@team-agent/installer 0.2.10 → 0.3.0

package/crates/team-agent/src/messaging/delivery.rs

@@ -0,0 +1,542 @@
+//! internal_delivery.py + delivery.py — coordinator/调度器侧 thin wrapper + 单条 tmux 注入投递
+//! + trust 有界重试 + turn-open arm (card §16/§65)。
+
+use std::path::Path;
+
+use rusqlite::{params, OptionalExtension};
+
+use crate::event_log::EventLog;
+use crate::message_store::MessageStore;
+use crate::model::enums::Provider;
+use crate::model::ids::TeamKey;
+use crate::transport::{InjectPayload, Key, PaneId, SessionName, Target, Transport, WindowName};
+
+use super::helpers::{message_exists, MessageStatusShadow};
+use super::{
+    DeliveryOutcome, DeliveryRefusal, DeliveryStage, DeliveryStatus, MessagingError,
+    PaneWidthQuery, TrustRetryPayload,
+};
+
+// ===========================================================================
+// internal_delivery.py — coordinator/调度器侧 thin wrapper (card §65)
+// ===========================================================================
+
+/// `deliver_stored_message` (`internal_delivery.py:16`):coordinator/调度器侧 team-scoped 单发
+/// (不重路由)。加 `_runtime_lock("send")`,直走 `_send_single_message_unlocked`。
+#[allow(clippy::too_many_arguments)]
+pub fn deliver_stored_message(
+    workspace: &Path,
+    target: Option<&str>,
+    content: &str,
+    task_id: Option<&crate::model::ids::TaskId>,
+    sender: &str,
+    requires_ack: bool,
+    wait_visible: bool,
+    timeout: f64,
+    team: Option<&TeamKey>,
+) -> Result<DeliveryOutcome, MessagingError> {
+    let _ = (wait_visible, timeout);
+    let recipient = target.unwrap_or("leader");
+    let store = MessageStore::open(workspace)?;
+    let message_id = store.create_message(
+        task_id.map(crate::model::ids::TaskId::as_str),
+        sender,
+        recipient,
+        content,
+        None,
+        requires_ack,
+        team.map(TeamKey::as_str),
+    )?;
+    Ok(DeliveryOutcome {
+        ok: true,
+        status: DeliveryStatus::Queued,
+        message_status: MessageStatusShadow("accepted".to_string()),
+        message_id: Some(message_id),
+        verification: None,
+        stage: None,
+        reason: None,
+        channel: None,
+    })
+}
+
+// ===========================================================================
+// delivery.py — 单条 tmux 注入投递 + trust 有界重试 + turn-open arm (card §16)
+// ===========================================================================
+
+/// `_tmux_pane_width` (`delivery.py:20`):查询 pane 列宽。**fail-safe** (bug-064/082):失败
+/// 返回 [`PaneWidthQuery::Failed`],**绝不**给默认宽度。借 step 9 transport 的 query。
+pub fn tmux_pane_width(transport: &dyn Transport, target: &Target) -> PaneWidthQuery {
+    let queried = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
+        transport.query(target, crate::transport::PaneField::PaneWidth)
+    }));
+    let result = match queried {
+        Ok(result) => result,
+        Err(_) => {
+            return PaneWidthQuery::Failed {
+                error: "tmux_query_failed:panic".to_string(),
+            };
+        }
+    };
+    match result {
+        Ok(Some(raw)) => match raw.trim().parse::<u32>() {
+            Ok(pane_width) if pane_width > 0 => PaneWidthQuery::Ok { pane_width },
+            Ok(_) => PaneWidthQuery::Failed { error: "non_positive_width".to_string() },
+            Err(_) => PaneWidthQuery::Failed { error: "unparseable_output".to_string() },
+        },
+        Ok(None) => PaneWidthQuery::Failed { error: "empty_output".to_string() },
+        Err(err) => PaneWidthQuery::Failed { error: format!("tmux_query_failed:{err}") },
+    }
+}
+
+/// `_deliver_pending_message` (`delivery.py:63`):对一条消息做 tmux 注入投递 (含 trust 提示
+/// 自动应答 + turn-open arm + first_send_at 戳)。daemon-path → Result。
+pub fn deliver_pending_message(
+    workspace: &Path,
+    store: &MessageStore,
+    transport: &dyn Transport,
+    message_id: &str,
+    event_log: &EventLog,
+    state: &serde_json::Value,
+) -> Result<DeliveryOutcome, MessagingError> {
+    if !message_exists(store, message_id)? {
+        return Ok(DeliveryOutcome {
+            ok: false,
+            status: DeliveryStatus::Failed,
+            message_status: MessageStatusShadow("failed".to_string()),
+            message_id: Some(message_id.to_string()),
+            verification: None,
+            stage: None,
+            reason: None,
+            channel: None,
+        });
+    }
+    let message = message_for_delivery(store, message_id)?;
+    if !store.claim_for_delivery(message_id)? {
+        return Ok(DeliveryOutcome {
+            ok: false,
+            status: DeliveryStatus::Refused,
