diff --git a/Cargo.lock b/Cargo.lock
index 2504aea..e2e2d24 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -141,7 +141,7 @@ version = "1.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "40c48f72fd53cd289104fc64099abca73db4166ad86ea0b4341abe65af83dadc"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]
@@ -152,7 +152,7 @@ checksum = "291e6a250ff86cd4a820112fb8898808a366d8f9f58ce16d1f538353ad55747d"
 dependencies = [
  "anstyle",
  "once_cell_polyfill",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]
@@ -1761,7 +1761,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
 dependencies = [
  "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -2113,9 +2113,9 @@ dependencies = [
 
 [[package]]
 name = "h2"
-version = "0.4.15"
+version = "0.4.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6cb093c84e8bd9b188d4c4a8cb6579fc016968d14c99882163cd3ff402a4f155"
+checksum = "a9f37a958b41b3b19ee2707c06439c0e9e547e847223eb791ecb0cb821c65e27"
 dependencies = [
  "atomic-waker",
  "bytes",
@@ -2380,7 +2380,7 @@ dependencies = [
  "libc",
  "percent-encoding",
  "pin-project-lite",
- "socket2 0.6.4",
+ "socket2 0.5.10",
  "tokio",
  "tower-service",
  "tracing",
@@ -3103,7 +3103,7 @@ version = "0.50.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -3822,7 +3822,7 @@ dependencies = [
  "quinn-udp",
  "rustc-hash 2.1.2",
  "rustls",
- "socket2 0.6.4",
+ "socket2 0.5.10",
  "thiserror 2.0.18",
  "tokio",
  "tracing",
@@ -3860,7 +3860,7 @@ dependencies = [
  "cfg_aliases",
  "libc",
  "once_cell",
- "socket2 0.6.4",
+ "socket2 0.5.10",
  "tracing",
  "windows-sys 0.60.2",
 ]
@@ -4421,14 +4421,14 @@ dependencies = [
  "errno",
  "libc",
  "linux-raw-sys 0.12.1",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
 name = "rustls"
-version = "0.23.40"
+version = "0.23.45"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ef86cd5876211988985292b91c96a8f2d298df24e75989a43a3c73f2d4d8168b"
+checksum = "0d41d731c7d2f962d1ccc364cec258de3c0e93b38c2fb3ba97ac74513048d634"
 dependencies = [
  "log",
  "once_cell",
@@ -4463,9 +4463,9 @@ dependencies = [
 
 [[package]]
 name = "rustls-webpki"
-version = "0.103.13"
+version = "0.103.15"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "61c429a8649f110dddef65e2a5ad240f747e85f7758a6bccc7e5777bd33f756e"
+checksum = "f3c3cf1d8b1e7d4927e2d154c3fcb02979afb9939629c62cd9048d4f07b60ac2"
 dependencies = [
  "ring",
  "rustls-pki-types",
@@ -4909,7 +4909,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "52d1cfed4120b4d927bf7c0f86d2087a4a7d6027c906d9f9d525a80573b9be51"
 dependencies = [
  "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]
@@ -5445,7 +5445,7 @@ dependencies = [
  "getrandom 0.3.4",
  "once_cell",
  "rustix 1.1.4",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -6252,7 +6252,7 @@ version = "0.1.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.48.0",
 ]
 
 [[package]]
diff --git a/crates/agent-runtime/src/backend.rs b/crates/agent-runtime/src/backend.rs
index 83c313c..86042ef 100644
--- a/crates/agent-runtime/src/backend.rs
+++ b/crates/agent-runtime/src/backend.rs
@@ -250,6 +250,11 @@ impl GenericCliBackend {
             timeout,
         )
     }
+    pub fn deepcode(timeout: Duration) -> Self {
+        // Headless one-shot; --json keeps the output machine-readable. The prompt
+        // is appended as the final argv by GenericCliBackend::execute.
+        Self::from_env("deepcode", "DEEPCODE_CMD", "deepcode exec --json", timeout)
+    }
 }
 
 #[async_trait]
diff --git a/crates/agent-runtime/src/main.rs b/crates/agent-runtime/src/main.rs
index b330bf3..cce3d20 100644
--- a/crates/agent-runtime/src/main.rs
+++ b/crates/agent-runtime/src/main.rs
@@ -117,6 +117,7 @@ fn backend_for(executor: &str, task_timeout: Duration) -> Arc<dyn CliAgentBacken
         "CODEX" => Arc::new(GenericCliBackend::codex(task_timeout)),
         "OPENCODE" => Arc::new(GenericCliBackend::opencode(task_timeout)),
         "CODEBUDDY" => Arc::new(GenericCliBackend::codebuddy(task_timeout)),
+        "DEEPCODE" => Arc::new(GenericCliBackend::deepcode(task_timeout)),
         _ => Arc::new(ClaudeBackend::new(task_timeout)),
     }
 }
diff --git a/crates/delivery/src/adapters/local.rs b/crates/delivery/src/adapters/local.rs
index 34b46e9..c31af2e 100644
--- a/crates/delivery/src/adapters/local.rs
+++ b/crates/delivery/src/adapters/local.rs
@@ -36,9 +36,12 @@ impl LocalCommandAgentExecutor {
         match kind {
             ExecutorKind::ClaudeCode => &self.claude_code,
             ExecutorKind::Codex => &self.codex,
-            // The local path only configures two argvs; OpenCode/CodeBuddy deliberately
-            // fall back to the claude argv (real routing lives in crates/agent-runtime).
-            ExecutorKind::OpenCode | ExecutorKind::CodeBuddy => &self.claude_code,
+            // The local path only configures two argvs; OpenCode/CodeBuddy/DeepCode
+            // deliberately fall back to the claude argv (real routing lives in
+            // crates/agent-runtime).
+            ExecutorKind::OpenCode | ExecutorKind::CodeBuddy | ExecutorKind::DeepCode => {
+                &self.claude_code
+            }
         }
     }
 }
diff --git a/crates/delivery/src/adapters/queue.rs b/crates/delivery/src/adapters/queue.rs
index 63caff4..99370ea 100644
--- a/crates/delivery/src/adapters/queue.rs
+++ b/crates/delivery/src/adapters/queue.rs
@@ -22,7 +22,7 @@ use crate::ports::{
     RuntimeInfo, WorkQueue, WorkSpec,
 };
 
