niceeval 0.13.2-canary.35 → 0.13.2

package/dist/runner/feedback/human.cjs

@@ -47,8 +47,7 @@ function panelCapabilityForFeedback(io) {
 const HUMAN_FAILURE_CAP = 10;
 /** live 面板 FAILURES 分节滚动保留的条数(见 cli.md「运行中的 live 面板」)。 */
 const LIVE_FAILURES_VISIBLE = 5;
-/** 结束反馈 FAILURES 面板按失败形态聚合后,展开的组数上限;超出收进
- *  `+K more kinds — niceeval view` 尾行(见 cli.md「人看的结束反馈」)。 */
+/** 结束反馈中断言失败形态的展开上限；execution error 必须逐条保留自己的 message，不受此上限折叠。 */
 const FAILURE_GROUPS_CAP = 10;
 /** 非 TTY / `--json` 没有可依赖的终端宽度,失败单行投影用固定预算(见
  *  docs/feature/assertions/library/display.md「一条摘要怎样排版」)。 */
@@ -59,6 +58,7 @@ const NON_TTY_HEARTBEAT_IDLE_MS = 30000;
  *  这里只需要给「下一帧」留出一行余地,不需要额外的表头/尾行预留)。 */
 const DASHBOARD_ROW_RESERVE = 1;
 const HUMAN_RESULTS_EXPERIMENT_CAP = 5;
+const HUMAN_ERROR_TEXT_MAX_BYTES = 240;
 function freshHumanResult(result) {
     const score = result.scoreResult;
     const scoreState = score === undefined
@@ -145,14 +145,21 @@ function buildResultsPanelRows(items) {
     }
     return rows;
 }
-function buildPreAttemptErrorRows(results) {
+function buildPreAttemptErrorRows(results, runIdsByExperiment, panel) {
     const errored = results.filter((result) => result.verdict === "errored" && result.locator === undefined && result.error !== undefined);
     if (errored.length === 0)
         return undefined;
     const groups = new Map();
     for (const result of errored) {
         const error = result.error;
-        const failureId = error.origin.scope === "run" ? error.origin.timingNodeId : `${error.code}:${error.message}`;
+        const runId = result.experimentId === undefined
+            ? undefined
+            : runIdsByExperiment.get(result.experimentId);
+        // Run timing node IDs are local identities. The same `n1` in another Run
+        // names a different physical failure and must not absorb its message.
+        const failureId = error.origin.scope === "run"
+            ? `${runId ?? result.experimentId ?? "default"}\u0000${error.origin.timingNodeId}`
+            : `${result.experimentId ?? "default"}\u0000${error.code}:${error.message}`;
         const members = groups.get(failureId);
         if (members)
             members.push(result);
@@ -160,28 +167,34 @@ function buildPreAttemptErrorRows(results) {
             groups.set(failureId, [result]);
     }
     const rows = [];
-    for (const [failureId, members] of groups) {
+    const contentWidth = (0, panel_ts_1.panelContentWidth)(panel.width, panel.mode);
+    for (const members of groups.values()) {
         const representative = members[0];
         const error = representative.error;
-        const codeMatch = /\bERR_[A-Z0-9_]+\b/u.exec(error.message);
-        const errorCode = codeMatch?.[0] ?? error.code;
-        const focusedMessage = codeMatch === null ? error.message.slice(-600) : error.message.slice(codeMatch.index);
-        rows.push({ kind: "line", text: `${FAILURE_SYMBOL} ${representative.experimentId ?? "default"} · sandbox provisioning failed` });
-        rows.push({ kind: "line", text: `  phase: ${error.origin.scope === "run" ? "sandbox.image.build" : error.origin.phase}` });
-        rows.push({ kind: "line", text: `  affected: ${members.map((member) => member.id).join(", ")}` });
-        rows.push({ kind: "line", text: `  shared failure: ${failureId}` });
-        rows.push({ kind: "line", text: `  ${errorCode}` });
-        for (const line of focusedMessage.split(/\r?\n/u).slice(0, 6)) {
