supercode-harness 0.4.12

//! UX-30 dev/02 — `Agent::set_model` actually changes what's sent on the
//! wire, not just a label somewhere. A mock `Provider` records the
//! `req.model` it was handed on each call; this drives two real `send()`
//! turns through the same agent, switching the model in between via
//! `set_model`, and asserts the SECOND request carried the NEW model while
//! the first carried the original — proving `run_loop` reads
//! `Config::model` fresh per request rather than baking it in at
//! construction. Same `ScriptedProvider` idiom as `tests/agent_loop.rs`.

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

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

struct RecordingProvider {
    calls: AtomicUsize,
    seen_models: Arc<Mutex<Vec<String>>>,
}

#[async_trait]
impl Provider for RecordingProvider {
    async fn complete(
        &self,
        req: &ChatRequest,
        _on_delta: &(dyn for<'a> Fn(&'a str) + Send + Sync),
    ) -> supercode_harness::Result<(ChatMessage, Usage)> {
        self.calls.fetch_add(1, Ordering::SeqCst);
        self.seen_models.lock().unwrap().push(req.model.clone());
        Ok((ChatMessage::assistant("ok"), Usage::default()))
    }
}

#[tokio::test]
async fn set_model_changes_the_model_on_the_very_next_request() {
    let config = Config::builder().model("vendor/model-a").build();
    let seen = Arc::new(Mutex::new(Vec::new()));
    let provider = Box::new(RecordingProvider {
        calls: AtomicUsize::new(0),
        seen_models: seen.clone(),
    });
    let mut agent = Agent::with_provider(config, provider);

    assert_eq!(agent.model(), "vendor/model-a");
    agent.send("first").await.unwrap();

    agent.set_model("vendor/model-b");
    assert_eq!(agent.model(), "vendor/model-b");
    agent.send("second").await.unwrap();

    let seen = seen.lock().unwrap();
    assert_eq!(
        seen.as_slice(),
        &["vendor/model-a".to_string(), "vendor/model-b".to_string()],
        "the second request must carry the switched model, not a stale copy"
    );
}

#[tokio::test]
async fn set_model_before_any_request_is_honored_from_the_first_send() {
    let config = Config::builder().model("vendor/model-a").build();
    let seen = Arc::new(Mutex::new(Vec::new()));
    let provider = Box::new(RecordingProvider {
        calls: AtomicUsize::new(0),
        seen_models: seen.clone(),
    });
    let mut agent = Agent::with_provider(config, provider);

    agent.set_model("vendor/model-c");
    agent.send("hi").await.unwrap();

    assert_eq!(
        seen.lock().unwrap().as_slice(),
        &["vendor/model-c".to_string()]
    );
}

// =============================================================================
// P4c (COMPOSABLE-HARNESS-DESIGN.md §5.2 "P4" core NEW-significant item,
// §1.10/§3.1 `core.model_switch.allow_switch`, dep 8): `Agent::switch_model`
// — the persisted `model_change` record + reasoning-artifact filtering on
// top of the UX-30 dev/02 mechanics proven above.
// =============================================================================

use supercode_harness::model_change::ModelChangeRecord;
use supercode_harness::session::Session;

/// A `Session` carrying one assistant message with model-A's reasoning
/// artifacts attached — both as `metadata` (the shape `Session`'s real
/// Claude Code/Codex importers actually produce, per `session.rs`) and as a
/// `content_parts` reasoning block (the defensive/future-provider shape) —
/// so a single fixture exercises both `filter_reasoning_artifacts` branches.
fn session_with_reasoning_artifacts() -> Session {
    let mut session = Session::from_claude_code_str("").unwrap();
    let mut assistant = ChatMessage::assistant("here's my answer");
    assistant.metadata.insert(
        "thinking".to_string(),
        "model-A's private chain of thought".to_string(),
    );
    assistant
        .metadata
        .insert("thinking_signature".to_string(), "sig-abc".to_string());
    assistant.content_parts = Some(vec![
        serde_json::json!({"type": "text", "text": "here's my answer"}),
        serde_json::json!({"type": "thinking", "text": "model-A's private chain of thought"}),
    ]);
    session.messages = vec![ChatMessage::user("question"), assistant];
    session
}

