@sema-agent/server 2.0.0 → 3.0.0

package/dist/hooks/hook-llm.js

@@ -1,3 +1,4 @@
+import { matchCatalogModel } from "../model-select.js";
 export function createHookLlm(deps) {
     const { config, metrics } = deps;
     // anthropic 路由自己的 baseUrl(与 core createAnthropicBrain 同一解析链:model.baseUrl || config.baseUrl
@@ -40,6 +41,14 @@ export function createHookLlm(deps) {
         }
         if (config.models[requested])
             return shape(requested);
+        // 复审 2026-07-29 D1:目录键是 **name** 形,而调用点会传 **provider id** 形——`http/routes/tasks.ts`
+        // 的 titler 腿([1992]②「标题跟 turn」)对对象形 spec 传的就是 `spec.model.id`。registry 热应用后
+        // name 与 id 常态分叉(目录键是运维起的短名,`Model.id` 是 provider 侧的真名),单键索引下这类请求全落到下面那条
+        // 「不在目录里」⇒ 用户每选一次非主模型,标题就 llm_failed(对话本身正常),与 [1992] 同族的病形。
+        // 双键索引沿用同仓先例 [865]② `matchCatalogModel`:name 优先、id 扫目录值、跳过 `default` 别名键。
+        const byId = matchCatalogModel(requested, config.models);
+        if (byId !== undefined)
+            return shape(byId);
         if (requested === config.model.id)
             return shapeFrom(requested, config.model, requested);
         return { ok: false, error: `hook model "${requested}" is not in this deployment's catalog` };

package/dist/http/routes/approvals-assistant.js

@@ -746,7 +746,7 @@ export async function streamApprovals(req, res, cs, scope, pollMs = APPROVALS_ST
         }
         prev = cur;
         if (Date.now() - start > MAX_MS) {
-            res.write(`event: error\ndata: ${JSON.stringify({ type: "error", code: "STREAM_MAX_DURATION", message: "approvals stream reached its 15-minute cap — reconnect to continue" })}\n\n`);
+            res.write(`event: error\ndata: ${JSON.stringify({ type: "error", errorCode: "STREAM_MAX_DURATION", message: "approvals stream reached its 15-minute cap — reconnect to continue" })}\n\n`);
             break;
         }
         if (Date.now() - lastBeat > 15_000) {

package/dist/http/routes/attachments.js

@@ -61,7 +61,7 @@ async function handleAttachmentsBody(req, res, url, ctx, miss) {
                 }
                 const allow = deps.config.attachmentMimeAllowlist;
                 if (allow && !allow.some((a) => a === mime || (a.endsWith("/*") && mime.startsWith(a.slice(0, -1))))) {
-                    sendError(res, 415, "attachment.mime_not_allowed", `mime '${mime}' is not in this deployment's allowlist`, { code: "attachment.mime_not_allowed", allowed: allow });
+                    sendError(res, 415, "attachment.mime_not_allowed", `mime '${mime}' is not in this deployment's allowlist`, { allowed: allow });
                     return;
                 }
                 const max = deps.config.attachmentMaxBytes ?? 32 * 1024 * 1024;
@@ -71,7 +71,7 @@ async function handleAttachmentsBody(req, res, url, ctx, miss) {
                 }
                 catch (err) {
                     if (err instanceof HttpError && err.status === 413) {
-                        sendError(res, 413, "attachment.too_large", `attachment exceeds the per-file limit (${max} bytes)`, { code: "attachment.too_large", maxBytes: max });
+                        sendError(res, 413, "attachment.too_large", `attachment exceeds the per-file limit (${max} bytes)`, { maxBytes: max });
                         return;
                     }
                     throw err;

