orch/engine/runner/
agent.rs

1//! Agent command building + tmux invocation.
2//!
3//! Supports Claude, Codex, OpenCode, Grok (plus Kimi/MiniMax as Claude aliases).
4//! Runs the agent via a runner.sh script executed as the tmux session shell.
5
6use crate::engine::runner::agents::shell_single_quote;
7use crate::template::render_template_str;
8use crate::tmux::TmuxManager;
9use std::collections::BTreeSet;
10use std::collections::HashMap;
11use std::path::PathBuf;
12
13const AGENT_SYSTEM_TEMPLATE: &str = include_str!("../../../prompts/agent_system.md");
14const AGENT_MESSAGE_TEMPLATE: &str = include_str!("../../../prompts/agent_message.md");
15const AGENT_MEMORY_ENTRY_TEMPLATE: &str = include_str!("../../../prompts/agent_memory_entry.md");
16const ALLOWED_TOOLS_TEMPLATE: &str = include_str!("../../../prompts/allowed_tools.md");
17const REVIEW_PROMPT_TEMPLATE: &str = include_str!("../../../prompts/review_task.md");
18const REVIEW_SYSTEM_TEMPLATE: &str = include_str!("../../../prompts/review_system.md");
19
20/// Whether the agent tests through shipcheck: the repo opts in and the shipcheck on PATH
21/// has `commit`. An older shipcheck gets the plain workflow step
22#[derive(Clone, Copy, Debug, PartialEq, Eq)]
23pub enum Shipcheck {
24    Absent,
25    NotInstalled,
26    Outdated,
27    Available,
28}
29
30impl Shipcheck {
31    pub async fn detect(work_dir: &std::path::Path) -> Self {
32        static SUPPORTS_COMMIT: tokio::sync::OnceCell<Option<bool>> =
33            tokio::sync::OnceCell::const_new();
34        let installed = SUPPORTS_COMMIT
35            .get_or_init(|| async {
36                let bin = which::which("shipcheck").ok()?;
37                Some(supports(&bin, "commit").await)
38            })
39            .await;
40        Self::classify(work_dir.join("shipcheck.toml").exists(), *installed)
41    }
42
43    fn classify(has_config: bool, supports_commit: Option<bool>) -> Self {
44        match (has_config, supports_commit) {
45            (false, _) => Shipcheck::Absent,
46            (true, None) => Shipcheck::NotInstalled,
47            (true, Some(false)) => Shipcheck::Outdated,
48            (true, Some(true)) => Shipcheck::Available,
49        }
50    }
51
52    /// What the prompt offers. shipcheck cannot start its own sandbox inside the OS sandbox, so a
53    /// sandboxed agent runs the checks itself and the push hook runs shipcheck outside
54    pub fn for_prompt(self, sandboxed: bool) -> Self {
55        if sandboxed && self == Shipcheck::Available {
56            return Shipcheck::Absent;
57        }
58
59        self
60    }
61
62    fn insert_vars(self, vars: &mut HashMap<String, String>) {
63        let flags: &[&str] = match self {
64            Shipcheck::Available => &["SHIPCHECK"],
65            Shipcheck::NotInstalled => &["NO_SHIPCHECK", "SHIPCHECK_MISSING"],
66            Shipcheck::Absent | Shipcheck::Outdated => &["NO_SHIPCHECK"],
67        };
68        for flag in flags {
69            vars.insert(flag.to_string(), "1".to_string());
70        }
71    }
72}
73
74/// The `orch progress` step appears only when a `cluster:` section turns the cluster on.
75/// Separate flags, because the template engine does not nest `{{#if}}` blocks
76fn insert_cluster_vars(shipcheck: Shipcheck, vars: &mut HashMap<String, String>) {
77    if !crate::cluster::enabled() {
78        return;
79    }
80
81    let commit = if shipcheck == Shipcheck::Available {
82        "CLUSTER_SHIPCHECK"
83    } else {
84        "CLUSTER_GIT"
85    };
86    for flag in ["CLUSTER", commit] {
87        vars.insert(flag.to_string(), "1".to_string());
88    }
89}
90
91async fn run_shipcheck(
92    bin: &std::path::Path,
93    work_dir: &std::path::Path,
94    args: &[&str],
95) -> anyhow::Result<std::process::Output> {
96    let output = tokio::process::Command::new(bin)
97        .args(args)
98        .current_dir(work_dir)
99        .stdin(std::process::Stdio::null())
100        .kill_on_drop(true)
101        .output();
102    Ok(tokio::time::timeout(std::time::Duration::from_secs(30), output).await??)
103}
104
105// `<sub> --help` exits 0 for any word, so read the subcommands the usage line lists
106async fn supports(bin: &std::path::Path, subcommand: &str) -> bool {
107    let Ok(out) = run_shipcheck(bin, &std::env::temp_dir(), &["--help"]).await else {
108        return false;
109    };
110    out.status.success()
111        && String::from_utf8_lossy(&out.stdout)
112            .split_whitespace()
113            .any(|word| word == subcommand)
114}
115
116/// Runs `shipcheck install` so every agent CLI gets shipcheck's hooks in this worktree.
