supercode-harness 0.4.10

//! Acceptance tests for TR-10 (`ReductionKind::ToolInputElided`): the
//! assistant-side twin of A7 (tool RESULTS)/A8 (stale file READS) — eliding
//! the disk-persisted payload argument of an already-executed, SUCCESSFUL
//! tool_use call, leaving every other argument (e.g. `path`) verbatim.
//!
//! Mirrors `reduce_projection.rs`'s idiom throughout: hand-built
//! `ChatMessage` vectors run through the REAL `project_messages`/`invert`/
//! `export_session` pipeline (never a re-implementation of the reduction
//! logic), with `session_of(msgs)` wrapping a message vector in a bare
//! `Session` to serve as the `sidecar_session` `invert` needs (the sidecar
//! IS the unreduced message list — SPEC.md A6).

use std::path::PathBuf;

use supercode_harness::reduce::{
    invert, is_tool_error, mark_tool_error, probe_tool_input_fresh, project_messages, reduction_id,
    tool_input_escalation_action, EscalationAction, ReductionKind, ReductionLog, ReductionPolicy,
    REDUCTION_SENTINEL,
};
use supercode_harness::{ChatMessage, FunctionCall, Role, Session, ToolCall};

// ---- shared helpers ---------------------------------------------------------

/// An empty (Claude-Code-origin) `Session` whose `messages` is then
/// overwritten — the same idiom `reduce_projection.rs::session_of` uses, so
/// a hand-built message vector can serve as `invert`'s `sidecar_session`.
fn session_of(msgs: Vec<ChatMessage>) -> Session {
    let mut session = Session::from_claude_code_str("").unwrap();
    session.messages = msgs;
    session
}

/// Build an assistant message issuing one `write_file(path, content)` tool
/// call with the given call id — every other field left at its default (no
/// leading text, one tool call).
fn write_call(id: &str, path: &str, content: &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: "write_file".to_string(),
                arguments: serde_json::json!({ "path": path, "content": content }).to_string(),
            },
        }]),
        tool_call_id: None,
        name: None,
        metadata: Default::default(),
    }
}

/// A successful `write_file` result message paired to `id`.
fn write_ok_result(id: &str, path: &str, bytes: usize) -> ChatMessage {
    ChatMessage::tool_result(id, "write_file", format!("Wrote {bytes} bytes to {path}"))
}

/// An ERRORED `write_file` result message paired to `id` — the TR-10/TR-6
/// boundary case: marked via [`mark_tool_error`], the only way `ChatMessage`
/// carries "this call failed" (see `reduce.rs::TOOL_ERROR_METADATA_KEY`'s
/// doc comment for why there's no other structural slot).
fn write_err_result(id: &str) -> ChatMessage {
    let mut m = ChatMessage::tool_result(id, "write_file", "Error: disk full");
    mark_tool_error(&mut m);
    m
}

/// A deterministic all-ASCII filler string of exactly `len` bytes.
fn filler(len: usize) -> String {
    (0..len).map(|i| (b'a' + (i % 26) as u8) as char).collect()
}

/// The tool_call's raw `arguments` string at `msg_index`'s first tool call.
fn args_of(msgs: &[ChatMessage], msg_index: usize) -> String {
    msgs[msg_index].tool_calls()[0].function.arguments.clone()
}

fn parsed_args(msgs: &[ChatMessage], msg_index: usize) -> serde_json::Value {
    serde_json::from_str(&args_of(msgs, msg_index)).unwrap()
}

// ---- dev/01: successful large Write elided, invert byte-exact --------------

#[test]
fn dev01_large_successful_write_is_elided_stub_visible_and_inverts_byte_exact() {
    let big = filler(20_000);
    let msgs = vec![
        ChatMessage::user("please write the report"),
        write_call("w1", "reports/out.txt", &big),
        write_ok_result("w1", "reports/out.txt", big.len()),
        ChatMessage::assistant("done"),
    ];
    let asst_idx = 1;

    let (view, log) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());

