Fix Go Logger safe-output accounting and failure diagnostics

.github/workflows/go-logger.md

@@ -244,9 +244,11 @@ After adding logging to **all selected files**, validate your changes before cre
 
 After validating your changes:
 
-1. The safe-outputs create-pull-request will automatically create a PR
-2. Ensure your changes follow the guidelines above
-3. The PR title will automatically have the "[log] " prefix
+1. Choose exactly one terminal outcome: `create_pull_request` after successful changes, `noop` when no changes are needed, or `report_incomplete` when a blocking failure prevents completion.
+2. Call the chosen safe-output command exactly once, as your final action. Do not call any other safe-output command before or after it.
+3. Do not probe safe outputs with `which`, `type`, `--help`, or schema-inspection commands.
+4. If the safe-output gateway rejects the call, stop immediately and surface its exact rejection message. Do not retry the call or switch to another terminal safe output.
+5. The PR title will automatically have the "[log] " prefix.
 
 ## Quality Checklist
 

pkg/cli/audit_mcp_tool_usage_test.go

@@ -56,6 +56,17 @@ func TestExtractMCPToolUsageData(t *testing.T) {
 			wantToolCalls: 2,
 			wantErr:       false,
 		},
+		{
+			name: "tool discovery is not tool usage",
+			// Discovery traffic identifies the contacted server but does not constitute tool usage.
+			logContent: `{"timestamp":"2024-01-12T10:00:00Z","level":"info","type":"request","event":"rpc_call","server_name":"safeoutputs","method":"tools/list","duration":50.0,"status":"success"}
+{"timestamp":"2024-01-12T10:00:01Z","level":"info","type":"request","event":"request","server_name":"safeoutputs","method":"tools/list","duration":50.0,"status":"success"}
+`,
+			wantServers:   1,
+			wantTools:     0,
+			wantToolCalls: 0,
+			wantErr:       false,
+		},
 		{
 			name:          "no gateway.jsonl file",
 			logContent:    "",

pkg/cli/audit_report.go

@@ -845,7 +845,7 @@ func extractGHErrorLines(filePath string) []string {
 	for line := range strings.SplitSeq(string(content), "\n") {
 		if strings.Contains(line, "##[error]") {
 			stripped := stripGHALogTimestamps(line)
-			if stripped != "" {
+			if stripped != "" && !isAgentToolResultAnnotation(stripped) {
 				errorLines = append(errorLines, stripped)
 			}
 		}
@@ -854,6 +854,33 @@ func extractGHErrorLines(filePath string) []string {
 	return errorLines
 }
 
+func isAgentToolResultAnnotation(line string) bool {
+	_, payload, found := strings.Cut(line, "##[error]")
+	if !found {
+		return false
+	}
+
+	var event struct {
+		Type    string `json:"type"`
+		Message struct {
+			Content []struct {
+				Type string `json:"type"`
+			} `json:"content"`
+		} `json:"message"`
+	}
+	if err := json.Unmarshal([]byte(strings.TrimSpace(payload)), &event); err != nil || event.Type != "user" {
+		return false
+	}
+	hasToolResult := false
+	for _, content := range event.Message.Content {
+		if content.Type != "tool_result" {
+			return false
+		}
+		hasToolResult = true
+	}
+	return hasToolResult
+}
+
 func extractAgentFailureError(agentRan bool, agentStdioPath string, maxMessageLen int) []ValidationIssue {
 	if !agentRan {
 		return nil

pkg/cli/audit_report_test.go

@@ -1448,6 +1448,35 @@ func TestExtractPreAgentStepErrors(t *testing.T) {
 		assert.Contains(t, errors[0].Message, "Lockdown mode is enabled", "Should include actionable ##[error] text")
 	})
 
+	t.Run("ignores annotated tool results and surfaces the runner failure", func(t *testing.T) {
+		dir := testutil.TempDir(t, "audit-step-*")
+		require.NoError(t, os.WriteFile(filepath.Join(dir, "agent-stdio.log"), []byte("agent output"), 0600))
+		workflowLogsDir := filepath.Join(dir, "workflow-logs", "agent")
+		require.NoError(t, os.MkdirAll(workflowLogsDir, 0755))
+		toolResult := `{"type":"user","message":{"content":[{"type":"tool_result","content":"raw Go source"}]}}`
+		logContent := "2026-08-13T04:03:11Z ##[error]" + toolResult + "\n" +
+			"2026-08-13T04:11:07Z ##[error]The action 'Execute Claude Code CLI' has timed out after 15 minutes."
+		require.NoError(t, os.WriteFile(filepath.Join(workflowLogsDir, "10_Execute Claude Code CLI.txt"), []byte(logContent), 0600))
+
+		errors := extractPreAgentStepErrors(dir)
+		require.Len(t, errors, 1)
+		assert.Contains(t, errors[0].Message, "timed out after 15 minutes")
+		assert.NotContains(t, errors[0].Message, "raw Go source")
+	})
+
+	t.Run("preserves mixed tool result annotations", func(t *testing.T) {
+		dir := testutil.TempDir(t, "audit-step-*")
+		workflowLogsDir := filepath.Join(dir, "workflow-logs", "agent")
+		require.NoError(t, os.MkdirAll(workflowLogsDir, 0755))
+		mixedContent := `{"type":"user","message":{"content":[{"type":"tool_result","content":"raw Go source"},{"type":"text","text":"runner failure"}]}}`
+		logContent := "2026-08-13T04:03:11Z ##[error]" + mixedContent
+		require.NoError(t, os.WriteFile(filepath.Join(workflowLogsDir, "10_Execute Claude Code CLI.txt"), []byte(logContent), 0600))
+
+		errors := extractPreAgentStepErrors(dir)
+		require.Len(t, errors, 1)
+		assert.Contains(t, errors[0].Message, "runner failure")
+	})
+
 	t.Run("returns nil when workflow-logs directory missing", func(t *testing.T) {
 		dir := testutil.TempDir(t, "audit-step-*")
 		// No agent-stdio.log and no workflow-logs directory

