niceeval 0.9.1 → 0.10.2

package/dist/report/components/attempt-detail/compute.js

@@ -0,0 +1,297 @@
+// Attempt 详情组件族的计算函数(docs/feature/reports/library/attempt-detail.md)。每个
+// `attempt*Data(evidence)` 都是纯同步派生:evidence 已经由 loadAttemptEvidence 一次性
+// 装配好全部证据,这里只做适合展示与序列化的取舍,不读文件、不 fetch、不重复调用
+// attempt.events() / trace() / diff()。
+//
+// 与 compute.ts(Scope → *Data)不同,这一族的输入恒为单个 AttemptEvidence,函数签名
+// 因此不是 async——没有 IO 就没有理由返回 Promise。
+import { attemptCostUSD } from "../../model/metrics.js";
+import { failureSummaryOf } from "../entity-lists/compute.js";
+// ───────────────────────── AttemptSummary(恒非空) ─────────────────────────
+export function attemptSummaryData(evidence) {
+    const { result } = evidence;
+    return {
+        locator: evidence.locator,
+        identity: evidence.identity,
+        verdict: result.verdict,
+        startedAt: result.startedAt,
+        durationMs: result.durationMs,
+        costUSD: attemptCostUSD(result),
+        capabilities: evidence.capabilities,
+    };
+}
+// ───────────────────────── AttemptError ─────────────────────────
+export function attemptErrorData(evidence) {
+    const err = evidence.result.error;
+    return err ?? null;
+}
+// ───────────────────────── AttemptAssertions ─────────────────────────
+export function attemptAssertionsData(evidence) {
+    const assertions = evidence.result.assertions;
+    if (!assertions || assertions.length === 0)
+        return null;
+    const attention = assertions.filter((a) => a.outcome !== "passed");
+    const passed = assertions.filter((a) => a.outcome === "passed");
+    const groups = new Map();
+    for (const a of passed) {
+        const key = a.groupPath?.join(" > ") ?? "";
+        const list = groups.get(key);
+        if (list)
+            list.push(a);
+        else
+            groups.set(key, [a]);
+    }
+    return { attention, passedGroups: [...groups.entries()].map(([group, items]) => ({ group, items })) };
+}
+// ───────────────────────── AttemptSource ─────────────────────────
+export function attemptSourceData(evidence) {
+    if (!evidence.capabilities.source || evidence.evalSource === null)
+        return null;
+    const { sourcePath, lines, unmapped, summary } = evidence.evalSource;
+    const projectedLines = lines.map((line) => ({
+        ...line,
+        turns: line.sends.map((send) => ({
+            label: send.label,
+            status: send.status,
+            ...(send.durationMs === undefined ? {} : { durationMs: send.durationMs }),
+            sentText: "",
+            replies: [],
+        })),
+    }));
+    const usedTurns = new Map();
+    const unlocatedTurns = [];
+    const conversation = attemptConversationData(evidence);
+    for (const [roundIndex, round] of (conversation?.rounds ?? []).entries()) {
+        const status = round.replies.some((reply) => reply.kind === "error" ||
+            ((reply.kind === "tool" || reply.kind === "subagent") && reply.status === "failed"))
+            ? "failed"
+            : round.replies.some((reply) => reply.kind === "input")
+                ? "waiting"
+                : "completed";
+        const fallback = {
+            label: `t${roundIndex + 1}`,
+            status,
+            sentText: round.sentText,
+            replies: round.replies,
+        };
+        const loc = round.loc;
+        if (!loc || loc.file !== sourcePath || loc.line < 1 || loc.line > projectedLines.length) {
+            unlocatedTurns.push(fallback);
+            continue;
+        }
+        const line = projectedLines[loc.line - 1];
+        const turnIndex = usedTurns.get(loc.line) ?? 0;