    assert_eq!(log.reductions.len(), 1, "exactly one reduction expected");
    let r = &log.reductions[0];
    let (original_bytes, path, content_hash, call_id, field) = match &r.kind {
        ReductionKind::ToolInputElided {
            original_bytes,
            path,
            content_hash,
            call_id,
            field,
        } => (
            *original_bytes,
            path.clone(),
            content_hash.clone(),
            call_id.clone(),
            field.clone(),
        ),
        other => panic!("expected ToolInputElided, got {other:?}"),
    };
    assert_eq!(original_bytes, big.len());
    assert_eq!(path, Some(PathBuf::from("reports/out.txt")));
    assert_eq!(
        content_hash,
        supercode_harness::reduce::content_hash(big.as_bytes())
    );
    assert_eq!(call_id, "w1");
    assert_eq!(field, "content");
    assert_eq!(r.ptr.span, None, "whole-field elision carries no byte span");
    assert_eq!(r.ptr.addr.index, asst_idx);
    assert_eq!(r.ptr.addr.role, Role::Assistant);

    // The stub is visible in the view's tool_call arguments (not `content` —
    // this kind's slot is the assistant's tool_use arguments).
    let reduced = parsed_args(&view, asst_idx);
    let content_val = reduced.get("content").unwrap().as_str().unwrap();
    assert!(content_val.contains(REDUCTION_SENTINEL), "{content_val}");
    assert!(content_val.contains("tool-input"), "{content_val}");
    assert!(content_val.contains(&r.id), "{content_val}");
    assert!(content_val.contains("write_file"), "{content_val}");
    assert!(content_val.contains("20,000"), "{content_val}");
    assert!(
        content_val.contains("reports/out.txt"),
        "path should be named in the summary: {content_val}"
    );

    // Non-payload argument (`path`) stays verbatim.
    assert_eq!(
        reduced.get("path").unwrap().as_str().unwrap(),
        "reports/out.txt"
    );

    // The message carries the reduction id (metadata never hits the wire).
    assert_eq!(reduction_id(&view[asst_idx]), Some(r.id.as_str()));

    // `invert` restores the EXACT original arguments string, byte for byte.
    let sidecar = session_of(msgs.clone());
    let inverted = invert(&view, &log, &sidecar).unwrap();
    assert_eq!(
        args_of(&inverted, asst_idx),
        args_of(&msgs, asst_idx),
        "invert must restore the original tool_call arguments byte-for-byte"
    );
    assert_eq!(reduction_id(&inverted[asst_idx]), None);

    // Every other message is untouched.
    for i in [0usize, 2, 3] {
        assert_eq!(
            inverted[i].content, msgs[i].content,
            "message {i} must be untouched"
        );
    }

    // Determinism + prefix stability: re-projecting from the same prior log
    // reproduces the SAME id and byte-identical stub.
    let (view2, log2) = project_messages(&msgs, &ReductionPolicy::default(), &log);
    assert_eq!(log2.reductions, log.reductions);
    assert_eq!(args_of(&view2, asst_idx), args_of(&view, asst_idx));
}

// ---- dev/02: errored / pending calls are never input-elided -----------------

#[test]
fn dev02_errored_write_call_is_never_input_elided() {
    let big = filler(20_000);
    let msgs = vec![write_call("w1", "out.txt", &big), write_err_result("w1")];
    assert!(is_tool_error(&msgs[1]));

    let (view, log) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());
    assert!(
        log.reductions.is_empty(),
        "an errored call's oversized input must never be elided: {:?}",
        log.reductions
    );
    assert_eq!(args_of(&view, 0), args_of(&msgs, 0));
}

#[test]
fn dev02_still_pending_write_call_is_never_input_elided() {
    let big = filler(20_000);
    // No paired tool result at all — the call hasn't finished (or the
    // transcript is mid-turn).
    let msgs = vec![
        ChatMessage::user("write it"),
        write_call("w1", "out.txt", &big),
    ];

    let (view, log) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());
    assert!(
        log.reductions.is_empty(),
        "a still-pending call's oversized input must never be elided: {:?}",
        log.reductions
    );
    assert_eq!(args_of(&view, 1), args_of(&msgs, 1));
}

