supercode-harness 0.4.5

//! Acceptance tests for SPEC.md TR-1 (T12 — model-invocable rehydration):
//! the `expand_reduction`/`sidecar_search` agent intrinsics.
//!
//! Each test is labeled with the TR-1.md acceptance-criterion id it proves.
//! Follows the ScriptedProvider/Recorder patterns of `tests/agent_loop.rs`,
//! `tests/tool_deferral.rs` (B6), and `tests/reduce_loop.rs` (A7/A9/A10).

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

use async_trait::async_trait;
use supercode_harness::reduce::rehydrate::{
    expand_reduction, sidecar_search, ExpandOutcome, SidecarSearchResult,
};
use supercode_harness::reduce::{
    export_session, project_messages, reduction_id, ReductionKind, ReductionLog, ReductionPolicy,
    REDUCTION_SENTINEL,
};
use supercode_harness::session::{Session, SessionFormat};
use supercode_harness::sidecar::SidecarWriter;
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-rehydrate-{tag}-{}-{}",
        std::process::id(),
        N.fetch_add(1, Ordering::SeqCst)
    ));
    std::fs::create_dir_all(&dir).unwrap();
    dir
}

/// Build an assistant message issuing one tool call, matching the
/// `tool_deferral.rs`/`reduce_loop.rs` inline-construction idiom (no public
/// `ChatMessage` constructor exists for an assistant tool call).
fn tool_call_msg(id: &str, name: &str, args: serde_json::Value) -> 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: name.to_string(),
                arguments: args.to_string(),
            },
        }]),
        tool_call_id: None,
        name: None,
        metadata: Default::default(),
    }
}

/// A deterministic 20,000-byte tool output whose first 4,096 bytes are `Q`
/// (the A7 kept prefix) and which carries a distinctive `needle` string well
/// past that prefix (in the portion A7 hides).
fn big_output_with_needle(needle: &str) -> String {
    let mut s = "Q".repeat(4096);
    s.push_str(&"x".repeat(10_000));
    s.push_str(needle);
    s.push_str(&"y".repeat(20_000 - s.len()));
    debug_assert_eq!(s.len(), 20_000);
    s
}

/// A `list_dir`-shaped tool returning a fixed big output — same idiom as
/// `reduce_loop.rs::BigOutputTool`.
struct BigOutputTool(String);
#[async_trait]
impl supercode_harness::tools::Tool for BigOutputTool {
    fn name(&self) -> &str {
        "list_dir"
    }
    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())
    }
}

// ---------------------------------------------------------------------------
// dev/01 + dev/04: schemas advertised; sidecar_search finds hidden content;
// expand_reduction returns byte-exact content, whole and ranged.
// ---------------------------------------------------------------------------

/// Turn 0: call `list_dir` (the big-output tool). Turn 1: `sidecar_search`
/// for a needle buried in the hidden tail. Turn 2: `expand_reduction` (whole).
/// Turn 3: `expand_reduction` with a `byte_range`. Turn 4: plain answer.
struct SearchThenExpand {
    calls: AtomicUsize,
    found_id: Mutex<Option<String>>,
}
#[async_trait]
impl Provider for SearchThenExpand {
    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);
        match n {
            0 => Ok((
                tool_call_msg("c1", "list_dir", serde_json::json!({})),
                Usage::default(),
            )),
            1 => {
                // dev/01: a reduced session advertises both intrinsics.
                assert!(
                    req.tools.iter().any(|t| t.name == "expand_reduction"),
                    "expand_reduction must be advertised once a ReductionPolicy is installed"
                );
                assert!(
                    req.tools.iter().any(|t| t.name == "sidecar_search"),
                    "sidecar_search must be advertised once a ReductionPolicy is installed"
                );
                // The needle must not be visible in the wire body (it's past
                // the A7 kept prefix).
                let body = serde_json::to_string(&req.messages).unwrap();
                assert!(body.contains(REDUCTION_SENTINEL));
                assert!(!body.contains("NEEDLE-XYZ-123"));
                Ok((
                    tool_call_msg(
                        "s1",
                        "sidecar_search",
                        serde_json::json!({"query": "NEEDLE-XYZ-123"}),
                    ),
                    Usage::default(),
                ))
            }
            2 => {
                // dev/04: sidecar_search's result names the hiding reduction.
                let last = req.messages.last().unwrap();
                assert_eq!(last.role, Role::Tool);
                let result: SidecarSearchResult =
                    serde_json::from_str(last.content.as_deref().unwrap()).unwrap();
                assert_eq!(
                    result.matches.len(),
                    1,
                    "expected exactly one match: {result:?}"
                );
                assert!(!result.truncated);
                assert_eq!(result.matches[0].kind, "tool-output");
                assert!(result.matches[0].snippet.contains("NEEDLE-XYZ-123"));
                *self.found_id.lock().unwrap() = Some(result.matches[0].reduction_id.clone());
                Ok((
                    tool_call_msg(
                        "e1",
                        "expand_reduction",
                        serde_json::json!({"reduction_id": result.matches[0].reduction_id}),
                    ),
                    Usage::default(),
                ))
            }
            3 => {
                // Next call: fetch a `byte_range` slice of the same reduction
                // (byte-exactness of both calls is checked afterward, against
                // `agent.history()` — see the test body's comment on why that
                // beats inspecting THIS request's view: a further oversized,
                // just-created tool result is itself an immediate A7
                // candidate under `protect_last_n_tool_results: 0`, which is
                // the dev/05 behavior, not what this test is checking).
                let id = self.found_id.lock().unwrap().clone().unwrap();
                Ok((
                    tool_call_msg(
                        "e2",
                        "expand_reduction",
                        serde_json::json!({"reduction_id": id, "byte_range": [0, 4]}),
                    ),
                    Usage::default(),
                ))
            }
            _ => Ok((ChatMessage::assistant("done"), Usage::default())),
        }
    }
}