117/// Returns whether install ran. It never fails the task
118pub async fn install_shipcheck_hooks(
119    work_dir: &std::path::Path,
120    shipcheck: Shipcheck,
121    bin: &std::path::Path,
122) -> bool {
123    if shipcheck != Shipcheck::Available {
124        return false;
125    }
126
127    // A shipcheck without `hook` only installs git hooks, and orch skips it
128    if !supports(bin, "hook").await {
129        return false;
130    }
131
132    match run_shipcheck(bin, work_dir, &["install"]).await {
133        Ok(out) if out.status.success() => {}
134        Ok(out) => tracing::warn!(
135            work_dir = %work_dir.display(),
136            stderr = %String::from_utf8_lossy(&out.stderr),
137            "shipcheck install failed"
138        ),
139        Err(e) => tracing::warn!(
140            work_dir = %work_dir.display(),
141            error = %e,
142            "shipcheck install failed"
143        ),
144    }
145    true
146}
147
148const SHIPCHECK_RULE_PATH: &str = ".codex/rules/orch-shipcheck.rules";
149
150// Only the commands the agent prompt uses, never clean, post, login or daemon
151const SHIPCHECK_RULES: &str = r#"prefix_rule(pattern=["shipcheck", "run"], decision="allow")
152prefix_rule(pattern=["shipcheck", "commit"], decision="allow")
153prefix_rule(pattern=["shipcheck", "status"], decision="allow")
154prefix_rule(pattern=["shipcheck", "logs"], decision="allow")
155prefix_rule(pattern=["shipcheck", "doctor"], decision="allow")
156"#;
157
158// Codex runs a command a rule allows outside its sandbox, so shipcheck reaches Docker and its
159// results dir while the agent's own commands still cannot write there
160async fn install_shipcheck_rule(work_dir: &std::path::Path) -> anyhow::Result<()> {
161    // The worktree holds repository content, so never follow a symlink it plants on the path
162    let mut dir = work_dir.to_path_buf();
163    for part in [".codex", "rules"] {
164        dir.push(part);
165        match tokio::fs::symlink_metadata(&dir).await {
166            Ok(meta) if meta.is_dir() => {}
167            Ok(_) => anyhow::bail!("{} is not a directory", dir.display()),
168            Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
169                tokio::fs::create_dir(&dir).await?;
170            }
171            Err(e) => return Err(e.into()),
172        }
173    }
174
175    // remove_file drops a symlink itself, and create_new never follows one
176    let rule = work_dir.join(SHIPCHECK_RULE_PATH);
177    match tokio::fs::remove_file(&rule).await {
178        Err(e) if e.kind() != std::io::ErrorKind::NotFound => return Err(e.into()),
179        _ => {}
180    }
181
182    let mut file = tokio::fs::OpenOptions::new()
183        .write(true)
184        .create_new(true)
185        .open(&rule)
186        .await?;
187    tokio::io::AsyncWriteExt::write_all(&mut file, SHIPCHECK_RULES.as_bytes()).await?;
188
189    // Keep the rule out of the agent's commits
190    let output = tokio::process::Command::new("git")
191        .args(["rev-parse", "--git-path", "info/exclude"])
192        .current_dir(work_dir)
193        .output()
194        .await?;
195    if !output.status.success() {
196        return Ok(());
197    }
198
199    let exclude = work_dir.join(String::from_utf8_lossy(&output.stdout).trim());
200    let line = format!("/{SHIPCHECK_RULE_PATH}");
201    let current = tokio::fs::read_to_string(&exclude)
202        .await
203        .unwrap_or_default();
204    if current.lines().any(|l| l == line) {
205        return Ok(());
206    }
207
208    if let Some(dir) = exclude.parent() {
209        tokio::fs::create_dir_all(dir).await?;
210    }
211
212    let sep = if current.is_empty() || current.ends_with('\n') {
213        ""
214    } else {
215        "\n"
216    };
217    tokio::fs::write(&exclude, format!("{current}{sep}{line}\n")).await?;
218    Ok(())
219}
220
221/// Run `git rev-parse <arg>` from `work_dir` and canonicalize the
222/// result. Returns `Ok(None)` when git is not available or the command fails.
223pub(crate) async fn resolve_git_dir_arg(
224    work_dir: &std::path::Path,
225    arg: &str,
226) -> anyhow::Result<Option<PathBuf>> {
227    let output = tokio::process::Command::new("git")
228        .args(["rev-parse", arg])
229        .current_dir(work_dir)
230        .output()
231        .await?;
232
233    if !output.status.success() {
234        return Ok(None);
235    }
236
237    let raw = String::from_utf8_lossy(&output.stdout).trim().to_string();
238    if raw.is_empty() {
239        return Ok(None);
240    }
241
242    // The path may be relative to work_dir; canonicalize to absolute.
243    let candidate = if std::path::Path::new(&raw).is_absolute() {
244        PathBuf::from(&raw)
245    } else {
246        work_dir.join(&raw)
247    };
248
249    let canonical = tokio::fs::canonicalize(&candidate)
250        .await
251        .unwrap_or(candidate);
252    Ok(Some(canonical))
253}
254
255fn render_prompt_template(template: &str, vars: HashMap<String, String>) -> String {
256    match render_template_str(template, &vars) {
257        Ok(rendered) => rendered,
258        Err(err) => {
259            tracing::error!(error = %err, "failed to render prompt template");
260            template.to_string()
261        }
262    }
263}
264
265/// Agent invocation configuration.
266pub struct AgentInvocation {
267    /// Agent name (claude, codex, opencode, grok, kimi, minimax)
268    pub agent: String,
269    /// Model override (e.g., "claude-sonnet-4-6", "o3", "gpt-4.1")
270    pub model: Option<String>,
271    /// Working directory
272    pub work_dir: PathBuf,
273    /// System prompt content
274    pub system_prompt: String,
275    /// Agent message (task prompt)
276    pub agent_message: String,
277    /// Task ID
278    pub task_id: String,
279    /// Disallowed tools pattern
280    pub disallowed_tools: Vec<String>,
281    /// Git author name
282    pub git_author_name: String,
283    /// Git author email
284    pub git_author_email: String,
285    /// Output file path for agent response
286    pub output_file: PathBuf,
287    /// Timeout in seconds (0 = no timeout)
288    pub timeout_seconds: u64,
289    /// Repository slug (owner/repo) for per-repo state isolation
290    pub repo: String,
291    /// Whether the agent runs in the OS sandbox, from `workflow.sandbox`
292    pub sandbox: bool,
293    /// The project's main checkout, which the sandbox keeps read-only but for its `.git`
294    pub main_project_dir: PathBuf,
295    /// The task's branch, from orch's records, the one ref the sandbox lets the agent move
296    pub branch: String,
297    /// Current attempt number (1-indexed)
298    pub attempt: u32,
299}
300
301/// Build the runner script content (saved to attempt dir), plus the small
302/// helper script it execs the agent command through under a pty. Returns
303/// `(runner_script, agent_cmd_script)`.
304fn build_tmux_runner_scripts(
305    inv: &AgentInvocation,
306    agent_cmd: &str,
307    attempt_dir: &std::path::Path,
308    sandbox: Option<&[std::ffi::OsString]>,
309) -> anyhow::Result<(String, String)> {
310    let status_file = attempt_dir.join("exit.txt");
311    let stderr_file = attempt_dir.join("stderr.txt");
312    let agent_cmd_script_path = attempt_dir.join("agent_cmd.sh");
313    let sq_git_name = shell_single_quote(&inv.git_author_name);
314    let sq_git_email = shell_single_quote(&inv.git_author_email);
315    let sq_task_id = shell_single_quote(&inv.task_id);
316    let sq_output_file = shell_single_quote(&inv.output_file.display().to_string());
317    let sq_work_dir = shell_single_quote(&inv.work_dir.display().to_string());
318    let sq_status_file = shell_single_quote(&status_file.display().to_string());
319    let sq_stderr_file = shell_single_quote(&stderr_file.display().to_string());
320    let sq_agent_cmd_script = shell_single_quote(&agent_cmd_script_path.display().to_string());
321
322    // Runs the agent with stderr split off to a real file before the pty layer
323    // below merges stdout+stderr, exactly as the previous plain-pipe invocation did.