-            const text = line.trimEnd();
-            if (text.length > 0)
-                rows.push({ kind: "line", text: `  ${text}` });
-        }
-        if (errorCode === "ERR_PNPM_IGNORED_BUILDS") {
-            rows.push({ kind: "line", text: "  fix: Review these dependencies and configure pnpm allowBuilds." });
-        }
+        rows.push({
+            kind: "line",
+            text: `${FAILURE_SYMBOL} ${representative.experimentId ?? "default"} · ${preAttemptErrorTitle(error.origin)}`,
+        });
+        rows.push({
+            kind: "line",
+            text: `  affected: ${members.map((member) => `${member.experimentId ?? "default"}/${member.id}`).join(", ")}`,
+        });
+        rows.push(...labelledWrappedRows("error", boundedHumanError(error.message), contentWidth));
     }
     return { rows, slots: errored.length };
 }
+function preAttemptErrorTitle(origin) {
+    if (origin.scope === "run")
+        return "Sandbox image build failed";
+    const phase = origin.phase;
+    switch (phase) {
+        case "judge.precheck": return "Judge precheck failed";
+        case "experiment.setup": return "Experiment setup failed";
+        case "sandbox.queue": return "Sandbox queue failed";
+        case "sandbox.create": return "Sandbox provisioning failed";
+        default: return `${phaseLabel(phase)} failed`;
+    }
+}
 /** 按 `io.stderr.isTTY` 派发:构造时选一次,运行期不再切换。 */
 function createHumanRenderer(options) {
     const { io, command } = options;
@@ -305,7 +318,7 @@ function renderDurableLines(event, state, panel) {
         case "summary":
             return buildSummaryLines(event, state, panel);
         case "receipt":
-            return buildReceiptLines(event, panel);
+            return buildReceiptLines(event, state, panel);
         default: {
             // 穷尽性检查:新增 DurableFeedbackEvent 变体时这里编译期报错提醒补上对应分支。
             const exhaustive = event;
@@ -381,20 +394,9 @@ function buildFailureFactLine(failure, maxWidth) {
     const budget = Math.max(0, maxWidth - (0, text_layout_ts_1.stringWidth)(prefix));
     const info = failure.fact
         ? (0, display_ts_1.fitCompactAssertionSummary)(failure.fact, budget)
-        : buildErroredInfo(failure.phase, failure.code, failure.reason);
+        : `error: ${boundedHumanError(failure.reason)}`;
     return prefix + clipDisplayWidth(info, budget);
 }
-/**
- * errored 且没有主 Fact 摘要（真正的结构化执行错误）时的信息段:
- * `errored · <phase> · <code>`,余量够再接 `: <message>`。`phase`/`code` 用未翻译的原始字面量
- * (`LifecyclePhase` 枚举值 / `AttemptError.code`),不走 `phaseLabel()` 的人读短语 —— 这一段
- * 是给读者对照 `--json` `error` 事件、`result.json` 与 `show` 里同一个字面量用的,不是散文。
- */
-function buildErroredInfo(phase, code, reason) {
-    const parts = ["errored", ...(phase ? [phase] : []), ...(code ? [code] : [])];
-    const base = parts.join(" · ");
-    return reason ? `${base}: ${reason}` : base;
-}
 /** 按显示列硬夹紧,超宽尾部截断补 `…`。`fitCompactAssertionSummary` 的截断口径是字符数,不是
  *  显示列(见该函数注释「显示宽度的精确裁剪仍归渲染面」);这里补上那道显示宽度的硬夹紧,
  *  堵住 CJK 内容"按字符数收口仍超显示列"的缝,保证渲染行恒不超过预算。 */
@@ -423,6 +425,11 @@ function buildDiagnosticLines(event, state) {
     // 跨用例撞上同一类状况时,读者需要的是「这类事发生了 N 次」,不是 N 段几乎相同的文字;
     // 逐 code 的次数与首条 message 由结束反馈的 WARNINGS / RECOVERY 面板汇总。
     const code = event.code ?? event.key;
+    // Judge precheck already has a live lifecycle row and a terminal pre-Attempt
+    // error block. Repeating its machine code in Human output adds no fact and
+    // exposes machine vocabulary in the default profile.
