ag-agent 0.12.2

Agentty is an ADE (Agentic Development Environment) for structured, controllable AI-assisted software development.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
use std::path::Path;
use std::process::{Command, Stdio};

use ag_protocol::agent_response_output_schema_json;

use super::backend::{AgentBackend, AgentBackendError, BuildCommandRequest};
use super::prompt::{CliPromptAccessRootMode, append_cli_prompt_access_directories};

/// Lists the Claude tools Agentty enables for unattended sessions.
///
/// `Bash` remains available for Claude workflows that need shell commands,
/// while `WebSearch` and `WebFetch` let Claude answer current-information
/// prompts without an interactive permission grant.
const CLAUDE_ALLOWED_TOOLS: &str =
    "Bash,Edit,MultiEdit,Write,WebSearch,WebFetch,EnterPlanMode,ExitPlanMode";

/// Backend implementation for the Claude CLI.
///
/// Commands are built with `--strict-mcp-config` so provider-level MCP
/// connector defaults (for example Claude.ai account connectors) are ignored
/// unless explicitly configured by Agentty. Claude runs in `stream-json` mode
/// so progress and tool-use events can surface live while the final turn still
/// honors native schema validation.
pub(super) struct ClaudeBackend;

impl AgentBackend for ClaudeBackend {
    fn setup(&self, _folder: &Path) -> Result<(), AgentBackendError> {
        // Claude Code needs no config files
        Ok(())
    }

    fn build_command<'request>(
        &'request self,
        request: BuildCommandRequest<'request>,
    ) -> Result<Command, AgentBackendError> {
        let BuildCommandRequest {
            attachments,
            folder,
            main_checkout_root,
            model,
            request_kind,
            prompt: _prompt,
            replay_transcript: _replay_transcript,
            reasoning_level,
        } = request;
        let mut command = Command::new("claude");

        if request_kind.is_resume() {
            command.arg("-c");
        }

        append_cli_prompt_access_directories(
            &mut command,
            folder,
            attachments,
            CliPromptAccessRootMode::AttachmentsOnly,
        );

        command.arg("-p");
        command.arg("--allowedTools").arg(CLAUDE_ALLOWED_TOOLS);
        append_claude_workspace_settings(&mut command, folder, main_checkout_root);
        command.arg("--input-format").arg("text");
        command.arg("--strict-mcp-config");
        command.arg("--verbose");
        command.arg("--effort").arg(reasoning_level.claude());
        command.arg("--output-format").arg("stream-json");
        command
            .arg("--json-schema")
            .arg(agent_response_output_schema_json());
        command
            .env("ANTHROPIC_MODEL", model)
            .current_dir(folder)
            .stdout(Stdio::piped())
            .stderr(Stdio::piped());

        Ok(command)
    }
}

/// Appends per-turn Claude Code settings that keep known non-session
/// checkouts read-only.
fn append_claude_workspace_settings(
    command: &mut Command,
    workspace_folder: &Path,
    main_checkout_root: Option<&Path>,
) {
    let mut deny_rules = Vec::new();
    let mut deny_write_paths = Vec::new();
    if let Some(main_checkout_root) = main_checkout_root {
        let main_checkout_rule_path = claude_absolute_permission_path(main_checkout_root);
        deny_rules.push(format!("Edit({main_checkout_rule_path}/**)"));
        deny_write_paths.push(main_checkout_root.to_string_lossy().into_owned());
    }

    let settings = serde_json::json!({
        "permissions": {
            "deny": deny_rules,
        },
        "sandbox": {
            "enabled": true,
            "filesystem": {
                "allowWrite": [workspace_folder.to_string_lossy().into_owned()],
                "denyWrite": deny_write_paths,
            }
        }
    });

    command.arg("--settings").arg(settings.to_string());
}

/// Returns a Claude Code permission-rule absolute path using the `//`
/// prefix required by `Read` and `Edit` path rules.
fn claude_absolute_permission_path(path: &Path) -> String {
    let path = path.to_string_lossy().replace('\\', "/");
    let path_without_root = path.trim_start_matches('/');

    format!("//{path_without_root}")
}

#[cfg(test)]
mod tests {
    use std::ffi::OsStr;
    use std::path::PathBuf;