324    let agent_cmd = match sandbox {
325        Some(argv) => {
326            let prefix: Vec<String> = argv
327                .iter()
328                .map(|a| shell_single_quote(&a.to_string_lossy()))
329                .collect();
330            // The tmux and ssh-agent sockets are out of reach, and their variables go too
331            let tmpdir = shell_single_quote(
332                &super::sandbox::run_temp_dir(attempt_dir)
333                    .display()
334                    .to_string(),
335            );
336            format!(
337                "env -u TMUX -u TMUX_PANE -u SSH_AUTH_SOCK TMPDIR={tmpdir} {} /bin/bash -c {}",
338                prefix.join(" "),
339                shell_single_quote(agent_cmd)
340            )
341        }
342        None => agent_cmd.to_string(),
343    };
344    let agent_cmd_script = format!("#!/usr/bin/env bash\n{agent_cmd} 2>{sq_stderr_file}\n");
345
346    let runner_script = format!(
347        r#"#!/usr/bin/env bash
348set -euo pipefail
349
350status_file={sq_status_file}
351stderr_file={sq_stderr_file}
352trap 'status=$?; printf "%s\n" "$status" > "$status_file"' EXIT
353
354# Environment — ~/.path and ~/.private are loaded by orch at startup into the process env
355export PATH="/opt/homebrew/bin:/usr/local/bin:$PATH"
356export GIT_AUTHOR_NAME={sq_git_name}
357export GIT_COMMITTER_NAME={sq_git_name}
358export GIT_AUTHOR_EMAIL={sq_git_email}
359export GIT_COMMITTER_EMAIL={sq_git_email}
360# Env config overrides the operator's global commit.gpgsign, the signing agent
361# (1Password etc) is often locked in unattended sessions and blocks `git commit`
362export GIT_CONFIG_COUNT="1"
363export GIT_CONFIG_KEY_0="commit.gpgsign"
364export GIT_CONFIG_VALUE_0="false"
365export TASK_ID={sq_task_id}
366export OUTPUT_FILE={sq_output_file}
367unset CLAUDECODE  # allow nested claude invocations from orch
368
369cd {sq_work_dir} || {{
370    printf '%s\n' "worktree directory does not exist: {sq_work_dir}" > "$stderr_file"
371    exit 1
372}}
373
374# Run agent under a pty (via `script`) so its stdout is a tty, not a plain pipe —
375# many CLI runtimes silently switch to full block-buffering when stdout isn't a
376# tty, which can delay pane output long enough to trip silence detection even
377# while the agent is actively streaming. Tee to both output file and terminal
378# (tmux pane) for live streaming.
379set +e
380script -q /dev/null /bin/bash {sq_agent_cmd_script} | tee {sq_output_file}
381CMD_STATUS=${{PIPESTATUS[0]:-0}}
382set -e
383
384exit $CMD_STATUS
385"#,
386    );
387
388    Ok((runner_script, agent_cmd_script))
389}
390
391/// Spawn the agent in a tmux session.
392///
393/// Returns the tmux session name.
394pub async fn spawn_in_tmux(tmux: &TmuxManager, inv: &AgentInvocation) -> anyhow::Result<String> {
395    // Prepare attempt directory and prompt files (system + message).
396    let attempt_dir =
397        crate::home::task_attempt_dir_async(&inv.repo, &inv.task_id, inv.attempt).await?;
398
399    let sys_file = attempt_dir.join("prompt-sys.md");
400    let msg_file = attempt_dir.join("prompt-msg.md");
401    // Build unified permission rules and sys content
402    let mut permissions = super::agents::PermissionRules::from_config();
403    if !inv.disallowed_tools.is_empty() {
404        for tool in &inv.disallowed_tools {
405            if !permissions.disallowed_tools.contains(tool) {
406                permissions.disallowed_tools.push(tool.clone());
407            }
408        }
409    }
410    permissions.allowed_edit_paths.push(inv.work_dir.clone());
411
412    // Codex cannot start its own Seatbelt inside the OS sandbox, so the OS sandbox replaces it
413    let sandbox = if inv.sandbox {
414        permissions.sandbox = super::agents::SandboxLevel::Os;
415        let temp = super::sandbox::run_temp_dir(&attempt_dir);
416        tokio::fs::create_dir_all(&temp).await?;
417
418        // The policy makes only paths that exist read-only, so the files orch writes later in the
419        // attempt dir must exist now, or the agent could create them first
420        for name in [
421            "runner.sh",
422            "agent_cmd.sh",
423            "exit.txt",
424            "stderr.txt",
425            "prompt-sys.md",
426            "prompt-msg.md",
427            "result.json",
428        ] {
429            let path = attempt_dir.join(name);
430
431            if !path.exists() {
432                tokio::fs::write(&path, "").await?;
433            }
434        }
435
436        if !inv.output_file.exists() && inv.output_file.parent().is_some_and(|p| p.is_dir()) {
437            tokio::fs::write(&inv.output_file, "").await?;
438        }
439
440        Some(super::sandbox::argv(inv, &temp).await?)
441    } else {
442        None
443    };
444
445    // For Codex running in workspace-write sandbox mode, grant write access to
446    // git metadata dirs when they live outside the worktree root. We include
447    // both --git-common-dir and --git-dir because index.lock can be created
448    // under the worktree-specific git dir.
449    if inv.agent == "codex" {
450        let mut extra_dirs = BTreeSet::new();
451        for (arg, label) in [
452            ("--git-common-dir", "git common dir"),
453            ("--git-dir", "git dir"),
454        ] {
455            match resolve_git_dir_arg(&inv.work_dir, arg).await {
456                Ok(Some(dir)) if !dir.starts_with(&inv.work_dir) => {
457                    extra_dirs.insert(dir);
458                }
459                Ok(_) => {}
460                Err(e) => {
461                    tracing::debug!(
462                        work_dir = %inv.work_dir.display(),
463                        arg,
464                        error = %e,
465                        "failed to resolve {}; skipping --add-dir",
466                        label
467                    );
468                }
469            }
470        }
471        permissions.extra_writable_dirs.extend(extra_dirs);
472
473        if Shipcheck::detect(&inv.work_dir).await == Shipcheck::Available {
474            if let Err(e) = install_shipcheck_rule(&inv.work_dir).await {
475                tracing::warn!(
476                    work_dir = %inv.work_dir.display(),
477                    error = %e,
478                    "failed to install the shipcheck codex rule"
479                );
480            }
481        }
482    }
483
484    let sys_content = if !permissions.allowed_tools.is_empty() {
485        let tools_list = permissions
486            .allowed_tools
487            .iter()
488            .map(|t| format!("- {t}"))
489            .collect::<Vec<_>>()
490            .join("\n");
491        let mut vars = HashMap::new();
492        vars.insert("TOOLS_LIST".to_string(), tools_list);
493        let tools_prompt = render_prompt_template(ALLOWED_TOOLS_TEMPLATE, vars);
494        format!("{}\n\n{}", inv.system_prompt, tools_prompt)
495    } else {
496        inv.system_prompt.clone()
497    };
498
499    tokio::fs::write(&sys_file, &sys_content).await?;
500    tokio::fs::write(&msg_file, &inv.agent_message).await?;
501
502    // Build agent command using per-agent runner (used in non-PTY path below)
503    let runner = super::agents::get_runner(&inv.agent);
504
505    // Prepare environment map for tmux session (does not write GH token to disk)
506    let mut env = std::collections::HashMap::new();
507    env.insert("GIT_AUTHOR_NAME".to_string(), inv.git_author_name.clone());
508    env.insert(
509        "GIT_COMMITTER_NAME".to_string(),
510        inv.git_author_name.clone(),
511    );
512    env.insert("GIT_AUTHOR_EMAIL".to_string(), inv.git_author_email.clone());
513    env.insert(
514        "GIT_COMMITTER_EMAIL".to_string(),
515        inv.git_author_email.clone(),
516    );
517    env.insert("TASK_ID".to_string(), inv.task_id.clone());
518    env.insert(
519        "OUTPUT_FILE".to_string(),
520        inv.output_file.display().to_string(),
521    );
522
523    // Write runner script and execute it in a detached tmux session.
