *mockMCPStore) ListServers(ctx context.Context) ([]store.MCPServerData, error) { return nil, nil }
+func (m *mockMCPStore) ListServers(ctx context.Context) ([]store.MCPServerData, error) {
+	return nil, nil
+}
 func (m *mockMCPStore) UpdateServer(ctx context.Context, id uuid.UUID, updates map[string]any) error {
 	return nil
 }
-func (m *mockMCPStore) DeleteServer(ctx context.Context, id uuid.UUID) error        { return nil }
+func (m *mockMCPStore) DeleteServer(ctx context.Context, id uuid.UUID) error           { return nil }
 func (m *mockMCPStore) GrantToAgent(ctx context.Context, g *store.MCPAgentGrant) error { return nil }
 func (m *mockMCPStore) RevokeFromAgent(ctx context.Context, serverID, agentID uuid.UUID) error {
 	return nil
@@ -132,6 +136,60 @@ func TestStoreGrantChecker_InvalidateClearsCache(t *testing.T) {
 	}
 }
 
+func TestStoreGrantChecker_BusInvalidationSurvivesOtherMCPSubscribers(t *testing.T) {
+	initialServerID := uuid.New()
+	newServerID := uuid.New()
+	agentID := uuid.New()
+	userID := "system"
+
+	mockStore := &mockMCPStore{
+		accessible: []store.MCPAccessInfo{
+			{Server: store.MCPServerData{BaseModel: store.BaseModel{ID: initialServerID}, Name: "initial-server"}},
+		},
+	}
+
+	msgBus := bus.New()
+	gc := NewStoreGrantChecker(mockStore, msgBus)
+	ctx := context.Background()
+
+	if allowed, reason := gc.IsAllowed(ctx, agentID, userID, initialServerID, "tool"); !allowed {
+		t.Fatalf("expected initial server allowed, reason=%q", reason)
+	}
+	if mockStore.callCount != 1 {
+		t.Fatalf("expected 1 DB call after warm cache, got %d", mockStore.callCount)
+	}
+
+	// The gateway also subscribes to MCP cache events. Subscriber IDs must be
+	// unique; otherwise this later subscription replaces the grant checker and
+	// its stale cache survives grant changes.
+	otherSubscriberCalled := false
+	msgBus.Subscribe(bus.TopicCacheMCP, func(event bus.Event) {
+		if event.Name == protocol.EventCacheInvalidate {
+			otherSubscriberCalled = true
+		}
+	})
+
+	mockStore.accessible = []store.MCPAccessInfo{
+		{Server: store.MCPServerData{BaseModel: store.BaseModel{ID: initialServerID}, Name: "initial-server"}},
+		{Server: store.MCPServerData{BaseModel: store.BaseModel{ID: newServerID}, Name: "new-server"}},
+	}
+
+	msgBus.Broadcast(bus.Event{
+		Name:    protocol.EventCacheInvalidate,
+		Payload: bus.CacheInvalidatePayload{Kind: bus.CacheKindMCP},
+	})
+	if !otherSubscriberCalled {
+		t.Fatal("expected other MCP subscriber to receive invalidation")
+	}
+
+	if allowed, reason := gc.IsAllowed(ctx, agentID, userID, newServerID, "tool"); !allowed {
+		t.Fatalf("expected grant checker cache to refresh after bus invalidation, reason=%q", reason)
+	}
+	if mockStore.callCount != 2 {
+		t.Fatalf("expected cache miss after bus invalidation, got callCount=%d", mockStore.callCount)
+	}
+}
+
 func TestStoreGrantChecker_DenyReasonSurfaced(t *testing.T) {
 	serverID := uuid.New()
 	agentID := uuid.New()

