@team-agent/installer 0.2.11 → 0.3.1

package/crates/team-agent/src/provider/startup_prompt.rs

@@ -0,0 +1,423 @@
+//! codex startup-prompt recognizer — workspace-trust + update-skip screen detection.
+//!
+//! Golden (READ-ONLY truth `team-agent-public` v0.2.11): `provider_cli/codex.py`
+//!   - `CodexAdapter.handle_startup_prompts` (:142-182)
+//!   - `maybe_skip_update_prompt` (:262-268)
+//!
+//! recognizer-class (Gap 29 — burned 4 Mac minis): a NAIVE substring port gets the RECENCY命门 wrong.
+//! A prompt is acted on ONLY when its `rfind` position is GREATER than the ready marker's `rfind`
+//! position (i.e. it appears LATER / more recently in the captured scrollback). A stale prompt ABOVE an
+//! already-ready marker is left alone — ready wins and polling stops. RED-first skeleton; porter-d
+//! implements GREEN black-box against golden codex.py.
+
+use std::time::Duration;
+
+use crate::transport::{CaptureRange, Key, Target, Transport};
+
+const TRUST_MARKERS: &[&str] = &[
+    "Do you trust the contents of this directory?",
+    "Do you trust the files in this folder?",
+    "Do you trust this folder?",
+];
+const UPDATE_MARKERS: &[&str] = &["Update available!", "Update now"];
+/// Plain ready markers (not the bare `›` glyph — that glyph also indicates a
+/// numbered-menu selector and is handled by [`rightmost_input_prompt_glyph`] with
+/// shape gating per N15 / CR-063: detect by SHAPE, not a single Unicode codepoint).
+const READY_MARKERS: &[&str] = &["OpenAI Codex", "codex>"];
+
+/// Per-poll decision for the codex startup screen. Golden order each iteration (codex.py:160-181):
+/// update-skip is checked FIRST, then workspace-trust, then ready (stop), else keep polling.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum StartupScreenDecision {
+    /// `maybe_skip_update_prompt` matched: update_pos >= 0 && update_pos > ready_pos (codex.py:262-267).
+    SkipUpdatePrompt,
+    /// workspace-trust: trust_pos >= 0 && trust_pos > ready_pos (codex.py:166-174).
+    AnswerWorkspaceTrust,
+    /// ready_pos >= 0 with no actionable prompt above it (codex.py:178) -> stop polling.
+    Ready,
+    /// none of the above (codex.py:180) -> sleep + keep polling.
+    KeepPolling,
+}
+
+/// One handled startup prompt — an entry of golden's `handled` list.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct HandledPrompt {
+    pub prompt: String,
+    pub action: String,
+}
+
+/// PURE recognizer (codex.py:160-181 + maybe_skip_update_prompt :262-268): captured scrollback ->
+/// decision. NO IO. The RECENCY命门: a prompt is acted on ONLY when its `rfind` position is strictly
+/// GREATER than the ready marker's `rfind` position. update is evaluated before trust.
+///   trust strings (rfind max of):  "Do you trust the contents of this directory?" /
+///       "Do you trust the files in this folder?" / "Do you trust this folder?"
+///   update strings (rfind max of): "Update available!" / "Update now"
+///   ready markers (rfind max of):  "OpenAI Codex" / "›" / "codex>"
+pub fn classify_codex_startup_screen(output: &str) -> StartupScreenDecision {
+    // CR-063 / subroot real-machine residual: actionable-shape override BEFORE recency.
+    // The recency model ("prompt above ready = stale-scrolled, ignore") assumes the
+    // active state is the LATEST byte on screen. Real Codex breaks that assumption:
+    // while a trust modal is still active, Codex pre-renders the Update box, the
+    // OpenAI Codex banner, AND a bottom `› Find and fix a bug…` input-prompt indicator
+    // BELOW the trust menu — so recency would mark the screen Ready and the trust
+    // menu would never be answered. When the captured text has the actionable trust
+    // shape (`Do you trust …` phrase + a `› <digit>. ` numbered-menu line, N15),
+    // the modal IS the live state regardless of what comes after it. Return early.