-const KNOWN_CAPS: [ExecutorKind; 4] = ExecutorKind::ALL;
+const KNOWN_CAPS: [ExecutorKind; 5] = ExecutorKind::ALL;
 
 #[derive(Default)]
 pub struct InMemoryWorkQueue {
diff --git a/crates/delivery/src/adapters/redis_queue.rs b/crates/delivery/src/adapters/redis_queue.rs
index 86db1d8..42ae4d6 100644
--- a/crates/delivery/src/adapters/redis_queue.rs
+++ b/crates/delivery/src/adapters/redis_queue.rs
@@ -25,7 +25,7 @@ const ACKMAP: &str = "fleet:ackmap";
 const RT_INDEX: &str = "fleet:rt:index";
 const SEP: char = '\u{1f}';
 
-fn known_caps() -> [ExecutorKind; 4] {
+fn known_caps() -> [ExecutorKind; 5] {
     ExecutorKind::ALL
 }
 
diff --git a/crates/delivery/src/domain/delivery.rs b/crates/delivery/src/domain/delivery.rs
index c6b6771..d003ce3 100644
--- a/crates/delivery/src/domain/delivery.rs
+++ b/crates/delivery/src/domain/delivery.rs
@@ -12,13 +12,14 @@ pub enum ExecutorKind {
     Codex,
     OpenCode,
     CodeBuddy,
+    DeepCode,
 }
 
 impl ExecutorKind {
     /// Single authoritative list of executor kinds; queue adapters and tests consume
     /// it, so a new variant only touches this file.
-    pub const ALL: [ExecutorKind; 4] =
-        [Self::ClaudeCode, Self::Codex, Self::OpenCode, Self::CodeBuddy];
+    pub const ALL: [ExecutorKind; 5] =
+        [Self::ClaudeCode, Self::Codex, Self::OpenCode, Self::CodeBuddy, Self::DeepCode];
 
     pub fn as_str(&self) -> &'static str {
         match self {
@@ -26,6 +27,7 @@ impl ExecutorKind {
             Self::Codex => "CODEX",
             Self::OpenCode => "OPENCODE",
             Self::CodeBuddy => "CODEBUDDY",
+            Self::DeepCode => "DEEPCODE",
         }
     }
 