+        usedTurns.set(loc.line, turnIndex + 1);
+        const annotated = line.turns[turnIndex];
+        if (annotated) {
+            annotated.sentText = round.sentText;
+            annotated.replies = round.replies;
+        }
+        else {
+            line.turns.push(fallback);
+        }
+    }
+    return { locator: evidence.locator, sourcePath, lines: projectedLines, unmapped, unlocatedTurns, summary };
+}
+// ───────────────────────── AttemptFixPrompt ─────────────────────────
+/** 单条 attempt 版的批量修复 prompt(与 CopyFixPrompt 的多条版本同一份步骤文案)。 */
+export function attemptFixPromptData(evidence) {
+    const { result, identity } = evidence;
+    if (result.verdict !== "failed" && result.verdict !== "errored")
+        return null;
+    const { summary, more } = failureSummaryOf(result);
+    if (summary === null)
+        return null;
+    const reason = more > 0 ? `${summary} (+${more} more failures)` : summary;
+    const prompt = [
+        "Fix the failing eval from this niceeval run.",
+        "",
+        "## Failure",
+        `eval "${identity.evalId}" [experiment ${identity.experimentId}] — ${result.verdict}`,
+        `  reason: ${reason}`,
+        `  inspect: niceeval show ${evidence.locator}`,
+        "",
+        "## Steps",
+        "1. niceeval is NOT in your training data. Read the relevant guide in `node_modules/niceeval/docs-site/` (English at the top level, Chinese under `zh/`) before changing anything.",
+        "2. Run the inspect command above with `--source`, `--execution`, `--timing`, and `--diff` to see the assertions, transcript, timing, and workspace diff.",
+        "3. Decide which side the defect is on: the program under test, or the eval itself (over-tight assertion, wrong fixture, missing setup). Fix that side; do not weaken assertions just to turn the run green.",
+        `4. Re-run: \`npx niceeval exp ${identity.experimentId} ${identity.evalId}\`. Already-passing evals are skipped by the fingerprint cache; pass \`--force\` to re-run everything.`,
+        "5. Run `npx niceeval show` and confirm this failure is gone.",
+    ].join("\n");
+    return { prompt };
+}
+// ───────────────────────── AttemptTimeline ─────────────────────────
+/** 收尾段的阶段名(见 docs/feature/results/architecture.md);两面渲染都把这些单列在主链之后,不计入主链总耗时。 */
+export const TIMELINE_CLOSING_PHASES = new Set([
+    "eval.teardown",
+    "agent.teardown",
+    "sandbox.teardown",
+    "sandbox.suspend",
+    "sandbox.stop",
+]);
+export function attemptTimelineData(evidence) {
+    const phases = evidence.result.phases;
+    if (!phases || phases.length === 0)
+        return null;
+    return { locator: evidence.locator, phases, trace: evidence.trace };
+}
+// ───────────────────────── AttemptConversation ─────────────────────────
+/**
+ * 标准事件流按 `loc` 分轮(docs/feature/reports/library/attempt-detail.md「Attempt 详情组件」):
+ * 带 loc 的 user 消息开一轮;无 loc 的 user 消息不开新轮——与当前轮 sent 同文本的回显直接
+ * 吃掉,其它(stop-hook 反馈、skill 注入等轮内注入)作为回复条目留在当前轮。流首出现无 loc
+ * 的 user 消息(没有当前轮可归入)时退化开一条 loc 缺省的兜底轮,不丢弃。未识别的事件类型
+ * 包成 `raw` 条目原样呈现,不吞没其余事件——StreamEvent 是随 artifact 版本演进的开放词表,
+ * 这份纯函数不能假设自己认识每一种将来会出现的 type。
+ */
+export function attemptConversationData(evidence) {
+    const events = evidence.events;
+    if (!events || events.length === 0)
+        return null;
+    const rounds = [];
+    const toolByCallId = new Map();
+    const subagentByCallId = new Map();
+    let current = null;
+    for (const ev of events) {
+        if (ev.type === "message" && ev.role === "user") {
+            if (!ev.loc && current) {