+    if has_actionable_trust_shape(output) {
+        return StartupScreenDecision::AnswerWorkspaceTrust;
+    }
+    // N15/CR-063 root-cause (recency lane): the bare `›` glyph is BOTH the Codex
+    // input-prompt indicator AND the numbered-menu selector on a real trust pane
+    // (`› 1. Yes, continue`). Detect by SHAPE: `›` is a ready marker only when its
+    // tail is NOT a `<digit>. ` menu item.
+    let ready_pos = max_two(
+        max_rfind(output, READY_MARKERS),
+        rightmost_input_prompt_glyph(output),
+    );
+    if is_more_recent(max_rfind(output, UPDATE_MARKERS), ready_pos) {
+        return StartupScreenDecision::SkipUpdatePrompt;
+    }
+    if is_more_recent(max_rfind(output, TRUST_MARKERS), ready_pos) {
+        return StartupScreenDecision::AnswerWorkspaceTrust;
+    }
+    if ready_pos.is_some() {
+        StartupScreenDecision::Ready
+    } else {
+        StartupScreenDecision::KeepPolling
+    }
+}
+
+/// Actionable trust shape (N15): the captured text contains a trust phrase AND a
+/// numbered-menu selector line `› <digit>. `. This is the modal-still-active signal
+/// that survives Codex's pre-rendering of the banner/input prompt below the menu.
+/// Does NOT match a single-screen "trust phrase + bare `›`" (e.g. plain Ready
+/// follow-up text), so historical "trust ABOVE ready" recency tests keep passing
+/// — those fixtures do not include a `› <digit>. ` menu line.
+fn has_actionable_trust_shape(output: &str) -> bool {
+    if !TRUST_MARKERS.iter().any(|marker| output.contains(marker)) {
+        return false;
+    }
+    contains_numbered_menu_glyph(output)
+}
+
+/// `true` iff any `›` in the output is followed by a numbered-menu selector
+/// (` <digit>. `). The shape pairs the glyph with a digit-dot line item — the
+/// Codex trust/update menu printing convention.
+fn contains_numbered_menu_glyph(output: &str) -> bool {
+    let glyph = '›';
+    let glyph_len = glyph.len_utf8();
+    let mut start = 0;
+    while let Some(rel) = output[start..].find(glyph) {
+        let abs = start + rel;
+        let tail_start = abs + glyph_len;
+        if tail_start > output.len() {
+            break;
+        }
+        if is_numbered_menu_tail(&output[tail_start..]) {
+            return true;
+        }
+        start = tail_start;
+    }
+    false
+}
+
+/// Rightmost `›` whose tail is NOT a numbered-menu selector (` <digit>. `). A bare
+/// `›` followed by free text or whitespace is the Codex main-input prompt indicator;
+/// a `›` followed by `1. Yes, continue` is part of the trust/update menu and is NOT
+/// a ready signal.
+fn rightmost_input_prompt_glyph(output: &str) -> Option<usize> {
+    let glyph = '›';
+    let glyph_len = glyph.len_utf8();
+    let mut best = None;
+    let bytes = output.as_bytes();
+    let mut start = 0;
+    while let Some(rel) = output[start..].find(glyph) {
+        let abs = start + rel;
+        let tail_start = abs + glyph_len;
+        if tail_start <= bytes.len() && !is_numbered_menu_tail(&output[tail_start..]) {
+            best = Some(abs);
+        }
+        start = tail_start;
+        if start > output.len() {
+            break;
+        }
+    }
+    best
+}
+
+fn is_numbered_menu_tail(tail: &str) -> bool {
+    let trimmed = tail.trim_start_matches(' ');
+    let mut chars = trimmed.chars();
+    matches!(
+        (chars.next(), chars.next()),
+        (Some(d), Some('.')) if d.is_ascii_digit()
+    )
+}
+
+fn max_two(a: Option<usize>, b: Option<usize>) -> Option<usize> {
+    match (a, b) {
+        (Some(x), Some(y)) => Some(x.max(y)),
+        (Some(x), None) | (None, Some(x)) => Some(x),
+        (None, None) => None,
+    }
+}
+
+/// Capture-poll loop (codex.py:142-182) over the `transport.capture()` seam (NOT a raw subprocess, so
+/// it stays unit-testable). On `AnswerWorkspaceTrust` -> send `Enter` + push
+/// {prompt:"codex_workspace_trust", action:"sent_enter"}; on `SkipUpdatePrompt` -> send `Down`,`Enter`
+/// + push {prompt:"codex_update_available", action:"sent_skip"}; on `Ready` -> stop. Loops up to
+///   `checks` (golden default 30), `sleep_s` (golden 0.5) between iterations. Returns the ordered
+///   `handled` list. Capture is full scrollback (golden `tmux capture-pane -p -S - -t <target>`).