internal/pipeline/compaction_pending_preservation_test.go

@@ -0,0 +1,131 @@
+package pipeline
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"github.com/nextlevelbuilder/goclaw/internal/providers"
+)
+
+// TestPruneStage_Compaction_PreservesPending is the regression guard for the
+// mid-loop compaction tool-pairing bug. ThinkStage appends the current
+// iteration's assistant(tool_calls) message to the pending buffer, then
+// PruneStage runs next. If compaction fires, CompactMessages only receives
+// History() — pending is NOT part of the compaction input — yet ReplaceHistory
+// clears pending. Without preservation, the assistant(tool_calls) message is
+// dropped, ToolStage later appends orphaned tool_result messages, and the next
+// provider call rejects the unpaired tool_result (OpenAI/DeepSeek 400).
+//
+// This test asserts the pending assistant(tool_calls) message survives a
+// mid-loop compaction so tool_calls -> tool_result pairing stays intact.
+func TestPruneStage_Compaction_PreservesPending(t *testing.T) {
+	t.Parallel()
+	// budget = ContextWindow(1000) - overhead(0) - MaxTokens(100) = 900.
+	// History of 50 msgs * 100 tokens = 5000 > 900 forces the compaction path.
+	deps := &PipelineDeps{
+		Config: PipelineConfig{
+			ContextWindow: 1000,
+			MaxTokens:     100,
+		},
+		TokenCounter: &mockTokenCounter{countPerMessage: 100},
+		PruneMessages: func(msgs []providers.Message, _ int) ([]providers.Message, PruneStats) {
+			return msgs, PruneStats{} // no reduction — forces fall-through to compaction
+		},
+		CompactMessages: func(_ context.Context, msgs []providers.Message, _ string) ([]providers.Message, error) {
+			// Compaction summarizes history into a single message. Crucially it
+			// only ever sees History(), never the pending assistant(tool_calls).
+			return []providers.Message{{Role: "user", Content: "[compacted summary]"}}, nil
+		},
+	}
+	stage := NewPruneStage(deps, nil)
+	state := defaultState()
+
+	history := make([]providers.Message, 50)
+	for i := range history {
+		history[i] = providers.Message{Role: "user", Content: "msg"}
+	}
+	state.Messages.SetHistory(history)
+
+	// The in-flight assistant(tool_calls) message ThinkStage placed in pending
+	// this iteration — it must outlive ReplaceHistory.
+	assistantToolCall := providers.Message{
+		Role:      "assistant",
+		ToolCalls: []providers.ToolCall{{ID: "tc-keep", Name: "read_file"}},
+	}
+	state.Messages.AppendPending(assistantToolCall)
+
+	if err := stage.Execute(context.Background(), state); err != nil {
+		t.Fatalf("Execute() error: %v", err)
+	}
+
+	if !state.Prune.MidLoopCompacted {
+		t.Fatal("MidLoopCompacted = false, want true (compaction should have fired)")
+	}
+	if state.Compact.CompactionCount != 1 {
+		t.Errorf("CompactionCount = %d, want 1", state.Compact.CompactionCount)
+	}
+
+	// The assistant(tool_calls) message must still be pending after compaction.
+	pending := state.Messages.Pending()
+	if len(pending) != 1 {
+		t.Fatalf("Pending() len = %d, want 1 (assistant tool_calls preserved across ReplaceHistory)", len(pending))
+	}
+	if pending[0].Role != "assistant" || len(pending[0].ToolCalls) != 1 || pending[0].ToolCalls[0].ID != "tc-keep" {
+		t.Errorf("pending[0] = %+v, want the preserved assistant tool_calls message (id tc-keep)", pending[0])
+	}
+
+	// History should be the compacted summary, not carrying a duplicate of the
+	// preserved pending message.
+	hist := state.Messages.History()
+	if len(hist) != 1 || hist[0].Content != "[compacted summary]" {
+		t.Errorf("History() = %+v, want single compacted summary message", hist)
+	}
+}
+
+// TestThinkStage_EmergencyCompaction_PreservesPending guards the second
+// ReplaceHistory call site. When CallLLM reports a context-overflow error,
+// ThinkStage runs an emergency compaction and retries the iteration. Pending
+// messages already staged this iteration (e.g. iteration-budget nudges from
+// maybeInjectNudge) must survive the ReplaceHistory that emergency compaction
+// performs, for the same reason as the prune path.
+func TestThinkStage_EmergencyCompaction_PreservesPending(t *testing.T) {
+	t.Parallel()
+	deps := &PipelineDeps{
+		Config: PipelineConfig{MaxIterations: 10, MaxTokens: 1000},
+		CallLLM: func(_ context.Context, _ *RunState, _ providers.ChatRequest) (*providers.ChatResponse, error) {
+			// "context length exceeded" is matched by IsContextOverflowMessage.
+			return nil, errors.New("context length exceeded for this request")
+		},
+		CompactMessages: func(_ context.Context, msgs []providers.Message, _ string) ([]providers.Message, error) {
+			return []providers.Message{{Role: "user", Content: "[compacted summary]"}}, nil
+		},
+	}
+	stage := NewThinkStage(deps)
+	state := defaultState()
+	state.Messages.SetHistory([]providers.Message{{Role: "user", Content: "old turn"}})
+
+	// A pending message staged earlier in this iteration (an iteration nudge).
+	nudge := providers.Message{Role: "user", Content: "[System] wrap up your work"}
+	state.Messages.AppendPending(nudge)
+
+	if err := stage.Execute(context.Background(), state); err != nil {
+		t.Fatalf("Execute() error: %v", err)
+	}
+
+	// Emergency compaction retries the iteration (Continue) and records one retry.
+	if stage.Result() != Continue {
+		t.Errorf("Result() = %v, want Continue (retry after emergency compaction)", stage.Result())
+	}
+	if state.Think.OverflowRetries != 1 {
+		t.Errorf("OverflowRetries = %d, want 1", state.Think.OverflowRetries)
+	}
+
+	pending := state.Messages.Pending()
+	if len(pending) != 1 {
+		t.Fatalf("Pending() len = %d, want 1 (nudge preserved across emergency ReplaceHistory)", len(pending))
+	}
+	if pending[0].Content != "[System] wrap up your work" {
+		t.Errorf("pending[0] = %+v, want the preserved nudge message", pending[0])
+	}
+}

internal/pipeline/deps.go

@@ -90,6 +90,9 @@ type PipelineDeps struct {
 	ExecuteToolRaw func(ctx context.Context, tc providers.ToolCall) (providers.Message, any, error)
 	// ProcessToolResult processes a raw tool result with state mutation (sequential only).
 	ProcessToolResult func(ctx context.Context, state *RunState, tc providers.ToolCall, rawMsg providers.Message, rawData any) []providers.Message
+	// AuthorizeToolCall validates whether a tool call is allowed to execute.
+	// Used by ToolStage as a runtime guard against out-of-policy tool calls.
+	AuthorizeToolCall func(ctx context.Context, state *RunState, tc providers.ToolCall) (bool, string)
 	// SequentialToolCall returns true for tools that must preserve same-response order.
 	// When any tool call in a batch matches, ToolStage uses ExecuteToolCall for the
 	// whole batch instead of parallel raw execution.