+                if (current.replies.length === 0 && (ev.text || "") === current.sentText)
+                    continue;
+                current.replies.push({ kind: "user", text: ev.text || "" });
+                continue;
+            }
+            current = { loc: ev.loc, sentText: ev.text || "", replies: [] };
+            rounds.push(current);
+            continue;
+        }
+        if (!current) {
+            // 流首没有开轮的用户消息就先来了其它事件(旧 artifact 的边界情况):开一条无 loc 兜底轮。
+            current = { sentText: "", replies: [] };
+            rounds.push(current);
+        }
+        current.replies.push(...conversationReplyOf(ev, toolByCallId, subagentByCallId));
+    }
+    return { locator: evidence.locator, rounds };
+}
+/** 单条事件 → 0 或 1 条回复条目;action.result/subagent.completed 只更新已有条目,不新增。 */
+function conversationReplyOf(ev, toolByCallId, subagentByCallId) {
+    switch (ev.type) {
+        case "message":
+            // role === "user" 已在主循环处理(开轮 / 回显吃掉 / 轮内注入),这里只剩 assistant。
+            return [{ kind: "assistant", text: ev.text }];
+        case "thinking":
+            return [{ kind: "thinking", text: ev.text }];
+        case "error":
+            return [{ kind: "error", text: ev.message }];
+        case "skill.loaded":
+            return [{ kind: "skill", skill: ev.skill }];
+        case "input.requested":
+            return [{ kind: "input", request: ev.request }];
+        case "compaction":
+            return [{ kind: "compaction", reason: ev.reason }];
+        case "action.called": {
+            const reply = {
+                kind: "tool",
+                callId: ev.callId,
+                name: ev.name,
+                tool: ev.tool,
+                input: ev.input,
+            };
+            toolByCallId.set(ev.callId, reply);
+            return [reply];
+        }
+        case "action.result": {
+            const tool = toolByCallId.get(ev.callId);
+            if (tool) {
+                tool.output = ev.output;
+                tool.status = ev.status;
+            }
+            return [];
+        }
+        case "subagent.called": {
+            const reply = {
+                kind: "subagent",
+                callId: ev.callId,
+                name: ev.name,
+                remoteUrl: ev.remoteUrl,
+            };
+            subagentByCallId.set(ev.callId, reply);
+            return [reply];
+        }
+        case "subagent.completed": {
+            const subagent = subagentByCallId.get(ev.callId);
+            if (subagent) {
+                subagent.output = ev.output;
+                subagent.status = ev.status;
+            }
+            return [];
+        }
+        default: {
+            // 穷尽性检查在这里刻意不做:StreamEvent 是随 artifact 版本演进的开放词表,未识别的
+            // 事件(将来的新 type,或第三方 harness 的自定义变体)包成 raw 原样呈现,不静默丢弃、
+            // 也不因为一个不认识的条目让整个装配失败。
+            return [{ kind: "raw", raw: ev }];
+        }
+    }
+}
+// ───────────────────────── AttemptDiagnostics ─────────────────────────
+export function attemptDiagnosticsData(evidence) {
+    const diagnostics = evidence.result.diagnostics;
+    if (!diagnostics || diagnostics.length === 0)
+        return null;
+    const groups = new Map();
+    for (const d of diagnostics) {
+        const list = groups.get(d.phase);
+        if (list)
+            list.push(d);
+        else
+            groups.set(d.phase, [d]);
+    }
+    return { groups: [...groups.entries()].map(([phase, items]) => ({ phase, items })) };
+}
+// ───────────────────────── AttemptUsage ─────────────────────────
+export function attemptUsageData(evidence) {
+    const usage = evidence.result.usage;