+pub fn codex_handle_startup_prompts(
+    transport: &dyn Transport,
+    target: &Target,
+    checks: usize,
+    sleep_s: f64,
+) -> Vec<HandledPrompt> {
+    let mut handled = Vec::new();
+    for _ in 0..checks {
+        let screen = match transport.capture(target, CaptureRange::Full) {
+            Ok(captured) => captured.text,
+            Err(_) => String::new(),
+        };
+        match classify_codex_startup_screen(&screen) {
+            StartupScreenDecision::SkipUpdatePrompt => {
+                let _ = transport.send_keys(target, &[Key::Down, Key::Enter]);
+                handled.push(HandledPrompt {
+                    prompt: "codex_update_available".to_string(),
+                    action: "sent_skip".to_string(),
+                });
+                sleep_between_polls(sleep_s);
+            }
+            StartupScreenDecision::AnswerWorkspaceTrust => {
+                let _ = transport.send_keys(target, &[Key::Enter]);
+                handled.push(HandledPrompt {
+                    prompt: "codex_workspace_trust".to_string(),
+                    action: "sent_enter".to_string(),
+                });
+                sleep_between_polls(sleep_s);
+            }
+            StartupScreenDecision::Ready => break,
+            StartupScreenDecision::KeepPolling => sleep_between_polls(sleep_s),
+        }
+    }
+    handled
+}
+
+fn max_rfind(output: &str, needles: &[&str]) -> Option<usize> {
+    needles.iter().filter_map(|needle| output.rfind(needle)).max()
+}
+
+fn is_more_recent(prompt_pos: Option<usize>, ready_pos: Option<usize>) -> bool {
+    match (prompt_pos, ready_pos) {
+        (Some(prompt), Some(ready)) => prompt > ready,
+        (Some(_), None) => true,
+        _ => false,
+    }
+}
+
+fn sleep_between_polls(sleep_s: f64) {
+    let millis = (sleep_s * 1000.0).round();
+    if millis.is_finite() && millis > 0.0 && millis <= u64::MAX as f64 {
+        std::thread::sleep(Duration::from_millis(millis as u64));
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
+    use super::*;
+    use crate::model::enums::PaneLiveness;
+    use crate::transport::{
+        AttachOutcome, BackendKind, CaptureRange, CapturedText, InjectPayload, InjectReport, Key,
+        PaneField, PaneId, PaneInfo, SessionName, SetEnvOutcome, SpawnResult, Target, TransportError,
+        WindowName,
+    };
+    use std::collections::BTreeMap;
+    use std::path::Path;
+    use std::sync::Mutex;
+
+    // ── EXACT golden strings (provider_cli/codex.py). Do not paraphrase — recognizer-class. ──────────
+    const TRUST_DIR: &str = "Do you trust the contents of this directory?";
+    const TRUST_FILES: &str = "Do you trust the files in this folder?";
+    const TRUST_FOLDER: &str = "Do you trust this folder?";
+    const UPDATE_AVAIL: &str = "Update available!";
+    const UPDATE_NOW: &str = "Update now";
+    const READY_BANNER: &str = "OpenAI Codex";
+    const READY_PROMPT: &str = "›"; // U+203A
+    const READY_BARE: &str = "codex>";
+
+    // ── ① + ② RED核心 — workspace-trust MORE RECENT than ready -> answer it ──────────────────────────
+    #[test]
+    fn trust_more_recent_than_ready_answers_workspace_trust() {
+        // ready banner appears early; the trust prompt appears LATER; no ready marker after it
+        // => trust_pos > ready_pos => answer.