internal/pipeline/prune_stage.go

@@ -180,11 +180,21 @@ func (s *PruneStage) Execute(ctx context.Context, state *RunState) error {
 		return nil // no compaction available
 	}
 
+	// Save pending before compaction — ReplaceHistory clears pending,
+	// but pending may contain the current iteration's assistant(tool_calls)
+	// message which must be preserved to maintain tool_calls → tool pairing.
+	savedPending := state.Messages.Pending()
+
 	compacted, err := s.deps.CompactMessages(ctx, state.Messages.History(), state.Model)
 	if err != nil {
 		return fmt.Errorf("compact messages: %w", err)
 	}
 	state.Messages.ReplaceHistory(compacted)
+
+	// Restore pending messages that were cleared by ReplaceHistory.
+	for _, msg := range savedPending {
+		state.Messages.AppendPending(msg)
+	}
 	state.Prune.MidLoopCompacted = true
 	state.Compact.CompactionCount++
 	state.Compact.MemoryFlushedThisCycle = false // reset for next cycle

internal/pipeline/substates.go

@@ -52,6 +52,9 @@ type PruneState struct {
 
 // ToolState: owned by ToolStage.
 type ToolState struct {
+	// AllowedTools is the per-iteration execution allowlist built from tool
+	// definitions sent to the provider. Nil means "no runtime restriction".
+	AllowedTools   map[string]bool
 	LoopDetector   any // concrete type toolLoopState lives in agent; Phase 5 defines LoopDetector interface
 	TotalToolCalls int
 	AsyncToolCalls []string      // tool names that executed async (spawn)

internal/pipeline/think_stage.go

@@ -44,6 +44,13 @@ func (s *ThinkStage) Execute(ctx context.Context, state *RunState) error {
 		if err != nil {
 			return fmt.Errorf("build tools: %w", err)
 		}
+		allowed := make(map[string]bool, len(toolDefs))
+		for _, td := range toolDefs {
+			allowed[td.Function.Name] = true
+		}
+		state.Tool.AllowedTools = allowed
+	} else {
+		state.Tool.AllowedTools = nil
 	}
 
 	// 3. Construct ChatRequest
@@ -158,12 +165,16 @@ func (s *ThinkStage) tryEmergencyCompaction(ctx context.Context, state *RunState
 	}
 
 	originalLen := len(state.Messages.History())
+	savedPending := state.Messages.Pending()
 	compacted, compactErr := s.deps.CompactMessages(ctx, state.Messages.History(), state.Model)
 	if compactErr != nil {
 		slog.Warn("emergency_compaction_failed", "reason", reason, "error", compactErr)
 		return false
 	}
 	state.Messages.ReplaceHistory(compacted)
+	for _, msg := range savedPending {
+		state.Messages.AppendPending(msg)
+	}
 	slog.Info("emergency_compaction_triggered",
 		"run_id", state.RunID,
 		"reason", reason,

internal/pipeline/tool_stage.go

@@ -149,6 +149,16 @@ func (s *ToolStage) preflightToolCalls(ctx context.Context, state *RunState, too
 }
 
 func (s *ToolStage) preflightToolCall(ctx context.Context, state *RunState, tc providers.ToolCall) (providers.ToolCall, *providers.Message) {
+	if s.deps.AuthorizeToolCall != nil {
+		if ok, reason := s.deps.AuthorizeToolCall(ctx, state, tc); !ok {
+			return tc, &providers.Message{
+				Role:       "tool",
+				Content:    reason,
+				ToolCallID: tc.ID,
+				IsError:    true,
+			}
+		}
+	}
 	r, _ := s.deps.FireHook(ctx, hooks.Event{
 		EventID:   uuid.NewString(),
 		SessionID: state.Input.SessionKey,