+    if (!usage)
+        return null;
+    return { usage, costUSD: attemptCostUSD(evidence.result) };
+}
+// ───────────────────────── AttemptTrace ─────────────────────────
+export function attemptTraceData(evidence) {
+    const spans = evidence.trace;
+    if (!spans || spans.length === 0)
+        return null;
+    return { locator: evidence.locator, spans };
+}
+// ───────────────────────── AttemptDiff ─────────────────────────
+/** 有界行 diff(公共前后缀修剪):对单区域编辑精确,复杂编辑给出上界近似;与 `niceeval show --diff` 同一算法。 */
+function lineDelta(before, after) {
+    const a = before === undefined ? [] : before.split("\n");
+    const b = after === undefined ? [] : after.split("\n");
+    let prefix = 0;
+    while (prefix < a.length && prefix < b.length && a[prefix] === b[prefix])
+        prefix++;
+    let suffix = 0;
+    while (suffix < a.length - prefix && suffix < b.length - prefix && a[a.length - 1 - suffix] === b[b.length - 1 - suffix])
+        suffix++;
+    return { added: b.length - prefix - suffix, deleted: a.length - prefix - suffix };
+}
+export function attemptDiffData(evidence) {
+    if (!evidence.capabilities.diff || evidence.diff === null)
+        return null;
+    const diff = evidence.diff;
+    const files = [];
+    for (const [path, summary] of Object.entries(diff.files).sort(([a], [b]) => a.localeCompare(b))) {
+        if (summary.net === "none")
+            continue;
+        const windows = diff.windows.filter((w) => w.changes[path] !== undefined).map((w) => w.window);
+        if (summary.binary) {
+            files.push({ path, net: summary.net, lines: { added: 0, deleted: 0 }, binary: true, windows });
+            continue;
+        }
+        const before = summary.net === "added" ? undefined : diff.windows.find((w) => w.changes[path]?.before !== undefined)?.changes[path]?.before;
+        const after = summary.net === "deleted" ? undefined : diff.get(path);
+        files.push({ path, net: summary.net, lines: lineDelta(before, after), windows });
+    }
+    return files.length > 0 ? { locator: evidence.locator, files } : null;
+}

package/dist/report/components/attempt-detail/faces.d.ts

@@ -0,0 +1,14 @@
+import type { AttemptAssertionsData, AttemptConversationData, AttemptDiagnosticsData, AttemptDiffData, AttemptErrorData, AttemptFixPromptData, AttemptSourceData, AttemptSummaryData, AttemptTimelineData, AttemptTraceData, AttemptUsageData } from "../../model/types.ts";
+import type { TextContext } from "../../definition/tree.ts";
+export declare function attemptSummaryText(data: AttemptSummaryData, ctx: TextContext): string;
+export declare function attemptErrorText(data: AttemptErrorData | null, _ctx: TextContext): string;
+export declare function attemptAssertionsText(data: AttemptAssertionsData | null, _ctx: TextContext): string;
+export declare function attemptSourceText(data: AttemptSourceData | null, ctx: TextContext): string;
+/** text 面零输出:终端已有 attemptSummaryText 里的 locator,直接跑 `niceeval show @<locator>` 即可;这里不重复整段 prompt。 */
+export declare function attemptFixPromptText(_data: AttemptFixPromptData | null, _ctx: TextContext): string;
+export declare function attemptTimelineText(data: AttemptTimelineData | null, ctx: TextContext): string;
+export declare function attemptConversationText(data: AttemptConversationData | null, ctx: TextContext): string;
+export declare function attemptDiagnosticsText(data: AttemptDiagnosticsData | null, _ctx: TextContext): string;
+export declare function attemptUsageText(data: AttemptUsageData | null, _ctx: TextContext): string;
