@tea-agent/loop-agent 0.4.0 → 0.5.0

package/dist/commands/loop-benchmark.js

@@ -37,17 +37,17 @@ export function parseLoopBenchmarkArgs(args) {
     return { json, markdown, outputPath };
 }
 export function printLoopBenchmarkUsage() {
-    console.log(`usage: loop-benchmark [options]
-
-Deterministic loop-agent loop benchmark baseline (no live Pi/Cursor calls).
-
-Options:
-  --json              Emit JSON (default when no format flag is set)
-  --markdown          Emit Markdown report
-  --output <path>     Write Markdown report to a repo-relative or absolute path
-  -h, --help          Show this help
-
-Control groups: single-repair, 3-pass-convergence, 3-pass-convergence+quota.
+    console.log(`usage: loop-benchmark [options]
+
+Deterministic loop-agent loop benchmark baseline (no live Pi/Cursor calls).
+
+Options:
+  --json              Emit JSON (default when no format flag is set)
+  --markdown          Emit Markdown report
+  --output <path>     Write Markdown report to a repo-relative or absolute path
+  -h, --help          Show this help
+
+Control groups: single-repair, 3-pass-convergence, 3-pass-convergence+quota.
 Recommendation never changes convergence.enabled default.`);
 }
 export async function runLoopBenchmark(repoRoot, rawArgs) {

package/dist/commands/pi-reuse-benchmark.js

@@ -106,22 +106,22 @@ export function parsePiReuseBenchmarkArgs(args) {
     };
 }
 export function printPiReuseBenchmarkUsage() {
-    console.log(`usage: pi-reuse-benchmark [options]
-
-Deterministic Pi runtime reuse benchmark/decision summary (no live Pi calls).
-
-Options:
-  --report <path>         Benchmark report markdown (approval status)
-  --approval <path>       Explicit approval JSON artifact
-  --off-executor <path>   Baseline executor.jsonl (reuse off)
-  --on-executor <path>    Treatment executor.jsonl (reuse on)
-  --off-task <task-id>    Resolve baseline from .harness/tasks/<id>/logs/executor.jsonl
-  --on-task <task-id>     Resolve treatment from .harness/tasks/<id>/logs/executor.jsonl
-  --json                  Emit JSON (default when no format flag is set)
-  --markdown              Emit Markdown summary
-  -h, --help              Show this help
-
-Recommendations: defer | maintain-opt-in | eligible-for-human-review
+    console.log(`usage: pi-reuse-benchmark [options]
+
+Deterministic Pi runtime reuse benchmark/decision summary (no live Pi calls).
+
+Options:
+  --report <path>         Benchmark report markdown (approval status)
+  --approval <path>       Explicit approval JSON artifact
+  --off-executor <path>   Baseline executor.jsonl (reuse off)
+  --on-executor <path>    Treatment executor.jsonl (reuse on)
+  --off-task <task-id>    Resolve baseline from .harness/tasks/<id>/logs/executor.jsonl
+  --on-task <task-id>     Resolve treatment from .harness/tasks/<id>/logs/executor.jsonl
+  --json                  Emit JSON (default when no format flag is set)
+  --markdown              Emit Markdown summary
+  -h, --help              Show this help
+
+Recommendations: defer | maintain-opt-in | eligible-for-human-review
 Never changes CODE_AGENT_PI_REUSE_RUNTIME default (off).`);
 }
 function resolveRepoRelative(repoRoot, filePath) {

package/dist/executors/config-core.js

@@ -1,5 +1,5 @@
 import { DEFAULT_CURSOR_MODEL } from './cursor-executor.js';
-import { DEFAULT_DAG_MODELS } from './model-routing.js';
+import { resolveExecutorModelMatrix } from './model-routing.js';
 export const TASK_COMPLEXITY_TO_DAG = {
     small: 'LOW',
     medium: 'MED',
@@ -13,7 +13,8 @@ export function resolveCursorModelForTaskConfig(taskConfig, cursorConfig) {
         return taskConfig.cursorModel.trim();
     }
     const dagLevel = TASK_COMPLEXITY_TO_DAG[taskConfig.complexity ?? 'medium'];
-    return DEFAULT_DAG_MODELS[dagLevel] ?? resolveCursorModel({}, cursorConfig);
+    const matrix = resolveExecutorModelMatrix('cursor', cursorConfig);
+    return matrix[dagLevel] ?? resolveCursorModel({}, cursorConfig);
 }
 export function resolveTaskExecutor(taskConfig, override) {
     if (override)