+        let screen = format!("{READY_BANNER} v1.2\nwelcome\n\n{TRUST_DIR}\n  hit enter ");
+        assert_eq!(
+            classify_codex_startup_screen(&screen),
+            StartupScreenDecision::AnswerWorkspaceTrust
+        );
+    }
+
+    // ── ② 命门 CORE — a STALE trust prompt ABOVE the ready marker is NOT answered (ready wins) ────────
+    #[test]
+    fn stale_trust_above_ready_is_not_answered_ready_wins() {
+        // trust prompt FIRST, then a ready marker LATER => trust_pos < ready_pos => do NOT answer.
+        // This is the positional-recency命门 a naive substring port gets wrong (would re-send Enter).
+        let screen = format!("{TRUST_DIR}\n[trusted earlier]\n{READY_BANNER} ready\n{READY_PROMPT} ");
+        assert_eq!(
+            classify_codex_startup_screen(&screen),
+            StartupScreenDecision::Ready,
+            "RECENCY命门: a trust prompt ABOVE the ready marker is stale; ready wins, NO Enter sent"
+        );
+    }
+
+    #[test]
+    fn each_trust_string_recognized_when_more_recent() {
+        for s in [TRUST_DIR, TRUST_FILES, TRUST_FOLDER] {
+            let screen = format!("{READY_BANNER}\n...banner...\n{s}\n");
+            assert_eq!(
+                classify_codex_startup_screen(&screen),
+                StartupScreenDecision::AnswerWorkspaceTrust,
+                "trust string {s:?} after ready must answer"
+            );
+        }
+    }
+
+    // ── ③ sibling — update-skip recognizer (maybe_skip_update_prompt), same recency命门 ──────────────
+    #[test]
+    fn update_more_recent_than_ready_skips_update() {
+        for s in [UPDATE_AVAIL, UPDATE_NOW] {
+            let screen = format!("{READY_BANNER}\nblah\n{s}\n");
+            assert_eq!(
+                classify_codex_startup_screen(&screen),
+                StartupScreenDecision::SkipUpdatePrompt,
+                "update string {s:?} after ready must skip"
+            );
+        }
+    }
+
+    #[test]
+    fn stale_update_above_ready_is_not_skipped_ready_wins() {
+        let screen = format!("{UPDATE_AVAIL}\n{READY_BANNER}\n{READY_PROMPT} ");
+        assert_eq!(classify_codex_startup_screen(&screen), StartupScreenDecision::Ready);
+    }
+
+    // ── golden ORDER — update is checked BEFORE trust (both more recent) -> SkipUpdatePrompt wins ─────
+    #[test]
+    fn update_checked_before_trust() {
+        // both update + trust appear after the ready marker; golden runs maybe_skip_update_prompt
+        // FIRST each iteration -> the screen resolves to SkipUpdatePrompt, not AnswerWorkspaceTrust.
+        let screen = format!("{READY_BANNER}\n{TRUST_DIR}\n{UPDATE_AVAIL}\n");
+        assert_eq!(classify_codex_startup_screen(&screen), StartupScreenDecision::SkipUpdatePrompt);
+    }
+
+    // ── ready-only / neither ─────────────────────────────────────────────────────────────────────────
+    #[test]
+    fn each_ready_marker_alone_is_ready() {
+        for m in [READY_BANNER, READY_PROMPT, READY_BARE] {
+            let screen = format!("booting...\n{m} ");
+            assert_eq!(
+                classify_codex_startup_screen(&screen),
+                StartupScreenDecision::Ready,
+                "ready marker {m:?} alone must be Ready"
+            );
+        }
+    }
+
+    #[test]
+    fn no_prompt_no_ready_keeps_polling() {
+        assert_eq!(
+            classify_codex_startup_screen("loading spinner...\nstill starting\n"),
+            StartupScreenDecision::KeepPolling
+        );
+    }
+
+    // ── ④ transport.capture() SEAM — the loop answers trust then breaks on ready, via the seam ───────
+    /// Scripted transport: `capture` pops the next canned screen; `send_keys` records the keys. All
+    /// other methods are unreachable by the startup-prompt loop.
