diff --git a/src-tauri/Cargo.toml b/src-tauri/Cargo.toml
index 2eee213b55..66dd64bdcb 100644
--- a/src-tauri/Cargo.toml
+++ b/src-tauri/Cargo.toml
@@ -22,6 +22,13 @@ crate-type = ["rlib"]
 
 [features]
 default = ["tauri-runtime", "browser-child"]
+# Tauri's release switch: with it, `tauri` embeds `frontendDist` and the app
+# loads its bundled assets; without it, `DEP_TAURI_DEV=true` and the app asks
+# `devUrl` (http://localhost:3000) for a dev server that a plain
+# `cargo build --release` never starts — an `ERR_CONNECTION_REFUSED` window.
+# `pnpm tauri build` turns this on; a hand-run cargo build must name it.
+custom-protocol = ["tauri/custom-protocol"]
+
 tauri-runtime = [
     "dep:tauri",
     "dep:tauri-plugin-opener",
diff --git a/src-tauri/src/acp/connection.rs b/src-tauri/src/acp/connection.rs
index e0717356fd..2a85fa9a44 100644
--- a/src-tauri/src/acp/connection.rs
+++ b/src-tauri/src/acp/connection.rs
@@ -3556,6 +3556,113 @@ async fn admit_permission(
     reply: PermissionReply,
     card: QueuedPermission,
 ) {
+    // Auto-approve: answer straight away instead of raising a card, for
+    // connections a codeg-driven launch marked with `CODEG_AUTO_APPROVE=1`.
+    // That env var is inherited by every ACP child codeg spawns, so a
+    // delegated sub-agent can never park on a card nobody is looking at.
+    //
+    // The motivating case is `deepseek-acp` 0.9.0: its ACP bridge has no
+    // approval MODE at all (`set_mode` / `availableModes` appear nowhere in
+    // the package), and the only options it ever offers are allow-once and
+    // reject-once (hard-coded in its `approval/request` handler). A write
+    // tool call therefore parked forever with its parent delegation waiting
+    // on it — twice for over 8 hours in 2026-11.
+    //
+    // Deliberately NOT gated on an agent type: the flag is the whole policy.
+    // `allow_once` only — never `allow_always`, which would write a standing
+    // rule into the user's own agent settings (see
+    // [`PermissionOptionInfo::allow_once_in`]). A request offering no such
+    // option falls through to the card exactly as before.
+    //
+    // A `cli_run` connection is NEVER auto-approved. That tag marks a headless
+    // `codeg-cli` run — its root connection, or a delegation child that
+    // inherited it at spawn — and its `--permissions` policy is authoritative
+    // there: `deny`, `reads`, `fail` and `wait` all mean something that
+    // answering `allow_once` here would silently overrule, along with the
+    // run's own permission reporting. Those connections keep the card/event
+    // path, which is what applies their policy.
+    if let Some(option_id) = crate::acp::auto_approve_option(
+        crate::acp::auto_approve_enabled(),
+        state.read().await.cli_run.is_some(),
+        &card.options,
+    ) {
+            let scope = permission_log_scope(state).await;
+            tracing::info!(
+                "[ACP] permission {} auto-approved {scope} (option={option_id}, CODEG_AUTO_APPROVE)",
+                card.request_id
+            );
+            let tool_call_id = crate::acp::session_state::extract_tool_call_id(&card.tool_call);
+            let kind = PermissionOptionInfo::allow_once_in(&card.options)
+                .map(|o| o.kind.clone())
+                .unwrap_or_default();
+            let name = PermissionOptionInfo::allow_once_in(&card.options)
+                .map(|o| o.name.clone())
+                .unwrap_or_default();
+            // The agent may have withdrawn the request while the `state` read
+            // above was waiting (or before it ever got here). Answering
+            // `allow_once` then would grant a call the agent already cancelled,
+            // and the `ToolApproval` record would claim a decision nobody made.
+            // Answer the withdrawal instead, and record nothing — exactly what
+            // `watch_permission_withdrawal` does for a card.
