@tea-agent/loop-agent 0.38.3 → 0.38.4

package/dist/workflows/dag/rerun-run.js

@@ -2,6 +2,7 @@ import { createHash } from "node:crypto";
 import { access, copyFile, mkdir, readdir, readFile } from "node:fs/promises";
 import path from "node:path";
 import { writeJsonAtomic } from "../../infrastructure/harness/atomic-write.js";
+import { composeDagPromptOverridePrompt, isDagRerunPromptOverrideEnabled, } from "../../shared/dag-prompt-override.js";
 import { resolveRunningControllerIdentity } from "../../shared/package-metadata.js";
 import { readControllerIdentityArtifact, captureControllerIdentity } from "./controller-identity.js";
 import { DAG_RUNS_DIR, getDagRunDir, isDagNonExecutionLifecycle, locateDagRun, readDagRunSpec, readDagRunState, } from "./lifecycle.js";
@@ -190,6 +191,8 @@ export async function planDagRerun(input) {
             parentSpec: nonCompletedSpec,
             parentState: nonCompletedState,
             parentLifecycle,
+            promptOverrides: input.promptOverrides,
+            promptOverrideEnabled: isDagRerunPromptOverrideEnabled(),
             workerManaged: nonCompletedWorkerHints.workerManaged,
             workerAssociation: nonCompletedWorkerHints.workerManaged
                 ? {
@@ -229,6 +232,8 @@ export async function planDagRerun(input) {
         parentTerminalWorkspace,
         currentWorkspace,
         skillSnapshotOk,
+        promptOverrides: input.promptOverrides,
+        promptOverrideEnabled: isDagRerunPromptOverrideEnabled(),
         workerManaged: workerHints.workerManaged,
         workerAssociation: workerHints.workerManaged
             ? {
@@ -384,15 +389,57 @@ function buildPendingNodeRecord(task) {
         ...(task.outputMode ? { outputMode: task.outputMode } : {}),
     };
 }
+/**
+ * Stage the continuation spec with operator prompt overrides applied. The
+ * parent spec stays frozen; every override must match the plan entry (sha256
+ * of the composed final prompt and of the parent base prompt) so the
+ * two-phase planHash remains the single source of truth for what changed.
+ * "append" entries keep the frozen base and append the operator addendum in a
+ * marked block; "replace" entries swap the whole prompt.
+ */
+function applyPromptOverridesToSpec(input) {
+    const planned = new Map((input.plan.promptOverrides ?? []).map((entry) => [entry.nodeId, entry]));
+    const spec = structuredClone(input.spec);
+    for (const override of input.overrides) {
+        const entry = planned.get(override.nodeId);
+        if (!entry) {
+            throw new Error(`prompt override is not present in the rerun plan: ${override.nodeId}`);
+        }
+        const task = spec.tasks.find((candidate) => candidate.id === override.nodeId);
+        if (!task) {
+            throw new Error(`prompt override node missing from spec: ${override.nodeId}`);
+        }
+        if (sha256Hex(task.subtask_prompt) !== entry.basePromptSha256) {
+            throw new Error(`prompt override base prompt drifted from plan: ${override.nodeId}`);
+        }
+        const finalPrompt = composeDagPromptOverridePrompt({
+            basePrompt: task.subtask_prompt,