+    struct ScriptedTransport {
+        screens: Mutex<Vec<String>>,
+        sent: Mutex<Vec<Vec<Key>>>,
+    }
+    impl Transport for ScriptedTransport {
+        fn kind(&self) -> BackendKind {
+            BackendKind::Tmux
+        }
+        fn spawn_first(&self, _s: &SessionName, _w: &WindowName, _a: &[String], _c: &Path, _e: &BTreeMap<String, String>) -> Result<SpawnResult, TransportError> {
+            unimplemented!("not reached by startup-prompt loop")
+        }
+        fn spawn_into(&self, _s: &SessionName, _w: &WindowName, _a: &[String], _c: &Path, _e: &BTreeMap<String, String>) -> Result<SpawnResult, TransportError> {
+            unimplemented!("not reached by startup-prompt loop")
+        }
+        fn inject(&self, _t: &Target, _p: &InjectPayload, _submit: Key, _b: bool) -> Result<InjectReport, TransportError> {
+            unimplemented!("not reached")
+        }
+        fn send_keys(&self, _t: &Target, keys: &[Key]) -> Result<(), TransportError> {
+            self.sent.lock().unwrap().push(keys.to_vec());
+            Ok(())
+        }
+        fn capture(&self, _t: &Target, range: CaptureRange) -> Result<CapturedText, TransportError> {
+            let mut q = self.screens.lock().unwrap();
+            let text = if q.is_empty() { String::new() } else { q.remove(0) };
+            Ok(CapturedText { text, range })
+        }
+        fn query(&self, _t: &Target, _f: PaneField) -> Result<Option<String>, TransportError> {
+            Ok(None)
+        }
+        fn liveness(&self, _p: &PaneId) -> Result<PaneLiveness, TransportError> {
+            unimplemented!("not reached")
+        }
+        fn list_targets(&self) -> Result<Vec<PaneInfo>, TransportError> {
+            unimplemented!("not reached")
+        }
+        fn has_session(&self, _s: &SessionName) -> Result<bool, TransportError> {
+            Ok(true)
+        }
+        fn list_windows(&self, _s: &SessionName) -> Result<Vec<WindowName>, TransportError> {
+            unimplemented!("not reached")
+        }
+        fn set_session_env(&self, _s: &SessionName, _k: &str, _v: &str) -> Result<SetEnvOutcome, TransportError> {
+            unimplemented!("not reached")
+        }
+        fn kill_session(&self, _s: &SessionName) -> Result<(), TransportError> {
+            unimplemented!("not reached")
+        }
+        fn kill_window(&self, _t: &Target) -> Result<(), TransportError> {
+            unimplemented!("not reached")
+        }
+        fn attach_session(&self, _s: &SessionName) -> Result<AttachOutcome, TransportError> {
+            unimplemented!("not reached")
+        }
+    }
+
+    #[test]
+    fn loop_answers_trust_then_breaks_on_ready_via_capture_seam() {
+        let t = ScriptedTransport {
+            screens: Mutex::new(vec![
+                // iter 1: trust prompt more recent than ready -> answer (send Enter) + continue.
+                format!("{READY_BANNER}\n{TRUST_DIR}\n"),
+                // iter 2: ready marker, no actionable prompt above it -> break.
+                format!("{READY_BANNER} ready\n{READY_PROMPT} "),
+            ]),
+            sent: Mutex::new(Vec::new()),
+        };
+        let target = Target::Pane(PaneId::new("%1"));
+
+        let handled = codex_handle_startup_prompts(&t, &target, 5, 0.0);
+
+        assert_eq!(
+            handled,
+            vec![HandledPrompt {
+                prompt: "codex_workspace_trust".to_string(),
+                action: "sent_enter".to_string(),
+            }],
+            "the loop must answer the workspace-trust prompt exactly once, then break on ready"
+        );
+        let sent = t.sent.lock().unwrap();
+        assert!(
+            sent.iter().any(|keys| keys.as_slice() == [Key::Enter]),
+            "on workspace-trust the loop must send Enter via the transport.capture() seam; got {sent:?}"
+        );
+    }
+}

package/crates/team-agent/src/provider/tests/adapter.rs

@@ -0,0 +1,239 @@
+#[test]
+fn abnormal_dedup_key_uses_signature_and_optional_turn_id() {
+    // probe(/tmp/probe_idle.py read_fault_facts): the C8 dedup key is
+    // (Signature, Option<TurnId>). Golden facts below are the exact extraction.