+export declare function attemptTraceText(data: AttemptTraceData | null, ctx: TextContext): string;
+export declare function attemptDiffText(data: AttemptDiffData | null, ctx: TextContext): string;

package/dist/report/components/attempt-detail/faces.js

@@ -0,0 +1,239 @@
+// Attempt 详情组件族的 text 面(docs/feature/reports/library/attempt-detail.md「在 show 与
+// view 怎样渲染」):与 web 面共享同一次 resolve 产出的 data 事实(verdict、计数、能力位、
+// 引用),允许把大块内容折成摘要 + 专用证据命令,但不得改变判定、计数或引用。
+// 零 react、零 IO、纯同步——text 宿主不需要 react-dom 的那一半。
+import { localeText } from "../../model/locale.js";
+import { formatDurationMs, formatMetricValue, formatReportDateTime, formatUSD, verdictMark } from "../../model/format.js";
+import { TIMELINE_CLOSING_PHASES } from "./compute.js";
+import { summaryText } from "../../../scoring/display.js";
+/**
+ * 证据切面下钻命令:`ctx.attemptCommand` 由宿主按当前 report 是否有 attempt-input page
+ * 注入(见 report.ts DEFAULT_ATTEMPT_COMMAND);没有时不生成假命令,调用方直接省略这一段
+ * (与 traceWaterfallText 同一套退化规则,architecture.md「Attempt 详情是一张参数化 page」)。
+ */
+function evidenceCommand(ctx, locator, flag) {
+    return ctx.attemptCommand ? `${ctx.attemptCommand(locator)} ${flag}` : undefined;
+}
+// ───────────────────────── AttemptSummary ─────────────────────────
+export function attemptSummaryText(data, ctx) {
+    const locale = ctx.locale;
+    const parts = [
+        data.locator,
+        data.identity.evalId,
+        data.identity.experimentId,
+        localeText(locale, "attemptSummary.attempt", { n: data.identity.attempt + 1 }),
+        `${verdictMark(data.verdict)} ${localeText(locale, `verdict.${data.verdict}`)}`,
+    ];
+    if (data.startedAt !== undefined)
+        parts.push(formatReportDateTime(data.startedAt, locale));
+    parts.push(formatDurationMs(data.durationMs));
+    if (data.costUSD !== null)
+        parts.push(formatUSD(data.costUSD));
+    return parts.join(" · ");
+}
+// ───────────────────────── AttemptError ─────────────────────────
+export function attemptErrorText(data, _ctx) {
+    if (data === null)
+        return "";
+    // message/cause 折单行加上限(同 assertionLine 的 summaryText 规则);stack 是唯一没有替代
+    // 查看入口的自由文本,原样保留多行,不折。
+    const lines = [`error: ${data.code}`, `  phase: ${data.phase}`, `  message: ${summaryText(data.message)}`];
+    if (data.cause) {
+        const causeMessage = summaryText(data.cause.message);
+        lines.push(`  cause: ${data.cause.name ? `${data.cause.name} · ${causeMessage}` : causeMessage}`);
+    }
+    const stack = data.stack?.replace(/\n+$/, "");
+    return stack ? `${lines.join("\n")}\n\n${stack}` : lines.join("\n");
+}
+// ───────────────────────── AttemptAssertions ─────────────────────────
+/** `loc` 是断言在 eval 源码里的调用点,独立于整份源码是否被捕获(`AttemptSource` 的能力位);
+ *  失败断言只要带 loc 就给可复制定位的源码锚,不必等 AttemptSource 可用才显示。 */
+function locAnchor(loc) {
+    return `${loc.file}:${loc.line}${loc.column ? `:${loc.column}` : ""}`;
+}
+function assertionLine(a) {
+    const group = a.groupPath && a.groupPath.length > 0 ? `${a.groupPath.join(" > ")} · ` : "";
+    if (a.outcome === "unavailable")
+        return `◌ unavailable · ${group}${a.name} — ${a.reason}`;
+    const mark = a.outcome === "passed" ? "✓" : "✗";
+    const detail = a.detail && a.detail !== a.name ? `: ${a.detail}` : "";
+    // expected/received 折单行 + 加字符上限(与 scoring 摘要面同一条 summaryText 规则):
+    // received 常常就是被检查的整份文本(如 includes() 对全文匹配),不收口会把整份源码/工具
+    // 输出灌进这一行;完整值仍在 web 面(可展开的完整结构化细节)与 --source 里原样可查。
+    const evidence = [
+        a.expected !== undefined ? `expected: ${summaryText(a.expected)}` : undefined,
+        a.received !== undefined ? `received: ${summaryText(a.received)}` : undefined,