    use ag_protocol::ProtocolSchemaInstructionMode;
    use serde_json::Value;
    use tempfile::tempdir;

    use super::*;
    use crate::agent::prompt as shared_prompt;
    use crate::channel::AgentRequestKind;
    use crate::model::agent::ReasoningLevel;
    use crate::model::turn_prompt::TurnPromptAttachment;

    fn session_start_request_kind() -> AgentRequestKind {
        AgentRequestKind::SessionStart
    }

    fn utility_request_kind() -> AgentRequestKind {
        AgentRequestKind::UtilityPrompt
    }

    fn settings_argument(command: &Command) -> Value {
        let args = command
            .get_args()
            .map(|arg| arg.to_string_lossy().into_owned())
            .collect::<Vec<_>>();
        let settings_position = args
            .iter()
            .position(|arg| arg == "--settings")
            .expect("--settings flag should be present");

        serde_json::from_str(&args[settings_position + 1]).expect("settings JSON should parse")
    }

    #[test]
    /// Verifies Claude permission-rule paths use slash separators for glob
    /// matching even when given a Windows-style checkout path.
    fn test_claude_absolute_permission_path_normalizes_windows_separators() {
        // Arrange
        let path = Path::new(r"C:\Users\dev\project");

        // Act
        let rule_path = claude_absolute_permission_path(path);

        // Assert
        assert_eq!(rule_path, "//C:/Users/dev/project");
        assert!(!rule_path.contains('\\'));
    }

    #[test]
    /// Verifies Claude sessions allow Agentty's required edit and web tools.
    fn test_claude_auto_edit_mode_uses_write_capable_allowed_tools() {
        // Arrange
        let temp_directory = tempdir().expect("failed to create temp dir");
        let main_checkout_root = temp_directory.path().join("main");
        let backend = ClaudeBackend;

        // Act
        let command = AgentBackend::build_command(
            &backend,
            BuildCommandRequest {
                attachments: &[],
                folder: temp_directory.path(),
                main_checkout_root: Some(main_checkout_root.as_path()),
                replay_transcript: None,
                model: "claude-sonnet-5",
                prompt: "Plan prompt",
                reasoning_level: ReasoningLevel::default(),
                request_kind: &session_start_request_kind(),
            },
        )
        .expect("command should build");
        let debug_command = format!("{command:?}");
        let args = command
            .get_args()
            .map(|arg| arg.to_string_lossy().into_owned())
            .collect::<Vec<_>>();

        // Assert
        assert!(debug_command.contains("--allowedTools"));
        assert!(debug_command.contains(CLAUDE_ALLOWED_TOOLS));
        assert!(debug_command.contains("Bash"));
        assert!(debug_command.contains("MultiEdit"));
        assert!(debug_command.contains("Write"));
        assert!(debug_command.contains("WebSearch"));
        assert!(debug_command.contains("WebFetch"));
        assert!(debug_command.contains("--strict-mcp-config"));
        assert!(debug_command.contains("--settings"));
        assert!(debug_command.contains("--effort"));
        assert!(debug_command.contains("--output-format"));
        assert!(debug_command.contains("stream-json"));
        assert!(!debug_command.contains("--permission-mode"));
        assert!(!args.iter().any(String::is_empty));

        let settings = settings_argument(&command);
        let deny_rules = settings
            .pointer("/permissions/deny")
            .and_then(Value::as_array)
            .expect("deny rules should be present");
        assert!(deny_rules.iter().any(|rule| {
            rule.as_str()
                .is_some_and(|rule| rule.starts_with("Edit(//") && rule.ends_with("/main/**)"))
        }));
        assert_eq!(
            settings
                .pointer("/sandbox/enabled")
                .and_then(Value::as_bool),
            Some(true)
        );
    }

    #[test]
    /// Verifies Claude commands pass the selected Opus 4.8 model through the
    /// Claude Code model environment variable.
    fn test_claude_command_sets_anthropic_model_to_claude_opus_48() {
        // Arrange
        let temp_directory = tempdir().expect("failed to create temp dir");
        let backend = ClaudeBackend;