+
+    // claude api_error → signature=api_error, turn_id=sess-9 (sessionId), kind=error.
+    let api_err = [serde_json::json!(
+        {"type":"system","subtype":"api_error","level":"error","sessionId":"sess-9"}
+    )];
+    let f = read_fault_facts(&api_err, Provider::Claude);
+    assert_eq!(f.len(), 1);
+    assert_eq!(f[0].signature, Signature::new("api_error"));
+    assert_eq!(f[0].turn_id, Some(TurnId::new("sess-9")));
+    assert_eq!(f[0].kind, FactKind::Error);
+
+    // claude tool_result is_error → signature=tool_result_is_error,
+    // turn_id=parentUuid ("p-1"), kind=error.
+    let tool_err = [serde_json::json!(
+        {"type":"user","uuid":"u","parentUuid":"p-1",
+         "message":{"content":[{"type":"tool_result","is_error":true}]}}
+    )];
+    let f = read_fault_facts(&tool_err, Provider::Claude);
+    assert_eq!(f.len(), 1);
+    assert_eq!(f[0].signature, Signature::new("tool_result_is_error"));
+    assert_eq!(f[0].turn_id, Some(TurnId::new("p-1")));
+
+    // claude api_error with NO ids → key = (api_error, None) — Option must hold None.
+    let api_err_noids = [serde_json::json!(
+        {"type":"system","subtype":"api_error","level":"error"}
+    )];
+    let f = read_fault_facts(&api_err_noids, Provider::Claude);
+    assert_eq!(f.len(), 1);
+    assert_eq!(f[0].signature, Signature::new("api_error"));
+    assert_eq!(f[0].turn_id, None, "api_error with no ids dedups on (api_error, None)");
+
+    // codex turn_failed → signature=turn_failed, turn_id=ct4, kind=failed.
+    let codex_failed = [serde_json::json!(
+        {"jsonrpc":"2.0","method":"turn/completed","params":{"turn":{"id":"ct4","status":"failed"}}}
+    )];
+    let f = read_fault_facts(&codex_failed, Provider::Codex);
+    assert_eq!(f.len(), 1);
+    assert_eq!(f[0].signature, Signature::new("turn_failed"));
+    assert_eq!(f[0].turn_id, Some(TurnId::new("ct4")));
+    assert_eq!(f[0].kind, FactKind::Failed);
+
+    // codex approval → signature=approval_required, turn_id=ct5, kind=approval.
+    let codex_approval = [serde_json::json!(
+        {"jsonrpc":"2.0","method":"session/requestApproval","params":{"turnId":"ct5"}}
+    )];
+    let f = read_fault_facts(&codex_approval, Provider::Codex);
+    assert_eq!(f.len(), 1);
+    assert_eq!(f[0].signature, Signature::new("approval_required"));
+    assert_eq!(f[0].turn_id, Some(TurnId::new("ct5")));
+    assert_eq!(f[0].kind, FactKind::Approval);
+}
+
+// ---- (e) session capture payload + bug-085 fallback (confidence=low) ----
+
+#[test]
+fn capture_session_id_success_payload_fields() {
+    // claude.py:73-106 success → CapturedSession high confidence, fs_watch,
+    // session_id Some, rollout_path Some. Drive the real fn against a temp cwd
+    // that contains a discoverable transcript; assert ALL 5 typed fields.