+        a.outcome === "failed" && a.loc ? `source: ${locAnchor(a.loc)}` : undefined,
+    ].filter((part) => part !== undefined);
+    const evidenceSuffix = evidence.length > 0 ? ` · ${evidence.join(" · ")}` : "";
+    return `${mark} ${a.severity} · ${group}${a.name}${detail}${evidenceSuffix}`;
+}
+export function attemptAssertionsText(data, _ctx) {
+    if (data === null)
+        return "";
+    const lines = data.attention.map(assertionLine);
+    for (const { group, items } of data.passedGroups) {
+        lines.push(`✓ passed · ${group || "(ungrouped)"} · ${items.length}`);
+    }
+    return lines.join("\n");
+}
+// ───────────────────────── AttemptSource ─────────────────────────
+export function attemptSourceText(data, ctx) {
+    if (data === null)
+        return "";
+    const command = evidenceCommand(ctx, data.locator, "--source");
+    const headerParts = [`${data.sourcePath} · ${data.summary.annotatedLines}/${data.summary.totalLines} lines annotated`];
+    if (command)
+        headerParts.push(command);
+    const hasConversation = data.unlocatedTurns.length > 0 || data.lines.some((line) => line.turns.length > 0);
+    const executionCommand = hasConversation ? evidenceCommand(ctx, data.locator, "--execution") : null;
+    if (executionCommand)
+        headerParts.push(executionCommand);
+    // 源码锚由 assertionLine 自己按 a.loc 拼(与 AttemptAssertions 共用同一份逻辑);这里只负责
+    // 挑出非 passed 的条目,不重复算锚点。
+    const failed = data.lines.flatMap((line) => line.assertions.filter((a) => a.outcome !== "passed"));
+    const lines = failed.map((a) => `  ${assertionLine(a)}`);
+    return [headerParts.join(" · "), ...lines].join("\n");
+}
+// ───────────────────────── AttemptFixPrompt ─────────────────────────
+/** text 面零输出:终端已有 attemptSummaryText 里的 locator,直接跑 `niceeval show @<locator>` 即可;这里不重复整段 prompt。 */
+export function attemptFixPromptText(_data, _ctx) {
+    return "";
+}
+// ───────────────────────── AttemptTimeline ─────────────────────────
+function fmtMs(ms) {
+    if (ms >= 60_000)
+        return `${Math.floor(ms / 60_000)}m ${Math.round((ms % 60_000) / 1000)}s`;
+    if (ms >= 1000)
+        return `${(ms / 1000).toFixed(1)}s`;
+    return `${ms}ms`;
+}
+function countTimingNodes(nodes) {
+    return nodes.reduce((sum, n) => sum + 1 + countTimingNodes(n.children ?? []), 0);
+}
+export function attemptTimelineText(data, ctx) {
+    if (data === null)
+        return "";
+    const main = data.phases.filter((p) => !TIMELINE_CLOSING_PHASES.has(p.name));
+    const closing = data.phases.filter((p) => TIMELINE_CLOSING_PHASES.has(p.name));
+    const total = main.reduce((sum, p) => sum + p.durationMs, 0);
+    const anyFailed = data.phases.some((p) => p.failed);
+    const command = evidenceCommand(ctx, data.locator, "--timing");
+    const head = [`timing: ${fmtMs(total)}${anyFailed ? " ✗" : ""}`];
+    if (command)
+        head.push(command);
+    const lines = [head.join(" · ")];
+    for (const p of main) {
+        const kids = p.children ?? [];
+        const suffix = kids.length > 0 ? ` (${countTimingNodes(kids)} children collapsed)` : "";
+        lines.push(`  ${p.failed ? "✗" : "·"} ${p.name} ${fmtMs(p.durationMs)}${suffix}`);
+    }
+    if (closing.length > 0) {
+        lines.push("  teardown:");
+        for (const p of closing)
+            lines.push(`    ${p.failed ? "✗" : "·"} ${p.name} ${fmtMs(p.durationMs)}`);
+    }
+    return lines.join("\n");
+}
+// ───────────────────────── AttemptConversation ─────────────────────────
+// 每条回复的自由文本(消息正文、发起请求的 prompt、未识别事件的原始 payload)都可能是任意大小的
