supercode-harness 0.4.13

//! Acceptance tests for TR-12 (`.volter/tracker/markdown/TR-12.md`): the
//! `cap_tool_output` supersession land-blocker fixed at
//! `Agent::run_loop`'s tool-result push site (SPEC.md D6/A7).
//!
//! THE DEFECT (pre-fix): under the default config
//! (`max_tool_output_bytes = Some(100_000)`), `cap_tool_output` ran
//! unconditionally at the history push site even with a recorder + policy
//! active. The recorder got the full bytes; `history` got the capped copy.
//! A10 `TurnsCleared` (and A7 `ToolOutputTruncated`) reduction hashes are
//! minted in `project_messages` from `history` — so any reduction covering a
//! tool output over 100KB was minted from an already-capped copy, which
//! could never recompute the same way once the sidecar was reloaded from
//! disk (`verify_log`/`invert`, the CLI's offline path). Existing
//! `invert_project_is_identity*` tests mint AND invert against the same
//! in-memory `Session`, so they structurally could not catch this.
//!
//! THE FIX: gate `cap_tool_output` off exactly when a recorder AND a
//! `ReductionPolicy` are both installed — the only combination under which a
//! reduction is ever minted over `history` with a durable sidecar behind it.
//! `history` then holds full bytes by construction, so `history[1..] ≡
//! sidecar.messages`, and every hash minted matches what a reloaded sidecar
//! recomputes.

use std::path::PathBuf;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::{Arc, Mutex};

use async_trait::async_trait;
use supercode_harness::reduce::{
    invert, project_messages, reduction_id, verify_log, ReductionPolicy,
};
use supercode_harness::session::Session;
use supercode_harness::sidecar::SidecarWriter;
use supercode_harness::store::SessionStore;
use supercode_harness::{
    Agent, ChatMessage, ChatRequest, Config, FunctionCall, Provider, Role, ToolCall, Usage,
};

fn temp_dir(tag: &str) -> PathBuf {
    static N: AtomicUsize = AtomicUsize::new(0);
    let dir = std::env::temp_dir().join(format!(
        "supercode-tr12-{tag}-{}-{}",
        std::process::id(),
        N.fetch_add(1, Ordering::SeqCst)
    ));
    std::fs::create_dir_all(&dir).unwrap();
    dir
}

/// A tool ("big_tool") that returns a fixed, distinctive payload of exactly
/// `bytes` length — big enough to trip `Config::max_tool_output_bytes`'s
/// default 100 KB cap.
struct BigTool(String);
#[async_trait]
impl supercode_harness::tools::Tool for BigTool {
    fn name(&self) -> &str {
        "big_tool"
    }
    fn description(&self) -> &str {
        "x"
    }
    fn parameters(&self) -> serde_json::Value {
        serde_json::json!({"type": "object"})
    }
    async fn execute(
        &self,
        _a: serde_json::Value,
        _c: &supercode_harness::tools::ToolContext,
    ) -> supercode_harness::Result<String> {
        Ok(self.0.clone())
    }
}

fn tool_call_msg(id: &str) -> ChatMessage {
    ChatMessage {
        role: Role::Assistant,
        content: None,
        content_parts: None,
        tool_calls: Some(vec![ToolCall {
            id: id.to_string(),
            kind: "function".to_string(),
            function: FunctionCall {
                name: "big_tool".to_string(),
                arguments: "{}".to_string(),
            },
        }]),
        tool_call_id: None,
        name: None,
        metadata: Default::default(),
    }
}

/// Turn 0: call `big_tool`. Every turn after: a plain text reply (no tool
/// calls), so `Agent::send` returns immediately and the caller drives how
/// many turns run by how many times it calls `send`. Every request's message
/// vector is captured so a test can inspect the wire body of any turn.
struct ToolThenPlainCapturing {
    calls: AtomicUsize,
    requests: Arc<Mutex<Vec<Vec<ChatMessage>>>>,
}
#[async_trait]
impl Provider for ToolThenPlainCapturing {
    async fn complete(
        &self,
        req: &ChatRequest,
        _on_delta: &(dyn for<'a> Fn(&'a str) + Send + Sync),
    ) -> supercode_harness::Result<(ChatMessage, Usage)> {
        let n = self.calls.fetch_add(1, Ordering::SeqCst);
        self.requests.lock().unwrap().push(req.messages.clone());
        if n == 0 {
            Ok((tool_call_msg("c1"), Usage::default()))
        } else {
            Ok((
                ChatMessage::assistant(format!("reply {n}")),
                Usage::default(),
            ))
        }
    }
}