#[tokio::test]
async fn dev01_dev04_search_then_expand_whole_and_ranged() {
    let dir = temp_dir("search-expand");
    let original = big_output_with_needle("NEEDLE-XYZ-123");

    let config = Config::builder().cwd(dir.clone()).build();
    let mut reg = supercode_harness::tools::ToolRegistry::new();
    reg.register(BigOutputTool(original.clone()));
    let mut agent = Agent::with_parts(
        config,
        Box::new(SearchThenExpand {
            calls: AtomicUsize::new(0),
            found_id: Mutex::new(None),
        }),
        reg,
    );
    // `protect_last_n_tool_results: 0` so `list_dir`'s own output — the ONLY
    // tool result at the time `sidecar_search` needs to find it — is reduced
    // immediately (dev/04 depends on this: nothing to search until it is).
    agent.set_reduction_policy(ReductionPolicy {
        tool_output_keep_bytes: 4096,
        tool_output_trigger_bytes: 8192,
        protect_last_n_tool_results: 0,
        ..ReductionPolicy::default()
    });

    let reply = agent.send("investigate").await.unwrap();
    assert_eq!(reply, "done");

    // dev/01 byte-exactness, checked against `agent.history()` — the tool
    // result's OWN content the instant `run_expand_reduction` produced it,
    // not a later wire view (which a further, unrelated A7 pass is free to
    // re-truncate on a subsequent projection — that's dev/05, a different
    // property, exercised in its own test below).
    let result_for = |call_id: &str| {
        agent
            .history()
            .iter()
            .find(|m| m.role == Role::Tool && m.tool_call_id.as_deref() == Some(call_id))
            .unwrap_or_else(|| panic!("no tool result for call `{call_id}` in history"))
            .content
            .clone()
            .unwrap_or_default()
    };
    assert_eq!(
        result_for("e1"),
        original,
        "expand_reduction with no byte_range must return the exact original bytes"
    );
    // A ranged expand carries a one-line provenance header naming the slice
    // and the TRUE total (F2: the model needs the total to plan its next
    // slice), then exactly the requested bytes.
    let ranged = result_for("e2");
    let (header, body) = ranged
        .split_once('\n')
        .expect("ranged expand result should have a header line");
    assert!(
        header.contains("bytes 0..4 of 20000"),
        "ranged expand header must name the slice and the total: {header}"
    );
    assert_eq!(
        body, "QQQQ",
        "expand_reduction with byte_range=[0,4] must return exactly that slice"
    );

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

// ---------------------------------------------------------------------------
// dev/02: stub ids in the projected VIEW are sufficient, on their own, to
// expand every kind (A7/A8/A10) — extracted purely from stub text, the way a
// model reading only the view would.
// ---------------------------------------------------------------------------

fn read_call(id: &str, path: &Path) -> 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: "read_file".to_string(),
                arguments: serde_json::json!({"path": path.to_string_lossy()}).to_string(),
            },
        }]),
        tool_call_id: None,
        name: None,
        metadata: Default::default(),
    }
}

/// All ids embedded in `[sc-reduced ...]` stub lines within `text` — exactly
/// what a model reading only the projected view could extract, with no
/// access to the `ReductionLog` struct itself.
fn stub_ids_in(text: &str) -> Vec<String> {
    let re = regex::Regex::new(r"r\d{4}-[0-9a-f]{4}").unwrap();
    re.find_iter(text).map(|m| m.as_str().to_string()).collect()
}

#[test]
fn dev02_stub_ids_are_sufficient_to_expand_every_kind() {
    let dir = temp_dir("dev02-stub-ids");
    let file_path = dir.join("f.txt");
    let fresh_content = "b".repeat(4096); // < 8192 trigger: never an A7 candidate.
    std::fs::write(&file_path, &fresh_content).unwrap();

    // 4 old (user, assistant) filler turns — the block A10 will clear.
    let mut msgs = Vec::new();
    for i in 0..4 {
        msgs.push(ChatMessage::user(format!("filler {i}")));
        msgs.push(ChatMessage::assistant(format!("filler reply {i}")));
    }
    let old_block_len = msgs.len(); // 8

    // Recent tail, never touched by A10: a fresh file read (A8 candidate)
    // and an oversized tool output (A7 candidate).
    msgs.push(read_call("rc1", &file_path));
    msgs.push(ChatMessage::tool_result(
        "rc1",
        "read_file",
        fresh_content.clone(),
    ));
    let big_output = "z".repeat(50_000);
    msgs.push(ChatMessage::tool_result(
        "bc1",
        "big_tool",
        big_output.clone(),
    ));
    msgs.push(ChatMessage::user("what did you find?"));
    msgs.push(ChatMessage::assistant("let me check"));
    let tail_len = msgs.len() - old_block_len; // 5

    // threshold/keep_recent chosen so the kept tail (5 messages) survives
    // whole and only the old block (8 messages) is cleared.
    let threshold = 10; // keep_recent = max(threshold/2, 2) = 5
    assert_eq!(tail_len, 5);

    let freshness = supercode_harness::reduce::probe_read_freshness(&msgs);
    let policy = ReductionPolicy {
        tool_output_keep_bytes: 4096,
        tool_output_trigger_bytes: 8192,
        protect_last_n_tool_results: 0,
        elide_stale_reads: true,
        read_freshness: freshness,
        clear_turns_older_than: Some(threshold),
        ..ReductionPolicy::default()
    };
    let (view, log) = project_messages(&msgs, &policy, &ReductionLog::default());

    // Sanity: exactly one of each kind present, none subsumed by A10.
    let kinds: Vec<&ReductionKind> = log.reductions.iter().map(|r| &r.kind).collect();
    assert!(kinds
        .iter()
        .any(|k| matches!(k, ReductionKind::ToolOutputTruncated { .. })));
    assert!(kinds
        .iter()
        .any(|k| matches!(k, ReductionKind::FileReadElided { .. })));
    assert!(kinds
        .iter()
        .any(|k| matches!(k, ReductionKind::TurnsCleared { .. })));
    assert_eq!(log.reductions.len(), 3, "{log:#?}");