internal/pipeline/tool_stage_authorize_test.go

@@ -0,0 +1,216 @@
+package pipeline
+
+import (
+	"context"
+	"testing"
+
+	"github.com/nextlevelbuilder/goclaw/internal/providers"
+)
+
+// makeAuthorizeGate builds a PipelineDeps.AuthorizeToolCall callback backed by
+// a static allowlist and an optional deferred-activator, mirroring the contract
+// that makeAuthorizeToolCall in loop_pipeline_callbacks.go provides.
+func makeAuthorizeGate(
+	allowed map[string]bool, // nil = allow-all
+	tryActivate func(name string) bool, // nil = no activator
+	isDenied func(name string) bool, // nil = never denied
+) func(ctx context.Context, state *RunState, tc providers.ToolCall) (bool, string) {
+	return func(_ context.Context, state *RunState, tc providers.ToolCall) (bool, string) {
+		a := state.Tool.AllowedTools
+		if a == nil {
+			return true, ""
+		}
+		if a[tc.Name] {
+			return true, ""
+		}
+		if tryActivate != nil && tryActivate(tc.Name) {
+			if isDenied != nil && isDenied(tc.Name) {
+				return false, "tool not allowed by policy: " + tc.Name
+			}
+			a[tc.Name] = true
+			return true, ""
+		}
+		return false, "tool not allowed by policy: " + tc.Name
+	}
+}
+
+// makeMinimalDeps returns a PipelineDeps just sufficient for ToolStage tests.
+func makeMinimalDeps(authorize func(context.Context, *RunState, providers.ToolCall) (bool, string)) *PipelineDeps {
+	return &PipelineDeps{
+		Config:            PipelineConfig{MaxToolCalls: 100},
+		AuthorizeToolCall: authorize,
+		ExecuteToolCall: func(_ context.Context, state *RunState, tc providers.ToolCall) ([]providers.Message, error) {
+			return []providers.Message{{Role: "tool", Content: "ok", ToolCallID: tc.ID}}, nil
+		},
+	}
+}
+
+// stateWithAllowedTools returns a minimal RunState with AllowedTools pre-set.
+func stateWithAllowedTools(allowed map[string]bool) *RunState {
+	st := defaultState()
+	st.Tool.AllowedTools = allowed
+	return st
+}
+
+func toolCallFor(name string) providers.ToolCall {
+	return providers.ToolCall{ID: "tc-" + name, Name: name}
+}
+
+// fakeResponse puts tool calls into the state's ThinkStage so ToolStage finds them.
+func setLastResponse(state *RunState, calls []providers.ToolCall) {
+	state.Think.LastResponse = &providers.ChatResponse{ToolCalls: calls}
+}
+
+// --- Tests ---
+
+// TestAuthorizeGate_NilAllowlist_AllowsAll verifies that a nil AllowedTools map
+// means no per-iteration restriction (allow-all semantics).
+func TestAuthorizeGate_NilAllowlist_AllowsAll(t *testing.T) {
+	t.Parallel()
+	st := defaultState()
+	st.Tool.AllowedTools = nil // explicit nil
+
+	gate := makeAuthorizeGate(nil, nil, nil)
+	tc := toolCallFor("exec")
+
+	ok, reason := gate(context.Background(), st, tc)
+	if !ok {
+		t.Errorf("nil allowlist must allow all tools; got blocked with reason %q", reason)
+	}
+}
+
+// TestAuthorizeGate_NamePresent_Allows checks that a tool explicitly in the
+// allowlist passes through.
+func TestAuthorizeGate_NamePresent_Allows(t *testing.T) {
+	t.Parallel()
+	st := stateWithAllowedTools(map[string]bool{"exec": true, "read_file": true})
+
+	gate := makeAuthorizeGate(nil, nil, nil)
+	for _, name := range []string{"exec", "read_file"} {
+		ok, reason := gate(context.Background(), st, toolCallFor(name))
+		if !ok {
+			t.Errorf("tool %q should be allowed; got blocked: %q", name, reason)
+		}
+	}
+}
+
+// TestAuthorizeGate_NameAbsent_NoActivator_Denies confirms fail-closed: a name
+// not in the allowlist with no activator is blocked.
+func TestAuthorizeGate_NameAbsent_NoActivator_Denies(t *testing.T) {
+	t.Parallel()
+	st := stateWithAllowedTools(map[string]bool{"read_file": true})
+
+	gate := makeAuthorizeGate(nil, nil, nil)
+	ok, reason := gate(context.Background(), st, toolCallFor("exec"))
+	if ok {
+		t.Error("tool absent from allowlist with no activator must be denied")
+	}
+	if reason == "" {
+		t.Error("denial must include a non-empty reason")
+	}
+}
+
+// TestAuthorizeGate_DeferredActivate_ThenDenied ensures that lazy activation
+// followed by a positive IsDenied check still blocks the tool.
+func TestAuthorizeGate_DeferredActivate_ThenDenied(t *testing.T) {
+	t.Parallel()
+	st := stateWithAllowedTools(map[string]bool{})
+
+	activated := false
+	tryActivate := func(name string) bool {
+		activated = true
+		return name == "mcp_svc__exec_cmd"
+	}
+	isDenied := func(name string) bool {
+		return name == "mcp_svc__exec_cmd" // explicitly denied
+	}
+
+	gate := makeAuthorizeGate(nil, tryActivate, isDenied)
+	ok, reason := gate(context.Background(), st, toolCallFor("mcp_svc__exec_cmd"))
+
+	if !activated {
+		t.Error("expected TryActivateDeferred to be called")
+	}
+	if ok {
+		t.Error("tool allowed by activator but denied by policy must still be blocked")
+	}
+	if reason == "" {
+		t.Error("denial must include a non-empty reason")
+	}
+	if st.Tool.AllowedTools["mcp_svc__exec_cmd"] {
+		t.Error("denied tool must not be added to AllowedTools")
+	}
+}
+
+// TestAuthorizeGate_PrefixedName_CanonicalLookup is the regression guard for the
+// toolCallPrefix bug: when an agent has toolCallPrefix set, the model emits
+// "proxy_exec" but AllowedTools is keyed by canonical name "exec". The gate must
+// resolve the name before lookup, otherwise every prefixed call is wrongly blocked.
+//
+// This test simulates the resolution by calling the gate with a *pre-resolved*
+// canonical name, matching what makeAuthorizeToolCall does after calling
+// resolveToolCallName(tc.Name). The pipeline layer always receives the resolved
+// name because AuthorizeToolCall in the real callback operates on the canonical
+// form.
+func TestAuthorizeGate_PrefixedName_CanonicalLookup(t *testing.T) {
+	t.Parallel()
+	// AllowedTools uses canonical name "exec" (as ThinkStage builds it from FilterTools).
+	st := stateWithAllowedTools(map[string]bool{"exec": true})
+
+	gate := makeAuthorizeGate(nil, nil, nil)
+
+	// Simulate what resolveToolCallName("proxy_exec") returns with prefix "proxy_":
+	// the canonical name "exec" is passed to the gate.
+	resolvedName := "exec" // prefix already stripped by makeAuthorizeToolCall
+	tc := providers.ToolCall{ID: "tc-prefix", Name: resolvedName}
+
+	ok, reason := gate(context.Background(), st, tc)
+	if !ok {
+		t.Errorf("canonical name after prefix resolution must be allowed; reason: %q", reason)
+	}
+}
+
+// TestToolStage_AllBlocked_CallsCheckExitConditions verifies that when all tool
+// calls in the parallel path are blocked, ToolStage still runs checkExitConditions
+// (so MaxToolCalls budget is enforced). Regression for early-return-nil bug.
+func TestToolStage_AllBlocked_CallsCheckExitConditions(t *testing.T) {
+	t.Parallel()
+
+	// Allow-all deps but with MaxToolCalls = 1 and AuthorizeToolCall that blocks.
+	denyAll := func(_ context.Context, _ *RunState, tc providers.ToolCall) (bool, string) {
+		return false, "blocked: " + tc.Name
+	}
+	deps := &PipelineDeps{
+		Config: PipelineConfig{MaxToolCalls: 1},
+		AuthorizeToolCall: denyAll,
+		ExecuteToolCall: func(_ context.Context, state *RunState, tc providers.ToolCall) ([]providers.Message, error) {
+			return nil, nil
+		},
+		ExecuteToolRaw: func(_ context.Context, tc providers.ToolCall) (providers.Message, any, error) {
+			return providers.Message{}, nil, nil
+		},
+		ProcessToolResult: func(_ context.Context, state *RunState, tc providers.ToolCall, rawMsg providers.Message, rawData any) []providers.Message {
+			return nil
+		},
+	}
+
+	stage := NewToolStage(deps)
+	state := defaultState()
+	state.Tool.AllowedTools = map[string]bool{} // non-nil allowlist → gate is active
+	state.Tool.TotalToolCalls = 1               // already at MaxToolCalls
+
+	// Two tool calls — parallel path is triggered (len > 1).
+	setLastResponse(state, []providers.ToolCall{
+		toolCallFor("exec"),
+		toolCallFor("write_file"),
+	})
+
+	if err := stage.Execute(context.Background(), state); err != nil {
+		t.Fatalf("Execute returned error: %v", err)
+	}
+
+	// checkExitConditions must have fired: TotalToolCalls >= MaxToolCalls → BreakLoop.
+	if stage.Result() != BreakLoop {
+		t.Errorf("expected BreakLoop after all-blocked + MaxToolCalls reached, got %v", stage.Result())
+	}
+}

