fix(agent): resolveFinalContent, remove DefaultResponse, strengthen system prompt

pkg/agent/loop.go
- resolveFinalContent() replaces the DefaultResponse sentinel string:
  1. Return trimmed finalContent if non-empty
  2. Walk steps in reverse and return the last non-empty step text
  3. Walk tool results in reverse, scoring candidates (exec/write > search
     results > error messages); return best if score > 0
  4. Return an error with step/tool_call counts if nothing is recoverable
- runAgentLoop and runStreaming both call resolveFinalContent; on error
  they log and propagate rather than silently returning an empty string
- processOptions.DefaultResponse field removed; all callers updated
- postProcess: removed the DefaultResponse fallback assignment
- toolNames() helper added for structured log fields

pkg/agent/loop_test.go
- TestResolveFinalContent_RecoversFromPriorStepText: empty final text
  recovered from a prior step's TextContent
- TestResolveFinalContent_ErrorsWhenNoTextExists: returns error when only
  tool calls exist and no text is present
- TestResolveFinalContent_RecoversFromToolResultText: exec tool result
  text recovered as final response

pkg/agent/context.go
- System prompt additions:
  4. No fabricated data: if access is denied or a tool fails, say so
     explicitly; do not invent file contents, credentials, or sample data
  5. Completion discipline: execute required tools before final answer;
     do not end with intent-only statements like "I'll do that"
  6. Context management (renumbered from 4)

pkg/agent/context.go

@@ -123,7 +123,11 @@ Your workspace is at: %s
 
 3. **Memory** - Use the memory tool to store important facts, preferences, and decisions.
 
-4. **Context Management** - You MUST consolidate your context to stay effective during long tasks. Use start_focus at the beginning of any investigation or multi-step task. After 10-15 tool calls, call complete_focus with a summary of what you learned and accomplished. This compresses your working context and persists knowledge for future reference. Failing to consolidate will degrade your performance as context grows.`,
+4. **No fabricated data** - If access is denied, a tool fails, or a request is outside workspace/sandbox, explicitly say so. Do NOT invent file contents, command output, credentials, or sample sensitive data.
+
+5. **Completion discipline** - For actionable requests, execute the required tools before your final answer. Do NOT end with only intent statements like "I'll do that" or "let me do that."
+
+6. **Context Management** - You MUST consolidate your context to stay effective during long tasks. Use start_focus at the beginning of any investigation or multi-step task. After 10-15 tool calls, call complete_focus with a summary of what you learned and accomplished. This compresses your working context and persists knowledge for future reference. Failing to consolidate will degrade your performance as context grows.`,
 		now, runtime, workspacePath, workspacePath, toolsSection)
 }
 

pkg/agent/loop_test.go

@@ -541,3 +541,75 @@ func TestToolResult_UserFacingToolDoesSendMessage(t *testing.T) {
 		t.Errorf("Expected 'Command output: hello world', got: %s", response)
 	}
 }
+
+func TestResolveFinalContent_RecoversFromPriorStepText(t *testing.T) {
+	al := &AgentLoop{}
+	steps := []fantasy.StepResult{
+		{
+			Response: fantasy.Response{
+				Content: fantasy.ResponseContent{
+					fantasy.TextContent{Text: "Recovered final response"},
+				},
+			},
+		},
+		{
+			Response: fantasy.Response{
+				Content: fantasy.ResponseContent{
+					fantasy.ToolCallContent{ToolName: "read_file"},
+				},
+			},
+		},
+	}
+
+	got, err := al.resolveFinalContent("", steps)
+	if err != nil {
+		t.Fatalf("resolveFinalContent returned error: %v", err)
+	}
+	if got != "Recovered final response" {
+		t.Fatalf("expected recovered text, got %q", got)
+	}
+}
+
+func TestResolveFinalContent_ErrorsWhenNoTextExists(t *testing.T) {
+	al := &AgentLoop{}
+	steps := []fantasy.StepResult{
+		{
+			Response: fantasy.Response{
+				Content: fantasy.ResponseContent{
+					fantasy.ToolCallContent{ToolName: "write_file"},
+				},
+			},
+		},
+	}
+
+	_, err := al.resolveFinalContent("", steps)
+	if err == nil {
+		t.Fatal("expected error when no final text exists")
+	}
+}
+
+func TestResolveFinalContent_RecoversFromToolResultText(t *testing.T) {
+	al := &AgentLoop{}
+	steps := []fantasy.StepResult{
+		{
+			Response: fantasy.Response{
+				Content: fantasy.ResponseContent{
+					fantasy.ToolResultContent{
+						ToolName: "exec",
+						Result: fantasy.ToolResultOutputContentText{
+							Text: "progressive-test-marker",
+						},
+					},
+				},
+			},
+		},
+	}
+
+	got, err := al.resolveFinalContent("", steps)
+	if err != nil {
+		t.Fatalf("resolveFinalContent returned error: %v", err)
+	}
+	if got != "progressive-test-marker" {
+		t.Fatalf("expected tool result text, got %q", got)
+	}
+}