/// Wire-serialized byte size of a request's message vector, the same
/// approximation `reduce_loop.rs`'s `wire_bytes` uses for its A7 acceptance
/// assertion.
fn wire_bytes(msgs: &[ChatMessage]) -> usize {
    serde_json::to_string(msgs).map(|s| s.len()).unwrap_or(0)
}

/// dev/01 (GUARANTOR REGRESSION, required verbatim by Control) + dev/04
/// (whole-suite invariant): a live agent run with a recorder + policy BOTH
/// active produces a >100KB tool output that lands inside an established
/// `clear_turns_older_than`/`TurnsCleared` range; `verify_log` AND `invert`
/// then run OFFLINE against the sidecar and reduction log **reloaded from
/// disk** (via `SessionStore`, the exact primitives
/// `cli/main.rs:1520,2864,3012`'s `show-reductions`/`convert`/`inspect` use)
/// — never the live `Agent`/in-memory `Session`. Both must pass clean, and
/// `invert` must restore the full original bytes byte-exact.
///
/// MUTATION CHECK (see the TR-12 builder report): this test was verified to
/// FAIL on the pre-fix gate (temporarily reverting `Agent::run_loop`'s
/// `for_history` gate to always call `self.cap_tool_output(output)`) with a
/// `verify_log`/hash-mismatch error, and to pass with the fix restored.
#[tokio::test]
async fn dev01_dev04_guarantor_regression_offline_verify_and_invert() {
    let dir = temp_dir("dev01-dev04");
    let store_dir = dir.join("store");
    let store = SessionStore::open(&store_dir).unwrap();
    let name = "tr12-dev01";
    let sidecar_path = store.sidecar_path(name);

    // A distinctive >100KB payload (150,000 bytes) with a needle so a
    // byte-exact restore can be checked precisely, not just by length.
    let mut original = "Q".repeat(120_000);
    original.push_str("TR12-NEEDLE-DEV01");
    original.push_str(&"y".repeat(150_000 - original.len()));
    assert_eq!(original.len(), 150_000);
    assert!(original.len() > 100_000, "must trip the default cap");

    let requests = Arc::new(Mutex::new(Vec::new()));
    let config = Config::builder()
        .cwd(dir.clone())
        .compact_after_messages(8)
        .build();
    let mut reg = supercode_harness::tools::ToolRegistry::new();
    reg.register(BigTool(original.clone()));
    let mut agent = Agent::with_parts(
        config,
        Box::new(ToolThenPlainCapturing {
            calls: AtomicUsize::new(0),
            requests: requests.clone(),
        }),
        reg,
    );

    let empty_session = Session::from_claude_code_str("").unwrap();
    let writer = SidecarWriter::create(&sidecar_path, &empty_session).unwrap();
    agent.set_recorder(writer);
    let policy = ReductionPolicy::default();
    agent.set_reduction_policy(policy.clone());

    // Turn 0 fires the big tool call. Turns 1..5 are plain text — enough for
    // `compact_after_messages(8)` to trigger a `TurnsCleared` range covering
    // the tool call/result from turn 0 (established once `history[1..].len()`
    // first exceeds 8, then reproduced verbatim forever after).
    for i in 0..6 {
        let reply = agent.send(format!("turn {i}")).await.unwrap();
        assert!(
            reply.starts_with("reply"),
            "unexpected reply at turn {i}: {reply}"
        );
    }

    let log = agent.reduction_log().clone();
    let cleared = log
        .reductions
        .iter()
        .find_map(|r| match r.kind {
            supercode_harness::reduce::ReductionKind::TurnsCleared { first, last, .. } => {
                Some((first, last))
            }
            _ => None,
        })
        .expect("a TurnsCleared reduction must have been established");
    // The tool call (index 1) and its tool result (index 2) in
    // `history[1..]` must fall inside the cleared range — otherwise this
    // test isn't exercising the scenario TR-12 requires.
    assert!(
        cleared.0 <= 1 && cleared.1 >= 2,
        "the big tool output must land inside the TurnsCleared range, got {cleared:?}"
    );

    // dev/04: `history[1..]` is byte-identical to the reloaded sidecar's
    // messages (direct assertion, not implied) — the property the D6/A7 gate
    // exists to guarantee.
    let sidecar_raw_live = std::fs::read_to_string(&sidecar_path).unwrap();
    let sidecar_live_reload = Session::from_native_str(&sidecar_raw_live).unwrap();
    assert_eq!(
        sidecar_live_reload.messages.len(),
        agent.history().len() - 1
    );
    for (a, b) in sidecar_live_reload
        .messages
        .iter()
        .zip(&agent.history()[1..])
    {
        assert_eq!(a.role, b.role);
        assert_eq!(
            a.content, b.content,
            "history[1..] must equal the sidecar, byte for byte"
        );
    }