        // Act
        let command = AgentBackend::build_command(
            &backend,
            BuildCommandRequest {
                attachments: &[],
                folder: temp_directory.path(),
                main_checkout_root: None,
                replay_transcript: None,
                model: "claude-opus-4-8",
                prompt: "Use Opus",
                reasoning_level: ReasoningLevel::default(),
                request_kind: &session_start_request_kind(),
            },
        )
        .expect("command should build");
        let anthropic_model = command
            .get_envs()
            .find(|(key, _value)| *key == OsStr::new("ANTHROPIC_MODEL"))
            .and_then(|(_key, value)| value)
            .map(|value| value.to_string_lossy().into_owned());

        // Assert
        assert_eq!(anthropic_model, Some("claude-opus-4-8".to_string()));
    }

    #[test]
    /// Verifies the `--effort` flag is passed to Claude with the correct value
    /// for each `ReasoningLevel`.
    fn test_claude_command_passes_effort_flag_for_each_reasoning_level() {
        // Arrange
        let temp_directory = tempdir().expect("failed to create temp dir");
        let backend = ClaudeBackend;
        let cases = [
            (ReasoningLevel::Low, "low"),
            (ReasoningLevel::Medium, "medium"),
            (ReasoningLevel::High, "high"),
            (ReasoningLevel::XHigh, "max"),
        ];

        for (reasoning_level, expected_effort) in cases {
            // Act
            let command = AgentBackend::build_command(
                &backend,
                BuildCommandRequest {
                    attachments: &[],
                    folder: temp_directory.path(),
                    main_checkout_root: None,
                    replay_transcript: None,
                    model: "claude-sonnet-5",
                    prompt: "Do work",
                    reasoning_level,
                    request_kind: &session_start_request_kind(),
                },
            )
            .expect("command should build");
            let args = command
                .get_args()
                .map(|arg| arg.to_string_lossy().into_owned())
                .collect::<Vec<_>>();

            // Assert
            let effort_pos = args
                .iter()
                .position(|arg| arg == "--effort")
                .expect("--effort flag should be present");
            assert_eq!(
                args[effort_pos + 1],
                expected_effort,
                "expected effort={expected_effort} for {reasoning_level:?}"
            );
        }
    }

    #[test]
    /// Verifies Claude turns grant filesystem access to pasted-image parent
    /// directories.
    fn test_claude_command_adds_attachment_access_directories() {
        // Arrange
        let temp_directory = tempdir().expect("failed to create temp dir");
        let backend = ClaudeBackend;
        let attachments = vec![
            TurnPromptAttachment {
                placeholder: "[Image #1]".to_string(),
                local_image_path: PathBuf::from("/tmp/agentty/images/one.png"),
            },
            TurnPromptAttachment {
                placeholder: "[Image #2]".to_string(),
                local_image_path: PathBuf::from("/tmp/agentty/images/two.png"),
            },
        ];

        // Act
        let command = AgentBackend::build_command(
            &backend,
            BuildCommandRequest {
                attachments: &attachments,
                folder: temp_directory.path(),
                main_checkout_root: None,
                replay_transcript: None,
                model: "claude-sonnet-5",
                prompt: "Inspect [Image #1] and [Image #2]",
                reasoning_level: ReasoningLevel::default(),
                request_kind: &session_start_request_kind(),
            },
        )
        .expect("command should build");
        let args = command
            .get_args()
            .map(|arg| arg.to_string_lossy().into_owned())
            .collect::<Vec<_>>();

        // Assert
        assert_eq!(
            args.iter()
                .filter(|arg| arg.as_str() == "--add-dir")
                .count(),
            1
        );
        assert!(args.contains(&"/tmp/agentty/images".to_string()));
    }

    #[test]
    /// Verifies Claude prompts include repo-root-relative path guidance.
    fn test_claude_prompt_stdin_payload_includes_repo_root_path_instructions() {
        // Arrange
        let temp_directory = tempdir().expect("failed to create temp dir");