+            if reply.cancellation().is_cancelled() {
+                tracing::info!(
+                    "[ACP] permission {} withdrawn before auto-approval {scope}",
+                    card.request_id
+                );
+                match reply {
+                    PermissionReply::Acp(responder) => {
+                        let _ = responder.respond(RequestPermissionResponse::new(
+                            RequestPermissionOutcome::Cancelled,
+                        ));
+                    }
+                    PermissionReply::CodexElicitation { responder, .. } => {
+                        let _ = responder.respond(
+                            serde_json::to_value(
+                                crate::acp::question::elicitation_cancel_response(),
+                            )
+                            .unwrap_or_default(),
+                        );
+                    }
+                }
+                return;
+            }
+            // Answer the way a user's click does — `respond_selected` is the
+            // same responder tail a card runs — but without parking the reply
+            // in the queue first: nothing will ever promote or withdraw it.
+            match reply {
+                PermissionReply::Acp(responder) => {
+                    let outcome = RequestPermissionOutcome::Selected(
+                        SelectedPermissionOutcome::new(option_id),
+                    );
+                    let _ = responder.respond(RequestPermissionResponse::new(outcome));
+                }
+                PermissionReply::CodexElicitation {
+                    responder,
+                    approval,
+                } => {
+                    let response =
+                        crate::acp::question::build_elicitation_approval_response(&approval, &option_id);
+                    let _ = responder.respond(serde_json::to_value(response).unwrap_or_default());
+                }
+            }
+            if !tool_call_id.is_empty() {
+                emit_with_state(
+                    state,
+                    emitter,
+                    AcpEvent::ToolApproval {
+                        approval: ToolApprovalInfo {
+                            tool_call_id,
+                            option_kind: kind,
+                            option_name: name,
+                            origin: crate::acp::AUTO_APPROVE_ORIGIN.to_string(),
+                            decided_at: chrono::Utc::now(),
+                        },
+                    },
+                )
+                .await;
+            }
+            return;
+    }
     let cancellation = reply.cancellation();
     let (settled_tx, settled_rx) = oneshot::channel();
     let request_id = card.request_id.clone();
diff --git a/src-tauri/src/acp/file_system_runtime.rs b/src-tauri/src/acp/file_system_runtime.rs
index 95f2f5d87e..40a3b023af 100644
--- a/src-tauri/src/acp/file_system_runtime.rs
+++ b/src-tauri/src/acp/file_system_runtime.rs
@@ -320,9 +320,18 @@ fn agent_data_roots(agent_type: AgentType, runtime_env: &BTreeMap<String, String
 ///   Measured on 2.4.1 with a 3,000,000-byte file: given the whole file, a
 ///   `Read` handed the model its first 958 lines, 99,918 characters ending in
 ///   Kimi's own "Next Read" hint, and one with `n_lines=10` the ten lines.
+///
+/// DeepSeek Harness (`dsh-acp` 0.9.0) is the third: its `FsClient` ships the
+/// `read` tool's `offset`/`limit` as its OWN page arguments and asks
+/// `fs/read_text_file` for the whole file, then pages the answer itself — so
+/// the `line`/`limit` hint never reaches it either, and every read of a file
+/// past 2 MiB fails with `Invalid params` the model cannot act on. Measured
+/// twice in 2026-11 (a 2.65 MiB single-line bundle), where the child then sat
+/// in `Prompting` until it was cancelled by hand. It carries no approval mode
+/// and no retry, so a read that errors is a read it never recovers from.
 fn read_response_limit(agent_type: AgentType) -> usize {
     match agent_type {
-        AgentType::Antigravity | AgentType::KimiCode => {
+        AgentType::Antigravity | AgentType::KimiCode | AgentType::DeepSeek => {
             usize::try_from(FS_MAX_FILE_SIZE_BYTES).unwrap_or(usize::MAX)
         }
         _ => FS_MAX_READ_RESPONSE_BYTES,
diff --git a/src-tauri/src/acp/mod.rs b/src-tauri/src/acp/mod.rs
index 89de586e1e..4aebc4af68 100644
--- a/src-tauri/src/acp/mod.rs
+++ b/src-tauri/src/acp/mod.rs
@@ -103,3 +103,185 @@ pub fn continued_session_ids(agent_type: crate::models::AgentType, session_id: &
     }
     crate::acp_transcript::continuation_ancestors(registry::registry_id_for(agent_type), session_id)
 }
+
+/// `CODEG_AUTO_APPROVE=1` turns codeg into the approver for every ACP child it
+/// launches: a `session/request_permission` is answered with its `allow_once`
+/// option instead of being raised as a card (#447).
