supercode-harness 0.4.10

//! P5-9 (COMPOSABLE-HARNESS-DESIGN.md §2 module 20 `checkpoint`): file
//! checkpointing / shadow-git, driven end to end through `Agent::send`.
//!
//! **LIVE-AGENT-TEST SAFETY.** Every test in this file drives `Agent`
//! through a scripted MOCK `Provider` (the same `ScriptedProvider` idiom
//! `tests/agent_loop.rs`/`tests/tools_background.rs` already use) — never a
//! real model endpoint.

use std::sync::atomic::{AtomicUsize, Ordering};

use async_trait::async_trait;
use supercode_harness::{
    Agent, ChatMessage, ChatRequest, Config, FunctionCall, Provider, Role, ToolCall, Usage,
};

fn tool_call(id: &str, name: &str, args: serde_json::Value) -> ToolCall {
    ToolCall {
        id: id.to_string(),
        kind: "function".to_string(),
        function: FunctionCall {
            name: name.to_string(),
            arguments: args.to_string(),
        },
    }
}

fn assistant_with_call(call: ToolCall) -> ChatMessage {
    ChatMessage {
        role: Role::Assistant,
        content: None,
        content_parts: None,
        tool_calls: Some(vec![call]),
        tool_call_id: None,
        name: None,
        metadata: Default::default(),
    }
}

fn tmp(tag: &str) -> std::path::PathBuf {
    let dir = std::env::temp_dir().join(format!(
        "supercode-checkpoint-engine-{tag}-{}-{}",
        std::process::id(),
        std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .unwrap()
            .as_nanos()
    ));
    std::fs::create_dir_all(&dir).unwrap();
    dir
}

/// Turn 1: `write_file` a new file. Turn 2: `edit_file` it, then
/// `write_file` a second, brand-new file. Turn 3: a final no-tool-call
/// answer. Exactly the same script every test in this file replays — so a
/// checkpoint-enabled agent's OBSERVABLE conversation (history, replies)
/// can be diffed against a checkpoint-DISABLED agent's for the C8
/// no-context-leak proof.
struct ScriptedWrites {
    calls: AtomicUsize,
}

#[async_trait]
impl Provider for ScriptedWrites {
    async fn complete(
        &self,
        req: &ChatRequest,
        _on_delta: &(dyn for<'a> Fn(&'a str) + Send + Sync),
    ) -> supercode_harness::Result<(ChatMessage, Usage)> {
        match self.calls.fetch_add(1, Ordering::SeqCst) {
            0 => Ok((
                assistant_with_call(tool_call(
                    "call_1",
                    "write_file",
                    serde_json::json!({"path": "a.txt", "content": "hello"}),
                )),
                Usage::default(),
            )),
            1 => Ok((ChatMessage::assistant("created a.txt"), Usage::default())),
            2 => Ok((
                assistant_with_call(tool_call(
                    "call_2",
                    "edit_file",
                    serde_json::json!({"path": "a.txt", "old_string": "hello", "new_string": "world"}),
                )),
                Usage::default(),
            )),
            3 => Ok((
                assistant_with_call(tool_call(
                    "call_3",
                    "write_file",
                    serde_json::json!({"path": "b.txt", "content": "brand new"}),
                )),
                Usage::default(),
            )),
            4 => Ok((
                ChatMessage::assistant("edited and added b.txt"),
                Usage::default(),
            )),
            n => panic!("unexpected provider call {n}: {:?}", req.messages.last()),
        }
    }
}

async fn run_script(config: Config) -> Agent {
    let mut agent = Agent::with_provider(
        config,
        Box::new(ScriptedWrites {
            calls: AtomicUsize::new(0),
        }),
    );
    let r1 = agent.send("create a.txt").await.unwrap();
    assert_eq!(r1, "created a.txt");
    let r2 = agent.send("edit it and add b.txt").await.unwrap();
    assert_eq!(r2, "edited and added b.txt");
    agent
}

fn history_signature(agent: &Agent) -> Vec<String> {
    agent.history().iter().map(|m| format!("{m:?}")).collect()
}

// ---------------------------------------------------------------------------
// 1. Default-off byte-identity: no shadow dir is ever created, no
//    `checkpoint_observer`, and the conversation is IDENTICAL to a run with
//    checkpoint never having existed at all.
// ---------------------------------------------------------------------------

#[tokio::test]
async fn checkpoint_off_by_default_touches_no_disk_and_is_a_true_no_op() {
    let project = tmp("off-project");
    let shadow = tmp("off-shadow");
    // Remove the shadow dir `tmp()` just created — we want to prove
    // `observer_for_config` never even calls `create_dir_all` on it.
    std::fs::remove_dir_all(&shadow).unwrap();

    let config = Config::builder()
        .cwd(project.clone())
        .checkpoint_dir(shadow.clone()) // set but NEVER consulted while disabled
        .build();
    assert!(!config.checkpoint_enabled, "checkpoint must default off");

    let agent = run_script(config).await;
    assert!(
        agent.checkpoint_observer().is_none(),
        "no observer should exist when the module is off"
    );
    assert!(
        !shadow.exists(),
        "checkpoint_enabled=false must never create the shadow directory"
    );

    std::fs::remove_dir_all(&project).ok();
}

#[tokio::test]
async fn checkpoint_on_or_off_produces_byte_identical_conversation_history() {
    // C8 proof: checkpoint is a pure side-channel to disk. Turning it on
    // must not add/remove/alter a single ChatMessage — no snapshot content,
    // no checkpoint id, nothing checkpoint-related ever reaches the model's
    // context or the recorded conversation.
    let project_off = tmp("parity-off-project");
    let project_on = tmp("parity-on-project");
    let shadow_on = tmp("parity-on-shadow");

    let config_off = Config::builder().cwd(project_off.clone()).build();
    let config_on = Config::builder()
        .cwd(project_on.clone())
        .checkpoint_enabled(true)
        .checkpoint_dir(shadow_on.clone())
        .build();

    let agent_off = run_script(config_off).await;
    let agent_on = run_script(config_on).await;

    assert_eq!(
        history_signature(&agent_off),
        history_signature(&agent_on),
        "checkpoint on/off must never change the observable conversation"
    );
    assert!(agent_on.checkpoint_observer().is_some());
    assert!(
        shadow_on.exists(),
        "checkpoint on DOES touch its own shadow dir"
    );

    std::fs::remove_dir_all(&project_off).ok();
    std::fs::remove_dir_all(&project_on).ok();
    std::fs::remove_dir_all(&shadow_on).ok();
}