    // The model's-eye view: render every message's visible text and extract
    // ids purely from stub TEXT (never touch `log` directly here).
    let rendered: String = view
        .iter()
        .map(|m| m.content.clone().unwrap_or_default())
        .collect::<Vec<_>>()
        .join("\n");
    let ids = stub_ids_in(&rendered);
    assert_eq!(ids.len(), 3, "expected one stub id per kind: {rendered}");

    for id in &ids {
        let outcome = expand_reduction(&log, &msgs, None, id, None)
            .unwrap_or_else(|e| panic!("expand_reduction({id}) failed: {e}"));
        let r = log.reductions.iter().find(|r| &r.id == id).unwrap();
        match &r.kind {
            ReductionKind::ToolOutputTruncated { .. } => {
                assert_eq!(outcome.content, big_output, "tool-output expand mismatch");
            }
            ReductionKind::FileReadElided { .. } => {
                assert_eq!(outcome.content, fresh_content, "file-read expand mismatch");
            }
            ReductionKind::TurnsCleared { .. } => {
                assert!(
                    outcome.content.contains("filler 0")
                        && outcome.content.contains("filler reply 3"),
                    "turns-cleared expand should render the cleared range: {}",
                    outcome.content
                );
            }
            other => panic!("unexpected kind: {other:?}"),
        }
    }

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

// ---------------------------------------------------------------------------
// dev/03: the expand_reduction call/result pair is honest transcript content
// — exports natively, no leak-guard special case, full-fidelity round trip.
// ---------------------------------------------------------------------------

/// Turn 0: call `list_dir`. Turn 1: `expand_reduction` on the resulting
/// stub. Turn 2: plain answer.
struct ExpandThenAnswer {
    calls: AtomicUsize,
}
#[async_trait]
impl Provider for ExpandThenAnswer {
    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);
        match n {
            0 => Ok((
                tool_call_msg("c1", "list_dir", serde_json::json!({})),
                Usage::default(),
            )),
            1 => {
                let body = serde_json::to_string(&req.messages).unwrap();
                let re = regex::Regex::new(r"r\d{4}-[0-9a-f]{4}").unwrap();
                let id = re.find(&body).expect("a stub id must be present").as_str();
                Ok((
                    tool_call_msg(
                        "e1",
                        "expand_reduction",
                        serde_json::json!({"reduction_id": id}),
                    ),
                    Usage::default(),
                ))
            }
            _ => Ok((ChatMessage::assistant("done"), Usage::default())),
        }
    }
}

#[tokio::test]
async fn dev03_expand_reduction_call_result_exports_honestly() {
    let dir = temp_dir("dev03-export");
    let sidecar_path = dir.join("sess.sidecar.jsonl");
    let original = "Q".repeat(20_000);

    let config = Config::builder().cwd(dir.clone()).build();
    let mut reg = supercode_harness::tools::ToolRegistry::new();
    reg.register(BigOutputTool(original.clone()));
    let mut agent = Agent::with_parts(
        config,
        Box::new(ExpandThenAnswer {
            calls: AtomicUsize::new(0),
        }),
        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()
    });

    let reply = agent.send("go").await.unwrap();
    assert_eq!(reply, "done");

    // `agent.history()` carries the expand_reduction call and its (byte-exact)
    // result as an ordinary tool call/result pair.
    let expand_call = agent
        .history()
        .iter()
        .find_map(|m| {
            m.tool_calls()
                .iter()
                .find(|c| c.function.name == "expand_reduction")
        })
        .expect("expand_reduction call must be in history");
    let expand_result = agent
        .history()
        .iter()
        .find(|m| {
            m.role == Role::Tool && m.tool_call_id.as_deref() == Some(expand_call.id.as_str())
        })
        .expect("matching expand_reduction result must be in history");
    assert_eq!(expand_result.content.as_deref(), Some(original.as_str()));

    let sidecar_jsonl = std::fs::read_to_string(&sidecar_path).unwrap();

    for format in [SessionFormat::ClaudeCode, SessionFormat::Codex] {
        let exported = export_session(&sidecar_jsonl, format)
            .unwrap_or_else(|e| panic!("export_session({format:?}) failed: {e}"));
        assert!(
            !exported.contains(REDUCTION_SENTINEL),
            "export_session({format:?}) leaked the reduction sentinel — the intrinsic pair \
             should need no leak-guard special case:\n{exported}"
        );

        let reloaded = Session::load_str(&exported, format)
            .unwrap_or_else(|e| panic!("reloading export_session({format:?}) failed: {e}"));
        let reloaded_call = reloaded
            .messages
            .iter()
            .find_map(|m| {
                m.tool_calls()
                    .iter()
                    .find(|c| c.function.name == "expand_reduction")
                    .cloned()
            })
            .unwrap_or_else(|| panic!("{format:?} reload lost the expand_reduction call"));
        let reloaded_result = reloaded
            .messages
            .iter()
            .find(|m| m.tool_call_id.as_deref() == Some(reloaded_call.id.as_str()))
            .unwrap_or_else(|| panic!("{format:?} reload lost the matching tool result"));
        assert_eq!(
            reloaded_result.content.as_deref(),
            Some(original.as_str()),
            "{format:?}: exported/reloaded expand result must still be byte-exact"
        );
    }

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

// ---------------------------------------------------------------------------
// dev/05: expand results are reduction-eligible — an oversized expand result
// is itself truncated-with-stub (A7) by the ordinary projection pass once it
// lands in history, and that fresh stub is re-expandable. No new machinery:
// this is exactly what SPEC.md TR-1's "no leak-guard special case needed"
// line implies (`reduce::rehydrate` never mints its own stubs).
// ---------------------------------------------------------------------------

#[test]
fn dev05_oversized_expand_result_is_itself_a7_truncated_and_reexpandable() {
    let big = "m".repeat(200_000);
    let msgs = vec![ChatMessage::tool_result("call_1", "bash", big.clone())];

    let policy = ReductionPolicy {
        tool_output_keep_bytes: 4096,
        tool_output_trigger_bytes: 8192,
        protect_last_n_tool_results: 0,
        ..ReductionPolicy::default()
    };
    let (_view1, log1) = project_messages(&msgs, &policy, &ReductionLog::default());
    assert_eq!(log1.reductions.len(), 1);
    let original_id = log1.reductions[0].id.clone();