@@ -35,6 +37,7 @@ impl ExecutorKind {
             "CODEX" => Some(Self::Codex),
             "OPENCODE" => Some(Self::OpenCode),
             "CODEBUDDY" => Some(Self::CodeBuddy),
+            "DEEPCODE" => Some(Self::DeepCode),
             _ => None,
         }
     }
diff --git a/crates/server/src/jev.rs b/crates/server/src/jev.rs
new file mode 100644
index 0000000..a0afb9b
--- /dev/null
+++ b/crates/server/src/jev.rs
@@ -0,0 +1,245 @@
+//! Jev (TypeSafe System One) verification judge: a fast typed-judgment model.
+//!
+//! Jev answers `noul` (yes/no probability) questions about a blob of state —
+//! no prose, no code generation. That makes it a drop-in, cheap alternative to
+//! the LLM judge for the delivery verification gate: one `noul` per acceptance
+//! criterion, all in a single HTTP call; the gate passes only when every
+//! probability reaches the configured threshold (default 0.8).
+//!
+//! Env vars: `SHEPHERD_JEV_API_KEY` (falls back to `TYPESAFE_API_KEY`),
+//! `SHEPHERD_JEV_URL` (default `https://api.typesafe.ai/v1/systemone`),
+//! `SHEPHERD_JEV_MODEL` (default `jev-latest`),
+//! `SHEPHERD_JEV_THRESHOLD` (default `0.8`).
+
+use std::collections::BTreeMap;
+use std::time::Duration;
+
+use async_trait::async_trait;
+use serde::Deserialize;
+use serde_json::json;
+
+use orchestrator::ports::{DeliverableView, Judge, Verdict};
+
+const DEFAULT_URL: &str = "https://api.typesafe.ai/v1/systemone";
+const DEFAULT_MODEL: &str = "jev-latest";
+const DEFAULT_THRESHOLD: f64 = 0.8;
+
+#[derive(Deserialize)]
+struct SystemOneResponse {
+    #[serde(default)]
+    answers: BTreeMap<String, NoulAnswer>,
+}
+
+#[derive(Deserialize)]
+struct NoulAnswer {
+    noul: f64,
+}
+
+pub struct JevJudge {
+    client: reqwest::Client,
+    url: String,
+    key: String,
+    model: String,
+    threshold: f64,
+}
+
+impl JevJudge {
+    pub fn from_env() -> Option<Self> {
+        let key = std::env::var("SHEPHERD_JEV_API_KEY")
+            .ok()
+            .filter(|s| !s.trim().is_empty())
+            .or_else(|| std::env::var("TYPESAFE_API_KEY").ok().filter(|s| !s.trim().is_empty()))?;
+        let url = env_or("SHEPHERD_JEV_URL", DEFAULT_URL);
+        let model = env_or("SHEPHERD_JEV_MODEL", DEFAULT_MODEL);
+        let threshold = std::env::var("SHEPHERD_JEV_THRESHOLD")
+            .ok()
+            .and_then(|v| v.parse().ok())
+            .filter(|t| (0.0..=1.0).contains(t))
+            .unwrap_or(DEFAULT_THRESHOLD);
+        let client = reqwest::Client::builder()
+            .no_proxy()
+            .timeout(Duration::from_secs(30))
+            .connect_timeout(Duration::from_secs(10))
+            .build()
+            .unwrap_or_else(|_| reqwest::Client::new());
+        Some(Self { client, url, key, model, threshold })
+    }
+
+    #[cfg(test)]
+    fn new(url: impl Into<String>, key: impl Into<String>, threshold: f64) -> Self {
+        Self {
+            client: reqwest::Client::new(),
+            url: url.into(),
+            key: key.into(),
+            model: DEFAULT_MODEL.into(),
+            threshold,
+        }
+    }
+
+    fn state(criteria: &[String], deliverable: &DeliverableView) -> String {
+        let criteria = criteria.iter().map(|c| format!("- {c}")).collect::<Vec<_>>().join("\n");
+        format!(
+            "Acceptance criteria:\n{criteria}\n\nDeliverable:\nkind: {}\nreference: {}\nsummary: {}",
+            deliverable.kind, deliverable.reference, deliverable.summary
+        )
+    }
+}
+
+#[async_trait]
+impl Judge for JevJudge {
+    async fn judge(&self, criteria: &[String], deliverable: &DeliverableView) -> Verdict {
+        if criteria.is_empty() {
+            return Verdict { passed: true, reason: "jev: no criteria to judge".into() };
+        }
+        let questions: BTreeMap<String, serde_json::Value> = criteria
+            .iter()
+            .enumerate()
+            .map(|(i, c)| {
+                (
+                    format!("c{i}"),
+                    json!({
+                        "type": "noul",
+                        "instructions": format!("Does the deliverable satisfy this acceptance criterion: {c}"),
+                    }),
+                )
+            })
+            .collect();
+        let body = json!({ "state": Self::state(criteria, deliverable), "model": self.model, "questions": questions });
+
+        let resp = match self.client.post(&self.url).bearer_auth(&self.key).json(&body).send().await
+        {
+            Ok(r) if r.status().is_success() => r,
+            Ok(r) => return Verdict { passed: false, reason: format!("jev HTTP {}", r.status()) },
+            Err(e) => return Verdict { passed: false, reason: format!("jev unreachable: {e}") },
+        };
+        let parsed: SystemOneResponse = match resp.json().await {
+            Ok(p) => p,
+            Err(e) => {
+                return Verdict {
+                    passed: false,
+                    reason: format!("failed to parse jev response: {e}"),
+                }
+            }
+        };
+
+        // Every criterion must clear the threshold; report the weakest one.
+        let mut weakest = (1.0_f64, String::new());
+        for (i, c) in criteria.iter().enumerate() {
+            let key = format!("c{i}");
+            let p = match parsed.answers.get(&key) {
+                Some(a) => a.noul,
+                None => {
+                    return Verdict {
+                        passed: false,
+                        reason: format!("jev response missing answer {key}"),
+                    }
+                }
+            };
+            if p < weakest.0 {
+                weakest = (p, c.clone());
+            }
+        }
+        if weakest.0 >= self.threshold {
+            Verdict {
+                passed: true,
+                reason: format!(
+                    "jev: min noul {:.2} >= threshold {:.2}",
+                    weakest.0, self.threshold
+                ),
+            }
+        } else {
+            Verdict {
+                passed: false,
+                reason: format!(
+                    "jev: noul {:.2} < threshold {:.2} for criterion: {}",
+                    weakest.0, self.threshold, weakest.1
+                ),
+            }
+        }
+    }
+}
+
+fn env_or(key: &str, default: &str) -> String {
+    std::env::var(key).ok().filter(|s| !s.trim().is_empty()).unwrap_or_else(|| default.to_string())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use axum::{routing::post, Json, Router};
+
+    async fn spawn(app: Router, path: &str) -> String {
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.expect("bind");
+        let addr = listener.local_addr().expect("addr");
+        tokio::spawn(async move { axum::serve(listener, app).await.expect("serve") });
+        format!("http://{addr}{path}")
+    }
+
+    fn dv() -> DeliverableView {
+        DeliverableView {
+            kind: "DIFF".into(),
+            reference: "branch:feat".into(),
+            summary: "done".into(),
+        }
+    }
+
+    #[tokio::test]
+    async fn jev_judge_passes_when_every_criterion_clears_threshold() {
+        let app = Router::new().route(
+            "/v1/systemone",
+            post(|| async {
+                Json(json!({
+                    "model": "jev-1.13.0",
+                    "answers": { "c0": { "type": "noul", "noul": 0.95 }, "c1": { "type": "noul", "noul": 0.9 } },
+                    "usage": { "input_tokens": 1, "output_tokens": 1 }
+                }))
+            }),
+        );
+        let url = spawn(app, "/v1/systemone").await;
+        let j = JevJudge::new(url, "k", 0.8);
+        let v = j.judge(&["login success".into(), "token issued".into()], &dv()).await;
+        assert!(v.passed, "got: {v:?}");
+        assert!(v.reason.contains("0.90"), "reason names the weakest: {}", v.reason);
+    }
+
+    #[tokio::test]
+    async fn jev_judge_fails_on_criterion_below_threshold() {
+        let app = Router::new().route(
+            "/v1/systemone",
+            post(|| async {
+                Json(json!({
+                    "model": "jev-1.13.0",
+                    "answers": { "c0": { "type": "noul", "noul": 0.99 }, "c1": { "type": "noul", "noul": 0.5 } },
+                    "usage": { "input_tokens": 1, "output_tokens": 1 }
+                }))
+            }),
+        );
+        let url = spawn(app, "/v1/systemone").await;
+        let j = JevJudge::new(url, "k", 0.8);
+        let v = j.judge(&["login success".into(), "token issued".into()], &dv()).await;
+        assert!(!v.passed);
+        assert!(
+            v.reason.contains("token issued"),
+            "reason names the failing criterion: {}",
+            v.reason
+        );
+    }
+
+    #[tokio::test]
+    async fn jev_judge_passes_with_no_criteria() {
+        let j = JevJudge::new("http://127.0.0.1:9", "k", 0.8);
+        let v = j.judge(&[], &dv()).await;
+        assert!(v.passed);
+        assert_eq!(v.reason, "jev: no criteria to judge");
+    }
+
+    #[tokio::test]
+    async fn jev_judge_fail_closed_on_bad_json() {
+        let app = Router::new().route("/v1/systemone", post(|| async { "not json at all" }));
+        let url = spawn(app, "/v1/systemone").await;
+        let j = JevJudge::new(url, "k", 0.8);
+        let v = j.judge(&["login success".into()], &dv()).await;
+        assert!(!v.passed);
+        assert!(v.reason.contains("parse"), "got: {}", v.reason);
+    }
+}
diff --git a/crates/server/src/judge.rs b/crates/server/src/judge.rs
index dec3253..ec8174d 100644
--- a/crates/server/src/judge.rs
+++ b/crates/server/src/judge.rs
@@ -41,6 +41,12 @@ impl Judge for HttpJudge {
 }
 