pkg/cli/gateway_logs_mcp.go

@@ -123,8 +123,8 @@ func extractToolCallsFromGatewayLog(gatewayLogPath string, mcpData *MCPToolUsage
 			continue // Skip malformed lines
 		}
 
-		// Only process tool call events
-		if entry.Event == "tool_call" || entry.Event == "rpc_call" || entry.Event == "request" {
+		// Only process actual tool invocations, not protocol requests such as tools/list.
+		if entry.Event == "tool_call" || entry.Method == "tools/call" {
 			toolName := entry.ToolName
 			if toolName == "" {
 				toolName = entry.Method

pkg/cli/gateway_logs_parsing.go

@@ -190,8 +190,8 @@ func processGatewayLogEntry(entry *GatewayLogEntry, metrics *GatewayMetrics, ver
 				metrics.TotalDuration += entry.Duration
 			}
 
-			// Track tool calls
-			if entry.ToolName != "" || entry.Method != "" {
+			// Track only actual tool invocations, not protocol requests such as tools/list.
+			if entry.Event == "tool_call" || entry.Method == "tools/call" {
 				toolName := entry.ToolName
 				if toolName == "" {
 					toolName = entry.Method

pkg/cli/gateway_logs_test.go

@@ -53,7 +53,7 @@ func TestParseGatewayLogs(t *testing.T) {
 `,
 			wantServers:   3,
 			wantRequests:  3,
-			wantToolCalls: 3,
+			wantToolCalls: 0,
 			wantErrors:    0,
 			wantErr:       false,
 		},
@@ -447,14 +447,14 @@ func TestGatewayLogsWithMethodField(t *testing.T) {
 
 	assert.Len(t, metrics.Servers, 1)
 	assert.Equal(t, 2, metrics.TotalRequests)
-	assert.Equal(t, 2, metrics.TotalToolCalls)
+	assert.Equal(t, 1, metrics.TotalToolCalls)
 
 	server := metrics.Servers["github"]
 	require.NotNil(t, server)
-	assert.Len(t, server.Tools, 2)
+	assert.Len(t, server.Tools, 1)
 
-	// Check that methods were tracked as tools
-	assert.Contains(t, server.Tools, "tools/list")
+	// Protocol discovery remains a request but is not counted as tool usage.
+	assert.NotContains(t, server.Tools, "tools/list")
 	assert.Contains(t, server.Tools, "tools/call")
 }
 

pkg/workflow/prompts_test.go

@@ -276,6 +276,34 @@ func TestDailyFunctionNamerUsesConcreteClaudeModelsForExperiment(t *testing.T) {
 	}
 }
 
+func TestGoLoggerDefinesSingleTerminalSafeOutputContract(t *testing.T) {
+	repoRoot, err := findRepoRoot()
+	if err != nil {
+		t.Fatalf("Failed to find repo root: %v", err)
+	}
+
+	workflowFile := filepath.Join(repoRoot, ".github", "workflows", "go-logger.md")
+	content, err := os.ReadFile(workflowFile)
+	if err != nil {
+		t.Fatalf("Failed to read workflow file: %v", err)
+	}
+
+	workflow := string(content)
+	for _, keyword := range []string{
+		"exactly one terminal outcome",
+		"`create_pull_request`",
+		"`noop`",
+		"`report_incomplete`",
+		"exactly once, as your final action",
+		"Do not probe safe outputs",
+		"Do not retry the call or switch to another terminal safe output",
+	} {
+		if !strings.Contains(workflow, keyword) {
+			t.Fatalf("Expected go-logger workflow to include safe-output contract keyword %q", keyword)
+		}
+	}
+}
+
 func TestDailyCavemanOptimizerUsesConcreteClaudeModelsForExperiment(t *testing.T) {
 	repoRoot, err := findRepoRoot()
 	if err != nil {