+            message_status: MessageStatusShadow("target_resolved".to_string()),
+            message_id: Some(message_id.to_string()),
+            verification: None,
+            stage: None,
+            reason: Some(DeliveryRefusal::MessageAlreadyClaimed),
+            channel: None,
+        });
+    }
+    let Some(message) = message else {
+        return Ok(DeliveryOutcome {
+            ok: false,
+            status: DeliveryStatus::Failed,
+            message_status: MessageStatusShadow("failed".to_string()),
+            message_id: Some(message_id.to_string()),
+            verification: None,
+            stage: None,
+            reason: Some(DeliveryRefusal::UnknownRecipient),
+            channel: None,
+        });
+    };
+    let target = resolve_inject_target(state, &message.recipient);
+    // Contract B / MUST-10 / N31/N32: physical paste+Enter into a startup trust/update
+    // menu is NOT provider delivery — the menu consumes the Enter and the task text
+    // is lost (PROBE-2 root-cause). Before injection, peek at the recipient's pane for
+    // a Codex actionable startup prompt; if present, mark the row `queued_until_trust`
+    // and DO NOT inject the task. The coordinator's startup-prompt phase will dismiss
+    // the trust prompt, and the SAME message_id is later replayed through this same
+    // delivery pipeline (no parallel side channel).
+    if recipient_pane_has_actionable_startup_prompt(transport, state, &message.recipient, &target) {
+        store.mark(message_id, "queued_until_trust", None)?;
+        event_log.write(
+            "delivery.deferred_startup_prompt",
+            serde_json::json!({
+                "message_id": message_id,
+                "recipient": message.recipient,
+                "reason": "actionable_startup_prompt",
+            }),
+        )?;
+        return Ok(DeliveryOutcome {
+            ok: false,
+            status: DeliveryStatus::RetryScheduled,
+            message_status: MessageStatusShadow("queued_until_trust".to_string()),
+            message_id: Some(message_id.to_string()),
+            verification: None,
+            stage: Some(DeliveryStage::TrustAutoAnswerDismissalWait),
+            reason: None,
+            channel: None,
+        });
+    }
+    let rendered = render_message(
+        &message.sender,
+        message.task_id.as_deref(),
+        &message.content,
+        message_id,
+    );
+    transport.inject(
+        &target,
+        &InjectPayload::Text(rendered),
+        Key::Enter,
+        true,
+    )?;
+    store.mark(message_id, "delivered", None)?;
+    event_log.write(
+        "message.delivered",
+        serde_json::json!({"message_id": message_id}),
+    )?;
+    let outcome = DeliveryOutcome {
+        ok: true,
+        status: DeliveryStatus::Delivered,
+        message_status: MessageStatusShadow("delivered".to_string()),
+        message_id: Some(message_id.to_string()),
+        verification: None,
+        stage: None,
+        reason: None,
+        channel: None,
+    };
+    stamp_first_send_at_if_leader_to_worker(workspace, state, &message.sender, &message.recipient)?;
+    record_turn_open_if_leader_to_worker(
+        workspace,
+        state,
+        &message.sender,
+        &message.recipient,
+        &outcome,
+        event_log,
+    )?;
+    Ok(outcome)
+}
+
+/// Render a message into the worker-facing protocol block (port of `rust_core.py:render_message`,
+/// golden-verified): `Team Agent message from {sender}[ for {task_id}]:\n\n{content}\n\n
+/// [team-agent-token:{message_id}]`. The worker (fake or real provider) only builds a result_envelope
+/// when it sees this block + extracts the token — the bare content gives WORKING but never a report
+/// (rt-host-a loop #4). token == message_id (exactly-once correlation).
+fn render_message(sender: &str, task_id: Option<&str>, content: &str, message_id: &str) -> String {
+    let mut header = format!("Team Agent message from {sender}");
+    if let Some(task_id) = task_id.filter(|t| !t.is_empty()) {
+        header.push_str(&format!(" for {task_id}"));
+    }
+    format!("{header}:\n\n{content}\n\n[team-agent-token:{message_id}]")
+}
+
+/// Resolve a recipient agent-id to a tmux-RESOLVABLE inject target: the persisted pane-id if present,
+/// else a session-qualified `SessionWindow` (state.session_name + the agent's window, defaulting to the
+/// id). NEVER the bare agent-id as a pane — a clientless coordinator cannot resolve that
+/// ("can't find pane: w1", rt-host-a loop #3). Mirrors `coordinator/tick.rs::capture_target`.