// ---- dev/03: threshold + field-level selectivity ----------------------------

#[test]
fn dev03_args_below_threshold_are_untouched() {
    let small = filler(100); // well under the default 8192-byte trigger
    let msgs = vec![
        write_call("w1", "out.txt", &small),
        write_ok_result("w1", "out.txt", 100),
    ];

    let (view, log) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());
    assert!(
        log.reductions.is_empty(),
        "a small payload must not be elided"
    );
    assert_eq!(args_of(&view, 0), args_of(&msgs, 0));
}

#[test]
fn dev03_non_payload_field_is_never_elided_even_when_oversized() {
    // `path` is (unrealistically) huge here too — only `content` is ever the
    // designated payload field for `write_file`, so `path` must survive
    // untouched regardless of its own size.
    let big_content = filler(20_000);
    let big_path = format!("dir/{}.txt", filler(9_000));
    let msg = ChatMessage {
        role: Role::Assistant,
        content: None,
        content_parts: None,
        tool_calls: Some(vec![ToolCall {
            id: "w1".to_string(),
            kind: "function".to_string(),
            function: FunctionCall {
                name: "write_file".to_string(),
                arguments: serde_json::json!({ "path": big_path, "content": big_content })
                    .to_string(),
            },
        }]),
        tool_call_id: None,
        name: None,
        metadata: Default::default(),
    };
    let msgs = vec![msg, write_ok_result("w1", "dir/x.txt", big_content.len())];

    let (view, log) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());
    assert_eq!(
        log.reductions.len(),
        1,
        "exactly the `content` field is elided"
    );
    match &log.reductions[0].kind {
        ReductionKind::ToolInputElided { field, .. } => assert_eq!(field, "content"),
        other => panic!("expected ToolInputElided, got {other:?}"),
    }
    let reduced = parsed_args(&view, 0);
    assert_eq!(
        reduced.get("path").unwrap().as_str().unwrap(),
        big_path,
        "the oversized non-payload `path` field must remain verbatim, never elided"
    );
    assert!(reduced
        .get("content")
        .unwrap()
        .as_str()
        .unwrap()
        .contains(REDUCTION_SENTINEL));
}

// ---- dev/04: freshness matrix (à la A8) -------------------------------------

fn a8_style_temp_dir(tag: &str) -> PathBuf {
    let dir = std::env::temp_dir().join(format!(
        "supercode-tr10-{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
}

#[test]
fn dev04_freshness_matrix_fresh_keeps_stub_stale_rehydrates_from_sidecar() {
    let dir = a8_style_temp_dir("freshness");
    let file_path = dir.join("f.txt");
    let content = filler(20_000);
    std::fs::write(&file_path, &content).unwrap();

    let msgs = vec![
        write_call("w1", file_path.to_str().unwrap(), &content),
        write_ok_result("w1", file_path.to_str().unwrap(), content.len()),
    ];
    let (view, log) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());
    assert_eq!(log.reductions.len(), 1);
    let r = &log.reductions[0];
    let content_hash = match &r.kind {
        ReductionKind::ToolInputElided { content_hash, .. } => content_hash.clone(),
        other => panic!("expected ToolInputElided, got {other:?}"),
    };

    // ---- (i) disk still matches -> KeepStub. ----
    assert!(probe_tool_input_fresh(&file_path, &content_hash));
    assert_eq!(
        tool_input_escalation_action(probe_tool_input_fresh(&file_path, &content_hash)),
        EscalationAction::KeepStub
    );
    // The stub is still exactly what a "keep" decision leaves in place.
    assert_eq!(args_of(&view, 0), args_of(&view, 0));

    // ---- (ii) disk changes since -> RehydrateFromSidecar, and rehydrating
    // (invert) really does restore the original from the sidecar. ----
    std::fs::write(&file_path, format!("{content}-modified-on-disk")).unwrap();
    assert!(!probe_tool_input_fresh(&file_path, &content_hash));
    assert_eq!(
        tool_input_escalation_action(probe_tool_input_fresh(&file_path, &content_hash)),
        EscalationAction::RehydrateFromSidecar
    );
    let sidecar = session_of(msgs.clone());
    let inverted = invert(&view, &log, &sidecar).unwrap();
    assert_eq!(
        args_of(&inverted, 0),
        args_of(&msgs, 0),
        "rehydrating from the sidecar must restore the exact original write, \
         independent of what disk looks like now"
    );