524    let timeout_cmd = if inv.timeout_seconds > 0 {
525        format!("timeout {}", inv.timeout_seconds)
526    } else {
527        String::new()
528    };
529    let agent_cmd = runner.build_command(
530        inv.model.as_deref(),
531        &timeout_cmd,
532        &sys_file.to_string_lossy(),
533        &msg_file.to_string_lossy(),
534        &permissions,
535    );
536
537    // Write runner script (and the agent-cmd helper script it execs under a pty)
538    // to the per-task attempt dir.
539    let script_path = attempt_dir.join("runner.sh");
540    let agent_cmd_script_path = attempt_dir.join("agent_cmd.sh");
541    let (script_content, agent_cmd_script_content) =
542        build_tmux_runner_scripts(inv, &agent_cmd, &attempt_dir, sandbox.as_deref())?;
543    tokio::fs::write(&agent_cmd_script_path, &agent_cmd_script_content).await?;
544    tokio::fs::write(&script_path, &script_content).await?;
545
546    // Make executable
547    #[cfg(unix)]
548    {
549        use std::os::unix::fs::PermissionsExt;
550        tokio::fs::set_permissions(&script_path, std::fs::Permissions::from_mode(0o755)).await?;
551        tokio::fs::set_permissions(
552            &agent_cmd_script_path,
553            std::fs::Permissions::from_mode(0o755),
554        )
555        .await?;
556    }
557
558    let command = format!("bash \"{}\"", script_path.display());
559
560    // Resolve GitHub token via the process-wide shared resolver (cached after first call)
561    let github_token = match crate::github::token::shared().get_token().await {
562        Ok(Some(token)) => {
563            tracing::debug!("Resolved GitHub token via TokenResolver for agent session");
564            Some(token)
565        }
566        Ok(None) => {
567            tracing::warn!(
568                "No GitHub token available; set GH_TOKEN, GITHUB_TOKEN, or gh.auth.token, or run gh auth login"
569            );
570            None
571        }
572        Err(e) => {
573            tracing::warn!(error = %e, "Failed to resolve GitHub token");
574            None
575        }
576    };
577
578    // Inject GH_TOKEN into env so it's passed via `tmux new-session -e` along
579    // with everything else. The `-e` flag sets variables in the *initial* pane
580    // environment, so the agent process sees it immediately on startup.
581    //
582    // IMPORTANT: Do NOT use create_session_detached + set_env + create_window
583    // to inject GH_TOKEN. That approach creates a 2-window session (window 0 =
584    // idle shell, window 1 = agent). All tmux monitoring (is_session_active,
585    // capture_pane, batch_session_active) targets the session without a window
586    // index, which defaults to the *active* window — the idle shell. This
587    // breaks completion detection (shell never exits → pane_dead always 0),
588    // output capture (reads the shell prompt, not agent output), and silence
589    // detection. See the revert of PR #2903.
590    if let Some(ref token) = github_token {
591        env.insert("GH_TOKEN".to_string(), token.clone());
592    }
593
594    let env_vec: Vec<(&str, &str)> = env.iter().map(|(k, v)| (k.as_str(), v.as_str())).collect();
595
596    let session = tmux
597        .create_session(
598            &inv.repo,
599            &inv.task_id,
600            &inv.work_dir.to_string_lossy(),
601            &command,
602            env_vec.as_slice(),
603        )
604        .await?;
605
606    tracing::info!(
607        task_id = inv.task_id,
608        agent = inv.agent,
609        session = %session,
610        "agent spawned in tmux"
611    );
612
613    Ok(session)
614}
615
616/// Build the system prompt for the agent.
617pub fn build_system_prompt(
618    _task: &crate::backends::ExternalTask,
619    context: &super::context::TaskContext,
620    route_result: Option<&crate::engine::router::RouteResult>,
621    default_branch: &str,
622    shipcheck: Shipcheck,
623) -> String {
624    let mut vars = HashMap::new();
625
626    vars.insert("DEFAULT_BRANCH".to_string(), default_branch.to_string());
627    shipcheck.insert_vars(&mut vars);
628    insert_cluster_vars(shipcheck, &mut vars);
629
630    if let Some(rr) = route_result {
631        vars.insert("ROLE".to_string(), rr.profile.role.clone());
632
633        if !rr.profile.constraints.is_empty() {
634            let constraints = rr
635                .profile
636                .constraints
637                .iter()
638                .map(|c| format!("- {c}"))
639                .collect::<Vec<_>>()
640                .join("\n");
641            vars.insert("CONSTRAINTS".to_string(), constraints);
642        }
643    }
644
645    if !context.project_instructions.is_empty() {
646        vars.insert(
647            "PROJECT_INSTRUCTIONS".to_string(),
648            context.project_instructions.clone(),
649        );
650    }
651
652    if !context.skills_docs.is_empty() {
653        vars.insert("SKILLS_DOCS".to_string(), context.skills_docs.clone());
654    }
655
656    if !context.repo_tree.is_empty() {
657        vars.insert("REPO_TREE".to_string(), context.repo_tree.clone());
658    }
659
660    render_prompt_template(AGENT_SYSTEM_TEMPLATE, vars)
661}
662
663/// Build the agent message (task prompt).
664pub fn build_agent_message(
665    task: &crate::backends::ExternalTask,
666    context: &super::context::TaskContext,
667    attempts: u32,
668) -> String {
669    let mut vars = HashMap::new();
670    vars.insert("TASK_ID".to_string(), task.id.0.clone());
671    vars.insert("TASK_TITLE".to_string(), task.title.clone());
672    vars.insert("TASK_BODY".to_string(), task.body.clone());
673
674    if !context.task_context.is_empty() {
675        vars.insert("TASK_CONTEXT".to_string(), context.task_context.clone());
676    }
677
678    if !context.parent_context.is_empty() {
679        vars.insert("PARENT_CONTEXT".to_string(), context.parent_context.clone());
680    }
681
682    if !context.issue_comments.is_empty() {
683        vars.insert("ISSUE_COMMENTS".to_string(), context.issue_comments.clone());
684    }
685
686    if !context.pr_review_context.is_empty() {
687        vars.insert(
688            "PR_REVIEW_CONTEXT".to_string(),
689            context.pr_review_context.clone(),
690        );
691    }
692
693    if attempts > 0 && !context.git_diff.is_empty() {
694        vars.insert("GIT_DIFF".to_string(), context.git_diff.clone());
695    }
696
697    if attempts > 0 {
698        vars.insert("ATTEMPT_NUMBER".to_string(), (attempts + 1).to_string());
699    }
700
701    if !context.memory.is_empty() {
702        let mut sections = Vec::new();
703        for entry in &context.memory {
704            let mut entry_vars = HashMap::new();
705            entry_vars.insert("ATTEMPT".to_string(), entry.attempt.to_string());
706            entry_vars.insert("AGENT".to_string(), entry.agent.clone());
707
708            if let Some(ref model) = entry.model {
709                entry_vars.insert("MODEL".to_string(), model.clone());
710            }
711
712            if !entry.approach.is_empty() {
713                entry_vars.insert("APPROACH".to_string(), entry.approach.clone());
714            }
715
716            if !entry.learnings.is_empty() {
717                let learnings = entry
718                    .learnings
719                    .iter()
720                    .map(|learning| format!("- {}", learning))
721                    .collect::<Vec<_>>()
722                    .join("\n");
723                entry_vars.insert("LEARNINGS".to_string(), learnings);
724            }
725
726            if let Some(ref error) = entry.error {
727                entry_vars.insert("ERROR".to_string(), error.clone());
728            }
729
730            if !entry.files_modified.is_empty() {
731                entry_vars.insert(
732                    "FILES_MODIFIED".to_string(),
733                    entry.files_modified.join(", "),
734                );
735            }
736
737            let rendered = render_prompt_template(AGENT_MEMORY_ENTRY_TEMPLATE, entry_vars);
738            sections.push(rendered.trim().to_string());
739        }
740
741        let memory_section = sections.join("\n\n");
742        vars.insert("MEMORY_SECTION".to_string(), memory_section);
743    }
744
745    render_prompt_template(AGENT_MESSAGE_TEMPLATE, vars)
746}
747
748/// Build the review prompt for the review agent.