package/dist/executors/cursor-executor.js

@@ -29,7 +29,7 @@ export function resolveArtifactWriteDir(options) {
 }
 export function buildArtifactPathPrompt(writeDir) {
     if (!writeDir)
-        return `After changes, write artifacts/修改记录.md and artifacts/验证结果.md with verification evidence.
+        return `After changes, write artifacts/修改记录.md and artifacts/验证结果.md with verification evidence.
 ${ARTIFACT_INSTRUCTIONS}`;
     return [
         `After changes, write the following files:`,

package/dist/executors/model-routing.js

@@ -1,9 +1,52 @@
+import { DEFAULT_DAG_EXECUTOR_MODELS, } from '../workflows/dag/types.js';
 export const DEFAULT_DAG_CURSOR_MODEL = "composer-2.5";
 export const DEFAULT_DAG_MODELS = {
     HIGH: "gpt-5.5",
     MED: "composer-2.5",
     LOW: "composer-2.5",
 };
+/**
+ * DAG executor model tier keys that may carry a per-complexity override.
+ */
+const EXECUTOR_MODEL_TIERS = ["LOW", "MED", "HIGH"];
+/**
+ * Resolve the DAG executor model matrix for a single executor from its harness
+ * `executors.<name>` config.
+ *
+ * Priority per tier (LOW/MED/HIGH):
+ *   1. execConfig[tier]  (truthy and !== "default" sentinel)
+ *   2. execConfig.defaultModel (truthy and !== "default" sentinel)
+ *   3. DEFAULT_DAG_EXECUTOR_MODELS[executor][tier]
+ *
+ * The "default" literal (injected by the schema `.default("default")`) and
+ * absent/undefined both mean "no override, fall through".
+ */
+export function resolveExecutorModelMatrix(executor, execConfig) {
+    const tierValue = (tier) => {
+        const tierOverride = execConfig?.[tier];
+        if (tierOverride && tierOverride !== "default")
+            return tierOverride;
+        const defaultModel = execConfig?.defaultModel;
+        if (defaultModel && defaultModel !== "default")
+            return defaultModel;
+        return DEFAULT_DAG_EXECUTOR_MODELS[executor][tier];
+    };
+    return {
+        LOW: tierValue("LOW"),
+        MED: tierValue("MED"),
+        HIGH: tierValue("HIGH"),
+    };
+}
+/**
+ * Resolve both pi and cursor DAG executor model matrices from a harness manifest.
+ * Single entry point for DAG generation and --strict-models baseline resolution.
+ */
+export function resolveExecutorModelMatrices(manifest) {
+    return {
+        pi: resolveExecutorModelMatrix("pi", manifest.executors?.pi),
+        cursor: resolveExecutorModelMatrix("cursor", manifest.executors?.cursor),
+    };
+}
 export function resolveModelSelection(manifest, taskConfig, step, options) {
     const retryAttempt = options?.retryAttempt ?? 0;
     if (Object.keys(manifest.modelProfiles ?? {}).length === 0) {

package/dist/governance/manifest-types.js

@@ -21,7 +21,15 @@ export const executorManifestSchema = z.object({
     description: z.string().optional(),
     enabled: z.boolean().optional(),
     defaultModel: z.string().optional().default("default"),
-    requiresApiKey: z.string().optional().default("CURSOR_API_KEY"),
+    /**
+     * Per-complexity DAG model overrides. When set (and not the "default" sentinel),
+     * these take priority over defaultModel for the matching DAG executor tier.
+     * "default" literal and absent/undefined both mean "no override, fall through".
+     */
+    LOW: z.string().optional(),
+    MED: z.string().optional(),
+    HIGH: z.string().optional(),
+    requiresApiKey: z.string().optional(),
 });
 export const workflowPolicyProfileNameSchema = z.enum([
     "minimal",