+fn resolve_inject_target(state: &serde_json::Value, recipient: &str) -> Target {
+    let agent = state.get("agents").and_then(|a| a.get(recipient));
+    if let Some(pane_id) = agent
+        .and_then(|a| a.get("pane_id"))
+        .and_then(serde_json::Value::as_str)
+        .filter(|s| !s.is_empty())
+    {
+        return Target::Pane(PaneId::new(pane_id));
+    }
+    let session = state
+        .get("session_name")
+        .and_then(serde_json::Value::as_str)
+        .unwrap_or_default();
+    let window = agent
+        .and_then(|a| a.get("window"))
+        .and_then(serde_json::Value::as_str)
+        .filter(|s| !s.is_empty())
+        .unwrap_or(recipient);
+    Target::SessionWindow {
+        session: SessionName::new(session),
+        window: WindowName::new(window),
+    }
+}
+
+/// `_deliver_pending_messages` (`delivery.py:484`):扫 pending 队列逐条投递;busy 收件人写
+/// `send.deferred_busy` 跳过 (**不丢**,card §131)。返回投递的 message_id 列表。
+pub fn deliver_pending_messages(
+    workspace: &Path,
+    state: &serde_json::Value,
+    transport: &dyn Transport,
+    event_log: &EventLog,
+) -> Result<Vec<String>, MessagingError> {
+    let store = MessageStore::open(workspace)?;
+    let message_ids = {
+        let conn = crate::db::schema::open_db(store.db_path())?;
+        let mut stmt = conn.prepare(
+            "select message_id from messages
+             where status in ('pending', 'accepted')
+             order by created_at, message_id",
+        )?;
+        let rows = stmt.query_map([], |row| row.get::<_, String>(0))?;
+        rows.collect::<Result<Vec<_>, _>>()?
+    };
+    let mut delivered = Vec::new();
+    for message_id in message_ids {
+        if let Some(message) = message_for_delivery(&store, &message_id)? {
+            if recipient_is_busy(state, &message.recipient) {
+                event_log.write(
+                    "send.deferred_busy",
+                    serde_json::json!({
+                        "message_id": message_id,
+                        "sender": message.sender,
+                        "recipient": message.recipient,
+                        "reason": "recipient_busy",
+                    }),
+                )?;
+                continue;
+            }
+        }
+        let outcome = deliver_pending_message(workspace, &store, transport, &message_id, event_log, state)?;
+        if outcome.ok {
+            delivered.push(message_id);
+        }
+    }
+    Ok(delivered)
+}
+
+struct PendingMessage {
+    sender: String,
+    recipient: String,
+    content: String,
+    task_id: Option<String>,
+}
+
+fn message_for_delivery(
+    store: &MessageStore,
+    message_id: &str,
+) -> Result<Option<PendingMessage>, MessagingError> {
+    let conn = crate::db::schema::open_db(store.db_path())?;
+    let message = conn
+        .query_row(
+            "select sender, recipient, content, task_id from messages where message_id = ?1",
+            params![message_id],
+            |row| {
+                Ok(PendingMessage {
+                    sender: row.get::<_, String>(0)?,
+                    recipient: row.get::<_, String>(1)?,
+                    content: row.get::<_, String>(2)?,
+                    task_id: row.get::<_, Option<String>>(3)?,
+                })
+            },
+        )
+        .optional()?;
+    Ok(message)
+}
+
+/// Pre-inject gate (Contract B): peek the recipient pane and answer "is there an
+/// actionable Codex startup prompt right now (trust menu or update prompt)" using
+/// the SHARED provider/startup_prompt recognizer — no second classifier, no provider
+/// API calls. Returns `false` if capture fails so non-Codex providers (or any pane
+/// without the trust-menu shape) keep flowing through normal delivery.
+fn recipient_pane_has_actionable_startup_prompt(
+    transport: &dyn Transport,
+    state: &serde_json::Value,
+    recipient: &str,
+    target: &Target,
+) -> bool {
+    let agent = state
+        .get("agents")
+        .and_then(serde_json::Value::as_object)
+        .and_then(|agents| agents.get(recipient));
+    let provider = agent
+        .and_then(|agent| agent.get("provider"))
+        .and_then(serde_json::Value::as_str);
+    if !matches!(provider, Some("codex")) {
+        return false;
+    }
+    // step2-retry/scrollback root-cause (rt binary 6c9c6c1c): once the agent's
+    // `startup_prompts` has been flipped to `handled`/`complete`, the trust modal
+    // has been answered and is the AUTHORITATIVE record of "no actionable startup
+    // prompt remains". A `tmux capture-pane -S -` Full capture STILL contains the
+    // dismissed modal text in scrollback ("Do you trust …" + `› 1. Yes, continue`),
+    // so the recognizer's actionable-shape override matches the residue and the
+    // delivery gate would loop forever (49-attempt no-deliver in real machine).