749///
750/// This is called when a task completes and the review agent needs to
751/// review the changes before auto-merge.
752pub fn build_review_prompt(
753    task: &crate::backends::ExternalTask,
754    agent_summary: &str,
755    git_diff: &str,
756    git_log: &str,
757    default_branch: &str,
758    pr_number: u64,
759    shipcheck: Shipcheck,
760) -> String {
761    let mut vars = HashMap::new();
762    shipcheck.insert_vars(&mut vars);
763    insert_cluster_vars(shipcheck, &mut vars);
764    vars.insert("TASK_ID".to_string(), task.id.0.clone());
765    vars.insert("TASK_TITLE".to_string(), task.title.clone());
766    vars.insert("TASK_BODY".to_string(), task.body.clone());
767    vars.insert("DEFAULT_BRANCH".to_string(), default_branch.to_string());
768    vars.insert("PR_NUMBER".to_string(), pr_number.to_string());
769
770    if !agent_summary.is_empty() {
771        vars.insert("AGENT_SUMMARY".to_string(), agent_summary.to_string());
772    }
773
774    if !git_diff.is_empty() {
775        vars.insert("GIT_DIFF".to_string(), git_diff.to_string());
776    }
777
778    if !git_log.is_empty() {
779        vars.insert("GIT_LOG".to_string(), git_log.to_string());
780    }
781
782    render_prompt_template(REVIEW_PROMPT_TEMPLATE, vars)
783}
784
785/// Build the runner shell script contents for testing and inspection.
786///
787/// IMPORTANT: this must not embed secrets (GH_TOKEN) into the script.
788#[allow(dead_code)]
789pub fn build_runner_script(inv: &AgentInvocation) -> anyhow::Result<String> {
790    // Prepare attempt directory and prompt files
791    let attempt_dir = crate::home::task_attempt_dir(&inv.repo, &inv.task_id, inv.attempt)?;
792    std::fs::create_dir_all(&attempt_dir)?;
793
794    let sys_file = attempt_dir.join("prompt-sys.md");
795    let msg_file = attempt_dir.join("prompt-msg.md");
796    std::fs::write(&sys_file, &inv.system_prompt)?;
797    std::fs::write(&msg_file, &inv.agent_message)?;
798
799    build_runner_script_in_dir(inv, &attempt_dir)
800}
801
802/// Inner script builder given a pre-created attempt directory.
803/// Separated from `build_runner_script` so tests can pass a temp dir and
804/// avoid racing with the parallel binary/lib test that shares `~/.orch/state/`.
805fn build_runner_script_in_dir(
806    inv: &AgentInvocation,
807    attempt_dir: &std::path::Path,
808) -> anyhow::Result<String> {
809    let sys_file = attempt_dir.join("prompt-sys.md");
810    let msg_file = attempt_dir.join("prompt-msg.md");
811
812    // Build unified permission rules and sys content (minimal translation)
813    let mut permissions = super::agents::PermissionRules::from_config();
814    if !inv.disallowed_tools.is_empty() {
815        for tool in &inv.disallowed_tools {
816            if !permissions.disallowed_tools.contains(tool) {
817                permissions.disallowed_tools.push(tool.clone());
818            }
819        }
820    }
821    permissions.allowed_edit_paths.push(inv.work_dir.clone());
822
823    // Build agent command using per-agent runner
824    let runner = super::agents::get_runner(&inv.agent);
825    let timeout_cmd = if inv.timeout_seconds > 0 {
826        format!("timeout {}", inv.timeout_seconds)
827    } else {
828        String::new()
829    };
830    let agent_cmd = runner.build_command(
831        inv.model.as_deref(),
832        &timeout_cmd,
833        &sys_file.to_string_lossy(),
834        &msg_file.to_string_lossy(),
835        &permissions,
836    );
837
838    let status_file = attempt_dir.join("exit.txt");
839    let stderr_file = attempt_dir.join("stderr.txt");
840
841    let sq_status = shell_single_quote(&status_file.display().to_string());
842    let sq_stderr = shell_single_quote(&stderr_file.display().to_string());
843    let sq_out = shell_single_quote(&inv.output_file.display().to_string());
844    let sq_work_dir = shell_single_quote(&inv.work_dir.to_string_lossy());
845    let script = format!(
846        "#!/usr/bin/env bash\nset -euo pipefail\nstatus_file={sq_status}\nstderr_file={sq_stderr}\ntrap 'status=$?; printf \"%s\\n\" \"$status\" > \"$status_file\"' EXIT\nunset CLAUDECODE\ncd {sq_work_dir} || {{\n  printf '%s\\n' \"worktree directory does not exist: {sq_work_dir}\" > \"$stderr_file\"\n  exit 1\n}}\nset +e\nRESPONSE=$({agent_cmd} 2> \"$stderr_file\")\nCMD_STATUS=$?\nset -e\nprintf '%s' \"$RESPONSE\" > {sq_out}\nexit $CMD_STATUS\n",
847        agent_cmd = agent_cmd,
848    );
849
850    Ok(script)
851}
852
853/// Build the system prompt for the review agent.
854pub fn review_system_prompt() -> String {
855    render_prompt_template(REVIEW_SYSTEM_TEMPLATE, HashMap::new())
856}
857
858#[cfg(test)]
859mod tests {
860    use super::*;
861    use std::path::PathBuf;
862
863    #[test]
864    fn review_system_prompt_checks_stray_files_and_trust_boundaries() {
865        let p = review_system_prompt();
866        assert!(p.contains("**Files**"));
867        assert!(p.contains("**Trust boundaries**"));
868    }
869
870    fn test_invocation(task_id: &str) -> AgentInvocation {
871        AgentInvocation {
872            agent: "claude".to_string(),
873            model: None,
874            work_dir: PathBuf::from("/tmp"),
875            system_prompt: "test system prompt".to_string(),
876            agent_message: "test message".to_string(),
877            task_id: task_id.to_string(),
878            disallowed_tools: vec![],
879            git_author_name: "Test Bot".to_string(),
880            git_author_email: "bot@example.com".to_string(),
881            output_file: PathBuf::from("/tmp/test-output.json"),
882            timeout_seconds: 0,
883            repo: "orch-runner-test/token-check".to_string(),
884            attempt: 1,
885            sandbox: false,
886            main_project_dir: std::path::PathBuf::from("/tmp"),
887            branch: String::new(),
888        }
889    }
890
891    #[test]
892    fn spawn_in_tmux_rejects_invalid_env() {
893        // This test verifies the function signature and basic behavior
894        // without actually spawning a tmux session.