/// Default off: `switch_model` with `allow_switch` unset (the default) is
/// byte-identical to `set_model` — no record, no filtering. This is the
/// item-9 default-off proof.
#[tokio::test]
async fn switch_model_default_off_is_byte_identical_to_set_model() {
    let config = Config::builder().model("vendor/model-a").build();
    assert!(!config.model_switch_allow_switch);
    let mut agent = Agent::with_provider(
        config,
        Box::new(RecordingProvider {
            calls: AtomicUsize::new(0),
            seen_models: Arc::new(Mutex::new(Vec::new())),
        }),
    );
    agent.load_session(session_with_reasoning_artifacts());
    let before = agent.history().to_vec_metadata_snapshot();

    agent.switch_model("vendor/model-b");

    assert_eq!(agent.model(), "vendor/model-b");
    assert!(
        agent.model_change_records().is_empty(),
        "allow_switch=false must never create a model_change record"
    );
    // Reasoning artifacts are UNTOUCHED — the filter never ran.
    assert_eq!(agent.history().to_vec_metadata_snapshot(), before);
}

/// Happy path + the explicit item-9 proof requirement: with `allow_switch`
/// on, switching models (1) persists a typed `ModelChangeRecord`, and (2)
/// filters model-A's reasoning artifacts out of `history` so they can never
/// reach model-B's context on the next request built from it.
#[tokio::test]
async fn switch_model_on_filters_reasoning_artifacts_and_records_the_switch() {
    let config = Config::builder()
        .model("vendor/model-a")
        .model_switch_allow_switch(true)
        .build();
    let seen = Arc::new(Mutex::new(Vec::new()));
    let mut agent = Agent::with_provider(
        config,
        Box::new(RecordingProvider {
            calls: AtomicUsize::new(0),
            seen_models: seen.clone(),
        }),
    );
    agent.load_session(session_with_reasoning_artifacts());

    // Sanity: the fixture really does carry reasoning artifacts before the
    // switch (metadata never reaches the wire regardless, but this proves
    // the filter has actual work to do, not a vacuous pass).
    let assistant_before = agent
        .history()
        .iter()
        .find(|m| m.content.as_deref() == Some("here's my answer"))
        .unwrap();
    assert!(assistant_before.metadata.contains_key("thinking"));
    assert!(assistant_before
        .content_parts
        .as_ref()
        .unwrap()
        .iter()
        .any(|p| p["type"] == "thinking"));

    agent.switch_model("vendor/model-b");

    // (1) Persisted record.
    let records = agent.model_change_records();
    assert_eq!(records.len(), 1);
    let r = &records[0];
    assert_eq!(r.from_model, "vendor/model-a");
    assert_eq!(r.to_model, "vendor/model-b");
    assert!(r.reasoning_filtered);
    assert!(
        r.reasoning_artifacts_filtered >= 1,
        "the fixture message should have been counted as touched"
    );

    // (2) model-A's reasoning never reaches model-B's context: `history` —
    // the exact source `run_loop` builds the NEXT request from — no longer
    // carries the metadata key OR the content_parts reasoning block.
    let assistant_after = agent
        .history()
        .iter()
        .find(|m| {
            m.content.as_deref() == Some("here's my answer")
                || m.content_parts
                    .as_ref()
                    .map(|p| p.iter().any(|x| x["type"] == "text"))
                    .unwrap_or(false)
        })
        .expect("the assistant message with reasoning stripped is still present");
    assert!(!assistant_after.metadata.contains_key("thinking"));
    assert!(!assistant_after.metadata.contains_key("thinking_signature"));
    let parts_after = assistant_after.content_parts.as_ref().unwrap();
    assert!(
        parts_after.iter().all(|p| p["type"] != "thinking"),
        "{parts_after:?}"
    );
    // The non-reasoning text part survives.
    assert!(parts_after.iter().any(|p| p["type"] == "text"));

    // Drive an actual request through model-B and confirm the (already-
    // wire-excluded) metadata point holds end-to-end too: the request the
    // NEW model receives carries the switched model id.
    agent.send("continue").await.unwrap();
    assert_eq!(
        seen.lock().unwrap().last(),
        Some(&"vendor/model-b".to_string())
    );
}