+    if (code === "judge-precheck-failed")
+        return [];
     const firstForCode = state.diagnostics.find((d) => (d.code ?? d.key) === code);
     const count = state.diagnostics.find((d) => d.key === event.key)?.count ?? 1;
     if (count > 1 || firstForCode?.key !== event.key)
@@ -511,11 +518,11 @@ function buildSummaryLines(event, state, panel) {
             mode: panel.mode,
         }));
     }
-    const preAttemptErrors = buildPreAttemptErrorRows(summary.results);
+    const preAttemptErrors = buildPreAttemptErrorRows(summary.results, state.runIdsByExperiment, panel);
     if (preAttemptErrors !== undefined) {
         blocks.push((0, panel_ts_1.renderPanel)({
             title: (0, index_ts_1.t)("feedback.human.errorsHeader"),
-            meta: `${preAttemptErrors.slots} slots`,
+            meta: `${preAttemptErrors.slots} attempt${preAttemptErrors.slots === 1 ? "" : "s"} not started`,
             rows: preAttemptErrors.rows,
             width: panel.width,
             mode: panel.mode,
@@ -554,7 +561,10 @@ function buildSummaryLines(event, state, panel) {
     // `WARNINGS`(有诊断才出现,位于 FAILURES 与 NEXT 之间,见 cli.md「人看的结束反馈」):
     // 按 code 汇总本次去重后的诊断,运行中同 code 只完整打印过第一条,这里是「这类事一共
     // 发生了几次」的权威答案。
-    const warningsRows = buildWarningsPanelRows(state.diagnostics);
+    const preAttemptCodes = new Set(summary.results.flatMap((result) => result.verdict === "errored" && result.locator === undefined && result.error !== undefined
+        ? [result.error.code]
+        : []));
+    const warningsRows = buildWarningsPanelRows(state.diagnostics.filter((diagnostic) => diagnostic.severity !== "error" || !preAttemptCodes.has(diagnostic.code ?? diagnostic.key)));
     if (warningsRows) {
         blocks.push((0, panel_ts_1.renderPanel)({ title: (0, index_ts_1.t)("feedback.human.warningsHeader"), rows: warningsRows, width: panel.width, mode: panel.mode }));
     }
@@ -574,7 +584,14 @@ function buildSummaryLines(event, state, panel) {
 function buildFailuresPanelRows(failures, panel) {
     const groups = groupFailuresByShape(failures);
     const contentWidth = (0, panel_ts_1.panelContentWidth)(panel.width, panel.mode);
-    const shown = groups.slice(0, FAILURE_GROUPS_CAP);
+    let shownAssertionKinds = 0;
+    const shown = groups.filter((group) => {
+        if (group.representative.fact === undefined)
+            return true;
+        shownAssertionKinds += 1;
+        return shownAssertionKinds <= FAILURE_GROUPS_CAP;
+    });
+    const omitted = groups.length - shown.length;
     const multi = shown.filter((g) => g.size > 1);
     const countWidth = Math.max(0, ...multi.map((g) => (0, text_layout_ts_1.stringWidth)(`${FAILURE_SYMBOL} ×${g.size}`)));
     const rows = [];
@@ -586,10 +603,10 @@ function buildFailuresPanelRows(failures, panel) {
             rows.push(buildMultiFailureGroupRow(group, countWidth));
         }
     }
-    if (groups.length > FAILURE_GROUPS_CAP) {
+    if (omitted > 0) {
         rows.push({
             kind: "line",
-            text: (0, index_ts_1.t)("feedback.human.moreFailureKinds", { count: groups.length - FAILURE_GROUPS_CAP }),
+            text: (0, index_ts_1.t)("feedback.human.moreFailureKinds", { count: omitted }),
         });
     }
     return { rows, meta: (0, index_ts_1.t)("feedback.human.failuresTotalKinds", { total: failures.length, kinds: groups.length }) };
@@ -610,18 +627,14 @@ function groupFailuresByShape(failures) {
         .map(([key, g]) => ({ key, ...g }))
         .sort((a, b) => b.size - a.size);
 }
-/** 组 key 与形态摘要文本共用同一个"剥掉 received 的断言摘要"投影(有主断言摘要时,`failed`