internal/providers/adapter_anthropic_test.go

@@ -150,6 +150,48 @@ func TestAnthropicAdapterToRequest_Thinking(t *testing.T) {
 	}
 }
 
+func TestAnthropicAdapterToRequest_SkipsTemperatureForClaude46(t *testing.T) {
+	adapter, _ := NewAnthropicAdapter(ProviderConfig{APIKey: "sk-test"})
+
+	for _, model := range []string{"claude-opus-4-6", "claude-sonnet-4-6", "claude-opus-4-7-20260501"} {
+		t.Run(model, func(t *testing.T) {
+			req := ChatRequest{
+				Model:    model,
+				Messages: []Message{{Role: "user", Content: "hi"}},
+				Options:  map[string]any{OptTemperature: 0.7},
+			}
+			data, _, err := adapter.ToRequest(req)
+			if err != nil {
+				t.Fatal(err)
+			}
+			var body map[string]any
+			if err := json.Unmarshal(data, &body); err != nil {
+				t.Fatal(err)
+			}
+			if _, hasTemp := body["temperature"]; hasTemp {
+				t.Errorf("model %q: temperature should be omitted", model)
+			}
+		})
+	}
+
+	req := ChatRequest{
+		Model:    "claude-sonnet-4-5-20250929",
+		Messages: []Message{{Role: "user", Content: "hi"}},
+		Options:  map[string]any{OptTemperature: 0.7},
+	}
+	data, _, err := adapter.ToRequest(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var body map[string]any
+	if err := json.Unmarshal(data, &body); err != nil {
+		t.Fatal(err)
+	}
+	if body["temperature"] != 0.7 {
+		t.Errorf("sonnet 4.5 should keep temperature, got %v", body["temperature"])
+	}
+}
+
 func TestAnthropicAdapterFromResponse_ToolCalls(t *testing.T) {
 	adapter, _ := NewAnthropicAdapter(ProviderConfig{APIKey: "sk-test"})
 

internal/providers/anthropic_request.go

@@ -213,7 +213,9 @@ func (p *AnthropicProvider) buildRequestBody(model string, req ChatRequest, stre
 		body["max_tokens"] = v
 	}
 	if v, ok := req.Options[OptTemperature]; ok {
-		body["temperature"] = v
+		if !anthropicSkipsTemperature(model) {
+			body["temperature"] = v
+		}
 	}
 
 	// Enable extended thinking if thinking_level is set
@@ -234,6 +236,30 @@ func (p *AnthropicProvider) buildRequestBody(model string, req ChatRequest, stre
 	return body
 }
 