+    let dir = std::env::temp_dir().join(format!(
+        "ta-cap-success-{}",
+        std::process::id()
+    ));
+    let _ = std::fs::create_dir_all(&dir);
+    let transcript = dir.join("session.jsonl");
+    let _ = std::fs::write(
+        &transcript,
+        r#"{"type":"user","sessionId":"sess-1","cwd":"PLACEHOLDER","message":{"content":"hi"}}"#,
+    );
+    let adapter = get_adapter(Provider::ClaudeCode);
+    let cs = adapter
+        .capture_session_id("agentX", &dir, 3)
+        .expect("capture ok")
+        .expect("transcript found → Some");
+    assert_eq!(cs.captured_via, CaptureVia::FsWatch);
+    assert_eq!(cs.attribution_confidence, Confidence::High);
+    assert!(cs.session_id.is_some(), "found transcript yields a session_id");
+    assert!(cs.rollout_path.is_some(), "found transcript yields a rollout_path");
+    let _ = std::fs::remove_dir_all(&dir);
+}
+
+#[test]
+fn capture_session_id_bug085_compatible_api_fallback_is_low_confidence_none_session() {
+    // probe(claude fallback): compatible_api transcript w/o a session_id →
+    // session_id=None, captured_via=fs_mtime_fallback, confidence=low,
+    // rollout_path SET. The half-state is LEGAL and must not panic (bug-085).
+    let dir = std::env::temp_dir().join(format!(
+        "ta-cap-fallback-{}",
+        std::process::id()
+    ));
+    let _ = std::fs::create_dir_all(&dir);
+    let transcript = dir.join("nosession.jsonl");
+    let _ = std::fs::write(&transcript, r#"{"type":"user","message":{"content":"hi"}}"#);
+    let adapter = get_adapter(Provider::ClaudeCode);
+    let cs = adapter
+        .capture_session_id("agentX", &dir, 3)
+        .expect("capture ok")
+        .expect("fallback transcript → Some half-state");
+    assert_eq!(cs.session_id, None, "bug-085: fallback session_id is None");
+    assert_eq!(cs.captured_via, CaptureVia::FsMtimeFallback);
+    assert_eq!(cs.attribution_confidence, Confidence::Low);
+    assert!(cs.rollout_path.is_some(), "fallback still pins rollout_path");
+    let _ = std::fs::remove_dir_all(&dir);
+}
+
+#[test]
+fn capture_session_id_no_match_returns_none() {
+    // claude.py:111-113 deadline with no match → Ok(None), explicitly NOT Err.
+    let empty = std::env::temp_dir().join(format!("ta-cap-empty-{}", std::process::id()));
+    let _ = std::fs::create_dir_all(&empty);
+    let adapter = get_adapter(Provider::ClaudeCode);
+    let res = adapter.capture_session_id("agentX", &empty, 0);
+    // ProviderError 非 PartialEq(携 Io)→ 用 matches! 而非 ==。
+    assert!(matches!(res, Ok(None)), "no transcript + timeout 0 → Ok(None), never Err");
+    let _ = std::fs::remove_dir_all(&empty);
+}
+
+// ---- resume / fork / build_command / caps — command-build core ----
+
+#[test]
+fn claude_caps_resume_and_fork_and_native_mcp() {
+    // doc §59 + claude.py: claude supports resume + fork (static cap true,
+    // runtime auth-gated) + native --mcp-config; does NOT write a global settings file.
+    let adapter = get_adapter(Provider::ClaudeCode);
+    let caps = adapter.caps();
+    assert_eq!(
+        caps,
+        ProviderCaps {
+            resume: true,
+            fork: true,
+            native_mcp_config: true,
+            writes_global_settings: false,
+        }
+    );
+}
+
+#[test]
+fn gemini_writes_global_settings_cap() {
+    // gemini.py:40-78 — install_mcp writes ~/.gemini/settings.json →
+    // caps.writes_global_settings=true, native_mcp_config=false.
+    let adapter = get_adapter(Provider::GeminiCli);
+    let caps = adapter.caps();
+    assert!(caps.writes_global_settings, "gemini writes a global settings file");
+    assert!(!caps.native_mcp_config, "gemini has no native --mcp-config flag");
+}
+
+#[test]
+fn claude_fork_blocked_under_compatible_api() {
+    // claude.py:54 supports_session_fork == (auth_mode != compatible_api).