+    // Trust the state (same source step1-idem uses) and skip the classify entirely.
+    let startup_prompts = agent
+        .and_then(|agent| agent.get("startup_prompts"))
+        .and_then(serde_json::Value::as_str);
+    if matches!(startup_prompts, Some("handled" | "complete")) {
+        return false;
+    }
+    let captured = match std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
+        transport.capture(target, crate::transport::CaptureRange::Full)
+    })) {
+        Ok(Ok(captured)) => captured.text,
+        _ => return false,
+    };
+    matches!(
+        crate::provider::classify_codex_startup_screen(&captured),
+        crate::provider::StartupScreenDecision::AnswerWorkspaceTrust
+            | crate::provider::StartupScreenDecision::SkipUpdatePrompt
+    )
+}
+
+fn recipient_is_busy(state: &serde_json::Value, recipient: &str) -> bool {
+    state
+        .get("agents")
+        .and_then(serde_json::Value::as_object)
+        .and_then(|agents| agents.get(recipient))
+        .and_then(|agent| agent.get("status"))
+        .and_then(serde_json::Value::as_str)
+        == Some("busy")
+}
+
+/// `_handle_trust_retry_needed` (`delivery.py:221`):trust 应答失败时调度有界退避重试
+/// (`attempt < MAX` → schedule;`>= MAX` → 终态 mark failed + `trust_auto_answer_exhausted`)。
+pub fn handle_trust_retry_needed(
+    store: &MessageStore,
+    payload: &TrustRetryPayload,
+    event_log: &EventLog,
+) -> Result<DeliveryOutcome, MessagingError> {
+    if payload.attempt >= payload.max_attempts {
+        let _ = store.mark(&payload.message_id, "failed", Some("trust_auto_answer_exhausted"));
+        event_log.write(
+            "leader_panes.trust_auto_answer_exhausted",
+            serde_json::json!({"message_id": payload.message_id, "attempt": payload.attempt}),
+        )?;
+        return Ok(DeliveryOutcome {
+            ok: false,
+            status: DeliveryStatus::TrustAutoAnswerExhausted,
+            message_status: MessageStatusShadow("failed".to_string()),
+            message_id: Some(payload.message_id.clone()),
+            verification: None,
+            stage: Some(DeliveryStage::TrustAutoAnswerDismissalWait),
+            reason: None,
+            channel: None,
+        });
+    }
+    let next_attempt = payload.attempt.saturating_add(1);
+    let backoff = super::TRUST_RETRY_BACKOFF_SECONDS
+        .iter()
+        .find_map(|(attempt, seconds)| (*attempt == next_attempt).then_some(*seconds))
+        .unwrap_or(30);
+    let due_at = (chrono::Utc::now() + chrono::Duration::seconds(i64::from(backoff))).to_rfc3339();
+    let conn = crate::db::schema::open_db(store.db_path())?;
+    conn.execute(
+        "insert into scheduled_events(owner_team_id, due_at, target, kind, payload_json, status, created_at)
+         values (null, ?1, ?2, 'trust_retry', ?3, 'pending', ?4)",
+        params![
+            due_at,
+            payload.first_target.as_str(),
+            serde_json::json!({
+                "message_id": payload.message_id,
+                "attempt": next_attempt,
+                "max_attempts": payload.max_attempts,
+                "first_target": payload.first_target.as_str(),
+            })
+            .to_string(),
+            chrono::Utc::now().to_rfc3339(),
+        ],
+    )?;
+    let _ = store.mark(&payload.message_id, "queued_until_trust", None);
+    event_log.write(
+        "leader_panes.trust_auto_answer_retry_scheduled",
+        serde_json::json!({"message_id": payload.message_id, "attempt": next_attempt, "due_at": due_at}),
+    )?;
+    Ok(DeliveryOutcome {
+        ok: false,
+        status: DeliveryStatus::RetryScheduled,
+        message_status: MessageStatusShadow("queued_until_trust".to_string()),
+        message_id: Some(payload.message_id.clone()),
+        verification: None,
+        stage: Some(DeliveryStage::TrustAutoAnswerDismissalWait),
+        reason: None,
+        channel: None,
+    })
+}
+