    // Persist the reduction log the way the CLI does (`sessions
    // show-reductions`/`convert`/`inspect` all call `store.save_reduction_log`
    // via the resume/chat path; here we do it explicitly so this test's
    // offline reload is genuine).
    store.save_reduction_log(name, &log).unwrap();

    // --- OFFLINE from here: fresh reads from disk, no live `Agent`/Session ---
    let sidecar_jsonl = store
        .load_sidecar(name)
        .unwrap()
        .expect("sidecar must exist on disk");
    let sidecar = Session::from_sidecar_str(&sidecar_jsonl).unwrap();
    let log_reloaded = store
        .load_reduction_log(name)
        .unwrap()
        .expect("reduction log must exist on disk");

    // The exact primitive `cli/main.rs:1520` (`show-reductions`),
    // `:2864` (`convert`), and `:3012` (`inspect`) all call before doing
    // anything else with a reduced session.
    verify_log(&log_reloaded, &sidecar)
        .expect("verify_log must pass clean against the reloaded-from-disk sidecar");

    // Reconstruct the final projected view exactly as it was established
    // (prefix stability, A5: passing `log_reloaded` as `prior` reapplies the
    // existing `TurnsCleared` range verbatim rather than recomputing a new
    // one against the reloaded sidecar's now-final length).
    let (final_view, reprojected_log) = project_messages(&sidecar.messages, &policy, &log_reloaded);
    assert_eq!(
        reprojected_log, log_reloaded,
        "re-projecting from the reloaded sidecar with its own log must not invent new reductions"
    );

    let inverted = invert(&final_view, &log_reloaded, &sidecar)
        .expect("invert must pass clean against the reloaded-from-disk sidecar");
    assert_eq!(inverted.len(), sidecar.messages.len());
    for (a, b) in inverted.iter().zip(&sidecar.messages) {
        assert_eq!(a.role, b.role);
        assert_eq!(a.content, b.content);
    }

    // Byte-exact restore of the big tool output specifically.
    let restored_tool_msg = inverted
        .iter()
        .find(|m| m.role == Role::Tool && m.tool_call_id.as_deref() == Some("c1"))
        .and_then(|m| m.content.clone())
        .expect("the big tool result must be present after invert");
    assert_eq!(
        restored_tool_msg, original,
        "invert must restore the full 150,000-byte original byte-exact"
    );

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

/// dev/02: gate correctness. With a recorder + policy both active, the
/// history message pushed for a >cap tool result carries the FULL bytes (no
/// cap notice). With no reduction machinery at all (no policy — the legacy
/// path), `cap_tool_output` still runs and the notice still appears,
/// byte-identical to before TR-12 (pre-existing legacy-cap tests —
/// `agent_loop.rs::oversized_tool_output_is_capped`,
/// `agent_records_full_fidelity_while_capping_view` — are left unmodified
/// and still pass; this test adds the explicit side-by-side contrast the AC
/// calls for).
#[tokio::test]
async fn dev02_gate_correctness_full_bytes_with_machinery_capped_without() {
    let big = "N".repeat(150_000);

    // With machinery: recorder + policy both active.
    {
        let dir = temp_dir("dev02-with-machinery");
        let sidecar_path = dir.join("sess.sidecar.jsonl");
        let config = Config::builder().cwd(dir.clone()).build();
        let mut reg = supercode_harness::tools::ToolRegistry::new();
        reg.register(BigTool(big.clone()));
        let requests = Arc::new(Mutex::new(Vec::new()));
        let mut agent = Agent::with_parts(
            config,
            Box::new(ToolThenPlainCapturing {
                calls: AtomicUsize::new(0),
                requests,
            }),
            reg,
        );
        let empty_session = Session::from_claude_code_str("").unwrap();
        let writer = SidecarWriter::create(&sidecar_path, &empty_session).unwrap();
        agent.set_recorder(writer);
        agent.set_reduction_policy(ReductionPolicy::default());

        agent.send("go").await.unwrap();

        let history_copy = agent
            .history()
            .iter()
            .find(|m| m.role == Role::Tool)
            .and_then(|m| m.content.clone())
            .unwrap();
        assert_eq!(
            history_copy, big,
            "with machinery, history must keep the FULL bytes"
        );
        assert!(
            !history_copy.contains("bytes total, showing first"),
            "with machinery, no cap notice may ever appear in history"
        );
        std::fs::remove_dir_all(&dir).ok();
    }