 pub fn build_judge() -> Arc<dyn Judge> {
+    // Jev first: it is the cheap typed-judgment alternative to the LLM judge.
+    // Setting SHEPHERD_JEV_API_KEY (or TYPESAFE_API_KEY) routes the verification
+    // gate to Jev even when SHEPHERD_LLM_URL is also set for planner/PRD use.
+    if let Some(j) = crate::jev::JevJudge::from_env() {
+        return Arc::new(j);
+    }
     if let Some(j) = crate::llm::judge() {
         return j;
     }
diff --git a/crates/server/src/main.rs b/crates/server/src/main.rs
index 9575e27..0ab2bc0 100644
--- a/crates/server/src/main.rs
+++ b/crates/server/src/main.rs
@@ -11,6 +11,7 @@ mod debug_send;
 mod decomposition_run;
 mod design_bridge;
 mod import_scheduler;
+mod jev;
 mod judge;
 mod llm;
 mod mcp_bus;
diff --git a/crates/server/src/mcp_tools.rs b/crates/server/src/mcp_tools.rs
index 17902ee..251efa2 100644
--- a/crates/server/src/mcp_tools.rs
+++ b/crates/server/src/mcp_tools.rs
@@ -485,7 +485,7 @@ pub fn router(
         .requires("TASK", "ADD"))
         .tool(Tool::new(
             "shepherd_dispatch_delivery",
-            "把任务派发给 AI 执行者(executor: CLAUDE_CODE | CODEX | OPENCODE | CODEBUDDY);自动驱动任务、过验证门并回灌验证。\
+            "把任务派发给 AI 执行者(executor: CLAUDE_CODE | CODEX | OPENCODE | CODEBUDDY | DEEPCODE);自动驱动任务、过验证门并回灌验证。\
              可直接传 skillIds(+projectId)自动组合行为规范注入,无需先调 compose。",
             obj(
                 json!({
@@ -494,7 +494,7 @@ pub fn router(
                     "title": { "type": "string" },
                     "description": { "type": "string" },
                     "acceptanceCriteria": { "type": "array", "items": { "type": "string" } },
-                    "executor": { "type": "string", "enum": ["CLAUDE_CODE", "CODEX", "OPENCODE", "CODEBUDDY"] },
+                    "executor": { "type": "string", "enum": ["CLAUDE_CODE", "CODEX", "OPENCODE", "CODEBUDDY", "DEEPCODE"] },
                     "projectId": { "type": "string", "description": "skillIds 非空时必填(用于组合)" },
                     "skillIds": { "type": "array", "items": { "type": "string" }, "description": "可选:自动组合这些 skill 为行为规范" },
                     "instructions": { "type": "string", "description": "可选:显式行为规范(与 skillIds 组合结果合并)" }
diff --git a/crates/shepherd-cli/src/client.rs b/crates/shepherd-cli/src/client.rs
index 14e0968..26f5d44 100644
--- a/crates/shepherd-cli/src/client.rs
+++ b/crates/shepherd-cli/src/client.rs
@@ -345,8 +345,9 @@ pub fn normalize_agent(t: &str) -> R<String> {
         "opencode" => Ok("OPENCODE".into()),
         // The brand is one word, but the claude-code spelling invites code-buddy; accept both.
         "codebuddy" | "code_buddy" => Ok("CODEBUDDY".into()),
+        "deepcode" | "deep_code" => Ok("DEEPCODE".into()),
         other => Err(format!(
-            "unknown agent type: {other} (supported: claude-code | codex | opencode | codebuddy)"
+            "unknown agent type: {other} (supported: claude-code | codex | opencode | codebuddy | deepcode)"
         )
         .into()),
     }
diff --git a/crates/shepherd-cli/src/commands/agent.rs b/crates/shepherd-cli/src/commands/agent.rs
index 2c11cdf..fb12cad 100644
--- a/crates/shepherd-cli/src/commands/agent.rs
+++ b/crates/shepherd-cli/src/commands/agent.rs
@@ -3,7 +3,7 @@ use clap::Subcommand;
 
 #[derive(Subcommand)]
 pub enum AgentCmd {
-    /// Connect an AI executor (claude-code | codex | opencode | codebuddy) and save it as the dispatch default.
+    /// Connect an AI executor (claude-code | codex | opencode | codebuddy | deepcode) and save it as the dispatch default.
     Connect {
         #[arg(long = "type")]
         kind: String,
diff --git a/crates/shepherd-cli/src/main.rs b/crates/shepherd-cli/src/main.rs
index 816265a..a2b4dba 100644
--- a/crates/shepherd-cli/src/main.rs
+++ b/crates/shepherd-cli/src/main.rs
@@ -323,6 +323,8 @@ mod tests {
         assert_eq!(normalize_agent("opencode").unwrap(), "OPENCODE");
         assert_eq!(normalize_agent("CodeBuddy").unwrap(), "CODEBUDDY");
         assert_eq!(normalize_agent("code-buddy").unwrap(), "CODEBUDDY");
+        assert_eq!(normalize_agent("deepcode").unwrap(), "DEEPCODE");
+        assert_eq!(normalize_agent("DeepCode").unwrap(), "DEEPCODE");
         assert!(normalize_agent("gpt").is_err());
     }
 
diff --git a/docs/EXECUTORS.md b/docs/EXECUTORS.md
index bbaa562..efaae00 100644
--- a/docs/EXECUTORS.md
+++ b/docs/EXECUTORS.md
@@ -1,6 +1,6 @@
 # Shepherd — Running AI Executors
 
-How to run each supported AI executor (Claude Code / Codex / OpenCode / CodeBuddy)
+How to run each supported AI executor (Claude Code / Codex / OpenCode / CodeBuddy / DeepCode)
 behind `agent-runtime`. For fleet architecture and server-side setup see
 [USAGE.md §7](./USAGE.md); for image builds and deployment see [DEPLOYMENT.md](./DEPLOYMENT.md).
 