package/dist/http/routes/memory-policy.js

@@ -76,11 +76,11 @@ async function handleMemoryPolicyBody(req, res, url, ctx, miss) {
             scope = decodeURIComponent(url.slice("/v1/memory/sync/".length));
         }
         catch {
-            sendError(res, 400, "memory_sync_invalid_scope", "malformed percent-encoding in :scope", { code: "memory_sync_invalid_scope" });
+            sendError(res, 400, "memory_sync_invalid_scope", "malformed percent-encoding in :scope");
             return;
         }
         if (scope.length === 0) {
-            sendError(res, 400, "memory_sync_invalid_scope", "missing :scope", { code: "memory_sync_invalid_scope" });
+            sendError(res, 400, "memory_sync_invalid_scope", "missing :scope");
             return;
         }
         if (scope !== formatUserScope(principal) && !explicitOperatorOk(principal, deps.config.operatorPrincipals)) {
@@ -90,7 +90,7 @@ async function handleMemoryPolicyBody(req, res, url, ctx, miss) {
         const body = await readJson(req); // 413/400 (oversize / bad JSON) via the typed top-level catch — auth already done
         const parsed = parseMemorySyncRequest(body, scope);
         if (!parsed.ok) {
-            sendError(res, 422, "memory_sync_invalid_body", parsed.error, { code: "memory_sync_invalid_body" });
+            sendError(res, 422, "memory_sync_invalid_body", parsed.error);
             return;
         }
         try {

package/dist/http/routes/runs.js

@@ -154,7 +154,7 @@ async function handleRunsBody(req, res, url, ctx, miss) {
             const taskId = clientTaskId ?? uuidv7();
             const created = await runStore.createRun(taskId, sessionId, prepared.auth?.principal ?? null, deps.instanceId ?? "default", runMeta(prepared, source));
             if (created.ok)
-                deps.sessionTitler?.maybeTitle(sessionId, prepared.spec.objective); // fire-and-forget session auto-title
+                deps.sessionTitler?.maybeTitle(sessionId, prepared.spec.objective, typeof prepared.spec.model === "string" ? prepared.spec.model : prepared.spec.model?.id); // fire-and-forget;model 跟 turn([1992]②)
             if (!created.ok) {
                 // race twin: two concurrent same-taskId submits (gateway retry racing the original) — the loser's
                 // session claim collides with ITS OWN taskId → idempotent 202 replay, not a 409 (the run exists, once).

package/dist/http/routes/session-sync.js

@@ -234,7 +234,7 @@ async function handleSessionSyncBody(req, res, url, ctx, miss) {
                 if (deps.snapshotBlobSqlCapBytes !== undefined && body.byteLength > deps.snapshotBlobSqlCapBytes) {
                     // 复审 D-F1/B-F2:归一漏网(多行体,旧门正则扫不到)——A6 迁移到 sendError 后此站点由
                     // errorCode 必填位置参数保结构(不再靠源级反射门兜底)。
-                    sendError(res, 413, "blob_too_large_for_sql", `blob of ${body.byteLength} bytes exceeds this deployment's SQL snapshot-blob cap (${deps.snapshotBlobSqlCapBytes} bytes — TiDB's txn-entry-size-limit is the low wall at 6MiB by default; the mysql-protocol packet limit sits above it) — configure MinIO object storage (MINIO_ENDPOINT/MINIO_ACCESS_KEY/MINIO_SECRET_KEY) for large snapshot blobs, or raise SNAPSHOT_BLOB_SQL_MAX_BYTES if your deployment lifted those limits`, { code: "blob_too_large_for_sql" });
+                    sendError(res, 413, "blob_too_large_for_sql", `blob of ${body.byteLength} bytes exceeds this deployment's SQL snapshot-blob cap (${deps.snapshotBlobSqlCapBytes} bytes — TiDB's txn-entry-size-limit is the low wall at 6MiB by default; the mysql-protocol packet limit sits above it) — configure MinIO object storage (MINIO_ENDPOINT/MINIO_ACCESS_KEY/MINIO_SECRET_KEY) for large snapshot blobs, or raise SNAPSHOT_BLOB_SQL_MAX_BYTES if your deployment lifted those limits`);
                     return;
                 }
                 if (createHash("sha256").update(body).digest("hex") !== hash) {
@@ -246,7 +246,7 @@ async function handleSessionSyncBody(req, res, url, ctx, miss) {
                 // blob as durably stored when it isn't.
                 const put = await fs.putBlob(hash, new Uint8Array(body));
                 if (!put.ok) {
-                    sendError(res, 502, put.error.code, "blob store write failed", { code: put.error.code });
+                    sendError(res, 502, put.error.code, "blob store write failed");
                     return;
                 }
                 res.writeHead(204).end(); // no body