/// Boundary: switching to the model already in effect is a no-op — no
/// record is created (nothing actually changed, nothing to filter for).
#[tokio::test]
async fn switch_model_to_the_same_model_is_a_no_op_boundary() {
    let config = Config::builder()
        .model("vendor/model-a")
        .model_switch_allow_switch(true)
        .build();
    let mut agent = Agent::with_provider(
        config,
        Box::new(RecordingProvider {
            calls: AtomicUsize::new(0),
            seen_models: Arc::new(Mutex::new(Vec::new())),
        }),
    );
    agent.switch_model("vendor/model-a");
    assert_eq!(agent.model(), "vendor/model-a");
    assert!(agent.model_change_records().is_empty());
}

/// "round-trip losslessly through the store" (item 9's explicit
/// requirement) proven at the `Agent` level, not just `model_change.rs`'s
/// own unit tests: records `switch_model` produces survive a real
/// save/load cycle through `SessionStore`.
#[tokio::test]
async fn switch_model_records_round_trip_through_the_store() {
    let config = Config::builder()
        .model("vendor/model-a")
        .model_switch_allow_switch(true)
        .build();
    let mut agent = Agent::with_provider(
        config,
        Box::new(RecordingProvider {
            calls: AtomicUsize::new(0),
            seen_models: Arc::new(Mutex::new(Vec::new())),
        }),
    );
    agent.load_session(session_with_reasoning_artifacts());
    agent.switch_model("vendor/model-b");
    agent.switch_model("vendor/model-c");
    assert_eq!(agent.model_change_records().len(), 2);

    let tmp = std::env::temp_dir().join(format!(
        "sc-p4c-model-change-store-{}-{}",
        std::process::id(),
        std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .unwrap()
            .as_nanos()
    ));
    let store = supercode_harness::SessionStore::open(&tmp).unwrap();
    store.save("sess", "t", "[]").unwrap();
    agent.save_model_change_log(&store, "sess").unwrap();
    let loaded: Vec<ModelChangeRecord> = store.load_model_change_log("sess").unwrap();
    assert_eq!(loaded, agent.model_change_records());
    let _ = std::fs::remove_dir_all(&tmp);
}

// =============================================================================
// P4c-review (MEDIUM, proven): end-to-end reasoning leak through export.
// `filter_reasoning_artifacts` missed the `"thinking_blocks"` metadata key —
// the Claude Code importer's per-block reasoning list (`session.rs:3475-
// 3480`), which the Claude Code EXPORTER prefers over the legacy singular
// `thinking`/`thinking_signature` fields whenever present (`session.rs:5654-
// 5676`). Left unstripped, a mid-session `switch_model` did NOT actually
// prevent model-A's full signed chain-of-thought from being resurrected —
// re-attributed to model-B's session — the moment that session was exported
// back to Claude Code format. This test drives the exact failure mode:
// import-shaped `thinking_blocks` metadata -> `switch_model` A->B with
// `allow_switch` on -> export to Claude Code -> assert the exported BYTES
// (not just in-memory `history`) never contain model-A's reasoning text or
// signature.
// =============================================================================

use supercode_harness::SessionFormat;

/// A model-A secret chain-of-thought + its signature, kept as constants so
/// the leak assertion below and the fixture construction can't drift apart.
const MODEL_A_SECRET_COT: &str = "MODEL-A-SECRET-CHAIN-OF-THOUGHT-7f3a";
const MODEL_A_SECRET_SIGNATURE: &str = "MODEL-A-SECRET-SIGNATURE-9c21";