@@ -10,6 +10,7 @@ behind `agent-runtime`. For fleet architecture and server-side setup see
 - [Codex](#codex)
 - [OpenCode](#opencode)
 - [CodeBuddy](#codebuddy)
+- [DeepCode](#deepcode)
 - [One runtime, multiple executors](#one-runtime-multiple-executors)
 - [Mock executor (no CLI)](#mock-executor-no-cli)
 - [Windows](#windows)
@@ -24,7 +25,7 @@ behind `agent-runtime`. For fleet architecture and server-side setup see
 ## How dispatch reaches a CLI
 
 Every delivery attempt carries an `executor` kind (`CLAUDE_CODE` / `CODEX` /
-`OPENCODE` / `CODEBUDDY`). The server enqueues the task per kind; an
+`OPENCODE` / `CODEBUDDY` / `DEEPCODE`). The server enqueues the task per kind; an
 `agent-runtime` long-polls with its `SHEPHERD_CAPS` and only claims kinds it
 declared. On claim, the runtime spawns the matching CLI in a dedicated git
 worktree, snapshots the resulting changes as a commit, and reports back.
@@ -148,6 +149,20 @@ SHEPHERD_CAPS=CODEBUDDY AGENT_WORKDIR=/repo … ./agent-runtime
 Override with `CODEBUDDY_CMD`, e.g. widen permissions for tasks that need to
 run shell commands: `CODEBUDDY_CMD="codebuddy -p --permission-mode bypassPermissions"`.
 
+## DeepCode
+
+Generic backend. The default invocation is `deepcode exec --json "<prompt>"` —
+`--json` keeps the output machine-readable for the result summary.
+
+```bash
+deepcode --version
+
+SHEPHERD_CAPS=DEEPCODE AGENT_WORKDIR=/repo … ./agent-runtime
+```
+
+Override with `DEEPCODE_CMD`, e.g. widen permissions for tasks that need to
+edit files headlessly: `DEEPCODE_CMD="deepcode exec --access full-access --json"`.
+
 ## One runtime, multiple executors
 
 A single runtime can claim several kinds — list them all and make sure every
@@ -169,7 +184,7 @@ canned output without spawning a CLI — useful to smoke-test the dispatch loop
 before installing real CLIs.
 
 ```bash
-AGENT_MOCK=1 SHEPHERD_CAPS=CLAUDE_CODE,CODEX,OPENCODE,CODEBUDDY … ./agent-runtime
+AGENT_MOCK=1 SHEPHERD_CAPS=CLAUDE_CODE,CODEX,OPENCODE,CODEBUDDY,DEEPCODE … ./agent-runtime
 ```
 
 ## Windows
@@ -183,7 +198,7 @@ The runtime runs natively on Windows (build with the MSVC toolchain:
   `codex.exe` take precedence); explicit paths via `CLAUDE_BIN` / `*_CMD` are
   used as-is. `CLAUDE_BIN=D:\nvm4w\nodejs\claude.cmd` works: the Claude
   backend feeds the prompt over stdin, so the cmd.exe hop is harmless.
-- **Generic backends (codebuddy/codex/opencode) cannot go through a `.cmd`
+- **Generic backends (codebuddy/codex/opencode/deepcode) cannot go through a `.cmd`
   shim**: they pass the prompt as an argv, prompts contain newlines, and Rust
   refuses newline arguments to batch files (cmd.exe cannot carry them safely).
   Bypass cmd.exe by invoking node directly — read the shim (`type
@@ -254,7 +269,7 @@ shepherd dispatch --decomp d1 --task t1 --executor CODEBUDDY
 # or bind a default once: shepherd agent connect --kind codebuddy
 ```
 
-The MCP tool `shepherd_dispatch_delivery` accepts the same four kinds.
+The MCP tool `shepherd_dispatch_delivery` accepts the same five kinds.
 
 ### Pinning a task to one runtime
 
@@ -323,5 +338,5 @@ animation look the same as local ones.
 |---|---|
 | Task stays queued (`ready` grows) | No online runtime declares that kind — compare `SHEPHERD_CAPS` with the task's `executor`, then `GET /agent/runtime` for liveness |
 | Attempt delivers with "no code change" | CLI ran but refused edits (permission mode) or the prompt didn't ask for file changes — read the deliverable summary, it contains the CLI's own output |
-| Spawn error in events | CLI not on `PATH` inside the runtime's environment — set `CLAUDE_BIN` / `CODEX_CMD` / `OPENCODE_CMD` / `CODEBUDDY_CMD` |
+| Spawn error in events | CLI not on `PATH` inside the runtime's environment — set `CLAUDE_BIN` / `CODEX_CMD` / `OPENCODE_CMD` / `CODEBUDDY_CMD` / `DEEPCODE_CMD` |
 | Attempt fails at timeout | Raise `AGENT_TASK_TIMEOUT_SECS`; generic backends get killed (process group) at the deadline |
diff --git a/docs/EXECUTORS.zh-CN.md b/docs/EXECUTORS.zh-CN.md
index 728862f..5712220 100644
--- a/docs/EXECUTORS.zh-CN.md
+++ b/docs/EXECUTORS.zh-CN.md
@@ -1,7 +1,7 @@
 # Shepherd — AI 执行者运行指南
 
 介绍如何在 `agent-runtime` 下运行各家 AI 执行者(Claude Code / Codex / OpenCode /
-CodeBuddy)。机群架构与服务端配置见 [USAGE.zh-CN.md §7](./USAGE.zh-CN.md),
+CodeBuddy / DeepCode)。机群架构与服务端配置见 [USAGE.zh-CN.md §7](./USAGE.zh-CN.md),
 镜像构建与部署见 [DEPLOYMENT.zh-CN.md](./DEPLOYMENT.zh-CN.md)。
 
 - [派发如何到达 CLI](#派发如何到达-cli)
@@ -10,6 +10,7 @@ CodeBuddy)。机群架构与服务端配置见 [USAGE.zh-CN.md §7](./USAGE.zh-C
 - [Codex](#codex)
 - [OpenCode](#opencode)
 - [CodeBuddy](#codebuddy)
+- [DeepCode](#deepcode)
 - [一个 runtime 承接多种执行者](#一个-runtime-承接多种执行者)
 - [mock 执行者(无需 CLI)](#mock-执行者无需-cli)
 - [Windows](#windows)
@@ -24,7 +25,7 @@ CodeBuddy)。机群架构与服务端配置见 [USAGE.zh-CN.md §7](./USAGE.zh-C
 ## 派发如何到达 CLI
 
 每个交付尝试都带 `executor` 类型(`CLAUDE_CODE` / `CODEX` / `OPENCODE` /
-`CODEBUDDY`)。服务端按类型入队;`agent-runtime` 以 `SHEPHERD_CAPS` 长轮询,
+`CODEBUDDY` / `DEEPCODE`)。服务端按类型入队;`agent-runtime` 以 `SHEPHERD_CAPS` 长轮询,
 只认领自己声明过的类型。认领后 runtime 在独立 git worktree 中拉起对应 CLI,
 把产生的改动快照成 commit 并回报。
 
@@ -140,6 +141,20 @@ SHEPHERD_CAPS=CODEBUDDY AGENT_WORKDIR=/repo … ./agent-runtime
 用 `CODEBUDDY_CMD` 覆盖,例如任务需要跑 shell 时放宽权限:
 `CODEBUDDY_CMD="codebuddy -p --permission-mode bypassPermissions"`。
 
+## DeepCode
+
+通用后端。默认调用 `deepcode exec --json "<prompt>"` —— `--json` 让输出保持
+机器可读,便于回填结果摘要。
+
+```bash
+deepcode --version
+
+SHEPHERD_CAPS=DEEPCODE AGENT_WORKDIR=/repo … ./agent-runtime
+```
+
+用 `DEEPCODE_CMD` 覆盖,例如 headless 下需要放开文件编辑权限时:
+`DEEPCODE_CMD="deepcode exec --access full-access --json"`。
+
 ## 一个 runtime 承接多种执行者
 
 单个 runtime 可以声明多种类型——全部列出,并保证对应 CLI 都装好且已登录:
@@ -158,7 +173,7 @@ runtime(能力任意组合)。服务端设了 `SHEPHERD_FLEET_REDIS` 时,跨机
 CLI——适合在装 CLI 之前先冒烟验证派发链路。
 