+// 块——system prompt、完整工具输出等——同一条 summaryText 规则折单行加上限;完整内容仍在
+// --execution 里原样可查。
+function replySummary(reply) {
+    switch (reply.kind) {
+        case "assistant":
+            return `assistant: ${summaryText(reply.text)}`;
+        case "user":
+            return `user: ${summaryText(reply.text)}`;
+        case "thinking":
+            return `thinking: ${summaryText(reply.text)}`;
+        case "error":
+            return `error: ${summaryText(reply.text)}`;
+        case "tool":
+            return `tool ${reply.name}${reply.status ? ` (${reply.status})` : ""}`;
+        case "skill":
+            return `skill loaded: ${reply.skill}`;
+        case "subagent":
+            return `subagent ${reply.name}${reply.status ? ` (${reply.status})` : ""}`;
+        case "input":
+            return `input requested${reply.request.prompt ? `: ${summaryText(reply.request.prompt)}` : ""}`;
+        case "compaction":
+            return `compaction${reply.reason ? `: ${summaryText(reply.reason)}` : ""}`;
+        case "raw":
+            return `unrecognized event: ${summaryText(JSON.stringify(reply.raw))}`;
+    }
+}
+export function attemptConversationText(data, ctx) {
+    if (data === null)
+        return "";
+    const command = evidenceCommand(ctx, data.locator, "--execution");
+    const head = [`conversation: ${data.rounds.length} round${data.rounds.length === 1 ? "" : "s"}`];
+    if (command)
+        head.push(command);
+    const lines = [head.join(" · ")];
+    data.rounds.forEach((round, i) => {
+        lines.push(`  round ${i + 1}${round.sentText ? `: ${summaryText(round.sentText)}` : ""}`);
+        for (const reply of round.replies)
+            lines.push(`    ${replySummary(reply)}`);
+    });
+    return lines.join("\n");
+}
+// ───────────────────────── AttemptDiagnostics ─────────────────────────
+export function attemptDiagnosticsText(data, _ctx) {
+    if (data === null)
+        return "";
+    const lines = [];
+    for (const { phase, items } of data.groups) {
+        lines.push(`${phase}:`);
+        for (const d of items) {
+            const count = d.count && d.count > 1 ? ` (${d.count} occurrences)` : "";
+            lines.push(`  ${d.level} · ${d.code} — ${summaryText(d.message)}${count}`);
+        }
+    }
+    return lines.join("\n");
+}
+// ───────────────────────── AttemptUsage ─────────────────────────
+export function attemptUsageText(data, _ctx) {
+    if (data === null)
+        return "";
+    const { usage } = data;
+    const total = usage.inputTokens + usage.outputTokens;
+    const parts = [`usage: ${formatMetricValue(total)} tokens (${formatMetricValue(usage.inputTokens)} in / ${formatMetricValue(usage.outputTokens)} out)`];
+    if (usage.cacheReadTokens !== undefined)
+        parts.push(`cache read ${formatMetricValue(usage.cacheReadTokens)}`);
+    if (usage.cacheWriteTokens !== undefined)
+        parts.push(`cache write ${formatMetricValue(usage.cacheWriteTokens)}`);
+    if (usage.requests !== undefined)
+        parts.push(`${usage.requests} requests`);
+    if (data.costUSD !== null)
+        parts.push(formatUSD(data.costUSD));
+    return parts.join(" · ");
+}
+// ───────────────────────── AttemptTrace ─────────────────────────
+export function attemptTraceText(data, ctx) {
+    if (data === null)
+        return "";
+    const command = evidenceCommand(ctx, data.locator, "--timing");
+    const head = [`trace: ${data.spans.length} span${data.spans.length === 1 ? "" : "s"}`];
+    if (command)
+        head.push(command);
+    return head.join(" · ");
+}
+// ───────────────────────── AttemptDiff ─────────────────────────