+// anthropicSkipsTemperature reports whether the Messages API rejects sampling
+// parameters for this model. Claude Opus/Sonnet 4.6+ and Opus 4.7+ return HTTP
+// 400 when temperature (and top_p/top_k) are included; omit them entirely.
+func anthropicSkipsTemperature(model string) bool {
+	m := strings.ToLower(model)
+	for _, family := range []string{"claude-opus-4-", "claude-sonnet-4-"} {
+		after, ok := strings.CutPrefix(m, family)
+		if !ok {
+			continue
+		}
+		minor := 0
+		for _, c := range after {
+			if c < '0' || c > '9' {
+				break
+			}
+			minor = minor*10 + int(c-'0')
+		}
+		if minor >= 6 {
+			return true
+		}
+	}
+	return false
+}
+
 // anthropicThinkingBudget maps a thinking level to a token budget.
 func anthropicThinkingBudget(level string) int {
 	switch level {

internal/providers/defaults.go

@@ -38,6 +38,11 @@ func NewDefaultTransport() *http.Transport {
 
 // NewDefaultHTTPClient returns an *http.Client backed by NewDefaultTransport.
 // No Client.Timeout is set — rely on ctx deadlines and Transport stage timeouts.
+//
+// SSRF protection for user-configured provider URLs is enforced at provider
+// create/update time by validateProviderURL (resolves the host and rejects
+// private/reserved IPs via security.IsBlocked). Dial-time DNS-rebinding
+// hardening is tracked as a follow-up.
 func NewDefaultHTTPClient() *http.Client {
 	return &http.Client{Transport: NewDefaultTransport()}
 }

internal/providers/embedding_openai.go

@@ -40,8 +40,11 @@ func NewOpenAIEmbeddingProvider(apiKey, apiBase, model string) *OpenAIEmbeddingP
 		apiKey:       apiKey,
 		apiBase:      strings.TrimRight(apiBase, "/"),
 		model:        model,
-		client:       &http.Client{Timeout: 60 * time.Second},
-		retry:        DefaultRetryConfig(),
+		// Use a fixed 60s timeout client. Embedding requests are short but can be
+		// batched. The apiBase is validated at provider-creation time by
+		// validateProviderURL in internal/http/providers.go.
+		client: &http.Client{Timeout: 60 * time.Second},
+		retry:  DefaultRetryConfig(),
 	}
 }
 

internal/providers/vertex.go

@@ -131,11 +131,15 @@ func NewVertexProvider(ctx context.Context, cfg VertexConfig) (*OpenAIProvider,
 	// then transparently fetches a fresh one. No extra work for callers.
 	cached := oauth2.ReuseTokenSource(nil, tokenSource)
 
+	// Use NewDefaultTransport as the oauth2 base transport so Vertex requests
+	// inherit the same per-stage timeouts as other providers. SSRF defense is
+	// provided by validateVertexAPIBaseOverride, which enforces the googleapis.com
+	// host constraint at provider-creation time.
 	client := &http.Client{
 		Timeout: DefaultHTTPTimeout,
 		Transport: &oauth2.Transport{
 			Source: cached,
-			Base:   http.DefaultTransport,
+			Base:   NewDefaultTransport(),
 		},
 	}
 

internal/sandbox/docker_test.go

@@ -172,31 +172,3 @@ func TestFsBridgePathWithinUsesPathBoundaries(t *testing.T) {
 		})
 	}
 }
-
-func TestFsBridgeWriteFileCommandPreservesOverwriteTruncation(t *testing.T) {
-	args := fsBridgeWriteDDArgs("/workspace/file.txt", false)
-	for _, arg := range args {
-		if arg == "conv=notrunc" || arg == "oflag=append" {
-			t.Fatalf("overwrite command must truncate, got append-only arg %q in %v", arg, args)
-		}
-	}
-}
-
-func TestFsBridgeWriteFileCommandUsesNoTruncOnlyForAppend(t *testing.T) {
-	args := fsBridgeWriteDDArgs("/workspace/file.txt", true)
-	if !containsString(args, "conv=notrunc") {
-		t.Fatalf("append command missing conv=notrunc: %v", args)
-	}
-	if !containsString(args, "oflag=append") {
-		t.Fatalf("append command missing oflag=append: %v", args)
-	}
-}
-
-func containsString(values []string, target string) bool {
-	for _, value := range values {
-		if value == target {
-			return true
-		}
-	}
-	return false
-}

internal/sandbox/fsbridge.go

@@ -80,8 +80,12 @@ func (b *FsBridge) WriteFile(ctx context.Context, path, content string, appendMo
 		}
 	}
 
-	ddArgs := fsBridgeWriteDDArgs(resolved, appendMode)
-	_, stderr, exitCode, err := b.dockerExec(ctx, []byte(content), ddArgs...)
+	// Write content via stdin without invoking a shell. Passing the resolved path
+	// as a discrete argv entry prevents shell metacharacters in filenames from
+	// being interpreted as commands inside the sandbox container.
+	// "--" terminates option parsing so paths starting with "-" are safe.
+	teeArgs := fsBridgeWriteTeeArgs(resolved, appendMode)
+	_, stderr, exitCode, err := b.dockerExec(ctx, []byte(content), teeArgs...)
 	if err != nil {
 		return fmt.Errorf("fsbridge write: %w", err)
 	}
@@ -92,11 +96,12 @@ func (b *FsBridge) WriteFile(ctx context.Context, path, content string, appendMo
 	return nil
 }
 