+///
+/// Read from the process env, which every ACP child inherits, so the flag is
+/// inherited rather than plumbed per-connection. Off unless the variable is
+/// exactly `1`, `true` or `yes` — an unset or empty value keeps the ordinary
+/// card flow.
+///
+/// This exists because some adapters cannot express an approval policy at all:
+/// `deepseek-acp` 0.9.0 has no `set_mode`, no `availableModes`, and a
+/// hard-coded allow-once / reject-once pair in its `approval/request` handler,
+/// so a write tool call parked forever and its parent delegation waited on it.
+pub fn auto_approve_enabled() -> bool {
+    parse_auto_approve(std::env::var(crate::acp::AUTO_APPROVE_ENV).ok().as_deref())
+}
+
+/// The flag's whole reading of its value, split out so it is testable without
+/// touching the process env (which is shared by every test in the binary).
+///
+/// `None` (unset) and any unrecognized string keep the card flow. Accepting
+/// `1`/`true`/`yes` in any case is deliberate: the var is set by hand, in a
+/// shell or an OS environment dialog, and a `TRUE` that silently did nothing
+/// would be worse than the extra leniency.
+fn parse_auto_approve(raw: Option<&str>) -> bool {
+    matches!(
+        raw.unwrap_or_default().trim().to_ascii_lowercase().as_str(),
+        "1" | "true" | "yes"
+    )
+}
+
+/// The env var [`auto_approve_enabled`] reads.
+pub const AUTO_APPROVE_ENV: &str = "CODEG_AUTO_APPROVE";
+
+/// The [`crate::acp::types::ToolApprovalInfo::origin`] an auto-approved
+/// permission records.
+pub const AUTO_APPROVE_ORIGIN: &str = "auto_approve";
+
+/// The whole "should this card be auto-approved?" decision, split out so its
+/// every branch is unit-testable without a live ACP connection.
+///
+/// `None` means "leave it on the card path"; `Some(id)` means "answer with this
+/// option". Three gates, in the order the caller checks them:
+///
+/// 1. the flag must be on;
+/// 2. the connection must NOT carry a `cli_run` tag — that tag marks a
+///    headless `codeg-cli` run (or a delegation child that inherited it) whose
+///    `--permissions` policy is authoritative, and answering here would
+///    overrule `deny` / `reads` / `fail` / `wait`;
+/// 3. the request must actually offer an `allow_once` option. Rejection
+///    deliberately comes first in some adapters (claude-agent-acp 0.91.0
+///    leads with it for `defaultToNo` calls), so this is a kind lookup, never
+///    a positional one. `allow_always` alone is NOT enough: picking it would
+///    write a standing rule into the user's own agent settings.
+pub fn auto_approve_option(
+    enabled: bool,
+    cli_run_tagged: bool,
+    options: &[crate::acp::types::PermissionOptionInfo],
+) -> Option<String> {
+    if !enabled || cli_run_tagged {
+        return None;
+    }
+    crate::acp::types::PermissionOptionInfo::allow_once_in(options)
+        .map(|o| o.option_id.clone())
+}
+
+#[cfg(test)]
+mod auto_approve_flag_tests {
+    use super::{auto_approve_option, parse_auto_approve, AUTO_APPROVE_ORIGIN};
+    use crate::acp::types::PermissionOptionInfo;
+
+    fn opt(id: &str, kind: &str) -> PermissionOptionInfo {
+        PermissionOptionInfo {
+            option_id: id.to_string(),
+            name: id.to_string(),
+            kind: kind.to_string(),
+            meta: None,
+        }
+    }
+
+    /// The enabled spellings. Case and surrounding whitespace are irrelevant
+    /// because the value is typed by hand into a shell or an OS env dialog.
+    #[test]
+    fn recognises_the_enabled_spellings() {
+        for raw in ["1", "true", "TRUE", "yes", " Yes ", "\ttrue\n"] {
+            assert!(parse_auto_approve(Some(raw)), "{raw:?} must enable it");
+        }
+    }
+
+    /// Everything else keeps the card flow. Unset is the important one: a
+    /// machine that never set the variable must behave exactly as before, and
+    /// an empty value is how a shell writes "set but blank".