package/dist/http/routes/sessions.js

@@ -796,7 +796,7 @@ async function handleSessionsBody(req, res, url, ctx, miss) {
                 if (typeof fs.blobSizes === "function") {
                     const known = (await fs.blobSizes([hash])).get(hash);
                     if (known !== undefined && known > cap) {
-                        sendError(res, 413, "workspace_file_too_large", "file exceeds the single-file read limit — download it via the archive endpoint", { code: "workspace_file_too_large", sizeBytes: known, limit: cap });
+                        sendError(res, 413, "workspace_file_too_large", "file exceeds the single-file read limit — download it via the archive endpoint", { sizeBytes: known, limit: cap });
                         return;
                     }
                 }
@@ -806,7 +806,7 @@ async function handleSessionsBody(req, res, url, ctx, miss) {
                     return;
                 }
                 if (bytes.byteLength > cap) {
-                    sendError(res, 413, "workspace_file_too_large", "file exceeds the single-file read limit — download it via the archive endpoint", { code: "workspace_file_too_large", sizeBytes: bytes.byteLength, limit: cap });
+                    sendError(res, 413, "workspace_file_too_large", "file exceeds the single-file read limit — download it via the archive endpoint", { sizeBytes: bytes.byteLength, limit: cap });
                     return;
                 }
                 // [1894]③:content-type 按扩展名(保守表,未知=octet-stream)+ 二进制判别头(首 8KiB NUL/控制

package/dist/http/routes/tasks.js

@@ -214,7 +214,7 @@ async function handleTasksBody(req, res, url, ctx, miss) {
                     const tid = earlyDurableTid ?? uuidv7(); // reuse the id already emitted as the X-Task-Id header (defensive mint if absent)
                     const created = await deps.runStore.createRun(tid, prepared.spec.sessionId, principal ?? null, deps.instanceId ?? "default", runMeta(prepared, source));
                     if (created.ok)
-                        deps.sessionTitler?.maybeTitle(prepared.spec.sessionId, prepared.spec.objective); // fire-and-forget; in-titler dedupe
+                        deps.sessionTitler?.maybeTitle(prepared.spec.sessionId, prepared.spec.objective, typeof prepared.spec.model === "string" ? prepared.spec.model : prepared.spec.model?.id); // fire-and-forget; in-titler dedupe;model 跟 turn([1992]②)
                     if (!created.ok) {
                         const conflict = { error: "session already has an active run — POST /v1/runs/{activeTaskId}/cancel stops it (same-instance interactive runs abort immediately)", errorCode: "conflict.session_active_run", activeTaskId: created.activeTaskId };
                         // Headers are already SSE — encode the rejection as the stream's terminal event; the 409 in
@@ -966,7 +966,7 @@ async function handleTasksBody(req, res, url, ctx, miss) {
                     const created = await deps.runStore.createRun(tid, prepared.spec.sessionId, principal ?? null, deps.instanceId ?? "default", runMeta(prepared, source));
                     if (!created.ok)
                         return { status: 409, body: { error: "session already has an active run — POST /v1/runs/{activeTaskId}/cancel stops it (same-instance interactive runs abort immediately)", errorCode: "conflict.session_active_run", activeTaskId: created.activeTaskId } };
-                    deps.sessionTitler?.maybeTitle(prepared.spec.sessionId, prepared.spec.objective); // fire-and-forget session auto-title
+                    deps.sessionTitler?.maybeTitle(prepared.spec.sessionId, prepared.spec.objective, typeof prepared.spec.model === "string" ? prepared.spec.model : prepared.spec.model?.id); // fire-and-forget;model 跟 turn([1992]②)
                     durableTaskId = tid;
                 }
                 if (deps.checkpointStore && prepared.spec.sessionId) {

package/dist/http/routes/trace-usage.js

@@ -269,13 +269,13 @@ async function streamTaskTrace(req, res, runStore, taskId, staleMs) {
         }
         const stale = Date.now() - new Date(run.updatedAt).getTime() > staleMs;
         if (stale && events.length === 0) {
-            res.write(`event: error\ndata: ${JSON.stringify({ code: "WORKER_DOWN", message: "run stalled (instance lost?)" })}\n\n`);
+            res.write(`event: error\ndata: ${JSON.stringify({ type: "error", errorCode: "WORKER_DOWN", message: "run stalled (instance lost?)" })}\n\n`);
             break;
         }
         if (Date.now() - start > MAX_MS) {
             // BL-16: the stream hit its 15-min cap but the RUN is still going — emit a terminal event so the client
             // can tell "reconnect to continue" from "task completed" (a silent close looked like completion).
-            res.write(`event: error\ndata: ${JSON.stringify({ code: "STREAM_MAX_DURATION", message: "trace stream reached its 15-minute cap — reconnect with Last-Event-ID to continue (the run is still active)" })}\n\n`);
+            res.write(`event: error\ndata: ${JSON.stringify({ type: "error", errorCode: "STREAM_MAX_DURATION", message: "trace stream reached its 15-minute cap — reconnect with Last-Event-ID to continue (the run is still active)" })}\n\n`);
             break;
         }
         if (events.length === 0) {