+            text: override.prompt,
+            mode: override.mode === "append" ? "append" : "replace",
+        });
+        if (sha256Hex(finalPrompt) !== entry.promptSha256) {
+            throw new Error(`prompt override text does not match the planned sha256: ${override.nodeId}`);
+        }
+        task.subtask_prompt = finalPrompt;
+    }
+    return spec;
+}
 async function stageContinuationRun(input) {
-    const { ranks } = topoSortToRanks(input.parentSpec);
-    const runId = await buildDagRunId(input.parentSpec, { cwd: input.cwd });
+    const specForRun = input.promptOverrides !== undefined && input.promptOverrides.length > 0
+        ? applyPromptOverridesToSpec({
+            spec: input.parentSpec,
+            plan: input.plan,
+            overrides: input.promptOverrides,
+        })
+        : input.parentSpec;
+    const { ranks } = topoSortToRanks(specForRun);
+    const runId = await buildDagRunId(specForRun, { cwd: input.cwd });
     const runDir = getDagRunDir(input.cwd, "active", runId);
     await prepareActiveRunDir(runDir);
-    await writeRunSpec(runDir, input.parentSpec);
+    await writeRunSpec(runDir, specForRun);
     const resetSet = new Set(input.plan.resetNodeIds);
     const importedSet = new Set(input.plan.importedNodeIds);
-    const tasksById = new Map(input.parentSpec.tasks.map((task) => [task.id, task]));
+    const tasksById = new Map(specForRun.tasks.map((task) => [task.id, task]));
     const importManifestNodes = [];
     const nodes = {};
     for (const nodeId of input.plan.importedNodeIds) {
@@ -433,6 +480,9 @@ async function stageContinuationRun(input) {
         reason: input.reason,
         requestId: input.requestId,
         planHash: input.plan.planHash,
+        ...(input.plan.promptOverrides
+            ? { promptOverrides: input.plan.promptOverrides }
+            : {}),
         createdAt,
     };
     const state = {
@@ -482,7 +532,7 @@ async function stageContinuationRun(input) {
     await writeJsonAtomic(path.join(runDir, IMPORT_MANIFEST_REL_PATH), importManifest);
     await writeJsonAtomic(path.join(runDir, RERUN_PLAN_REL_PATH), input.plan);
     await writeRunState(runDir, state);
-    return { runId, runDir, state };
+    return { runId, runDir, state, spec: specForRun };
 }
 export async function executeDagRerun(input) {
     const existing = await readExistingOperatorRequest(input.cwd, input.requestId);
@@ -493,6 +543,7 @@ export async function executeDagRerun(input) {
         cwd: input.cwd,
         parentRunId: input.parentRunId,
         selectedNodeId: input.selectedNodeId,
+        promptOverrides: input.promptOverrides,
     });
     if (plan.planHash !== input.planHash) {
         return {
@@ -530,10 +581,11 @@ export async function executeDagRerun(input) {
         plan,
         reason: input.reason,
         requestId: input.requestId,
+        promptOverrides: input.promptOverrides,
     });
     const continuationOptions = {
         cwd: input.cwd,
-        spec: parentSpec,
+        spec: staged.spec,
         state: staged.state,
         runDir: staged.runDir,
         ...(input.maxConcurrent !== undefined