+    #[test]
+    fn everything_else_leaves_the_card_flow_alone() {
+        for raw in [None, Some(""), Some("   "), Some("0"), Some("false"), Some("no"), Some("on")] {
+            assert!(
+                !parse_auto_approve(raw),
+                "{raw:?} must NOT enable auto-approval"
+            );
+        }
+    }
+
+    /// The origin string is a wire value the frontend's ApprovalChip matches on
+    /// (`src/components/message/activity/approval-chip.tsx`). Changing it here
+    /// without changing that file would silently render an unattended approval
+    /// as a person's click.
+    #[test]
+    fn origin_is_the_string_the_frontend_expects() {
+        assert_eq!(AUTO_APPROVE_ORIGIN, "auto_approve");
+    }
+
+    /// Regression for the review's "Preserve CLI permission policies" finding:
+    /// a `cli_run` connection (a headless `codeg-cli` run, or a delegation
+    /// child that inherited its tag) must keep the card path so its
+    /// `--permissions` policy stays authoritative.
+    #[test]
+    fn cli_run_tagged_connection_is_never_auto_approved() {
+        let opts = vec![opt("allow", "allow_once"), opt("deny", "reject_once")];
+        assert_eq!(
+            auto_approve_option(true, true, &opts),
+            None,
+            "the run's permission ceiling must win over CODEG_AUTO_APPROVE"
+        );
+    }
+
+    /// The flag off is the default state of every machine that never set it.
+    #[test]
+    fn disabled_flag_keeps_the_card_path() {
+        let opts = vec![opt("allow", "allow_once")];
+        assert_eq!(auto_approve_option(false, false, &opts), None);
+    }
+
+    /// claude-agent-acp 0.91.0 leads with the rejection for `defaultToNo`
+    /// calls. Selection is by kind, never by position.
+    #[test]
+    fn chooses_allow_once_even_when_rejection_comes_first() {
+        let opts = vec![
+            opt("reject-1", "reject_once"),
+            opt("allow-1", "allow_once"),
+        ];
+        assert_eq!(
+            auto_approve_option(true, false, &opts),
+            Some("allow-1".to_string())
+        );
+    }
+
+    /// `allow_always` alone must NOT be auto-picked: it would write a standing
+    /// rule into the user's own agent settings. Fall through to the card.
+    #[test]
+    fn only_allow_always_available_falls_through_to_the_card() {
+        let opts = vec![
+            opt("always", "allow_always"),
+            opt("reject", "reject_once"),
+        ];
+        assert_eq!(auto_approve_option(true, false, &opts), None);
+    }
+
+    /// An empty option list, and a list with no allow-kind at all, are both
+    /// ordinary card cases — no panic, no approval invented.
+    #[test]
+    fn no_usable_option_falls_through_to_the_card() {
+        assert_eq!(auto_approve_option(true, false, &[]), None);
+        assert_eq!(
+            auto_approve_option(true, false, &[opt("reject", "reject_once")]),
+            None
+        );
+    }
+
+    /// The happy path, once, to pin the returned id rather than just Some(_).
+    #[test]
+    fn enabled_non_cli_request_with_allow_once_is_answered() {
+        let opts = vec![
+            opt("allow-1", "allow_once"),
+            opt("reject-1", "reject_once"),
+        ];
+        assert_eq!(
+            auto_approve_option(true, false, &opts),
+            Some("allow-1".to_string())
+        );
+    }
+}
diff --git a/src-tauri/src/commands/conversations.rs b/src-tauri/src/commands/conversations.rs
index a21907626b..b728e3524e 100644
--- a/src-tauri/src/commands/conversations.rs
+++ b/src-tauri/src/commands/conversations.rs
@@ -1567,7 +1567,24 @@ pub async fn get_folder_conversation_core(
         let agent_env = agent_setting_service::load_agent_env(conn, at).await;
         tokio::task::spawn_blocking(move || -> Result<_, AppCommandError> {
             let parser = build_agent_parser(at, &agent_env);
-            match parser.get_conversation(&eid) {
+            let parse_started = std::time::Instant::now();
+            let parsed = parser.get_conversation(&eid);
+            // Duration of a full transcript parse is the quantity the turn
+            // window cannot bound: `apply_turn_window` slices AFTER this
+            // returns, so a 80 MB session is read and turned into messages
+            // whatever the window asks for. Logged for every detail load
+            // because the only way to know whether the window needs to move
+            // into the parsers is to see this number on real sessions.