package/dist/http/routes/workflows.js

@@ -485,8 +485,10 @@ async function streamWorkflowRun(req, res, runId, scope) {
         }
     }
     catch {
+        // A1/A6/B:归一 error 帧形(`type` + 机器码)。此前只发人话 message ⇒ 消费端要么锚文案、要么把
+        // "流断了"和"流正常结束"混为一谈;`workflow.stream_error` 是本帧的铸码(附录 A `*.stream_error` 族)。
         if (!closed && !res.writableEnded)
-            res.write(`event: error\ndata: ${JSON.stringify({ message: "workflow stream error" })}\n\n`);
+            res.write(`event: error\ndata: ${JSON.stringify({ type: "error", errorCode: "workflow.stream_error", message: "workflow stream error" })}\n\n`);
     }
     finally {
         clearInterval(hb);

package/dist/http/server.d.ts

@@ -197,7 +197,7 @@ export interface ServiceCoordinatorDeps {
     leaderEndpoint?: LeaderEndpoint;
     /** Fire-and-forget session auto-titler (first-submit hook; absent = feature dark). */
     sessionTitler?: {
-        maybeTitle(sessionId: string, objective: string): void;
+        maybeTitle(sessionId: string, objective: string, model?: string): void;
     };
 }
 /** **函数/构件缝**:main.ts 把跨多个子系统的动作编译成一个可调用面交给路由(路由不该自己