    // The model calls expand_reduction on the original — gets the full,
    // un-truncated 200 KB back (this function never caps on its own).
    let ExpandOutcome { content, .. } =
        expand_reduction(&log1, &msgs, None, &original_id, None).unwrap();
    assert_eq!(
        content, big,
        "expand_reduction must return the exact original bytes"
    );

    // Simulated agent behavior: the expand result lands as an ordinary new
    // tool result in history (exactly what `Agent::run_loop` does with any
    // tool's returned string).
    let mut msgs2 = msgs.clone();
    msgs2.push(ChatMessage::tool_result(
        "call_expand",
        "expand_reduction",
        content.clone(),
    ));
    let expand_result_idx = msgs2.len() - 1;

    // A LATER projection (protect_last_n_tool_results: 0, so nothing is
    // exempt for being "newest") re-truncates it exactly like any other
    // oversized tool output — no bespoke code path involved.
    let (view2, log2) = project_messages(&msgs2, &policy, &log1);
    assert_eq!(
        log2.reductions.len(),
        2,
        "the original reduction reapplies verbatim, plus a fresh one over the expand result"
    );
    let fresh = log2
        .reductions
        .iter()
        .find(|r| r.ptr.addr.index == expand_result_idx)
        .expect("a fresh reduction must cover the just-expanded, still-oversized tool result");
    assert_ne!(
        fresh.id, original_id,
        "the fresh reduction gets its own new id"
    );
    assert!(fresh.placeholder.starts_with(REDUCTION_SENTINEL));
    assert!(
        fresh.placeholder.contains(&fresh.id),
        "the fresh stub carries its own id: {}",
        fresh.placeholder
    );

    // No unbounded regrowth: the reduced VIEW's copy of the expand result is
    // small (bounded by `tool_output_keep_bytes` plus the stub), not 200 KB.
    let reduced_copy = view2[expand_result_idx].content.clone().unwrap();
    assert!(
        reduced_copy.len() < 10_000,
        "the re-truncated view must stay small: {} bytes",
        reduced_copy.len()
    );

    // And it, too, is re-expandable — using the FRESH id — recovering the
    // exact same 200 KB.
    let outcome2 = expand_reduction(&log2, &msgs2, None, &fresh.id, None).unwrap();
    assert_eq!(
        outcome2.content, big,
        "re-expanding the fresh stub must be byte-exact"
    );
}

// ---------------------------------------------------------------------------
// dev/06: Tier-2 demo — a scripted agent loop where the model hits a
// question answerable only from A10-cleared content, calls `sidecar_search`
// then `expand_reduction`, and answers correctly. Transcript archived below.
// ---------------------------------------------------------------------------

const SECRET: &str = "the deploy key is DK-771-ZQ";

/// Turn 0: user shares a secret, model acknowledges. Turns 1-3: plain filler
/// exchanges (grows history past the A10 `compact_after_messages` threshold,
/// clearing the secret's turn out of the view). Turn 4 (the question, model
/// call n=4): the model can't see the secret directly — it calls
/// `sidecar_search`, then (n=5) `expand_reduction` on the match, then (n=6)
/// answers using the recovered content.
struct Tier2Demo {
    calls: AtomicUsize,
}
#[async_trait]
impl Provider for Tier2Demo {
    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);
        match n {
            0 => Ok((
                ChatMessage::assistant("Noted, I'll remember that."),
                Usage::default(),
            )),
            1..=3 => Ok((ChatMessage::assistant(format!("ack {n}")), Usage::default())),
            4 => {
                assert!(req.tools.iter().any(|t| t.name == "sidecar_search"));
                assert!(req.tools.iter().any(|t| t.name == "expand_reduction"));
                let body = serde_json::to_string(&req.messages).unwrap();
                assert!(
                    body.contains(REDUCTION_SENTINEL),
                    "the secret's turn must have been A10-cleared by now: {body}"
                );
                assert!(
                    !body.contains("DK-771-ZQ"),
                    "the secret must not be directly visible in the reduced view"
                );
                Ok((
                    tool_call_msg(
                        "s1",
                        "sidecar_search",
                        serde_json::json!({"query": "deploy key"}),
                    ),
                    Usage::default(),
                ))
            }
            5 => {
                let last = req.messages.last().unwrap();
                assert_eq!(last.role, Role::Tool);
                let result: SidecarSearchResult =
                    serde_json::from_str(last.content.as_deref().unwrap()).unwrap();
                assert_eq!(result.matches.len(), 1, "{result:?}");
                Ok((
                    tool_call_msg(
                        "e1",
                        "expand_reduction",
                        serde_json::json!({"reduction_id": result.matches[0].reduction_id}),
                    ),
                    Usage::default(),
                ))
            }
            _ => {
                let last = req.messages.last().unwrap();
                assert_eq!(last.role, Role::Tool);
                let expanded = last.content.clone().unwrap_or_default();
                assert!(
                    expanded.contains("DK-771-ZQ"),
                    "expand_reduction must have recovered the secret: {expanded}"
                );
                Ok((
                    ChatMessage::assistant("The deploy key is DK-771-ZQ."),
                    Usage::default(),
                ))
            }
        }
    }
}

#[tokio::test]
async fn dev06_tier2_demo_sidecar_search_then_expand_answers_from_cleared_turns() {
    let dir = temp_dir("dev06-demo");
    let sidecar_path = dir.join("demo.sidecar.jsonl");

    let config = Config::builder()
        .cwd(dir.clone())
        .compact_after_messages(6) // keep_recent = max(6/2, 2) = 3
        .build();
    let mut agent = Agent::with_provider(
        config,
        Box::new(Tier2Demo {
            calls: AtomicUsize::new(0),
        }),
    );

    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(format!("Please remember this: {SECRET}"))
        .await
        .unwrap();
    for i in 0..3 {
        agent
            .send(format!("filler turn {i}, just say ack"))
            .await
            .unwrap();
    }
    // A10 must have cleared the secret's turn out of the log by now.
    assert!(
        agent
            .reduction_log()
            .reductions
            .iter()
            .any(|r| matches!(r.kind, ReductionKind::TurnsCleared { .. })),
        "expected a TurnsCleared reduction before the question turn: {:?}",
        agent.reduction_log()
    );

    let reply = agent.send("What was the deploy key?").await.unwrap();
    assert_eq!(reply, "The deploy key is DK-771-ZQ.");