package/dist/workflows/dag/retry-policy.js

@@ -165,6 +165,36 @@ export const WRITER_TRANSPORT_RETRY_POLICY = {
     maxDelayMs: 30000,
     retryCategories: [WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY],
 };
+/**
+ * Implicit runtime retry for canonical final verify shells. Not a DagRetryPolicy:
+ * `spawn-error` / `test-timeout` are runtime command categories and must not enter
+ * `dagRetryPolicySchema` / `ALL_DAG_RETRY_CATEGORIES`.
+ */
+export const VERIFY_SHELL_RETRY_CATEGORIES = [
+    "timeout",
+    "spawn-error",
+    "test-timeout",
+];
+export const VERIFY_SHELL_RETRY_POLICY = {
+    maxAttempts: 2,
+    backoff: "exponential",
+    initialDelayMs: 2000,
+    maxDelayMs: 30000,
+    retryCategories: VERIFY_SHELL_RETRY_CATEGORIES,
+};
+export const TARGET_TEMPLATE_TRANSIENT_RETRY_PROFILE = "standard-reviewed-supervised-v1";
+/**
+ * Target-template exclusive implement-pi transport retry. Same schema category
+ * as frontend (`writer-clean-timeout`) so DagSpec validation stays unchanged;
+ * only the delay cap is wider. Do not mutate {@link WRITER_TRANSPORT_RETRY_POLICY}.
+ */
+export const TARGET_TEMPLATE_WRITER_TRANSPORT_RETRY_POLICY = {
+    maxAttempts: 2,
+    backoff: "exponential",
+    initialDelayMs: 5000,
+    maxDelayMs: 60000,
+    retryCategories: [WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY],
+};
 /**
  * Backend-test generation writers: empty-diff once, plus bounded incomplete-write-set
  * recovery attempts driven by Completeness Gate (missing/broken target files).
@@ -238,6 +268,67 @@ export function isWriterTransportRetryCandidate(task) {
         task.writerOutcomePolicy?.type === "implementation-outcome-v1" &&
         task.writerOutcomePolicy.requireChangedFiles !== true);
 }
+/**
+ * Canonical standard exclusive `implement-pi` only. Frontend designers and
+ * completeness writers stay on the existing timeout-only transport candidate.
+ */
+export function isTargetTemplateImplementPi(task) {
+    return isWriterTransportRetryCandidate(task) && task.id === "implement-pi";
+}
+export function hasTargetTemplateTransientRetryProfile(task) {
+    return (task?.transientRetryProfile === TARGET_TEMPLATE_TRANSIENT_RETRY_PROFILE);
+}
+/**
+ * Dual-condition eligibility: canonical shape AND generated-path stamp.
+ * Custom / other-template nodes that only match id or shape stay ineligible.
+ */
+export function isTargetTemplateTransientRetryNode(task) {
+    if (!hasTargetTemplateTransientRetryProfile(task))
+        return false;
+    return (isCanonicalFinalVerifyShellRetryCandidate(task) ||
+        isTargetTemplateImplementPi(task));
+}
+/**
+ * Shape-based eligibility for implicit canonical final verify-shell retry.
+ * Does not consult node id or a template field.
+ */
+export function isCanonicalFinalVerifyShellRetryCandidate(task) {
+    if (task.executor !== "shell")
+        return false;
+    if (task.role !== "verifier")
+        return false;
+    if (task.writePolicy !== undefined &&
+        task.writePolicy !== "read-only" &&
+        task.writePolicy !== "none") {
+        return false;
+    }
+    if ((task.writeSet?.length ?? 0) > 0)
+        return false;
+    const shell = task.shell;
+    if (!shell || (shell.commands?.length ?? 0) === 0)
+        return false;
+    const evidence = shell.verifyEvidence;
+    if (!evidence)
+        return false;
+    if (evidence.phase !== "final")
+        return false;
+    if (evidence.quota !== "full")
+        return false;
+    if (evidence.commandSource !== "adapter")
+        return false;
+    if (evidence.finalFullRequired !== true)
+        return false;
+    if (shell.frontendPrewriteGate ||
+        shell.frontendVerificationBundle ||
+        shell.frontendBrowserToolPreflight ||
+        shell.backendTestPipeline ||
+        shell.jacocoCoverage ||
+        shell.repairArtifactGate ||
+        shell.verdictGate) {
+        return false;
+    }
+    return true;
+}
 export function isSafeReadOnlyPiRetryCandidate(task) {
     if (task.executor !== "pi")
         return false;

package/dist/workflows/dag/types.js

@@ -2,7 +2,7 @@ import { z } from "zod";
 import { isOpenspecSpecFilePath } from "../../shared/openspec-spec.js";
 import { campaignBudgetSchema, } from "../../application/evaluation/budget.js";
 import { assertDagPromptSourceRule } from "./prompt-source.js";
-import { dagRetryPolicySchema } from "./retry-policy.js";
+import { dagRetryPolicySchema, TARGET_TEMPLATE_TRANSIENT_RETRY_PROFILE, } from "./retry-policy.js";
 import { dagLivenessPolicySchema, } from "./liveness-policy.js";
 import { dagOutputProtocolSchema } from "./output-protocol.js";
 export const dagComplexitySchema = z.enum(["HIGH", "MED", "LOW"]);
@@ -849,6 +849,14 @@ export const dagTaskSchema = z.object({ id: z.string().regex(/^[a-z][a-z0-9-]*$/
     forbiddenPaths: z.array(z.string()).optional().default([]),
     decisionGate: dagDecisionGateSchema.optional(),
     retryPolicy: dagRetryPolicySchema.optional(),
+    /**
+     * Generation-path stamp for implicit same-run transient retry. Only the
+     * three target templates stamp canonical final verify shells and
+     * exclusive `implement-pi`. Zod-declared so parseDagSpec does not strip it.
+     */
+    transientRetryProfile: z
+        .literal(TARGET_TEMPLATE_TRANSIENT_RETRY_PROFILE)
+        .optional(),
     /** Optional per-node adaptive liveness override (merged over defaults). */
     livenessPolicy: dagLivenessPolicySchema,
     dynamicExpansion: dagDynamicExpansionSchema.optional(),