        // Act
        let prompt = String::from_utf8(
            shared_prompt::build_prompt_stdin_payload(
                BuildCommandRequest {
                    attachments: &[],
                    folder: temp_directory.path(),
                    main_checkout_root: None,
                    replay_transcript: None,
                    model: "claude-sonnet-5",
                    prompt: "Plan prompt",
                    reasoning_level: ReasoningLevel::default(),
                    request_kind: &session_start_request_kind(),
                },
                ProtocolSchemaInstructionMode::TransportSchema,
                "Claude",
            )
            .expect("prompt payload should build"),
        )
        .expect("prompt payload should be utf-8");

        // Assert
        assert!(prompt.contains("repository-root-relative POSIX paths"));
        assert!(prompt.contains("Allowed forms: `path`, `path:line`, `path:line:column`."));
        assert!(prompt.contains("summary"));
    }

    #[test]
    /// Verifies one-shot Claude prompts keep protocol JSON guidance while
    /// native schema enforcement carries the full response schema.
    fn test_claude_one_shot_command_enforces_json_schema_without_summary_prose() {
        // Arrange
        let temp_directory = tempdir().expect("failed to create temp dir");
        let backend = ClaudeBackend;

        // Act
        let command = AgentBackend::build_command(
            &backend,
            BuildCommandRequest {
                attachments: &[],
                folder: temp_directory.path(),
                main_checkout_root: None,
                replay_transcript: None,
                model: "claude-sonnet-5",
                prompt: "Generate title",
                reasoning_level: ReasoningLevel::default(),
                request_kind: &utility_request_kind(),
            },
        )
        .expect("command should build");
        let debug_command = format!("{command:?}");
        let prompt = String::from_utf8(
            shared_prompt::build_prompt_stdin_payload(
                BuildCommandRequest {
                    attachments: &[],
                    folder: temp_directory.path(),
                    main_checkout_root: None,
                    replay_transcript: None,
                    model: "claude-sonnet-5",
                    prompt: "Generate title",
                    reasoning_level: ReasoningLevel::default(),
                    request_kind: &utility_request_kind(),
                },
                ProtocolSchemaInstructionMode::TransportSchema,
                "Claude",
            )
            .expect("prompt payload should build"),
        )
        .expect("prompt payload should be utf-8");

        // Assert
        assert!(prompt.contains("Structured response protocol:"));
        assert!(prompt.contains("summary"));
        assert!(!prompt.contains("Authoritative JSON Schema:"));
        assert!(debug_command.contains("--output-format"));
        assert!(debug_command.contains("stream-json"));
        assert!(debug_command.contains("--json-schema"));
        assert!(debug_command.contains("--input-format"));
    }

    #[test]
    /// Verifies structured Claude commands include native JSON schema
    /// validation.
    fn test_claude_start_command_includes_json_schema() {
        // Arrange
        let temp_directory = tempdir().expect("failed to create temp dir");
        let backend = ClaudeBackend;

        // Act
        let command = AgentBackend::build_command(
            &backend,
            BuildCommandRequest {
                attachments: &[],
                folder: temp_directory.path(),
                main_checkout_root: None,
                replay_transcript: None,
                model: "claude-sonnet-5",
                prompt: "Return protocol response",
                reasoning_level: ReasoningLevel::default(),
                request_kind: &session_start_request_kind(),
            },
        )
        .expect("command should build");
        let debug_command = format!("{command:?}");
        let prompt = String::from_utf8(
            shared_prompt::build_prompt_stdin_payload(
                BuildCommandRequest {
                    attachments: &[],
                    folder: temp_directory.path(),
                    main_checkout_root: None,
                    replay_transcript: None,
                    model: "claude-sonnet-5",
                    prompt: "Return protocol response",
                    reasoning_level: ReasoningLevel::default(),
                    request_kind: &session_start_request_kind(),
                },
                ProtocolSchemaInstructionMode::TransportSchema,
                "Claude",
            )
            .expect("prompt payload should build"),
        )
        .expect("prompt payload should be utf-8");

        // Assert
        assert!(debug_command.contains("--json-schema"));
        assert!(debug_command.contains("AgentResponse"));
        assert!(prompt.contains("Structured response protocol:"));
        assert!(prompt.contains("summary"));
        assert!(!prompt.contains("Authoritative JSON Schema:"));
    }
}