 ```bash
-AGENT_MOCK=1 SHEPHERD_CAPS=CLAUDE_CODE,CODEX,OPENCODE,CODEBUDDY … ./agent-runtime
+AGENT_MOCK=1 SHEPHERD_CAPS=CLAUDE_CODE,CODEX,OPENCODE,CODEBUDDY,DEEPCODE … ./agent-runtime
 ```
 
 ## Windows
@@ -171,7 +186,7 @@ runtime 可在 Windows 原生运行(MSVC 工具链:`cargo build --release -p age
   (`codex.exe` 这类原生二进制优先);`CLAUDE_BIN` / `*_CMD` 里写的显式路径原样使用。
   `CLAUDE_BIN=D:\nvm4w\nodejs\claude.cmd` 没问题:Claude 后端的 prompt 走 stdin,
   经 cmd.exe 一跳无害。
-- **通用后端(codebuddy/codex/opencode)不能走 `.cmd` 垫片**:它们把 prompt 作为
+- **通用后端(codebuddy/codex/opencode/deepcode)不能走 `.cmd` 垫片**:它们把 prompt 作为
   命令行参数传,prompt 必含换行,而 Rust 拒绝给批处理文件传含换行的参数
   (cmd.exe 无法安全携带)。绕开 cmd.exe,直接用 node 调包入口——先
   `type codebuddy.cmd` 看垫片真正调用什么,然后照抄,如
@@ -233,7 +248,7 @@ shepherd dispatch --decomp d1 --task t1 --executor CODEBUDDY
 # 或先绑定默认:shepherd agent connect --kind codebuddy
 ```
 
-MCP 工具 `shepherd_dispatch_delivery` 同样接受这四种类型。
+MCP 工具 `shepherd_dispatch_delivery` 同样接受这五种类型。
 
 ### 定向到具体某台 runtime
 
@@ -293,5 +308,5 @@ cargo r -p pool-runner
 |---|---|
 | 任务一直排队(`ready` 增长) | 没有在线 runtime 声明该类型——核对 `SHEPHERD_CAPS` 与任务 `executor`,再看 `GET /agent/runtime` 的在线状态 |
 | 交付显示"无代码变动" | CLI 跑了但拒绝改文件(权限模式),或提示词本身没要求改文件——看交付摘要,里面是 CLI 的原始输出 |
-| 事件里报 spawn 错误 | CLI 不在 runtime 进程的 `PATH` 上——设 `CLAUDE_BIN` / `CODEX_CMD` / `OPENCODE_CMD` / `CODEBUDDY_CMD` |
+| 事件里报 spawn 错误 | CLI 不在 runtime 进程的 `PATH` 上——设 `CLAUDE_BIN` / `CODEX_CMD` / `OPENCODE_CMD` / `CODEBUDDY_CMD` / `DEEPCODE_CMD` |
 | 任务超时失败 | 调大 `AGENT_TASK_TIMEOUT_SECS`;通用后端到点会被整进程组杀掉 |
diff --git a/docs/USAGE.md b/docs/USAGE.md
index 610b285..86f9f7a 100644
--- a/docs/USAGE.md
+++ b/docs/USAGE.md
@@ -20,7 +20,7 @@ you file a requirement
         │                        │ dispatch
         ▼                        ▼
                           FLEET DISPATCH ──► agent-runtime executor
-                          (pull / long-poll)   (claude / codex / opencode / codebuddy)
+                          (pull / long-poll)   (claude / codex / opencode / codebuddy / deepcode)
                                  │ runs in a git worktree
                                  ▼
                           deliverable (diff / PR)
@@ -59,7 +59,7 @@ Company AI tools (Claude Code, Codex, …) run on internal dev machines or CI 
 | Quick start (Docker) | Docker + Docker Compose v2 |
 | From-source dev | Rust (stable, edition 2021; CI uses `rust:1.86`), Node.js 18+, a PostgreSQL 16 instance |
 | Multi-host fleet | Redis 7 |
-| Real AI executors | `git` plus the agent CLIs on `PATH` (`claude` / `codex` / `opencode` / `codebuddy`) |
+| Real AI executors | `git` plus the agent CLIs on `PATH` (`claude` / `codex` / `opencode` / `codebuddy` / `deepcode`) |
 
 PostgreSQL is required; the server **auto-applies migrations on startup**. Redis is required **only** for the multi-host fleet.
 
@@ -187,6 +187,23 @@ All server start-up switches are consolidated into a typed config (`crates/serve
 
 Advanced/lazy-read switches also exist for the pluggable AI touchpoints — `SHEPHERD_AGENT_URL` / `SHEPHERD_AGENT_CMD` / `SHEPHERD_AGENT_ASYNC` (executor routing), `SHEPHERD_LLM_URL`, `SHEPHERD_PLANNER_URL`, `SHEPHERD_JUDGE_URL`, `SHEPHERD_MAX_REVISIONS`. Defaults need none of them.
 