-func fsBridgeWriteDDArgs(resolved string, appendMode bool) []string {
-	args := []string{"dd", "bs=1048576", "status=none", "of=" + resolved}
+func fsBridgeWriteTeeArgs(resolved string, appendMode bool) []string {
+	args := []string{"tee"}
 	if appendMode {
-		args = append(args, "conv=notrunc", "oflag=append")
+		args = append(args, "-a")
 	}
+	args = append(args, "--", resolved)
 	return args
 }
 

internal/sandbox/fsbridge_test.go

@@ -0,0 +1,189 @@
+package sandbox
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// fakeDockerScript is a fake `docker` binary used in WriteFile tests.
+// It logs all argv to $DOCKER_LOG and, when invoked with a `tee` sub-command,
+// captures stdin to $DOCKER_STDIN.
+// `realpath -e -- <path>` echoes the path back (simulates existing path).
+// `mkdir` exits 0 silently.
+const fakeDockerScript = `#!/bin/sh
+{
+  echo CALL
+  i=0
+  for arg in "$@"; do
+    echo "ARG[$i]=$arg"
+    i=$((i + 1))
+  done
+} >> "$DOCKER_LOG"
+
+# realpath: echo the last arg back as the resolved path
+for j in $(seq 0 $#); do
+  eval "a=\${$j}"
+  if [ "$a" = "realpath" ]; then
+    # print the last argument (the path)
+    eval "echo \"\${$#}\""
+    exit 0
+  fi
+done
+
+# tee: capture stdin
+for arg in "$@"; do
+  if [ "$arg" = "tee" ]; then
+    if [ -n "$DOCKER_STDIN" ]; then
+      cat > "$DOCKER_STDIN"
+    else
+      cat > /dev/null
+    fi
+    exit 0
+  fi
+done
+
+exit 0
+`
+
+// installFakeDocker writes the fake docker script to tmp, prepends tmp to PATH,
+// and sets DOCKER_LOG and optionally DOCKER_STDIN env vars for the test.
+func installFakeDocker(t *testing.T, tmp, logPath, stdinPath string) {
+	t.Helper()
+	dockerPath := filepath.Join(tmp, "docker")
+	if err := os.WriteFile(dockerPath, []byte(fakeDockerScript), 0o755); err != nil {
+		t.Fatalf("write fake docker: %v", err)
+	}
+	t.Setenv("PATH", tmp+string(os.PathListSeparator)+os.Getenv("PATH"))
+	t.Setenv("DOCKER_LOG", logPath)
+	if stdinPath != "" {
+		t.Setenv("DOCKER_STDIN", stdinPath)
+	}
+}
+
+// TestFsBridgeWriteFileDoesNotInvokeShell asserts that WriteFile passes the
+// resolved path as a discrete argv entry to `tee -- <path>` and does NOT
+// build a shell command string (no sh / -c / shell metachar expansion).
+func TestFsBridgeWriteFileDoesNotInvokeShell(t *testing.T) {
+	tmp := t.TempDir()
+	logPath := filepath.Join(tmp, "docker.log")
+	stdinPath := filepath.Join(tmp, "stdin.txt")
+
+	installFakeDocker(t, tmp, logPath, stdinPath)
+
+	bridge := NewFsBridge("container-id", "/workspace")
+	maliciousPath := `nested/evil$(touch /tmp/goclaw-fsbridge-pwned);name.txt`
+	content := "safe content"
+
+	if err := bridge.WriteFile(context.Background(), maliciousPath, content, false); err != nil {
+		t.Fatalf("WriteFile returned error: %v", err)
+	}
+
+	logBytes, err := os.ReadFile(logPath)
+	if err != nil {
+		t.Fatalf("read docker log: %v", err)
+	}
+	log := string(logBytes)
+
+	// Must NOT have invoked sh or built a shell string.
+	if strings.Contains(log, "=sh") || strings.Contains(log, "=-c") || strings.Contains(log, "cat >") {
+		t.Fatalf("WriteFile invoked shell command path; log:\n%s", log)
+	}
+	// Must have called tee.
+	if !strings.Contains(log, "=tee") {
+		t.Fatalf("expected write command to use tee without shell; log:\n%s", log)
+	}
+	// Must have included -- to terminate option parsing.
+	if !strings.Contains(log, "=--") {
+		t.Fatalf("expected tee delimiter (--) before filename; log:\n%s", log)
+	}
+	// The malicious path must appear verbatim as a single argv entry (not split/executed).
+	resolved := "/workspace/" + maliciousPath
+	if !strings.Contains(log, "="+resolved) {
+		t.Fatalf("expected malicious filename to remain one argv entry; log:\n%s", log)
+	}
+
+	stdinBytes, err := os.ReadFile(stdinPath)
+	if err != nil {
+		t.Fatalf("read captured stdin: %v", err)
+	}
+	if string(stdinBytes) != content {
+		t.Fatalf("stdin content = %q, want %q", string(stdinBytes), content)
+	}
+}
+
+// TestFsBridgeWriteFileAppendUsesTeeAppendArg asserts that append mode passes
+// -a before -- in the tee argv, i.e. `tee -a -- <path>`.
+func TestFsBridgeWriteFileAppendUsesTeeAppendArg(t *testing.T) {
+	tmp := t.TempDir()
+	logPath := filepath.Join(tmp, "docker.log")
+
+	installFakeDocker(t, tmp, logPath, "")
+
+	bridge := NewFsBridge("container-id", "/workspace")
+	if err := bridge.WriteFile(context.Background(), "append.txt", "more", true); err != nil {
+		t.Fatalf("WriteFile append returned error: %v", err)
+	}
+
+	logBytes, err := os.ReadFile(logPath)
+	if err != nil {
+		t.Fatalf("read docker log: %v", err)
+	}
+	log := string(logBytes)
+
+	// Must have: tee, then -a, then --.
+	if !strings.Contains(log, "=tee") {
+		t.Fatalf("expected tee in argv; log:\n%s", log)
+	}
+	if !strings.Contains(log, "=-a") {
+		t.Fatalf("expected -a flag for append mode; log:\n%s", log)
+	}
+	if !strings.Contains(log, "=--") {
+		t.Fatalf("expected -- delimiter in argv; log:\n%s", log)
+	}
+}
+
+// TestFsBridgeWriteTeeArgs_Overwrite checks the tee args slice for overwrite mode.
+func TestFsBridgeWriteTeeArgs_Overwrite(t *testing.T) {
+	args := fsBridgeWriteTeeArgs("/workspace/file.txt", false)
+	if len(args) != 3 {
+		t.Fatalf("expected 3 args, got %d: %v", len(args), args)
+	}
+	if args[0] != "tee" {
+		t.Errorf("args[0] = %q, want tee", args[0])
+	}
+	if args[1] != "--" {
+		t.Errorf("args[1] = %q, want --", args[1])
+	}
+	if args[2] != "/workspace/file.txt" {
+		t.Errorf("args[2] = %q, want /workspace/file.txt", args[2])
+	}
+	// Must not contain -a in overwrite mode.
+	for _, a := range args {
+		if a == "-a" {
+			t.Errorf("overwrite mode must not contain -a: %v", args)
+		}
+	}
+}
+
+// TestFsBridgeWriteTeeArgs_Append checks the tee args slice for append mode.
+func TestFsBridgeWriteTeeArgs_Append(t *testing.T) {
+	args := fsBridgeWriteTeeArgs("/workspace/file.txt", true)
+	if len(args) != 4 {
+		t.Fatalf("expected 4 args, got %d: %v", len(args), args)
+	}
+	if args[0] != "tee" {
+		t.Errorf("args[0] = %q, want tee", args[0])
+	}
+	if args[1] != "-a" {
+		t.Errorf("args[1] = %q, want -a", args[1])
+	}
+	if args[2] != "--" {
+		t.Errorf("args[2] = %q, want --", args[2])
+	}
+	if args[3] != "/workspace/file.txt" {
+		t.Errorf("args[3] = %q, want /workspace/file.txt", args[3])
+	}
+}