package/dist/http/server.js

@@ -281,7 +281,7 @@ export function createHttpServer(rawDeps) {
             // 4xx, not a 500. Everything else is a genuine internal error: log the cause, never echo it (info disclosure).
             if (err instanceof HttpError) {
                 if (!res.headersSent)
-                    sendError(res, err.status, httpErrorCode(err.status, err.code), err.message, { ...(err.code ? { code: err.code } : {}), ...(err.extra ?? {}) });
+                    sendError(res, err.status, httpErrorCode(err.status, err.code), err.message, { ...(err.extra ?? {}) });
                 else
                     res.end();
                 return;
@@ -1091,7 +1091,7 @@ export function createHttpServer(rawDeps) {
             if (err instanceof HttpError) {
                 // Typed rejection: echo the stable `code` + any structured extras (e.g. the scenario
                 // allowlist) beside the human message — machine code on the wire, prose stays with the shell/web.
-                sendError(res, err.status, httpErrorCode(err.status, err.code), err.message, { ...(err.code ? { code: err.code } : {}), ...(err.extra ?? {}) });
+                sendError(res, err.status, httpErrorCode(err.status, err.code), err.message, { ...(err.extra ?? {}) });
                 return null;
             }
             throw err;
@@ -1181,6 +1181,27 @@ export function createHttpServer(rawDeps) {
         }
         if (!cp)
             return { status: 404, body: { error: "checkpoint not found", errorCode: "not_found.checkpoint" } };
+        // 🔴 GATE-KIND GUARD ([1995]①,`gate_not_resumable` / `gate_not_plan_review` 的**反方向镜像**):这条腿
+        // 只解 TOOL 审批门(core gateMatch:`human` / `irreversible_ask` ↔ outcome.gate "policy_ask")。此前它无
+        // 条件铸 outcome{gate:"policy_ask"} 交给 core —— 一个 plan_review / resource_limit / task_done park 被误
+        // 路由到 /decide(而 /v1/approvals 把它们与工具审批**混在同一个队列**里、toolName 同为 null)就会深入到
+        // core 的 `checkpoint.gate_mismatch`。那是 core 的 **pre-CAS** 拒绝(checkpoint 仍 pending),但 server 的
+        // CheckpointError 分类表把 gate_mismatch 归为 TERMINAL ⇒ 行被 setTerminal 成 failed + 释放 task_active,
+        // 于是「行说这活死了、park 却还活着」:claim 泄漏(同 session 可并发起第二个 run)、事后用正确的腿救能
+        // 200 真跑完却拿不到 taskId、run 行永停 failed(账本与现实分叉)。这里 pre-CAS 纯拒绝:不动 checkpoint、
+        // 不动 run 行、不烧模型腿,与另外两条腿的守卫同姿势(core 的 gate-match 仍是 fail-closed 兜底)。
+        // ⚠️ 门放在 parked 赎回腿**之前**:赎回腿同样只铸 policy_ask outcome,错门 cp 在那条链上只会更深地炸。
+        // D-D SLA deny-sweep 不受影响:它的取行查询本就只选 gate_kind IN ('human','irreversible_ask')。
+        const decideGateKind = cp.gate?.kind;
+        if (decideGateKind !== "human" && decideGateKind !== "irreversible_ask") {
+            return {
+                status: 409,
+                body: {
+                    error: `task is suspended on a '${decideGateKind ?? "unknown"}' gate, not a tool approval — a plan_review gate is resolved via POST /v1/assistant/tasks/:id/plan_review, a resource_limit suspension via POST /v1/assistant/tasks/:id/resume`,
+                    errorCode: "gate_not_tool_approval",
+                },
+            };
+        }
         // design/153 件3d parked 赎回腿([1574]/[1584]/[1588] core 裁定;src/parked-decide.ts):该 pending
         // checkpoint 若属于一个 parked 后台子代,decide 必须走 claim→裸 execute(reviveClaim) 链——legacy 腿的
         // resumeStream 会绕开 bg registry(无 consumeParkedFlip、行永 parked、生命周期分叉),故 parked-first。
@@ -1243,7 +1264,7 @@ export function createHttpServer(rawDeps) {
             if (e instanceof HttpError) {
                 return {
                     status: 409,
-                    body: { error: `resume blocked by an admission policy change: ${e.message}`, code: "resume_blocked_by_policy", errorCode: "resume_blocked_by_policy", ...(e.code ? { blockedBy: e.code } : {}), ...(e.extra ?? {}) },
+                    body: { error: `resume blocked by an admission policy change: ${e.message}`, errorCode: "resume_blocked_by_policy", ...(e.code ? { blockedBy: e.code } : {}), ...(e.extra ?? {}) },
                 };
             }
             throw e;
@@ -1871,7 +1892,6 @@ export function createHttpServer(rawDeps) {
                     status,
                     body: {
                         error: e.message,
-                        code: e.code,
                         // invalid_outcome means DIFFERENT things per gate: for policy_ask it IS the D-1 action-binding
                         // mismatch; for plan_review/dry_run_review it is a content-sanitize reject (editedPlan/reason carried a
                         // forbidden tag) — give those a gate-appropriate code so the UI routes to the right recovery, not the
@@ -2210,7 +2230,7 @@ export function createHttpServer(rawDeps) {
         res.setHeader("retry-after", String(adm.retryAfterSec));
         // E4 定稿:429 带 typed code + center 结构化字段 verbatim 透传(windowType/remaining/resetAt/
         // isEstimate/pool)——人话文案归壳/web(契约如此),error 串保留旧消费者兼容。
-        sendError(res, 429, "quota_exhausted", "budget exhausted (quota lease)", { code: "quota_exhausted", retryAfterSec: adm.retryAfterSec, ...(adm.detail ?? {}) });
+        sendError(res, 429, "quota_exhausted", "budget exhausted (quota lease)", { retryAfterSec: adm.retryAfterSec, ...(adm.detail ?? {}) });
         return true;
     }
     /** Apply the per-principal rate limit; responds 429 and returns true when the caller is over. */

package/dist/http/sse-log.js

@@ -55,7 +55,7 @@ export async function streamSseLog(req, res, provider, id, staleMs) {
         if (Date.now() - start > MAX_MS) {
             // BL-16: cap reached but the run continues — emit a terminal reconnect signal (not a silent close, which
             // a client couldn't distinguish from a completed run). Distinct from the `failed`/`done` run events.
-            res.write(`event: error\ndata: ${JSON.stringify({ type: "error", code: "STREAM_MAX_DURATION", message: "stream reached its 15-minute cap — reconnect with Last-Event-ID to continue (the run is still active)" })}\n\n`);
+            res.write(`event: error\ndata: ${JSON.stringify({ type: "error", errorCode: "STREAM_MAX_DURATION", message: "stream reached its 15-minute cap — reconnect with Last-Event-ID to continue (the run is still active)" })}\n\n`);
             break;
         }
         if (events.length === 0) {