    // Archive the demo transcript (git-ignored `target/`; referenced in the
    // TR-1 build report) — every message the scripted loop actually
    // exchanged, in order, proving the sidecar_search -> expand_reduction ->
    // correct-answer chain end to end.
    let workspace_target = Path::new(env!("CARGO_MANIFEST_DIR")).join("../../target");
    std::fs::create_dir_all(&workspace_target).ok();
    let transcript_path = workspace_target.join("tr1-demo-transcript.jsonl");
    agent.save_transcript(&transcript_path).unwrap();
    assert!(transcript_path.exists());
    eprintln!(
        "dev/06 demo transcript archived at {}",
        transcript_path.display()
    );

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

// ---------------------------------------------------------------------------
// B1 (fix pass) / TR-12: a tool output over `max_tool_output_bytes` (default
// 100 KB), with a live recorder + policy both active. Pre-TR-12, this
// diverged history from the sidecar (history kept a `cap_tool_output`-capped
// copy; the recorder kept the full bytes), and the intrinsics needed the
// `recorded`-upgrade path (`reduce::rehydrate`'s two-source resolution) to
// reach past the cap. TR-12's D6/A7 supersession gate (`Agent::run_loop`)
// means a recorder+policy pair now keeps `cap_tool_output` off entirely, so
// for a freshly-recorded session like this one `history` already holds the
// full bytes and never diverges from the sidecar — the `recorded` upgrade is
// exercised here but is a provable no-op (`recorded == minted`). The
// two-source resolution itself remains load-bearing as defense-in-depth for
// a LEGACY sidecar recorded before this gate existed (unit-tested directly,
// with a hand-built capped/full divergence, in
// `crates/harness/src/reduce/rehydrate.rs`'s
// `recorded_copy_supersedes_a_capped_minted_copy`).
// ---------------------------------------------------------------------------

/// A 150,000-byte output: 4,096 `Q`s (the A7 kept prefix), filler, then a
/// needle at byte 120,000 — past BOTH the A7 keep boundary and the 100 KB
/// `cap_tool_output` cap, so only the recorded (sidecar) copy contains it.
fn over_cap_output_with_needle(needle: &str) -> String {
    let mut s = "Q".repeat(4096);
    s.push_str(&"x".repeat(120_000 - s.len()));
    s.push_str(needle);
    s.push_str(&"y".repeat(150_000 - s.len()));
    s
}

/// Turn 0: call the 150 KB tool. Turn 1: `sidecar_search` for the needle
/// beyond the cap. Turn 2: `expand_reduction` on the match. Turn 3: answer.
struct OverCapSearchThenExpand {
    calls: AtomicUsize,
}
#[async_trait]
impl Provider for OverCapSearchThenExpand {
    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);
        match n {
            0 => Ok((
                tool_call_msg("c1", "list_dir", serde_json::json!({})),
                Usage::default(),
            )),
            1 => {
                let body = serde_json::to_string(&req.messages).unwrap();
                assert!(body.contains(REDUCTION_SENTINEL));
                assert!(
                    !body.contains("TAIL-NEEDLE-B1"),
                    "the needle lives past the cap; it must not be visible in the view"
                );
                Ok((
                    tool_call_msg(
                        "s1",
                        "sidecar_search",
                        serde_json::json!({"query": "TAIL-NEEDLE-B1"}),
                    ),
                    Usage::default(),
                ))
            }
            2 => {
                // The needle sits at byte 120,000 — beyond the 100 KB capped
                // history copy. Finding it proves search resolved the
                // recorder's full bytes, not history's capped copy.
                let last = req.messages.last().unwrap();
                assert_eq!(last.role, Role::Tool);
                let result: SidecarSearchResult =
                    serde_json::from_str(last.content.as_deref().unwrap()).unwrap();
                assert_eq!(
                    result.matches.len(),
                    1,
                    "search must reach past the cap into the recorded bytes: {result:?}"
                );
                assert!(result.matches[0].snippet.contains("TAIL-NEEDLE-B1"));
                Ok((
                    tool_call_msg(
                        "e1",
                        "expand_reduction",
                        serde_json::json!({"reduction_id": result.matches[0].reduction_id}),
                    ),
                    Usage::default(),
                ))
            }
            _ => Ok((ChatMessage::assistant("done"), Usage::default())),
        }
    }
}

#[tokio::test]
async fn b1_over_cap_output_expands_to_full_recorded_bytes() {
    let dir = temp_dir("b1-over-cap");
    let sidecar_path = dir.join("sess.sidecar.jsonl");
    let original = over_cap_output_with_needle("TAIL-NEEDLE-B1");
    assert_eq!(original.len(), 150_000);

    // DEFAULT config: `max_tool_output_bytes` stays at its 100 KB default —
    // pre-TR-12 this was exactly the regime where history and sidecar
    // diverged; post-TR-12, the D6/A7 gate keeps `cap_tool_output` off here
    // (recorder + policy both installed below), so they no longer do.
    let config = Config::builder().cwd(dir.clone()).build();
    let mut reg = supercode_harness::tools::ToolRegistry::new();
    reg.register(BigOutputTool(original.clone()));
    let mut agent = Agent::with_parts(
        config,
        Box::new(OverCapSearchThenExpand {
            calls: AtomicUsize::new(0),
        }),
        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()
    });

    let reply = agent.send("go").await.unwrap();
    assert_eq!(reply, "done");

    // history's copy of the 150 KB output is now the FULL original — the
    // D6/A7 gate (TR-12) keeps `cap_tool_output` off whenever a recorder and
    // a policy are both active, so `history` never diverges from the
    // sidecar for a freshly-recorded session like this one. No cap notice is
    // ever appended: `history` holds exactly what the tool returned.
    let history_copy = agent
        .history()
        .iter()
        .find(|m| m.role == Role::Tool && m.tool_call_id.as_deref() == Some("c1"))
        .and_then(|m| m.content.clone())
        .unwrap();
    assert_eq!(
        history_copy.len(),
        original.len(),
        "the gate must keep the FULL bytes in history, not a capped prefix"
    );
    assert_eq!(history_copy, original);
    assert!(
        !history_copy.contains("bytes total, showing first"),
        "no cap notice should ever be appended once the D6/A7 gate is active: {}",
        &history_copy[history_copy.len().saturating_sub(200)..]
    );