+            let parse_ms = parse_started.elapsed().as_millis() as u64;
+            tracing::info!(
+                agent = %at,
+                external_id = %eid,
+                parse_ms,
+                turns = parsed.as_ref().map(|d| d.turns.len()).unwrap_or(0),
+                slow = parse_ms > crate::commands::conversations::SLOW_PARSE_MS,
+                "[conversations] transcript parsed"
+            );
+            match parsed {
                 Ok(d) => {
                     // Claude `/clear` (and similar on-disk id changes) make
                     // the parser resolve a different uuid than we asked for.
@@ -2002,6 +2019,11 @@ fn apply_in_flight_message_id(
     Some(pending.message_id.clone())
 }
 
+/// A full transcript parse slower than this is logged as `slow`. Parsing runs
+/// before the turn window slices anything, so this is the number that decides
+/// whether windows have to move into the parsers themselves.
+pub const SLOW_PARSE_MS: u64 = 500;
+
 /// Resolve the raw `tailTurns` / `fromIndex` request fields into a window
 /// selector. `None` when neither is present (legacy full response); an error
 /// when both are (the two coordinate systems are mutually exclusive).
diff --git a/src/components/message/activity/approval-chip.tsx b/src/components/message/activity/approval-chip.tsx
index 88db10f0eb..674974d932 100644
--- a/src/components/message/activity/approval-chip.tsx
+++ b/src/components/message/activity/approval-chip.tsx
@@ -52,7 +52,9 @@ export function ApprovalChip({ approval }: { approval: ToolApproval }) {
       ? t("viaChatChannel")
       : approval.origin === "cli_policy"
         ? t("byPolicy")
-        : null
+        : approval.origin === "auto_approve"
+          ? t("byAutoApprove")
+          : null
   const title = [
     approval.optionName,
     who,
diff --git a/src/i18n/messages/ar.json b/src/i18n/messages/ar.json
index 092624ad9c..571cb5c5ab 100644
--- a/src/i18n/messages/ar.json
+++ b/src/i18n/messages/ar.json
@@ -4097,7 +4097,8 @@
           "allowAlways": "مسموح دائمًا",
           "rejected": "رُفض",
           "viaChatChannel": "عبر قناة دردشة",
-          "byPolicy": "بحسب سياسة أذونات التشغيل"
+          "byPolicy": "بحسب سياسة أذونات التشغيل",
+          "byAutoApprove": "تلقائيًا بواسطة CODEG_AUTO_APPROVE"
         }
       },
       "messageNav": {
diff --git a/src/i18n/messages/de.json b/src/i18n/messages/de.json
index 717bf1ec45..bfb325b97f 100644
--- a/src/i18n/messages/de.json
+++ b/src/i18n/messages/de.json
@@ -4097,7 +4097,8 @@
           "allowAlways": "immer erlaubt",
           "rejected": "abgelehnt",
           "viaChatChannel": "über einen Chat-Kanal",
-          "byPolicy": "durch die Berechtigungsrichtlinie des Laufs"
+          "byPolicy": "durch die Berechtigungsrichtlinie des Laufs",
+          "byAutoApprove": "automatisch durch CODEG_AUTO_APPROVE"
         }
       },
       "messageNav": {
diff --git a/src/i18n/messages/en.json b/src/i18n/messages/en.json
index 95c2505440..db9e036ef4 100644
--- a/src/i18n/messages/en.json
+++ b/src/i18n/messages/en.json
@@ -4097,7 +4097,8 @@
           "allowAlways": "always allowed",
           "rejected": "declined",
           "viaChatChannel": "via a chat channel",
-          "byPolicy": "by the run's permission policy"
+          "byPolicy": "by the run's permission policy",
+          "byAutoApprove": "automatically, by CODEG_AUTO_APPROVE"
         }
       },
       "messageNav": {
diff --git a/src/i18n/messages/es.json b/src/i18n/messages/es.json