package/dist/task/runtime.js

@@ -236,35 +236,35 @@ export function shouldInjectSubagentGuidance(step, mode) {
     return false;
 }
 /** Advisory guidance for `analyze-plan` mode. */
-export const SUBAGENT_GUIDANCE_STANDARD = `<subagent_guidance>
-You have access to the \`subagent\` tool for lightweight delegation within this step.
-Use it only for read-only tasks:
-- Parallel scout: dispatch multiple subagents to search/read different areas simultaneously
-- Chain: scout -> planner (one subagent scouts, another plans based on findings)
-- Reviewer: have a subagent review your analysis/plan before finalizing
-Do NOT use subagent for writing, editing, or executing commands.
-Subagent output is advisory only; always verify and incorporate findings into your own output.
-Do NOT treat subagent results as authoritative state or artifact sources.
+export const SUBAGENT_GUIDANCE_STANDARD = `<subagent_guidance>
+You have access to the \`subagent\` tool for lightweight delegation within this step.
+Use it only for read-only tasks:
+- Parallel scout: dispatch multiple subagents to search/read different areas simultaneously
+- Chain: scout -> planner (one subagent scouts, another plans based on findings)
+- Reviewer: have a subagent review your analysis/plan before finalizing
+Do NOT use subagent for writing, editing, or executing commands.
+Subagent output is advisory only; always verify and incorporate findings into your own output.
+Do NOT treat subagent results as authoritative state or artifact sources.
 </subagent_guidance>`;
 /** Strong guidance for `full` mode — prescriptive when to delegate. */
-export const SUBAGENT_GUIDANCE_STRONG = `<subagent_guidance>
-You have access to the \`subagent\` tool for lightweight delegation within this step.
-
-You SHOULD delegate to subagent scouts when:
-- The task requires scanning 3+ directories or comparing implementations across modules
-- You would otherwise need 5+ sequential read/grep calls to gather context
-- A reviewer subagent can independently catch scope drift before you finalize your output
-
-Delegation saves context tokens and produces better results.
-
-Allowed patterns:
-- Parallel scout: dispatch 2-3 subagents simultaneously to cover different file trees
-- Chain: scout -> planner (one subagent scouts, another plans based on findings)
-- Reviewer: have a subagent review your analysis/plan before finalizing
-
-Do NOT use subagent for writing, editing, or executing commands.
-Subagent output is advisory only; always verify and incorporate findings into your own output.
-Do NOT treat subagent results as authoritative state or artifact sources.
+export const SUBAGENT_GUIDANCE_STRONG = `<subagent_guidance>
+You have access to the \`subagent\` tool for lightweight delegation within this step.
+
+You SHOULD delegate to subagent scouts when:
+- The task requires scanning 3+ directories or comparing implementations across modules
+- You would otherwise need 5+ sequential read/grep calls to gather context
+- A reviewer subagent can independently catch scope drift before you finalize your output
+
+Delegation saves context tokens and produces better results.
+
+Allowed patterns:
+- Parallel scout: dispatch 2-3 subagents simultaneously to cover different file trees
+- Chain: scout -> planner (one subagent scouts, another plans based on findings)
+- Reviewer: have a subagent review your analysis/plan before finalizing
+
+Do NOT use subagent for writing, editing, or executing commands.
+Subagent output is advisory only; always verify and incorporate findings into your own output.
+Do NOT treat subagent results as authoritative state or artifact sources.
 </subagent_guidance>`;
 /** Preserved for backward compatibility (alias of STANDARD). */
 export const SUBAGENT_GUIDANCE = SUBAGENT_GUIDANCE_STANDARD;

package/dist/worker/pool/run-store.js

@@ -74,7 +74,15 @@ export async function readJsonlFile(filePath) {
 }
 async function appendJsonlFile(filePath, value) {
     await mkdir(path.dirname(filePath), { recursive: true });
-    await appendFile(filePath, `${JSON.stringify(value)}\n`, "utf-8");
+    try {
+        await appendFile(filePath, `${JSON.stringify(value)}\n`, "utf-8");
+    }
+    catch (error) {
+        const detail = error instanceof Error ? error.message : String(error);
+        // Include path so callers/tests can identify which JSONL file failed
+        // (Node EISDIR messages on Windows often omit the path).
+        throw new Error(`failed to append ${filePath}: ${detail}`, { cause: error });
+    }
 }
 async function ensurePoolDirs(repoRoot) {
     await mkdir(path.join(getTaskPoolRoot(repoRoot), "runs"), { recursive: true });