+/// `_execute_trust_retry` (`delivery.py:330`):trust_retry scheduled event 的消费者 ——
+/// 把行重置回 `accepted`,attempt 穿透,重跑 `_deliver_pending_message`。
+pub fn execute_trust_retry(
+    workspace: &Path,
+    store: &MessageStore,
+    transport: &dyn Transport,
+    payload: &TrustRetryPayload,
+    event_log: &EventLog,
+    owner_team_id: Option<&TeamKey>,
+) -> Result<DeliveryOutcome, MessagingError> {
+    let _ = owner_team_id;
+    let _ = store.mark(&payload.message_id, "accepted", None);
+    let state = crate::state::persist::load_runtime_state(workspace)?;
+    deliver_pending_message(workspace, store, transport, &payload.message_id, event_log, &state)
+}
+
+/// `_record_turn_open_if_leader_to_worker` (`delivery.py:430`):**take-over arm 来自真实投递**
+/// (card §121) —— 仅 leader→worker 注入**成功后**才调 `record_turn_open_after_delivery`,绝不凭空 arm。
+pub fn record_turn_open_if_leader_to_worker(
+    workspace: &Path,
+    state: &serde_json::Value,
+    sender: &str,
+    recipient: &str,
+    delivered: &DeliveryOutcome,
+    event_log: &EventLog,
+) -> Result<(), MessagingError> {
+    let _ = state;
+    if !delivered.ok || !matches!(sender, "leader" | "Leader") || recipient == "leader" {
+        return Ok(());
+    }
+    let mut state = crate::state::persist::load_runtime_state(workspace)?;
+    let Some(root) = state.as_object_mut() else {
+        return Ok(());
+    };
+    let coordinator = root
+        .entry("coordinator")
+        .or_insert_with(|| serde_json::json!({}));
+    if let Some(obj) = coordinator.as_object_mut() {
+        obj.insert(
+            "turn_open".to_string(),
+            serde_json::json!({"armed": true, "node_id": recipient, "turn_id": delivered.message_id}),
+        );
+    }
+    crate::state::persist::save_runtime_state(workspace, &state)?;
+    event_log.write(
+        "turn_open.armed_after_delivery",
+        serde_json::json!({"agent_id": recipient, "message_id": delivered.message_id}),
+    )?;
+    Ok(())
+}
+
+/// `_stamp_first_send_at_if_leader_to_worker` (`delivery.py:380`):首次 leader→worker 投递戳
+/// `first_send_at` (step 13 restart Route B atomicity 决策读它)。
+pub fn stamp_first_send_at_if_leader_to_worker(
+    workspace: &Path,
+    state: &serde_json::Value,
+    sender: &str,
+    recipient: &str,
+) -> Result<(), MessagingError> {
+    let _ = state;
+    if !matches!(sender, "leader" | "Leader") || recipient == "leader" {
+        return Ok(());
+    }
+    let mut state = crate::state::persist::load_runtime_state(workspace)?;
+    let now = chrono::Utc::now().to_rfc3339();
+    if let Some(agent) = state
+        .get_mut("agents")
+        .and_then(serde_json::Value::as_object_mut)
+        .and_then(|agents| agents.get_mut(recipient))
+        .and_then(serde_json::Value::as_object_mut)
+    {
+        if !agent.contains_key("first_send_at") || agent.get("first_send_at").is_some_and(serde_json::Value::is_null) {
+            agent.insert("first_send_at".to_string(), serde_json::Value::String(now));
+            crate::state::persist::save_runtime_state(workspace, &state)?;
+        }
+    }
+    Ok(())
+}
+
+/// `retry_injection_after_trust_auto_answer` (`trust_auto_answer.py`):leader 路径 trust 应答
+/// 后重注入 (查 pane_width fail-safe + attempt_trust_auto_answer + 等 dismissal + 重 inject)。
+pub fn retry_injection_after_trust_auto_answer(
+    workspace: &Path,
+    state: &serde_json::Value,
+    transport: &dyn Transport,
+    target: &Target,
+    text: &str,
+    provider: Provider,
+    event_log: &EventLog,
+) -> Result<DeliveryOutcome, MessagingError> {
+    let _ = (workspace, state, transport, target, text, provider, event_log);
+    Ok(DeliveryOutcome {
+        ok: false,
+        status: DeliveryStatus::RetryScheduled,
+        message_status: MessageStatusShadow("retry_scheduled".to_string()),
+        message_id: None,
+        verification: None,
+        stage: Some(DeliveryStage::TrustAutoAnswerDismissalWait),
+        reason: None,
+        channel: None,
+    })
+}