895        // The key security guarantee: tokens are passed via tmux -e flag,
896        // not written to any file on disk.
897        let inv = test_invocation("env-test");
898        // Just verify the invocation can be created without panicking
899        assert_eq!(inv.task_id, "env-test");
900    }
901
902    #[test]
903    fn build_runner_script_persists_exit_status_on_failure() {
904        // Use a temp dir so lib and bin compilation units don't race on ~/.orch/state/.
905        let tmp = tempfile::TempDir::new().expect("temp dir");
906        let inv = test_invocation("runner-status-test");
907        let script =
908            build_runner_script_in_dir(&inv, tmp.path()).expect("runner script should build");
909
910        assert!(script
911            .contains("trap 'status=$?; printf \"%s\\n\" \"$status\" > \"$status_file\"' EXIT"));
912        assert!(script.contains("set +e"));
913    }
914
915    #[test]
916    fn build_runner_script_reports_missing_worktree() {
917        // Use a temp dir so lib and bin compilation units don't race on ~/.orch/state/.
918        let tmp = tempfile::TempDir::new().expect("temp dir");
919        let inv = test_invocation("runner-missing-worktree-test");
920        let script =
921            build_runner_script_in_dir(&inv, tmp.path()).expect("runner script should build");
922
923        assert!(script.contains("worktree directory does not exist: '/tmp'"));
924        assert!(script.contains("stderr_file='"));
925    }
926
927    #[test]
928    fn tmux_runner_script_runs_agent_under_pty_not_a_plain_pipe() {
929        // Regression test for #3554: piping the agent binary's stdout straight
930        // into `tee` leaves it non-tty, which makes many CLI runtimes switch to
931        // full block-buffering and can starve the tmux pane long enough to trip
932        // silence detection even while the agent is actively working. The agent
933        // command must run under `script` (a real pty) before hitting `tee`.
934        let tmp = tempfile::TempDir::new().expect("temp dir");
935        let inv = test_invocation("pty-wrap-test");
936        let (runner_script, agent_cmd_script) =
937            build_tmux_runner_scripts(&inv, "echo hello", tmp.path(), None)
938                .expect("runner scripts should build");
939
940        assert!(
941            runner_script.contains("script -q /dev/null /bin/bash"),
942            "agent must be exec'd under a pty via `script`, got: {runner_script}"
943        );
944        assert!(
945            runner_script.contains("| tee"),
946            "output must still be tee'd to the output file, got: {runner_script}"
947        );
948        assert!(
949            !runner_script.contains("echo hello 2>"),
950            "the raw agent command must not be piped directly into tee anymore, got: {runner_script}"
951        );
952        assert!(
953            agent_cmd_script.contains("echo hello 2>"),
954            "the agent command should run (with stderr split to a file) inside the pty-wrapped helper script, got: {agent_cmd_script}"
955        );
956    }
957
958    #[test]
959    fn tmux_runner_script_captures_agent_exit_status_via_pipestatus() {
960        let tmp = tempfile::TempDir::new().expect("temp dir");
961        let inv = test_invocation("pty-exit-status-test");
962        let (runner_script, _) = build_tmux_runner_scripts(&inv, "echo hello", tmp.path(), None)
963            .expect("runner scripts should build");
964
965        assert!(runner_script.contains("CMD_STATUS=${PIPESTATUS[0]:-0}"));
966        assert!(runner_script.contains("exit $CMD_STATUS"));
967    }
968
969    #[test]
970    fn tmux_runner_script_disables_commit_signing() {
971        // Locked signing agent must not fail `git commit` inside agent sessions
972        let tmp = tempfile::TempDir::new().expect("temp dir");
973        let inv = test_invocation("gpgsign-test");
974        let (runner_script, _) = build_tmux_runner_scripts(&inv, "echo hello", tmp.path(), None)
975            .expect("runner scripts should build");
976
977        assert!(runner_script.contains("GIT_CONFIG_KEY_0=\"commit.gpgsign\""));
978        assert!(runner_script.contains("GIT_CONFIG_VALUE_0=\"false\""));
979    }
980
981    #[test]
982    fn tmux_runner_script_writes_executable_agent_cmd_helper() {
983        let tmp = tempfile::TempDir::new().expect("temp dir");
984        let inv = test_invocation("pty-helper-path-test");
985        let (runner_script, agent_cmd_script) =
986            build_tmux_runner_scripts(&inv, "echo hello", tmp.path(), None)
987                .expect("runner scripts should build");
988
989        let agent_cmd_script_path = tmp.path().join("agent_cmd.sh");
990        assert!(
991            runner_script.contains(&agent_cmd_script_path.display().to_string()),
992            "runner script must reference the agent_cmd.sh helper path, got: {runner_script}"
993        );
994        assert!(agent_cmd_script.starts_with("#!/usr/bin/env bash\n"));
995    }
996
997    #[test]
998    fn tmux_runner_script_runs_the_agent_inside_the_sandbox() {
999        let tmp = tempfile::TempDir::new().expect("temp dir");
1000        let inv = test_invocation("sandbox-test");
1001        let argv: Vec<std::ffi::OsString> = vec![
1002            "/usr/bin/sandbox-exec".into(),
1003            "-p".into(),
1004            "(version 1)".into(),
1005        ];
1006        let (_, agent_cmd_script) =
1007            build_tmux_runner_scripts(&inv, "claude -p 'hi'", tmp.path(), Some(&argv))
1008                .expect("runner scripts should build");
1009        assert!(
1010            agent_cmd_script.contains(
1011                "env -u TMUX -u TMUX_PANE -u SSH_AUTH_SOCK TMPDIR='"
1012            ) && agent_cmd_script.contains(
1013                "' '/usr/bin/sandbox-exec' '-p' '(version 1)' /bin/bash -c 'claude -p '\\''hi'\\''' 2>"
1014            ),
1015            "the agent must run inside the sandbox with stderr split outside it, got: {agent_cmd_script}"
1016        );
1017        assert!(
1018            agent_cmd_script.contains(&format!(
1019                "TMPDIR='{}'",
1020                crate::engine::runner::sandbox::run_temp_dir(tmp.path()).display()
1021            )),
1022            "the agent must get its own TMPDIR, got: {agent_cmd_script}"
1023        );
1024    }
1025
1026    #[test]
1027    fn a_sandboxed_prompt_offers_no_shipcheck() {
1028        assert_eq!(Shipcheck::Available.for_prompt(true), Shipcheck::Absent);
1029        assert_eq!(Shipcheck::Available.for_prompt(false), Shipcheck::Available);
1030        assert_eq!(
1031            Shipcheck::NotInstalled.for_prompt(true),
1032            Shipcheck::NotInstalled
1033        );
1034    }
1035
1036    #[test]
1037    fn build_runner_script_escapes_shell_injection_in_paths() {
1038        let tmp = tempfile::TempDir::new().expect("temp dir");
1039        let mut inv = test_invocation("injection-test");
1040        inv.work_dir = std::path::PathBuf::from("/tmp/$(whoami)/workspace");
1041        inv.output_file = std::path::PathBuf::from("/tmp/evil`id`.json");
1042        let script =
1043            build_runner_script_in_dir(&inv, tmp.path()).expect("runner script should build");
1044
1045        // Values must be single-quoted, neutralizing $() and backticks
1046        assert!(
1047            script.contains("'/tmp/$(whoami)/workspace'"),
1048            "work_dir with $() must be single-quoted: {script}"
1049        );
1050        assert!(
1051            script.contains("'/tmp/evil`id`.json'"),
1052            "output_file with backticks must be single-quoted: {script}"
1053        );
1054    }
1055
1056    #[test]
1057    fn review_prompt_forbids_ending_turn_on_pending_ci_status() {
1058        // Regression test for #3453: a review agent must not stop with a
1059        // prose "CI is still running" status update and leave the turn
1060        // without JSON. The prompt must explicitly route PENDING/NOT RUN CI
1061        // through the local-checks fallback and into the same-turn JSON output.