    // Without machinery: no policy at all (recorder absent too, the plainest
    // legacy configuration) — capping stays exactly as before TR-12.
    {
        let dir = temp_dir("dev02-legacy");
        let config = Config::builder().cwd(dir.clone()).build();
        let mut reg = supercode_harness::tools::ToolRegistry::new();
        reg.register(BigTool(big.clone()));
        let requests = Arc::new(Mutex::new(Vec::new()));
        let mut agent = Agent::with_parts(
            config,
            Box::new(ToolThenPlainCapturing {
                calls: AtomicUsize::new(0),
                requests,
            }),
            reg,
        );

        agent.send("go").await.unwrap();

        let history_copy = agent
            .history()
            .iter()
            .find(|m| m.role == Role::Tool)
            .and_then(|m| m.content.clone())
            .unwrap();
        assert!(
            history_copy.len() < big.len(),
            "without machinery, legacy capping must still run"
        );
        assert!(
            history_copy.contains("bytes total, showing first"),
            "without machinery, the legacy cap notice must still appear: {history_copy}"
        );
        assert!(
            history_copy.contains("full output not retained"),
            "without a recorder, the notice must honestly say so: {history_copy}"
        );
        std::fs::remove_dir_all(&dir).ok();
    }
}

/// dev/03: safety net. Even with the D6/A7 gate off (recorder + policy both
/// active, so `history`/the sidecar keep a `>>100KB` tool output in full),
/// the very next request's WIRE payload is still bounded — by A7's own
/// projection pass, in the same loop (the next `build_request_messages`
/// call, before the model ever sees the raw output) — and carries a
/// `reduction_id` stub (the TR-1 contract every rehydration intrinsic
/// depends on). ROADMAP P7's runaway-output/RSS concern is covered by this,
/// not by reintroducing lossy capping.
#[tokio::test]
async fn dev03_safety_net_wire_bounded_via_a7_with_gate_off() {
    let dir = temp_dir("dev03");
    let sidecar_path = dir.join("sess.sidecar.jsonl");

    // A full order of magnitude past the legacy 100KB cap.
    let huge = "H".repeat(1_000_000);
    assert!(huge.len() >= 100_000 * 10);

    let config = Config::builder().cwd(dir.clone()).build();
    let mut reg = supercode_harness::tools::ToolRegistry::new();
    reg.register(BigTool(huge.clone()));
    let requests = Arc::new(Mutex::new(Vec::new()));
    let mut agent = Agent::with_parts(
        config,
        Box::new(ToolThenPlainCapturing {
            calls: AtomicUsize::new(0),
            requests: requests.clone(),
        }),
        reg,
    );
    let empty_session = Session::from_claude_code_str("").unwrap();
    let writer = SidecarWriter::create(&sidecar_path, &empty_session).unwrap();
    agent.set_recorder(writer);
    agent.set_reduction_policy(ReductionPolicy {
        tool_output_keep_bytes: 4096,
        tool_output_trigger_bytes: 8192,
        protect_last_n_tool_results: 0,
        ..ReductionPolicy::default()
    });

    agent.send("go").await.unwrap();

    // history/the sidecar hold the full 1,000,000-byte output (the gate is
    // off — this is the accepted in-memory/durable-storage tradeoff).
    let history_copy = agent
        .history()
        .iter()
        .find(|m| m.role == Role::Tool)
        .and_then(|m| m.content.clone())
        .unwrap();
    assert_eq!(history_copy.len(), huge.len());

    // The SECOND request (built at the top of the next loop iteration, same
    // `send` call, immediately after the tool result was pushed) must carry
    // a bounded wire payload, not the raw 1,000,000-byte output.
    let reqs = requests.lock().unwrap().clone();
    let second_request = &reqs[1];
    let bytes = wire_bytes(second_request);
    assert!(
        bytes < 50_000,
        "the wire payload for the very next request must be bounded by A7, got {bytes} bytes"
    );
    let body = serde_json::to_string(second_request).unwrap();
    assert!(
        !body.contains(&huge),
        "the wire payload must never contain the raw huge output"
    );

    // A7's stub carries a `reduction_id` (TR-1 contract): every rehydration
    // intrinsic (`expand_reduction`/`sidecar_search`) locates a placeholder
    // by this id.
    assert!(
        second_request.iter().any(|m| reduction_id(m).is_some()),
        "the reduced view must carry a message with a reduction id"
    );

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