/// Mirrors exactly what `Session::from_claude_code_str`'s
/// `push_claude_assistant` writes at `session.rs:3475-3480` for a message
/// whose Claude Code record carried a `thinking` content block: a
/// `metadata["thinking_blocks"]` value that is a serialized JSON array of
/// `{"type":"thinking","thinking":<text>,"signature":<sig>}` objects — see
/// `session.rs:3336-3340`'s exact block shape. Deliberately does NOT also
/// set the legacy singular `"thinking"`/`"thinking_signature"` keys, so this
/// fixture isolates the `thinking_blocks` path: if `filter_reasoning_artifacts`
/// ever stops stripping `"thinking_blocks"` (e.g. it's reverted out of
/// `REASONING_METADATA_KEYS`), this is the ONLY metadata key carrying the
/// secret text, so the leak assertion below has nothing else masking it.
fn session_with_only_thinking_blocks_reasoning() -> Session {
    let mut session = Session::from_claude_code_str("").unwrap();
    let mut assistant = ChatMessage::assistant("here's my answer");
    assistant.metadata.insert(
        "thinking_blocks".to_string(),
        serde_json::json!([{
            "type": "thinking",
            "thinking": MODEL_A_SECRET_COT,
            "signature": MODEL_A_SECRET_SIGNATURE,
        }])
        .to_string(),
    );
    session.messages = vec![ChatMessage::user("question"), assistant];
    session
}

#[tokio::test]
async fn switch_model_then_claude_code_export_never_resurrects_model_a_thinking_blocks() {
    let config = Config::builder()
        .model("vendor/model-a")
        .model_switch_allow_switch(true)
        .build();
    let mut agent = Agent::with_provider(
        config,
        Box::new(RecordingProvider {
            calls: AtomicUsize::new(0),
            seen_models: Arc::new(Mutex::new(Vec::new())),
        }),
    );
    agent.load_session(session_with_only_thinking_blocks_reasoning());

    // Sanity: the fixture really carries the secret before the switch.
    let before = agent
        .history()
        .iter()
        .find(|m| m.content.as_deref() == Some("here's my answer"))
        .unwrap();
    assert_eq!(
        before
            .metadata
            .get("thinking_blocks")
            .map(|s| s.contains(MODEL_A_SECRET_COT)),
        Some(true)
    );

    agent.switch_model("vendor/model-b");
    let r = &agent.model_change_records()[0];
    assert!(r.reasoning_filtered);
    assert!(r.reasoning_artifacts_filtered >= 1);

    // Build a fresh Session from the post-switch history (exactly what a
    // real `/export` or cross-format translator would work from) and export
    // it to Claude Code — the format whose exporter specifically prefers
    // `thinking_blocks` over the legacy fields.
    let mut exported = Session::from_claude_code_str("").unwrap();
    exported.messages = agent.history().to_vec();
    let jsonl = exported.to_jsonl(SessionFormat::ClaudeCode).unwrap();

    assert!(
        !jsonl.contains(MODEL_A_SECRET_COT),
        "model-A's reasoning text leaked into the Claude Code export: {jsonl}"
    );
    assert!(
        !jsonl.contains(MODEL_A_SECRET_SIGNATURE),
        "model-A's reasoning signature leaked into the Claude Code export: {jsonl}"
    );

    // Contrast control: prove this assertion isn't vacuous by confirming the
    // secret WOULD have shown up in the export had the switch not run (i.e.
    // the CC exporter really does surface `thinking_blocks` when present).
    let mut unfiltered = session_with_only_thinking_blocks_reasoning();
    unfiltered.messages.remove(0); // drop the user turn, keep only the assistant message
    let unfiltered_jsonl = unfiltered.to_jsonl(SessionFormat::ClaudeCode).unwrap();
    assert!(
        unfiltered_jsonl.contains(MODEL_A_SECRET_COT),
        "contrast control failed: the CC exporter should surface an unfiltered \
         thinking_blocks message's reasoning text — {unfiltered_jsonl}"
    );
}

/// A helper extension trait so the "reasoning artifacts untouched" default-
/// off assertion above can compare a cheap, order-independent snapshot
/// instead of the whole `ChatMessage` (which isn't `PartialEq`).
trait MetadataSnapshot {
    fn to_vec_metadata_snapshot(&self) -> Vec<std::collections::BTreeMap<String, String>>;
}
impl MetadataSnapshot for [ChatMessage] {
    fn to_vec_metadata_snapshot(&self) -> Vec<std::collections::BTreeMap<String, String>> {
        self.iter().map(|m| m.metadata.clone()).collect()
    }
}