    // The expand result — resolved from `history[1..]` (now full, since the
    // gate is on) and cross-checked against the recorder's reload — is the
    // full 150 KB original, byte-exact, whether or not the `recorded` upgrade
    // path ever had to do anything (it doesn't, here — see the comment
    // above this test).
    let sidecar_raw = std::fs::read_to_string(&sidecar_path).unwrap();
    let recorded = Session::from_native_str(&sidecar_raw).unwrap();
    let expand_result = recorded
        .messages
        .iter()
        .find(|m| m.role == Role::Tool && m.tool_call_id.as_deref() == Some("e1"))
        .and_then(|m| m.content.clone())
        .expect("the expand result must be recorded in the sidecar");
    assert_eq!(
        expand_result, original,
        "expand_reduction must return the FULL recorded bytes, not the capped history copy"
    );

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

// ---------------------------------------------------------------------------
// B3 (fix pass): malformed byte_range shapes error loudly through the real
// agent dispatch — never a silent whole-content (or empty) return.
// ---------------------------------------------------------------------------

/// Issues one `expand_reduction` call per malformed `byte_range` shape,
/// asserting the previous call's tool result was a recoverable error naming
/// `byte_range` (and, where reachable, the true total).
struct MalformedRangeProbe {
    calls: AtomicUsize,
    found_id: Mutex<Option<String>>,
}

impl MalformedRangeProbe {
    /// Every malformed shape probed, in order: wrong arity (both ways),
    /// non-integers, negatives, floats — all shape errors — then a reversed
    /// range (well-formed shape, semantically invalid).
    fn shapes() -> Vec<serde_json::Value> {
        vec![
            serde_json::json!([1]),
            serde_json::json!([1, 2, 3]),
            serde_json::json!(["a", "b"]),
            serde_json::json!([-5, 10]),
            serde_json::json!([1.5, 2]),
            serde_json::json!([100, 5]),
        ]
    }
}

#[async_trait]
impl Provider for MalformedRangeProbe {
    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);
        let shapes = Self::shapes();
        if n == 0 {
            return Ok((
                tool_call_msg("c1", "list_dir", serde_json::json!({})),
                Usage::default(),
            ));
        }
        if n == 1 {
            // Pull the freshly-minted reduction id out of the request view's
            // stub text, the way a model would.
            let body = serde_json::to_string(&req.messages).unwrap();
            let re = regex::Regex::new(r"r\d{4}-[0-9a-f]{4}").unwrap();
            let id = re.find(&body).expect("a stub id must be present").as_str();
            *self.found_id.lock().unwrap() = Some(id.to_string());
        } else {
            // Every probe's reply must be a recoverable error naming
            // byte_range — and for these all-shape-error cases (probes
            // before the last), the true total too.
            let last = req.messages.last().unwrap();
            assert_eq!(last.role, Role::Tool);
            let content = last.content.as_deref().unwrap_or_default();
            assert!(
                content.starts_with("Error:"),
                "probe {} must produce an error result: {content}",
                n - 2
            );
            assert!(
                content.contains("byte_range"),
                "probe {} error must name byte_range: {content}",
                n - 2
            );
            let is_reversed_probe = n - 2 == shapes.len() - 1;
            if is_reversed_probe {
                assert!(
                    content.contains("reversed"),
                    "the reversed-range error must say so: {content}"
                );
            }
            assert!(
                content.contains("20000"),
                "probe {} error must name the true total: {content}",
                n - 2
            );
        }
        let probe = n - 1;
        if probe < shapes.len() {
            let id = self.found_id.lock().unwrap().clone().unwrap();
            Ok((
                tool_call_msg(
                    &format!("p{probe}"),
                    "expand_reduction",
                    serde_json::json!({"reduction_id": id, "byte_range": shapes[probe]}),
                ),
                Usage::default(),
            ))
        } else {
            Ok((ChatMessage::assistant("survived"), Usage::default()))
        }
    }
}

#[tokio::test]
async fn b3_malformed_byte_ranges_error_recoverably_through_dispatch() {
    let dir = temp_dir("b3-ranges");
    let config = Config::builder().cwd(dir.clone()).build();
    let mut reg = supercode_harness::tools::ToolRegistry::new();
    reg.register(BigOutputTool("Q".repeat(20_000)));
    let mut agent = Agent::with_parts(
        config,
        Box::new(MalformedRangeProbe {
            calls: AtomicUsize::new(0),
            found_id: Mutex::new(None),
        }),
        reg,
    );
    agent.set_reduction_policy(ReductionPolicy {
        tool_output_keep_bytes: 4096,
        tool_output_trigger_bytes: 8192,
        protect_last_n_tool_results: 0,
        ..ReductionPolicy::default()
    });

    // The loop must survive every malformed probe and reach the final
    // answer — errors are recoverable tool results, not aborts.
    let reply = agent.send("probe").await.unwrap();
    assert_eq!(reply, "survived");

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

// ---------------------------------------------------------------------------
// B2/F4 (fix pass): empty sidecar_search queries, unknown reduction ids, and
// unparseable arguments all surface as recoverable is_error tool results
// through the real dispatch — the loop continues every time.
// ---------------------------------------------------------------------------