+function netLetter(net) {
+    return net === "added" ? "A" : net === "deleted" ? "D" : "M";
+}
+export function attemptDiffText(data, ctx) {
+    if (data === null)
+        return "";
+    const command = evidenceCommand(ctx, data.locator, "--diff");
+    const head = [`changes: ${data.files.length} file${data.files.length === 1 ? "" : "s"} changed by agent`];
+    if (command)
+        head.push(command);
+    const lines = [head.join(" · ")];
+    for (const f of data.files) {
+        const delta = f.binary ? "binary" : `+${f.lines.added}/-${f.lines.deleted}`;
+        lines.push(`  ${netLetter(f.net)} ${f.path} (${delta})`);
+    }
+    return lines.join("\n");
+}

package/dist/report/components/attempt-detail/index.d.ts

@@ -0,0 +1,38 @@
+import { type ReportComponent } from "../../definition/tree.ts";
+import type { AttemptEvidence } from "../../../results/attempt-evidence.ts";
+import type { AttemptAssertionsData, AttemptConversationData, AttemptDiagnosticsData, AttemptDiffData, AttemptFixPromptData, AttemptSourceData, AttemptSummaryData, AttemptTimelineData, AttemptTraceData, AttemptUsageData } from "../../model/types.ts";
+export type AttemptSectionProps<Data> = {
+    input?: AttemptEvidence;
+    data?: never;
+    className?: string;
+} | {
+    data: Data;
+    input?: never;
+    className?: string;
+};
+export declare function validateSummaryData(data: unknown): string | null;
+export declare const AttemptSummary: ReportComponent<AttemptSectionProps<AttemptSummaryData>>;
+export declare function validateErrorData(data: unknown): string | null;
+export declare const AttemptError: ReportComponent<AttemptSectionProps<import("../../../types.ts").AttemptError>>;
+export declare function validateAssertionsData(data: unknown): string | null;
+export declare const AttemptAssertions: ReportComponent<AttemptSectionProps<AttemptAssertionsData>>;
+export declare function validateSourceData(data: unknown): string | null;
+export declare const AttemptSource: ReportComponent<AttemptSectionProps<AttemptSourceData>>;
+export declare function validateFixPromptData(data: unknown): string | null;
+export declare const AttemptFixPrompt: ReportComponent<AttemptSectionProps<AttemptFixPromptData>>;
+export declare function validateTimelineData(data: unknown): string | null;
+export declare const AttemptTimeline: ReportComponent<AttemptSectionProps<AttemptTimelineData>>;
+export declare function validateConversationData(data: unknown): string | null;
+export declare const AttemptConversation: ReportComponent<AttemptSectionProps<AttemptConversationData>>;
+export declare function validateDiagnosticsData(data: unknown): string | null;
+export declare const AttemptDiagnostics: ReportComponent<AttemptSectionProps<AttemptDiagnosticsData>>;
+export declare function validateUsageData(data: unknown): string | null;
+export declare const AttemptUsage: ReportComponent<AttemptSectionProps<AttemptUsageData>>;
+export declare function validateTraceData(data: unknown): string | null;
+export declare const AttemptTrace: ReportComponent<AttemptSectionProps<AttemptTraceData>>;
+export declare function validateDiffData(data: unknown): string | null;
+export declare const AttemptDiff: ReportComponent<AttemptSectionProps<AttemptDiffData>>;
+/** source / assertions fallback:有 source 放 AttemptSource,否则放 AttemptAssertions。 */
+export declare const AttemptAssessment: ReportComponent<Record<string, never>>;
+/** 内建排列顺序;有 source 时回复已按 loc 展开在 AttemptSource 行内，不再重复一份 round 卡。 */
+export declare const AttemptDetail: ReportComponent<Record<string, never>>;