index 8410c0165b..c9da907c5c 100644
--- a/src/i18n/messages/es.json
+++ b/src/i18n/messages/es.json
@@ -4097,7 +4097,8 @@
           "allowAlways": "siempre permitido",
           "rejected": "rechazado",
           "viaChatChannel": "desde un canal de chat",
-          "byPolicy": "por la política de permisos de la ejecución"
+          "byPolicy": "por la política de permisos de la ejecución",
+          "byAutoApprove": "automáticamente, por CODEG_AUTO_APPROVE"
         }
       },
       "messageNav": {
diff --git a/src/i18n/messages/fr.json b/src/i18n/messages/fr.json
index 1e03d8719c..fe1682106e 100644
--- a/src/i18n/messages/fr.json
+++ b/src/i18n/messages/fr.json
@@ -4097,7 +4097,8 @@
           "allowAlways": "toujours autorisé",
           "rejected": "refusé",
           "viaChatChannel": "depuis un canal de discussion",
-          "byPolicy": "par la politique d'autorisation de l'exécution"
+          "byPolicy": "par la politique d'autorisation de l'exécution",
+          "byAutoApprove": "automatiquement, par CODEG_AUTO_APPROVE"
         }
       },
       "messageNav": {
diff --git a/src/i18n/messages/ja.json b/src/i18n/messages/ja.json
index c4695f5207..6560bd423d 100644
--- a/src/i18n/messages/ja.json
+++ b/src/i18n/messages/ja.json
@@ -4097,7 +4097,8 @@
           "allowAlways": "常に許可",
           "rejected": "拒否",
           "viaChatChannel": "チャットチャネル経由",
-          "byPolicy": "実行の権限ポリシーによる"
+          "byPolicy": "実行の権限ポリシーによる",
+          "byAutoApprove": "CODEG_AUTO_APPROVE による自動承認"
         }
       },
       "messageNav": {
diff --git a/src/i18n/messages/ko.json b/src/i18n/messages/ko.json
index 412f820533..57fb23b21e 100644
--- a/src/i18n/messages/ko.json
+++ b/src/i18n/messages/ko.json
@@ -4097,7 +4097,8 @@
           "allowAlways": "항상 허용",
           "rejected": "거부됨",
           "viaChatChannel": "채팅 채널에서",
-          "byPolicy": "실행 권한 정책에 따름"
+          "byPolicy": "실행 권한 정책에 따름",
+          "byAutoApprove": "CODEG_AUTO_APPROVE에 의한 자동 승인"
         }
       },
       "messageNav": {
diff --git a/src/i18n/messages/pt.json b/src/i18n/messages/pt.json
index 503cba0504..6c0258f30b 100644
--- a/src/i18n/messages/pt.json
+++ b/src/i18n/messages/pt.json
@@ -4097,7 +4097,8 @@
           "allowAlways": "sempre permitido",
           "rejected": "recusado",
           "viaChatChannel": "por um canal de chat",
-          "byPolicy": "pela política de permissões da execução"
+          "byPolicy": "pela política de permissões da execução",
+          "byAutoApprove": "automaticamente, por CODEG_AUTO_APPROVE"
         }
       },
       "messageNav": {
diff --git a/src/i18n/messages/zh-CN.json b/src/i18n/messages/zh-CN.json
index b3bfccb921..84c9c27640 100644
--- a/src/i18n/messages/zh-CN.json
+++ b/src/i18n/messages/zh-CN.json
@@ -4097,7 +4097,8 @@
           "allowAlways": "总是允许",
           "rejected": "已拒绝",
           "viaChatChannel": "来自聊天频道",
-          "byPolicy": "由运行的权限策略决定"
+          "byPolicy": "由运行的权限策略决定",
+          "byAutoApprove": "由 CODEG_AUTO_APPROVE 自动批准"
         }
       },
       "messageNav": {
diff --git a/src/i18n/messages/zh-TW.json b/src/i18n/messages/zh-TW.json
index 5bc94c985c..dac919a63e 100644
--- a/src/i18n/messages/zh-TW.json
+++ b/src/i18n/messages/zh-TW.json
@@ -4097,7 +4097,8 @@
           "allowAlways": "一律允許",
           "rejected": "已拒絕",
           "viaChatChannel": "來自聊天頻道",
-          "byPolicy": "由執行的權限原則決定"
+          "byPolicy": "由執行的權限原則決定",
+          "byAutoApprove": "由 CODEG_AUTO_APPROVE 自動核准"
         }
       },
       "messageNav": {