+    // fork under CompatibleApi must Err (never silently empty) — capability path.
+    let adapter = get_adapter(Provider::ClaudeCode);
+    let sid = SessionId::new("src-1");
+    let res = adapter.fork(Some(&sid), AuthMode::CompatibleApi, None);
+    assert!(
+        matches!(
+            res,
+            Err(ProviderError::CapabilityUnsupported(_)) | Err(ProviderError::ResumeUnavailable(_))
+        ),
+        "fork under compatible_api must Err, got {res:?}"
+    );
+}
+
+#[test]
+fn build_resume_command_without_session_id_is_resume_unavailable() {
+    // claude.py:41-42 — no session_id → ResumeUnavailable, never a bare crash.
+    let adapter = get_adapter(Provider::ClaudeCode);
+    let res = adapter.build_resume_command(None, AuthMode::Subscription, None);
+    assert!(
+        matches!(res, Err(ProviderError::ResumeUnavailable(_))),
+        "resume without session_id must be ResumeUnavailable, got {res:?}"
+    );
+}
+
+#[test]
+fn session_is_resumable_none_session_is_false_not_panic() {
+    // claude.py:116-118 — session_id falsy → not resumable; bug-085 None穿透.
+    let adapter = get_adapter(Provider::ClaudeCode);
+    let res = adapter.session_is_resumable(None, AuthMode::CompatibleApi);
+    assert!(matches!(res, Ok(false)), "None session → Ok(false), never panic (bug-085)");
+}
+
+/// true iff `needle` (e.g. ["--model","opus"]) appears as a contiguous run in `hay`.
+fn argv_contains_adjacent(hay: &[String], needle: &[&str]) -> bool {
+    if needle.is_empty() {
+        return true;
+    }
+    hay.windows(needle.len())
+        .any(|w| w.iter().zip(needle).all(|(a, b)| a == b))
+}
+
+#[test]
+fn build_command_includes_model_and_system_prompt() {
+    // claude.py:175-199 — _base_command appends --model <m> + --append-system-prompt <p>;
+    // --strict-mcp-config only appears when mcp_config is present. With mcp_config=None,
+    // assert both flag pairs present adjacently AND --strict-mcp-config ABSENT.
+    let adapter = get_adapter(Provider::ClaudeCode);
+    let argv = adapter
+        .build_command(AuthMode::Subscription, None, Some("be helpful"), Some("opus"))
+        .expect("build_command ok");
+    assert!(
+        argv_contains_adjacent(&argv, &["--model", "opus"]),
+        "argv must contain `--model opus` adjacency: {argv:?}"
+    );
+    assert!(
+        argv_contains_adjacent(&argv, &["--append-system-prompt", "be helpful"]),
+        "argv must contain `--append-system-prompt 'be helpful'`: {argv:?}"
+    );
+    assert!(
+        !argv.iter().any(|a| a == "--strict-mcp-config"),
+        "no mcp_config → --strict-mcp-config must be absent: {argv:?}"
+    );
+}
+
+#[test]
+fn unsupported_provider_capability_error_message_shape() {
+    // unsupported.py:31 ProviderCapabilityError — placeholder providers reject
+    // on call. The skeleton Provider enum has no Copilot/Opencode variant, so
+    // the unsupported path is exercised through ProviderError::CapabilityUnsupported
+    // construction (message contract) here; full plug dispatch deferred.
+    let e = ProviderError::CapabilityUnsupported("opencode:start".to_string());
+    assert_eq!(
+        e.to_string(),
+        "provider capability unsupported: opencode:start"
+    );
+    let e2 = ProviderError::ResumeUnavailable("claude resume requires session_id".to_string());
+    assert_eq!(
+        e2.to_string(),
+        "resume unavailable: claude resume requires session_id"
+    );
+}
+
+// ═══════════════ P2 FIX-LOOP RED (复绿即对抗 cross-model findings) ═══════════════
+// Lock the CORRECT Python v0.2.11 behavior the strengthened contracts missed.
+// Golden re-probed via /tmp/probe_p2_provider.py vs team-agent-public @ 439bef8.
+