+#### 5.1.1 Verification judge: Jev
+
+The delivery verification gate (`Judge`) can run on [Jev](https://learnjev.com), TypeSafe's
+fast typed-judgment model (yes/no probability answers, no prose) instead of an LLM.
+Set the key and the gate asks one `noul` question per acceptance criterion in a single
+call; it passes only when every probability reaches the threshold.
+
+| Variable | Default | Meaning |
+|---|---|---|
+| `SHEPHERD_JEV_API_KEY` | *(falls back to `TYPESAFE_API_KEY`)* | TypeSafe API key — presence enables the Jev judge (takes priority over the LLM judge) |
+| `SHEPHERD_JEV_URL` | `https://api.typesafe.ai/v1/systemone` | Jev System One endpoint |
+| `SHEPHERD_JEV_MODEL` | `jev-latest` | Jev model alias |
+| `SHEPHERD_JEV_THRESHOLD` | `0.8` | Minimum `noul` probability per criterion to pass |
+
+Failures fail closed: any HTTP/parse error returns a not-passed verdict, so a broken
+Jev setup blocks delivery instead of silently passing it.
+
 ### 5.2 Agent-runtime (`agent-runtime`)
 
 | Variable | Default | Meaning |
@@ -203,6 +220,7 @@ Advanced/lazy-read switches also exist for the pluggable AI touchpoints — `SHE
 | `CODEX_CMD` | `codex exec` | Codex CLI invocation |
 | `OPENCODE_CMD` | `opencode run` | OpenCode CLI invocation |
 | `CODEBUDDY_CMD` | `codebuddy -p --permission-mode acceptEdits` | CodeBuddy CLI invocation |
+| `DEEPCODE_CMD` | `deepcode exec --json` | DeepCode CLI invocation |
 
 ---
 
@@ -275,6 +293,7 @@ The runtime picks a backend per task by its `executor` kind, unless `AGENT_MOCK=
 | `CODEX` | generic CLI | `codex exec` (`CODEX_CMD`) |
 | `OPENCODE` | generic CLI | `opencode run` (`OPENCODE_CMD`) |
 | `CODEBUDDY` | generic CLI | `codebuddy -p --permission-mode acceptEdits` (`CODEBUDDY_CMD`) |
+| `DEEPCODE` | generic CLI | `deepcode exec --json` (`DEEPCODE_CMD`) |
 | any (with `AGENT_MOCK=1`) | mock — returns canned output | none |
 
 Real backends need `git` and the CLI on `PATH` (or pointed at via the override env). Per-executor run recipes (login, permission modes, dispatch examples) are in [EXECUTORS.md](./EXECUTORS.md). Adding a new backend means implementing one `CliAgentBackend` (`async fn execute(prompt, cwd, sink)`) and registering an enum variant — see `crates/agent-runtime/src/backend.rs`.
@@ -344,7 +363,7 @@ Both honour `SHEPHERD_BASE` (default `http://127.0.0.1:9180`) and `SHEPHERD_USER
 | Web console shows blank / API 404 in dev | Vite proxy target mismatch. The dev proxy points at `:9180`; bind the server there or set `SHEPHERD_API` to your server URL. |
 | Tasks never get claimed | Server not in fleet mode (`SHEPHERD_AGENT_FLEET=1`), no runtime online, or capability mismatch — check `SHEPHERD_CAPS` vs the task's executor kind, and `GET /agent/work/stats`. |
 | Multi-host runtimes can't share work | `SHEPHERD_FLEET_REDIS` not set (or not the same Redis) on all server replicas → each falls back to its own in-process queue. |
-| Real agent does nothing / errors spawning | CLI not on `PATH`; set `CLAUDE_BIN` / `CODEX_CMD` / `OPENCODE_CMD` / `CODEBUDDY_CMD`, or run with `AGENT_MOCK=1` to confirm the loop. |
+| Real agent does nothing / errors spawning | CLI not on `PATH`; set `CLAUDE_BIN` / `CODEX_CMD` / `OPENCODE_CMD` / `CODEBUDDY_CMD` / `DEEPCODE_CMD`, or run with `AGENT_MOCK=1` to confirm the loop. |
 | API batch-run stuck `RUNNING` with no results | `SHEPHERD_RUNNER=noop` is set (demo placeholder). Unset it to use the native runner. |
 | New migration not applied | Restart the server — migrations run on boot; a new migration file needs a rebuild. |
 | OIDC endpoint 404 | The provider is only registered when **both** id and secret env vars are set. |
diff --git a/docs/USAGE.zh-CN.md b/docs/USAGE.zh-CN.md
index a8b242e..d3adead 100644
--- a/docs/USAGE.zh-CN.md
+++ b/docs/USAGE.zh-CN.md
@@ -20,7 +20,7 @@ Shepherd 是一个 AI 研发**监督**平台:从 idea 到交付,更安全、更
         │                  │ 派发
         ▼                  ▼
                     机群派发 ──► agent-runtime 执行者
-                   (拉取 / 长轮询)   (claude / codex / opencode / codebuddy)
+                   (拉取 / 长轮询)   (claude / codex / opencode / codebuddy / deepcode)
                           │ 在独立 git worktree 中运行
                           ▼
                     交付物(diff / PR)
@@ -59,7 +59,7 @@ Shepherd 是一个 AI 研发**监督**平台:从 idea 到交付,更安全、更
 | 快速开始(Docker) | Docker + Docker Compose v2 |
 | 源码开发 | Rust(stable,edition 2021;CI 用 `rust:1.86`)、Node.js 18+、一个 PostgreSQL 16 实例 |
 | 多机机群 | Redis 7 |
-| 真实 AI 执行者 | `git` + agent CLI 在 `PATH` 上(`claude` / `codex` / `opencode` / `codebuddy`) |
+| 真实 AI 执行者 | `git` + agent CLI 在 `PATH` 上(`claude` / `codex` / `opencode` / `codebuddy` / `deepcode`) |
 