package/docs/init-surface.manifest.json

@@ -50,6 +50,9 @@
     "scripts/kb-graph-materialize.mjs",
     "scripts/kb-graph-promote.mjs",
     "scripts/kb-graph-incremental-prepare.mjs",
+    "scripts/next-info.mjs",
+    "scripts/next-publish-gate.mjs",
+    "scripts/release-source-binding.mjs",
     "docs/templates/agent-dag.schema.json",
     "examples/example-dag.json",
     "skills/agent-worker/SKILL.md",
@@ -277,6 +280,7 @@
     "harness.json",
     "package.json",
     "src/commands/init.ts",
+    "src/commands/init-model-catalog.ts",
     "src/commands/client-recovery.ts",
     "src/commands/init-upgrade.ts",
     "src/cli.ts",
@@ -302,6 +306,7 @@
         "description": "Changes that may alter target-project initialization behavior, default DAG role skills, skill resolution, or package/init contracts.",
         "patterns": [
           "src/commands/init.ts",
+          "src/commands/init-model-catalog.ts",
           "src/commands/client-recovery.ts",
           "src/commands/init-upgrade.ts",
           "src/cli.ts",

package/docs/operations/README.md

@@ -7,7 +7,7 @@
 - [`local-development-environment.md`](local-development-environment.md)：本地环境、Windows 差异和常见排障。
 - [`branch-merge-guideline.md`](branch-merge-guideline.md)：按风险选择合并模式并留下 source-SHA 证据。
 - [`github-collaboration.md`](github-collaboration.md)：仓库内 GitHub 协作约定及发布通道。
-- [`version-and-release-management.md`](version-and-release-management.md)：版本发布管理——next / release-x.y / latest 渠道、workflow 与 promote 手册。
+- [`version-and-release-management.md`](version-and-release-management.md)：版本发布管理——next / release-x.y / latest 渠道、移动式 `npm-next` ref、`inspect next` 查询入口、workflow 与 promote 手册。
 - [`production-readiness.md`](production-readiness.md)：交付前的生产就绪判断。
 - [`agent-worker-production-readiness.md`](agent-worker-production-readiness.md)：agent-worker Feature/Task Pool 外层编排的 Production Readiness v1 操作合同。
 - [`backend-test-jacoco-coverage.md`](backend-test-jacoco-coverage.md)：Java 被测服务挂载 JaCoCo tcpserver agent 的 Maven、`java -jar` 与 Docker 操作手册。

package/docs/templates/init-managed-agents.md

@@ -80,9 +80,19 @@ loop-agent init update --repo-root . --apply-safe
 当用户要求初始化 loop-agent（“初始化 loop-agent”“loop agent 初始化”“loop agent初始化”“loop-agent 初始化”），或发现 `README.md` / `__LOOP_AGENT_GOVERNANCE_ROOT__/verification-matrix.md` 仍停留在通用初始化状态时，模型应在同一轮工作中完成初始化闭环，不要把补全文档和验证矩阵留给用户交互。
 