- *  与 Fact unavailable 造成的 `errored` 都走这条);没有主 Fact 摘要的 `errored`（真正的
- *  结构化执行错误)按 `phase · code` 分组,摘要文本复用 `buildErroredInfo`(不带 message)。 */
+/** 断言按稳定形态聚合；execution error 的 key 包含 locator，确保不同 Attempt 不会因共享
+ * phase/code 而吞掉各自的 Provider message。 */
 function failureShapeOf(failure) {
     if (failure.fact) {
         const shapeText = (0, display_ts_1.compactAssertionSummary)({ ...failure.fact, received: undefined, additionalFailures: 0 });
         return { key: `fact\u0000${failure.fact.title}\u0000${failure.fact.matcher ?? ""}`, shapeText };
     }
-    return {
-        key: `errored\u0000${failure.phase ?? "?"}\u0000${failure.code ?? "?"}`,
-        shapeText: buildErroredInfo(failure.phase, failure.code, ""),
-    };
+    return { key: `errored\u0000${failure.locator}`, shapeText: "error" };
 }
 function buildMultiFailureGroupRow(group, countWidth) {
     const countToken = (0, text_layout_ts_1.padStartDisplay)(`${FAILURE_SYMBOL} ×${group.size}`, countWidth);
@@ -636,13 +649,15 @@ function buildSingleFailureGroupRows(failure, contentWidth) {
     const identityLine = `${FAILURE_SYMBOL} ${failure.locator}  ${failure.identity.evalId}  [${failure.who}]`;
     const indent = (0, text_layout_ts_1.stringWidth)(`${FAILURE_SYMBOL} `);
     const budget = Math.max(0, contentWidth - indent);
-    const info = failure.fact
-        ? (0, display_ts_1.fitCompactAssertionSummary)(failure.fact, budget)
-        : buildErroredInfo(failure.phase, failure.code, failure.reason);
-    return [
-        { kind: "line", text: identityLine },
-        { kind: "line", text: `${" ".repeat(indent)}${clipDisplayWidth(info, budget)}` },
-    ];
+    const rows = [{ kind: "line", text: identityLine }];
+    if (failure.fact) {
+        rows.push({ kind: "line", text: `${" ".repeat(indent)}${clipDisplayWidth((0, display_ts_1.fitCompactAssertionSummary)(failure.fact, budget), budget)}` });
+    }
+    else {
+        rows.push(...labelledWrappedRows("error", boundedHumanError(failure.reason), contentWidth));
+    }
+    rows.push(...labelledWrappedRows("details", `niceeval show ${failure.locator}`, contentWidth));
+    return rows;
 }
 /** `WARNINGS` 面板:`state.diagnostics` 已经按去重 key 折叠,这里再按对外稳定词法 `code` 二次
  *  聚合。info 由自己的完成面板消费；没有 warning/error 时返回 undefined。 */
@@ -672,38 +687,98 @@ function buildWarningsPanelRows(diagnostics) {
     }));
 }
 function buildRecoveryPanelRows(diagnostics) {
-    let slots = 0;
     let locks = 0;
     for (const d of diagnostics) {
         if (d.severity !== "info" || d.code !== types_ts_1.COORDINATION_RECOVERED_CODE)
             continue;
-        if (d.data?.resource === "concurrency-slot")
-            slots += d.count;
         if (d.data?.resource === "case-lock")
             locks += d.count;
     }
-    if (slots === 0 && locks === 0)
+    if (locks === 0)
         return undefined;
-    const summary = (0, index_ts_1.t)("feedback.human.coordinationRecoveredSummary", {
-        slots: pluralUnit(slots, "feedback.human.unit.concurrencySlot", "feedback.human.unit.concurrencySlots"),
-        locks: pluralUnit(locks, "feedback.human.unit.caseLock", "feedback.human.unit.caseLocks"),
-    });
+    const summary = pluralUnit(locks, "feedback.human.unit.caseLock", "feedback.human.unit.caseLocks");
     return [{ kind: "line", text: `i ${types_ts_1.COORDINATION_RECOVERED_CODE}  ${summary}` }];
 }
 function warningCodeLabel(code, count) {
     return count > 1 ? `${code} ×${count}` : code;
 }
 /** `NEXT` 面板只由 receipt 的真实 Run ID 给出 Record 读取命令。 */
-function buildReceiptLines(event, panel) {