1062        let task = crate::backends::ExternalTask {
1063            id: crate::backends::ExternalId("155460".to_string()),
1064            title: "test task".to_string(),
1065            body: "test body".to_string(),
1066            state: "open".to_string(),
1067            labels: vec![],
1068            author: "tester".to_string(),
1069            created_at: "2026-01-01T00:00:00Z".to_string(),
1070            updated_at: "2026-01-01T00:00:00Z".to_string(),
1071            url: "https://example.com/1".to_string(),
1072        };
1073        let prompt = build_review_prompt(&task, "", "", "", "main", 1, Shipcheck::Absent);
1074
1075        assert!(
1076            prompt.contains("Pending CI is never a terminal review state"),
1077            "prompt must state pending CI is not terminal: {prompt}"
1078        );
1079        assert!(
1080            prompt.contains("Never end your turn with a CI-waiting status update"),
1081            "prompt must forbid ending the turn on a CI-waiting status update: {prompt}"
1082        );
1083    }
1084
1085    #[test]
1086    fn shell_single_quote_escapes_embedded_quotes() {
1087        use crate::engine::runner::agents::shell_single_quote;
1088
1089        assert_eq!(shell_single_quote("hello"), "'hello'");
1090        assert_eq!(shell_single_quote("O'Brien"), "'O'\\''Brien'");
1091        assert_eq!(shell_single_quote("$(rm -rf /)"), "'$(rm -rf /)'");
1092        assert_eq!(shell_single_quote("`whoami`"), "'`whoami`'");
1093        assert_eq!(shell_single_quote("a\"b"), "'a\"b'");
1094        assert_eq!(shell_single_quote("$HOME"), "'$HOME'");
1095    }
1096
1097    fn git_init(dir: &std::path::Path) {
1098        let out = std::process::Command::new("git")
1099            .args(["init", "-q"])
1100            .current_dir(dir)
1101            .output()
1102            .unwrap();
1103        assert!(out.status.success());
1104    }
1105
1106    #[tokio::test]
1107    async fn install_shipcheck_rule_writes_rule_and_keeps_tree_clean() {
1108        let dir = tempfile::tempdir().unwrap();
1109        git_init(dir.path());
1110
1111        install_shipcheck_rule(dir.path()).await.unwrap();
1112        install_shipcheck_rule(dir.path()).await.unwrap();
1113
1114        let rule = std::fs::read_to_string(dir.path().join(SHIPCHECK_RULE_PATH)).unwrap();
1115        assert_eq!(rule, SHIPCHECK_RULES);
1116        for cmd in ["run", "commit", "status", "logs", "doctor"] {
1117            assert!(
1118                rule.contains(&format!(r#"["shipcheck", "{cmd}"]"#)),
1119                "{cmd}"
1120            );
1121        }
1122        assert!(!rule.contains(r#"["shipcheck"]"#));
1123        for cmd in [
1124            "clean", "post", "login", "logout", "daemon", "init", "install",
1125        ] {
1126            assert!(!rule.contains(&format!(r#""{cmd}""#)), "{cmd}");
1127        }
1128
1129        let exclude = std::fs::read_to_string(dir.path().join(".git/info/exclude")).unwrap();
1130        assert_eq!(exclude.matches("orch-shipcheck.rules").count(), 1);
1131        let status = std::process::Command::new("git")
1132            .args(["status", "--porcelain", "--untracked-files=all"])
1133            .current_dir(dir.path())
1134            .output()
1135            .unwrap();
1136        assert!(status.stdout.is_empty(), "{:?}", status);
1137    }
1138
1139    #[tokio::test]
1140    async fn install_shipcheck_rule_refuses_symlinked_dirs() {
1141        for link in [".codex", ".codex/rules"] {
1142            let dir = tempfile::tempdir().unwrap();
1143            let target = tempfile::tempdir().unwrap();
1144            git_init(dir.path());
1145            std::fs::create_dir_all(dir.path().join(".codex")).unwrap();
1146            if link == ".codex" {
1147                std::fs::remove_dir(dir.path().join(".codex")).unwrap();
1148            }
1149            std::os::unix::fs::symlink(target.path(), dir.path().join(link)).unwrap();
1150
1151            assert!(install_shipcheck_rule(dir.path()).await.is_err(), "{link}");
1152            assert_eq!(
1153                std::fs::read_dir(target.path()).unwrap().count(),
1154                0,
1155                "{link}"
1156            );
1157        }
1158    }
1159
1160    #[tokio::test]
1161    async fn install_shipcheck_rule_replaces_a_symlinked_file_without_following_it() {
1162        let dir = tempfile::tempdir().unwrap();
1163        let target = tempfile::tempdir().unwrap();
1164        let outside = target.path().join("victim");
1165        std::fs::write(&outside, "keep").unwrap();
1166        git_init(dir.path());
1167        std::fs::create_dir_all(dir.path().join(".codex/rules")).unwrap();
1168        std::os::unix::fs::symlink(&outside, dir.path().join(SHIPCHECK_RULE_PATH)).unwrap();
1169
1170        install_shipcheck_rule(dir.path()).await.unwrap();
1171
1172        assert_eq!(std::fs::read_to_string(&outside).unwrap(), "keep");
1173        let meta = std::fs::symlink_metadata(dir.path().join(SHIPCHECK_RULE_PATH)).unwrap();
1174        assert!(meta.is_file());
1175    }
1176
1177    fn prompt_task() -> crate::backends::ExternalTask {
1178        crate::backends::ExternalTask {
1179            id: crate::backends::ExternalId("1".to_string()),
1180            title: "t".to_string(),
1181            body: "b".to_string(),
1182            state: "open".to_string(),
1183            labels: vec![],
1184            author: "a".to_string(),
1185            created_at: "2026-01-01T00:00:00Z".to_string(),
1186            updated_at: "2026-01-01T00:00:00Z".to_string(),
1187            url: "https://example.com/1".to_string(),
1188        }
1189    }
1190
1191    #[test]
1192    fn system_prompt_uses_shipcheck_only_when_available() {
1193        let ctx = super::super::context::TaskContext::default();
1194        let task = prompt_task();
1195        let prompt = |s| build_system_prompt(&task, &ctx, None, "main", s);
1196
1197        let available = prompt(Shipcheck::Available);
1198        assert!(available.contains("shipcheck run"));
1199        assert!(available.contains("shipcheck commit"));