// ---------------------------------------------------------------------------
// 2. Enabled: capture, turn-diff (D3), and restore (D4-adjacent) — driven
//    entirely through the real write tools, not by calling the store
//    directly (the unit tests in `crates/harness/src/checkpoint.rs` already
//    cover the store's own API in isolation).
// ---------------------------------------------------------------------------

#[tokio::test]
async fn enabled_checkpoint_captures_and_restore_undoes_a_whole_turn() {
    let project = tmp("restore-project");
    let shadow = tmp("restore-shadow");
    let config = Config::builder()
        .cwd(project.clone())
        .checkpoint_enabled(true)
        .checkpoint_dir(shadow.clone())
        .build();

    let agent = run_script(config).await;
    assert_eq!(
        std::fs::read_to_string(project.join("a.txt")).unwrap(),
        "world"
    );
    assert_eq!(
        std::fs::read_to_string(project.join("b.txt")).unwrap(),
        "brand new"
    );

    let cp = agent.checkpoint_observer().expect("checkpoint is enabled");
    let checkpoints = cp.list().unwrap();
    assert_eq!(checkpoints.len(), 2, "one checkpoint per turn");
    let turn2 = &checkpoints[0]; // newest first
    let turn1 = &checkpoints[1];

    // D3 turn-diff.
    let mut diff1 = cp.turn_diff(&turn1.id).unwrap();
    diff1.sort();
    assert_eq!(diff1, vec!["a.txt".to_string()]);
    let mut diff2 = cp.turn_diff(&turn2.id).unwrap();
    diff2.sort();
    assert_eq!(diff2, vec!["a.txt".to_string(), "b.txt".to_string()]);

    // D4-adjacent revert: undo turn 2 only — a.txt goes back to "hello"
    // (turn 2's pre-image), b.txt (created during turn 2) is deleted.
    let report = cp.restore(&turn2.id).unwrap();
    assert!(report.refused.is_empty(), "{:?}", report.refused);
    assert_eq!(
        std::fs::read_to_string(project.join("a.txt")).unwrap(),
        "hello"
    );
    assert!(!project.join("b.txt").exists());

    // Restoring turn 1 on top undoes the create entirely: a.txt goes away
    // too (it didn't exist before turn 1).
    let report1 = cp.restore(&turn1.id).unwrap();
    assert!(report1.refused.is_empty(), "{:?}", report1.refused);
    assert!(!project.join("a.txt").exists());

    std::fs::remove_dir_all(&project).ok();
    std::fs::remove_dir_all(&shadow).ok();
}

#[tokio::test]
async fn enabled_checkpoint_never_touches_the_projects_real_git() {
    let project = tmp("realgit-project");
    let shadow = tmp("realgit-shadow");
    let git_dir = project.join(".git");
    std::fs::create_dir_all(&git_dir).unwrap();
    std::fs::write(git_dir.join("HEAD"), "ref: refs/heads/main\n").unwrap();
    let head_before = std::fs::read(git_dir.join("HEAD")).unwrap();

    let config = Config::builder()
        .cwd(project.clone())
        .checkpoint_enabled(true)
        .checkpoint_dir(shadow.clone())
        .build();
    let agent = run_script(config).await;

    // The shadow store lives entirely under `shadow`, never under
    // `project/.git`.
    assert_eq!(
        std::fs::read(git_dir.join("HEAD")).unwrap(),
        head_before,
        "the project's real .git/HEAD must be byte-identical"
    );
    let mut git_entries: Vec<_> = std::fs::read_dir(&git_dir)
        .unwrap()
        .map(|e| e.unwrap().file_name())
        .collect();
    git_entries.sort();
    assert_eq!(
        git_entries,
        vec![std::ffi::OsString::from("HEAD")],
        "checkpoint must never write anything into the project's .git"
    );

    // The shadow store itself DID capture the turns (proving this isn't a
    // vacuous "checkpoint silently did nothing" pass).
    let cp = agent.checkpoint_observer().unwrap();
    assert_eq!(cp.list().unwrap().len(), 2);

    std::fs::remove_dir_all(&project).ok();
    std::fs::remove_dir_all(&shadow).ok();
}

#[tokio::test]
async fn enabled_checkpoint_prunes_old_checkpoints_past_the_retain_bound() {
    let project = tmp("prune-project");
    let shadow = tmp("prune-shadow");
    let config = Config::builder()
        .cwd(project.clone())
        .checkpoint_enabled(true)
        .checkpoint_dir(shadow.clone())
        .checkpoint_retain(1)
        .build();
    let agent = run_script(config).await; // 2 turns, retain=1
    let cp = agent.checkpoint_observer().unwrap();
    let checkpoints = cp.list().unwrap();
    assert_eq!(
        checkpoints.len(),
        1,
        "retain=1 must prune down to the single newest checkpoint"
    );

    std::fs::remove_dir_all(&project).ok();
    std::fs::remove_dir_all(&shadow).ok();
}