 1. 运行 `loop-agent init instructions --repo-root .`。
-2. 运行 `loop-agent init --repo-root . --profile full --merge`。
-3. 基于真实项目文件补全根 `README.md`、`__LOOP_AGENT_GOVERNANCE_ROOT__/verification-matrix.md`，必要时适配 `scripts/ci-tests.sh`。
-4. 运行 `loop-agent init doctor --repo-root .`、`loop-agent inspect`、`loop-agent docs audit` 并完成 quick verification。
+2. 运行 `loop-agent init model-catalog --repo-root . --json`。loop-agent 不直接打开或解析认证文件原文，只让 Pi/OpenCode 自己的 runtime 加载其配置，并输出白名单模型元数据、Pi setup 状态、Pi 兼容性与 credential-ready 布尔值；token、headers、URL、原始命令输出和认证路径不会进入 catalog。
+3. 先处理 `piSetup`：如果没有模型定义，向用户索取无法推断的 provider/model id、API/base URL、上下文窗口、最大输出、输入/推理能力、成本约束和凭据方式，帮助按 Pi 原生 schema 合并 `~/.pi/agent/models.json`；不得让用户把凭据值写入对话、harness 或报告，优先 Pi 登录/auth storage 或环境变量引用。如果已有定义但凭据未就绪，只修复认证。随后重跑 catalog，未出现 credential-ready Pi model 前不得继续初始化。
+4. 对已有 `harness.json`，逐档保留仍存在于 Pi 的 LOW/MED/HIGH 原值；只为已移除或缺失的档位重新选择。结合目标项目实际技术栈、模型能力、成本与上下文窗口，从 `eligibleForHarness=true` 的候选中自主选择，并直接运行带完整三档参数的 init/update；代码会拒绝改动仍有效的档位。三档可以相同。只有凭据、成本、合规或可用性存在无法推断的实质歧义时才询问用户；OpenCode-only 候选不能证明 Pi executor 可用。
+5. 基于真实项目文件补全根 `README.md`、`__LOOP_AGENT_GOVERNANCE_ROOT__/verification-matrix.md`，必要时适配 `scripts/ci-tests.sh`。
+6. 运行 `loop-agent init doctor --repo-root .`、`loop-agent inspect`、`loop-agent docs audit` 并完成 quick verification；doctor 未确认 Pi model matrix 时不得开始 DAG。
+
+```bash
+loop-agent init --repo-root . --profile full --merge \
+  --low-model <provider/model> --med-model <provider/model> --high-model <provider/model>
+# 已初始化项目只替换失效档位；仍有效档位在完整参数中保持当前引用
+loop-agent init update --repo-root . --apply-safe \
+  --low-model <current-or-selected> --med-model <current-or-selected> --high-model <current-or-selected>
+```
 
 信息不足时写明尚未识别到的事实；不要留下泛化占位符。
 

package/harness.json

@@ -83,8 +83,8 @@
     "pi": {
       "description": "Pi 负责规划、评审、诊断；当 DAG toolProfile=write 时也可做有界写入。模型按复杂度三档配置，支持 provider/model 字符串或带 thinking 的对象；斜杠前为 Pi provider，后为 modelId，勿只写裸 modelId。",
       "LOW": "wizard-local/minimax-m3",
-      "MED": "wizard-local/grok-4.6",
-      "HIGH": "wizard-local/gpt-5.6-sol"
+      "MED": "wizard-local/minimax-m3",
+      "HIGH": "wizard-local/minimax-m3"
     }
   }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tea-agent/loop-agent",
-  "version": "0.38.3",
+  "version": "0.38.4",
   "type": "module",
   "bin": {
     "loop-agent": "bin/loop-agent.js",
@@ -33,6 +33,9 @@
     "scripts/kb-graph-materialize.mjs",
     "scripts/kb-graph-promote.mjs",
     "scripts/kb-graph-incremental-prepare.mjs",
+    "scripts/next-info.mjs",
+    "scripts/next-publish-gate.mjs",
+    "scripts/release-source-binding.mjs",
     "README.md",
     "CHANGELOG.md"
   ],
@@ -58,6 +61,7 @@
     "test": "node scripts/run-tests.mjs",
     "test:host": "vitest run test/init-upgrade.test.ts --maxWorkers=1 -t \"records and reuses|controller-authorized semantic merge|authority-sensitive\"",
     "test:fast": "vitest run --config vitest.fast.config.ts",
+    "test:io-heavy": "vitest run --config vitest.io-heavy.config.ts",
     "test:integration": "vitest run --config vitest.integration.config.ts",
     "docs:dev": "npm --prefix website start",
     "docs:build": "npm --prefix website run build",