 PostgreSQL 必需;服务端**启动时自动跑迁移**。Redis 仅在多机机群时必需。
 
@@ -186,6 +186,22 @@ cargo clippy --workspace -- -D warnings
 
 可插拔 AI 触点还有若干进阶/懒读开关:`SHEPHERD_AGENT_URL` / `SHEPHERD_AGENT_CMD` / `SHEPHERD_AGENT_ASYNC`(执行者路由)、`SHEPHERD_LLM_URL`、`SHEPHERD_PLANNER_URL`、`SHEPHERD_JUDGE_URL`、`SHEPHERD_MAX_REVISIONS`。默认全都无需设置。
 
+#### 5.1.1 验证门 Judge:Jev
+
+交付验证门(`Judge`)可以跑在 [Jev](https://learnjev.com)(TypeSafe 的快速类型化判断
+模型,只回答是/否概率,不生成文字)上,替代 LLM judge。配好 key 后,验证门会为每条
+验收标准发一个 `noul` 问题(单次请求),所有概率都达到阈值才判过。
+
+| 变量 | 默认 | 含义 |
+|---|---|---|
+| `SHEPHERD_JEV_API_KEY` | *(缺省回退 `TYPESAFE_API_KEY`)* | TypeSafe API key——设了即启用 Jev judge(优先级高于 LLM judge) |
+| `SHEPHERD_JEV_URL` | `https://api.typesafe.ai/v1/systemone` | Jev System One 端点 |
+| `SHEPHERD_JEV_MODEL` | `jev-latest` | Jev 模型别名 |
+| `SHEPHERD_JEV_THRESHOLD` | `0.8` | 每条标准的最低 `noul` 通过概率 |
+
+失败一律 fail-closed:任何 HTTP/解析错误都判为不通过,即 Jev 配置坏了会阻断交付,
+而不是静默放行。
+
 ### 5.2 Agent-runtime(`agent-runtime`)
 
 | 变量 | 默认 | 含义 |
@@ -202,6 +218,7 @@ cargo clippy --workspace -- -D warnings
 | `CODEX_CMD` | `codex exec` | Codex CLI 调用 |
 | `OPENCODE_CMD` | `opencode run` | OpenCode CLI 调用 |
 | `CODEBUDDY_CMD` | `codebuddy -p --permission-mode acceptEdits` | CodeBuddy CLI 调用 |
+| `DEEPCODE_CMD` | `deepcode exec --json` | DeepCode CLI 调用 |
 
 ---
 
@@ -271,6 +288,7 @@ runtime 按任务的 `executor` 类型选后端,除非 `AGENT_MOCK=1` 强制走
 | `CODEX` | 通用 CLI | `codex exec`(`CODEX_CMD`) |
 | `OPENCODE` | 通用 CLI | `opencode run`(`OPENCODE_CMD`) |
 | `CODEBUDDY` | 通用 CLI | `codebuddy -p --permission-mode acceptEdits`(`CODEBUDDY_CMD`) |
+| `DEEPCODE` | 通用 CLI | `deepcode exec --json`(`DEEPCODE_CMD`) |
 | 任意(配 `AGENT_MOCK=1`) | mock——返回固定输出 | 无 |
 
 真实后端需 `git` 与 CLI 在 `PATH`(或经覆盖 env 指定)。各执行者的运行指南(登录、权限模式、派发示例)见 [EXECUTORS.zh-CN.md](./EXECUTORS.zh-CN.md)。新增后端只需实现一个 `CliAgentBackend`(`async fn execute(prompt, cwd, sink)`)并注册一个枚举变体——见 `crates/agent-runtime/src/backend.rs`。
@@ -340,7 +358,7 @@ python3 .agents/skills/openapi-bootstrap/scenarios_all.py
 | dev 下控制台空白 / API 404 | Vite 代理目标不匹配。dev 代理指向 `:9180`;把服务端绑到那里,或设 `SHEPHERD_API` 为你的服务端 URL。 |
 | 任务始终无人认领 | 服务端未开机群模式(`SHEPHERD_AGENT_FLEET=1`)、无 runtime 在线,或能力不匹配——核对 `SHEPHERD_CAPS` 与任务 executor 类型,以及 `GET /agent/work/stats`。 |
 | 多机 runtime 无法共享任务 | 各服务端副本未设(或非同一个)`SHEPHERD_FLEET_REDIS` → 各自回落到自己的进程内队列。 |
-| 真实 agent 不动作 / spawn 报错 | CLI 不在 `PATH`;设 `CLAUDE_BIN` / `CODEX_CMD` / `OPENCODE_CMD` / `CODEBUDDY_CMD`,或先用 `AGENT_MOCK=1` 验证链路。 |
+| 真实 agent 不动作 / spawn 报错 | CLI 不在 `PATH`;设 `CLAUDE_BIN` / `CODEX_CMD` / `OPENCODE_CMD` / `CODEBUDDY_CMD` / `DEEPCODE_CMD`,或先用 `AGENT_MOCK=1` 验证链路。 |
 | 接口批量运行卡在 `RUNNING` 无结果 | 设了 `SHEPHERD_RUNNER=noop`(演示占位)。取消它以走原生 runner。 |
 | 新迁移未生效 | 重启服务端——迁移在启动时跑;新迁移文件需重新构建。 |
 | OIDC 端点 404 | provider 仅在 id 与 secret 两个 env **都**设置时才注册。 |
diff --git a/web/src/api.ts b/web/src/api.ts
index aeb4f81..1181146 100644
--- a/web/src/api.ts
+++ b/web/src/api.ts
@@ -1154,6 +1154,7 @@ export const EXECUTOR_LABEL: Record<string, string> = {
   CODEX: 'Codex',
   OPENCODE: 'OpenCode',
   CODEBUDDY: 'CodeBuddy',
+  DEEPCODE: 'DeepCode',
 }
 
 /** One task-center row: system-wide delivery attempts aggregated by status/executor/result/completion rate. */
@@ -1166,7 +1167,7 @@ export interface TaskCenterItem {
   description: string
   /** Module (title of the owning requirement; empty string when none). */
   module: string
-  /** Executor: CLAUDE_CODE / CODEX / OPENCODE / CODEBUDDY. */
+  /** Executor: CLAUDE_CODE / CODEX / OPENCODE / CODEBUDDY / DEEPCODE. */
   executor: string
   /** Target runtime name for directed dispatch; empty = any runtime with the capability. */
   targetRuntime?: string | null