1200        assert!(!available.contains(".github/workflows/"));
1201        assert!(!available.contains("not installed"));
1202
1203        let outdated = prompt(Shipcheck::Outdated);
1204        assert!(!outdated.contains("shipcheck"));
1205        assert!(outdated.contains(".github/workflows/"));
1206
1207        let missing = prompt(Shipcheck::NotInstalled);
1208        assert!(!missing.contains("shipcheck run"));
1209        assert!(missing.contains(".github/workflows/"));
1210        assert!(missing.contains("shipcheck is not installed"));
1211
1212        let absent = prompt(Shipcheck::Absent);
1213        assert!(!absent.contains("shipcheck"));
1214        assert!(absent.contains(".github/workflows/"));
1215
1216        for p in [&available, &outdated, &missing, &absent] {
1217            assert!(!p.contains("{{"), "{p}");
1218            assert!(!p.contains("install shipcheck"), "{p}");
1219        }
1220    }
1221
1222    #[test]
1223    fn review_prompt_uses_shipcheck_only_when_available() {
1224        let task = prompt_task();
1225        let prompt = |s| build_review_prompt(&task, "", "", "", "main", 1, s);
1226        assert!(prompt(Shipcheck::Available).contains("shipcheck run"));
1227        for s in [
1228            Shipcheck::Outdated,
1229            Shipcheck::NotInstalled,
1230            Shipcheck::Absent,
1231        ] {
1232            let p = prompt(s);
1233            assert!(!p.contains("shipcheck"), "{p}");
1234            assert!(p.contains(".github/workflows/"), "{p}");
1235        }
1236    }
1237
1238    #[test]
1239    fn shipcheck_classify_needs_config_and_commit() {
1240        assert_eq!(Shipcheck::classify(false, Some(true)), Shipcheck::Absent);
1241        assert_eq!(Shipcheck::classify(false, None), Shipcheck::Absent);
1242        assert_eq!(Shipcheck::classify(true, None), Shipcheck::NotInstalled);
1243        assert_eq!(Shipcheck::classify(true, Some(false)), Shipcheck::Outdated);
1244        assert_eq!(Shipcheck::classify(true, Some(true)), Shipcheck::Available);
1245    }
1246
1247    #[tokio::test]
1248    async fn supports_reads_the_usage_line() {
1249        let new_dir = tempfile::tempdir().unwrap();
1250        let new = fake_shipcheck(new_dir.path(), true, true);
1251        assert!(supports(&new, "commit").await);
1252        assert!(supports(&new, "hook").await);
1253        let old_dir = tempfile::tempdir().unwrap();
1254        let old = fake_shipcheck(old_dir.path(), false, false);
1255        assert!(!supports(&old, "commit").await);
1256        assert!(!supports(&old, "hook").await);
1257        let missing = old_dir.path().join("missing");
1258        assert!(!supports(&missing, "commit").await);
1259    }
1260
1261    // Mimics shipcheck, whose `--help` prints usage and exits 0 after any subcommand
1262    fn fake_shipcheck(dir: &std::path::Path, commit: bool, hook: bool) -> PathBuf {
1263        use std::os::unix::fs::PermissionsExt;
1264        let mut usage = "usage: shipcheck run [--sha <rev>] | status | logs".to_string();
1265        if commit {
1266            usage.push_str(" | commit -m <message>");
1267        }
1268        if hook {
1269            usage.push_str(" | hook");
1270        }
1271        usage.push_str(" | install | doctor");
1272        let bin = dir.join("shipcheck");
1273        std::fs::write(
1274            &bin,
1275            format!(
1276                "#!/bin/sh\nfor a in \"$@\"; do [ \"$a\" = --help ] && {{ echo '{usage}'; exit 0; }}; done\n[ \"$1\" = install ] && touch installed\nexit 0\n"
1277            ),
1278        )
1279        .unwrap();
1280        std::fs::set_permissions(&bin, std::fs::Permissions::from_mode(0o755)).unwrap();
1281        bin
1282    }
1283
1284    #[tokio::test]
1285    async fn install_shipcheck_hooks_runs_install_when_supported() {
1286        let bin_dir = tempfile::tempdir().unwrap();
1287        let work = tempfile::tempdir().unwrap();
1288        let bin = fake_shipcheck(bin_dir.path(), true, true);
1289        assert!(install_shipcheck_hooks(work.path(), Shipcheck::Available, &bin).await);
1290        assert!(work.path().join("installed").exists());
1291    }
1292
1293    #[tokio::test]
1294    async fn install_shipcheck_hooks_skips_a_shipcheck_without_hook() {
1295        let bin_dir = tempfile::tempdir().unwrap();
1296        let work = tempfile::tempdir().unwrap();
1297        let bin = fake_shipcheck(bin_dir.path(), true, false);
1298        assert!(!install_shipcheck_hooks(work.path(), Shipcheck::Available, &bin).await);
1299        assert!(!work.path().join("installed").exists());
1300    }
1301
1302    #[tokio::test]
1303    async fn install_shipcheck_hooks_skips_when_not_available() {
1304        let bin_dir = tempfile::tempdir().unwrap();
1305        let work = tempfile::tempdir().unwrap();
1306        let bin = fake_shipcheck(bin_dir.path(), true, true);
1307        for s in [
1308            Shipcheck::Outdated,
1309            Shipcheck::NotInstalled,
1310            Shipcheck::Absent,
1311        ] {
1312            assert!(!install_shipcheck_hooks(work.path(), s, &bin).await);
1313        }
1314        assert!(!work.path().join("installed").exists());
1315    }
1316
1317    #[test]
1318    fn the_progress_step_appears_only_in_cluster_mode() {
1319        let mut vars = HashMap::new();
1320        insert_cluster_vars(Shipcheck::Available, &mut vars);
1321        assert!(vars.is_empty(), "no cluster section, no cluster flags");
1322        let plain = crate::template::render_template_str(AGENT_SYSTEM_TEMPLATE, &vars).unwrap();
1323        assert!(!plain.contains("orch progress"));
1324
1325        for flag in ["CLUSTER", "CLUSTER_SHIPCHECK"] {
1326            vars.insert(flag.to_string(), "1".to_string());
1327        }
1328        let cluster = crate::template::render_template_str(AGENT_SYSTEM_TEMPLATE, &vars).unwrap();
1329        assert!(cluster.contains("orch progress"));
1330        assert!(cluster.contains("commits only through `shipcheck commit`"));
1331        assert!(!cluster.contains("It commits anything you left"));
1332    }
1333}
orch/engine/runner/agent.rs

orch/engine/runner/
agent.rs