internal/security/ssrf.go

@@ -77,6 +77,16 @@ func isBlocked(ip net.IP) bool {
 	return false
 }
 
+// IsBlocked reports whether ip falls within any blocked CIDR (loopback,
+// link-local including cloud-metadata 169.254.169.254, RFC 1918 private,
+// multicast, and unspecified 0.0.0.0/:: ranges).
+//
+// Use this in provider-URL validation and dial-time guards to avoid
+// duplicating the CIDR list across packages.
+func IsBlocked(ip net.IP) bool {
+	return isBlocked(ip)
+}
+
 // redactURL strips query string and userinfo for safe logging.
 func redactURL(rawURL string) string {
 	u, err := url.Parse(rawURL)

internal/tools/policy.go

@@ -53,7 +53,7 @@ var toolProfiles = map[string][]string{
 }
 
 // Legacy tool aliases — migrated to Registry.RegisterAlias() at startup.
-// Kept as seed data only; resolveAlias() is no longer used.
+// resolveAlias() is used by IsDenied to expand names before deny-spec matching.
 var legacyToolAliases = map[string]string{
 	"bash":           "exec",
 	"apply-patch":    "apply_patch",
@@ -434,19 +434,34 @@ func unionWithSpec(reg *Registry, current []string, allTools []string, spec []st
 
 // IsDenied checks if a tool name is explicitly denied by global or agent policy.
 // Used to prevent lazy-activated deferred tools from bypassing the deny list.
+// Checks under all candidate names: the raw name, the legacy alias (e.g. bash→exec),
+// and the registry alias when available.
 func (pe *PolicyEngine) IsDenied(name string, agentPolicy *config.ToolPolicySpec) bool {
+	candidates := map[string]struct{}{name: {}}
+	// Keep legacy alias compatibility (e.g. bash -> exec).
+	candidates[resolveAlias(name)] = struct{}{}
+	// Include registry alias mapping when available.
 	pe.mu.RLock()
 	reg := pe.registry
 	pe.mu.RUnlock()
+	if reg != nil {
+		if canonical, ok := reg.Aliases()[name]; ok && canonical != "" {
+			candidates[canonical] = struct{}{}
+		}
+	}
 
 	if pe.globalPolicy != nil {
-		if matchDenySpec(reg, name, pe.globalPolicy.Deny) {
-			return true
+		for candidate := range candidates {
+			if matchDenySpec(reg, candidate, pe.globalPolicy.Deny) {
+				return true
+			}
 		}
 	}
 	if agentPolicy != nil {
-		if matchDenySpec(reg, name, agentPolicy.Deny) {
-			return true
+		for candidate := range candidates {
+			if matchDenySpec(reg, candidate, agentPolicy.Deny) {
+				return true
+			}
 		}
 	}
 	return false

tests/integration/git_adapter_ssh_test.go

@@ -160,6 +160,7 @@ func TestGitAdapter_SSH_CleanupOnExecFailure(t *testing.T) {
 // TestGitAdapter_SSH_HostMismatch_NoTmpfile: rejected host mismatch must
 // NOT leave an orphaned tmpfile in os.TempDir().
 func TestGitAdapter_SSH_HostMismatch_NoTmpfile(t *testing.T) {
+	t.Setenv("TMPDIR", t.TempDir())
 	cred := sshCredFor(t, "github.com")
 	adapter := tools.AdapterFor("git")
 

ui/web/nginx.conf

@@ -36,6 +36,7 @@ server {
 
     # API proxy
     location /v1/ {
+        client_max_body_size 500M;
         proxy_pass $upstream_backend;
         proxy_set_header Host $host;
         proxy_set_header X-Real-IP $remote_addr;