/// n=0: expand_reduction with an unknown id. n=1: expand_reduction with
/// arguments that aren't JSON at all. n=2: sidecar_search with no `query`.
/// n=3: sidecar_search with a whitespace query. n=4: final answer. Each
/// step first asserts the previous step's result was an `Error:` tool
/// result carrying the expected hint.
struct ErrorPathProbe {
    calls: AtomicUsize,
}
#[async_trait]
impl Provider for ErrorPathProbe {
    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);
        let last_error = |hint: &str| {
            let last = req.messages.last().unwrap();
            assert_eq!(last.role, Role::Tool);
            let content = last.content.as_deref().unwrap_or_default();
            assert!(
                content.starts_with("Error:"),
                "expected an error result, got: {content}"
            );
            assert!(
                content.contains(hint),
                "error should mention `{hint}`: {content}"
            );
        };
        match n {
            0 => Ok((
                tool_call_msg(
                    "x1",
                    "expand_reduction",
                    serde_json::json!({"reduction_id": "r9999-dead"}),
                ),
                Usage::default(),
            )),
            1 => {
                last_error("r9999-dead");
                // Arguments that fail JSON parsing entirely.
                let call = ChatMessage {
                    role: Role::Assistant,
                    content: None,
                    content_parts: None,
                    tool_calls: Some(vec![ToolCall {
                        id: "x2".to_string(),
                        kind: "function".to_string(),
                        function: FunctionCall {
                            name: "expand_reduction".to_string(),
                            arguments: "not json".to_string(),
                        },
                    }]),
                    tool_call_id: None,
                    name: None,
                    metadata: Default::default(),
                };
                Ok((call, Usage::default()))
            }
            2 => {
                last_error("expand_reduction");
                Ok((
                    tool_call_msg("x3", "sidecar_search", serde_json::json!({})),
                    Usage::default(),
                ))
            }
            3 => {
                last_error("query");
                Ok((
                    tool_call_msg("x4", "sidecar_search", serde_json::json!({"query": "   "})),
                    Usage::default(),
                ))
            }
            _ => {
                last_error("query");
                Ok((ChatMessage::assistant("recovered"), Usage::default()))
            }
        }
    }
}

#[tokio::test]
async fn b2_f4_error_paths_are_recoverable_through_the_loop() {
    let dir = temp_dir("error-paths");
    let config = Config::builder().cwd(dir.clone()).build();
    let mut agent = Agent::with_provider(
        config,
        Box::new(ErrorPathProbe {
            calls: AtomicUsize::new(0),
        }),
    );
    agent.set_reduction_policy(ReductionPolicy::default());

    let reply = agent.send("probe the error paths").await.unwrap();
    assert_eq!(reply, "recovered");

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

// ---------------------------------------------------------------------------
// B4 (fix pass): content living ONLY in tool-call arguments inside an
// A10-cleared range is both searchable and present in the expansion.
// ---------------------------------------------------------------------------

#[test]
fn b4_cleared_tool_call_arguments_are_searchable_and_expandable() {
    // An old block containing a write_file call whose file body exists
    // NOWHERE except the call's arguments, plus its result — then filler so
    // the block ages past the clear threshold.
    let mut msgs = vec![
        ChatMessage::user("write my notes file"),
        tool_call_msg(
            "w1",
            "write_file",
            serde_json::json!({
                "path": "notes.txt",
                "content": "ARGS-ONLY-PAYLOAD-77: the real content of the file"
            }),
        ),
        ChatMessage::tool_result("w1", "write_file", "ok, wrote notes.txt"),
        ChatMessage::assistant("Written."),
    ];
    for i in 0..4 {
        msgs.push(ChatMessage::user(format!("filler {i}")));
        msgs.push(ChatMessage::assistant(format!("filler reply {i}")));
    }

    // threshold 8 -> keep_recent = 4: the write_file block (first 4
    // messages) is cleared; the last 4 filler messages survive.
    let policy = ReductionPolicy {
        clear_turns_older_than: Some(8),
        ..ReductionPolicy::default()
    };
    let (_view, log) = project_messages(&msgs, &policy, &ReductionLog::default());
    let cleared = log
        .reductions
        .iter()
        .find(|r| matches!(r.kind, ReductionKind::TurnsCleared { .. }))
        .expect("the old block must have been cleared");

    // Searchable: the payload exists only in the cleared call's ARGUMENTS.
    let hits = sidecar_search(&log, &msgs, None, "ARGS-ONLY-PAYLOAD-77").unwrap();
    assert_eq!(hits.matches.len(), 1, "{hits:?}");
    assert_eq!(hits.matches[0].reduction_id, cleared.id);
    assert!(hits.matches[0].snippet.contains("ARGS-ONLY-PAYLOAD-77"));

    // Expandable: the expansion carries the tool call's id, name, full
    // arguments, and the tool result's attribution.
    let outcome = expand_reduction(&log, &msgs, None, &cleared.id, None).unwrap();
    assert!(
        outcome.content.contains("ARGS-ONLY-PAYLOAD-77"),
        "expansion must include argument-only content: {}",
        outcome.content
    );
    assert!(outcome.content.contains("write_file"));
    assert!(outcome.content.contains("w1"));
    assert!(outcome.content.contains("tool_call_id=w1"));
    assert!(outcome.content.contains("ok, wrote notes.txt"));
}

// ---------------------------------------------------------------------------
// B5 (fix pass): the intrinsics are advertised ONLY when a ReductionPolicy
// is installed — the negative case, so the gate can't silently regress to
// unconditional advertising.
// ---------------------------------------------------------------------------

#[tokio::test]
async fn b5_intrinsics_absent_without_a_reduction_policy() {
    // No policy: neither intrinsic may be advertised.
    let names = std::sync::Arc::new(Mutex::new(None));
    /// Records the tool names of the first request, then answers.
    struct Shared(std::sync::Arc<Mutex<Option<Vec<String>>>>);
    #[async_trait]
    impl Provider for Shared {
        async fn complete(
            &self,
            req: &ChatRequest,
            _on_delta: &(dyn for<'a> Fn(&'a str) + Send + Sync),
        ) -> supercode_harness::Result<(ChatMessage, Usage)> {
            let mut names = self.0.lock().unwrap();
            if names.is_none() {
                *names = Some(req.tools.iter().map(|t| t.name.clone()).collect());
            }
            Ok((ChatMessage::assistant("ok"), Usage::default()))
        }
    }

    let config = Config::builder().build();
    let mut agent = Agent::with_provider(config, Box::new(Shared(names.clone())));
    agent.send("hi").await.unwrap();
    let seen = names.lock().unwrap().clone().unwrap();
    assert!(
        !seen.iter().any(|n| n == "expand_reduction"),
        "expand_reduction must not be advertised without a policy: {seen:?}"
    );
    assert!(
        !seen.iter().any(|n| n == "sidecar_search"),
        "sidecar_search must not be advertised without a policy: {seen:?}"
    );