+function buildReceiptLines(event, state, panel) {
     const rows = [];
+    const published = new Set(event.receipt.runIds);
+    const emitted = new Set();
+    for (const [experimentId, runId] of state.runIdsByExperiment) {
+        if (!published.has(runId))
+            continue;
+        rows.push({ kind: "line", text: experimentId });
+        rows.push(...labelledWrappedRows("details", `niceeval show --run ${runId}`, (0, panel_ts_1.panelContentWidth)(panel.width, panel.mode)));
+        emitted.add(runId);
+    }
     for (const runId of event.receipt.runIds) {
-        rows.push({ kind: "line", text: `niceeval show --run ${runId}` });
+        if (!emitted.has(runId))
+            rows.push({ kind: "line", text: `details: niceeval show --run ${runId}` });
     }
     if (rows.length === 0) {
         rows.push({ kind: "line", text: "No published Record Runs are available for this invocation." });
     }
     return (0, panel_ts_1.renderPanel)({ title: (0, index_ts_1.t)("feedback.human.nextHeader"), rows, width: panel.width, mode: panel.mode });
 }
+function boundedHumanError(message) {
+    const safe = (0, display_ts_1.stripControl)(message).replace(/\s+/gu, " ").trim();
+    const signalIndexes = [
+        safe.search(/\b[1-5]\d{2}\s+[A-Z][A-Za-z -]{2,40}(?:\s+[—·:-]\s+[A-Z][A-Z0-9_-]+)?/u),
+        safe.search(/\b[A-Z][A-Z0-9]+(?:_[A-Z0-9]+){1,}\b/u),
+    ].filter((index) => index >= 0);
+    // Wrappers often prepend a long command or transport path before the
+    // Provider's safe status/code. Focus the Human budget at that first stable
+    // signal; the full structured error remains available through details.
+    const focused = signalIndexes.length === 0 ? safe : safe.slice(Math.min(...signalIndexes));
+    if (utf8Bytes(focused) <= HUMAN_ERROR_TEXT_MAX_BYTES)
+        return focused;
+    const separator = " … ";
+    const available = HUMAN_ERROR_TEXT_MAX_BYTES - utf8Bytes(separator);
+    const head = Math.ceil(available / 2);
+    const tail = available - head;
+    return `${utf8Prefix(focused, head)}${separator}${utf8Suffix(focused, tail)}`;
+}
+function utf8Bytes(value) {
+    return new TextEncoder().encode(value).byteLength;
+}
+function utf8Prefix(value, maximumBytes) {
+    let bytes = 0;
+    let result = "";
+    for (const character of value) {
+        const nextBytes = utf8Bytes(character);
+        if (bytes + nextBytes > maximumBytes)
+            break;
+        result += character;
+        bytes += nextBytes;
+    }
+    return result;
+}
+function utf8Suffix(value, maximumBytes) {
+    let bytes = 0;
+    const result = [];
+    const characters = Array.from(value);
+    for (let index = characters.length - 1; index >= 0; index -= 1) {
+        const character = characters[index];
+        const nextBytes = utf8Bytes(character);
+        if (bytes + nextBytes > maximumBytes)
+            break;
+        result.push(character);
+        bytes += nextBytes;
+    }
+    return result.reverse().join("");
+}
+function labelledWrappedRows(label, text, contentWidth) {
+    const prefix = `  ${label}: `;
+    const continuation = " ".repeat((0, text_layout_ts_1.stringWidth)(prefix));
+    const lines = (0, text_layout_ts_1.wrapDisplay)(text, Math.max(4, contentWidth - (0, text_layout_ts_1.stringWidth)(prefix)));
+    return lines.map((line, index) => ({
+        kind: "line",
+        text: `${index === 0 ? prefix : continuation}${line}`,
+    }));
+}
 /** estimated cost 展示:金额前加 `~`,明确是价目表估算(observed 只留在 result.usage.costUSD)。 */
 function formatEstimatedCost(value) {
     const cost = (0, format_ts_1.formatCost)(value);