    // ---- (iii) file deleted -> never fresh either (fails closed). ----
    std::fs::remove_file(&file_path).unwrap();
    assert!(!probe_tool_input_fresh(&file_path, &content_hash));

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

// ---- dev/05: export purity (A11 leak guard) ---------------------------------

#[test]
fn dev05_export_never_leaks_the_stub_original_args_present() {
    let big = filler(20_000);
    let msgs = vec![
        ChatMessage::user("write it"),
        write_call("w1", "out.txt", &big),
        write_ok_result("w1", "out.txt", big.len()),
    ];
    // Sanity: this really does mint a ToolInputElided reduction (otherwise
    // the "no sentinel" assertion below would be vacuously true).
    let (_, log) = project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());
    assert_eq!(log.reductions.len(), 1);

    // Export ALWAYS reads the sidecar (the unreduced messages), never the
    // reduced view or the log — build the sidecar JSONL the same way
    // `export_sidecar.rs`/`rehydrate.rs` do: an empty imported prefix plus
    // every message as an APPENDED `NativeTurn` record (`to_native_jsonl_v2`
    // — unlike `to_native_jsonl`, which serializes `raw`, not `messages`).
    let empty = Session::from_claude_code_str("").unwrap();
    let native = empty.to_native_jsonl_v2(&msgs);

    let exported =
        supercode_harness::reduce::export_session(&native, supercode_harness::SessionFormat::Codex)
            .expect("export must succeed on a genuinely unreduced sidecar");
    assert_eq!(
        exported.matches(REDUCTION_SENTINEL).count(),
        0,
        "exported transcript must contain zero stubs"
    );
    assert!(
        exported.contains(&big),
        "exported transcript must contain the ORIGINAL full write content"
    );

    let exported_spliced = supercode_harness::reduce::export_session_spliced(
        &native,
        supercode_harness::SessionFormat::Codex,
        None,
    )
    .expect("spliced export must succeed on a genuinely unreduced sidecar");
    assert_eq!(exported_spliced.matches(REDUCTION_SENTINEL).count(), 0);
    assert!(exported_spliced.contains(&big));
}

// ---- MCP per-tool opt-in -----------------------------------------------------

#[test]
fn mcp_tool_only_elided_when_opted_in_via_policy() {
    let big = filler(20_000);
    let mcp_call = ChatMessage {
        role: Role::Assistant,
        content: None,
        content_parts: None,
        tool_calls: Some(vec![ToolCall {
            id: "m1".to_string(),
            kind: "function".to_string(),
            function: FunctionCall {
                name: "mcp__blobstore__put".to_string(),
                arguments: serde_json::json!({ "key": "asset/1", "data": big }).to_string(),
            },
        }]),
        tool_call_id: None,
        name: None,
        metadata: Default::default(),
    };
    let msgs = vec![
        mcp_call,
        ChatMessage::tool_result("m1", "mcp__blobstore__put", "ok"),
    ];

    // Not opted in: default policy never touches it.
    let (_, log_default) =
        project_messages(&msgs, &ReductionPolicy::default(), &ReductionLog::default());
    assert!(
        log_default.reductions.is_empty(),
        "an MCP tool must never be elided without an explicit opt-in"
    );

    // Opted in: the same call is now a candidate.
    let mut policy = ReductionPolicy::default();
    policy
        .tool_input_elidable_fields
        .insert("mcp__blobstore__put".to_string(), "data".to_string());
    let (view, log) = project_messages(&msgs, &policy, &ReductionLog::default());
    assert_eq!(log.reductions.len(), 1);
    let reduced = parsed_args(&view, 0);
    assert!(reduced
        .get("data")
        .unwrap()
        .as_str()
        .unwrap()
        .contains(REDUCTION_SENTINEL));
    assert_eq!(reduced.get("key").unwrap().as_str().unwrap(), "asset/1");
}