    // Same agent shape WITH a policy: both appear (the positive control,
    // proving this test would catch a regression in either direction).
    let names2 = std::sync::Arc::new(Mutex::new(None));
    let config = Config::builder().build();
    let mut agent = Agent::with_provider(config, Box::new(Shared(names2.clone())));
    agent.set_reduction_policy(ReductionPolicy::default());
    agent.send("hi").await.unwrap();
    let seen = names2.lock().unwrap().clone().unwrap();
    assert!(seen.iter().any(|n| n == "expand_reduction"), "{seen:?}");
    assert!(seen.iter().any(|n| n == "sidecar_search"), "{seen:?}");
}

// ---------------------------------------------------------------------------
// F1 (fix pass): an ImageRedacted reduction's id is reachable from the
// model's-eye view (image stubs land in `content_parts` as a text part) and
// expands to the original data: URL.
// ---------------------------------------------------------------------------

#[test]
fn f1_image_reduction_id_reachable_from_view_and_expands() {
    // A 20 KB data: URL — over the default `image_redact_min_bytes` (8192).
    let data_url = format!("data:image/png;base64,{}", "A".repeat(20_000));
    let msgs = vec![
        ChatMessage::user_with_images("look at this screenshot", std::slice::from_ref(&data_url)),
        ChatMessage::assistant("Looking."),
    ];

    // Default policy: `redact_images` is on by default.
    let policy = ReductionPolicy::default();
    let (view, log) = project_messages(&msgs, &policy, &ReductionLog::default());
    assert_eq!(log.reductions.len(), 1, "{log:#?}");
    assert!(matches!(
        log.reductions[0].kind,
        ReductionKind::ImageRedacted { .. }
    ));

    // The model's-eye rendering MUST include `content_parts` text — that is
    // where image stubs land (the redacted part becomes a `{"type":"text"}`
    // part carrying the stub). Extract the id from that rendering alone.
    let rendered: String = view
        .iter()
        .map(|m| {
            let mut s = m.content.clone().unwrap_or_default();
            for part in m.content_parts.as_deref().unwrap_or_default() {
                if let Some(text) = part.get("text").and_then(|t| t.as_str()) {
                    s.push('\n');
                    s.push_str(text);
                }
            }
            s
        })
        .collect::<Vec<_>>()
        .join("\n");
    assert!(
        !rendered.contains(&data_url),
        "the data URL itself must be gone from the view"
    );
    let ids = stub_ids_in(&rendered);
    assert_eq!(
        ids.len(),
        1,
        "the image stub id must be visible in the model's view: {rendered}"
    );
    assert_eq!(ids[0], log.reductions[0].id);

    // Expanding that id returns the original data: URL, byte-exact.
    let outcome = expand_reduction(&log, &msgs, None, &ids[0], None).unwrap();
    assert_eq!(outcome.content, data_url);
    assert_eq!(outcome.total_bytes, data_url.len());
}

// ---------------------------------------------------------------------------
// TR-3 dev/06: a `FileReadDiffed` stub carries its `reduction_id`, visible in
// the model's-eye view purely from stub text (same idiom as dev/02, above),
// and `expand_reduction` on that id returns the full, verbatim re-read — not
// the diff, not the base — byte for byte.
// ---------------------------------------------------------------------------

#[test]
fn tr3_dev06_file_read_diffed_stub_carries_id_and_expands_to_full_re_read() {
    use std::fmt::Write as _;

    let file_path = PathBuf::from("/workspace/src/tr3_dev06.rs");
    let mut base_content = String::with_capacity(5000);
    for i in 0..500 {
        writeln!(base_content, "line {i:04}").unwrap();
    }
    let mut new_content = String::with_capacity(base_content.len());
    for (i, line) in base_content.lines().enumerate() {
        if (250..252).contains(&i) {
            writeln!(new_content, "CHANGED {i}").unwrap();
        } else {
            writeln!(new_content, "{line}").unwrap();
        }
    }
    assert_ne!(base_content, new_content);

    let msgs = vec![
        ChatMessage::user("read the file"),
        read_call("rd1", &file_path),
        ChatMessage::tool_result("rd1", "read_file", base_content.clone()),
        ChatMessage::user("small edit"),
        ChatMessage::assistant("done"),
        read_call("rd2", &file_path),
        ChatMessage::tool_result("rd2", "read_file", new_content.clone()),
    ];
    let new_idx = 6;

    let policy = ReductionPolicy {
        tool_output_trigger_bytes: usize::MAX, // isolate TR-3 from A7
        protect_last_n_tool_results: 0,
        ..ReductionPolicy::default() // diff_rereads: true (default)
    };
    let (view, log) = project_messages(&msgs, &policy, &ReductionLog::default());

    assert_eq!(log.reductions.len(), 1, "{log:#?}");
    let r = &log.reductions[0];
    assert!(
        matches!(r.kind, ReductionKind::FileReadDiffed { .. }),
        "expected FileReadDiffed, got {:?}",
        r.kind
    );

    // The model's-eye view: the stub id is recoverable from stub TEXT alone.
    let rendered: String = view
        .iter()
        .map(|m| m.content.clone().unwrap_or_default())
        .collect::<Vec<_>>()
        .join("\n");
    let ids = stub_ids_in(&rendered);
    assert_eq!(ids.len(), 1, "{rendered}");
    assert_eq!(ids[0], r.id);
    assert_eq!(reduction_id(&view[new_idx]), Some(r.id.as_str()));

    // expand_reduction returns the full VERBATIM re-read -- not the diff
    // text, not the base's content.
    let outcome = expand_reduction(&log, &msgs, None, &ids[0], None).unwrap();
    assert_eq!(outcome.content, new_content);
    assert_eq!(outcome.total_bytes, new_content.len());
    assert_ne!(outcome.content, base_content);
}