fix(workflows): stop safeStringify node budget from silently truncating artifacts

[onychen]

Problem

Fixes #558. The four safeStringify callers in extensions/workflows/artifacts.ts that write workflow.json, transcripts.json, and result.json pass only maxBytes and inherit serialization.ts's DEFAULT_MAX_NODES = 20_000. That node budget is a global counter over the whole value tree, so it is reached long before the 1–2 MiB byte budget. When it is, toSerializable inserts a "[truncated: node limit]" marker and breaks — producing a structurally valid file that reads back as a clean completed run while trailing data is silently gone. No reader consumes the marker.

Reachable on default config: 12 agents each with ~202 transcript entries (the upstream ceiling TRANSCRIPT_MAX_ENTRIES = 200) writes only 11 of 12 transcripts, at 28% of the 2 MiB byte cap; the 12th agent hydrates as an empty transcript in /workflows with no notice. Every other safeStringify call in the repo already passes an explicit maxNodes (model.ts:604, progress-projection.ts:65, sandbox.ts:164,378), and persistWorkflowAgentResult uses fail-closed encodeCompleteJson — these four sites were simply missed. journal.ts:281-289 already documents this exact hazard for a sibling artifact.

Value

Terminal workflow artifacts stop losing data silently — the failure mode AGENTS.md's "runtime owns … exact terminal evidence" contract is meant to prevent. Within the valid configured range the artifacts are written in full; genuine overflow is reported, not hidden.

Approach

Make the byte budget the binding limit instead of the incidental 20k-node default. A value dense enough to reach N nodes always serializes to more than N bytes, so pinning maxNodes to each artifact's byte budget guarantees the honest, reported byte cap is what binds.

• transcripts.json — new boundedTranscriptsArtifact helper keeps whole agents in index order until the next would exceed the 2 MiB budget (binary-searched for exactness), and reports the dropped tail via a new transcriptsOmitted: { agents, entries } field rather than letting the serializer silently drop it. Empty-transcript hydration was the most user-hostile symptom, so partial retention + a reported count is warranted here.
• workflow.json (terminal + final manifest) and result.json — pass an explicit maxNodes equal to the byte budget. The manifest is bounded-shape metadata that never truncates within the valid agent range (delivery/status/transcriptArtifact preserved); result.json holds arbitrary script data, so genuine overflow falls back to the existing honest {truncated, reason, preview} stub → read-back downgrades to uncertain. This is the issue's second suggested direction.
• Read/display — transcriptsOmitted is mirrored on WorkflowDetails (like logsDropped), validated on untrusted read-back in normalizePersistedWorkflowDetails, and surfaced in buildWorkflowReport and completion alerts.

Out of scope (per the issue): serialization.ts defaults are unchanged, and retention.ts's in-memory settled projection keeps its own omitted accounting — the canonical on-disk workflow.json carries transcriptsOmitted, so a re-read restores it.

Validation

• bunx biome format / biome lint / tsc --noEmit — clean on all changed files. Config-, docs-, and discipline-contract checks pass (internal budget, not a package-owned user choice, so no setup/README updates required).
• New tests/extensions/workflows/artifact-node-budget.test.ts (5 tests, all pass): the reported 12-agent bug; partial-retention + reported count on 2 MiB overflow; the boundedTranscriptsArtifact unit contract; 1024-agent manifest surviving without losing delivery/status/transcriptArtifact; honest result.json overflow.
• All 347 workflow tests pass (2 platform-skipped), including artifacts, retention, dashboard, narrator, and completion-projection.
• Pre-existing, unrelated failures in the full suite (web/ui deps not installed → Vitest + check:web; plan-mode ANSI rendering; an installed pi-coding-agent 0.84.1 vs 0.85.1 version-drift assertion) were verified to reproduce on clean main with this change stashed.

Impact

• User-visible: /workflows now shows an "N of M agent transcript(s) omitted" notice instead of silently rendering empty transcripts; large runs no longer read back as clean completed with missing data.
• Model-visible context/tools: none.
• Runtime/lifecycle: the manifest no longer drops the delivery receipt at high agent counts, so a pending background completion is redelivered on restart as intended.
• Persisted config/data: adds an optional transcriptsOmitted field to persisted workflow.json; older records without it read back unchanged.
• Compatibility/risk: low. Behavior changes only at the previously-silent truncation boundary; normal-range workflow output format is unchanged.

🤖 Generated with Claude Code

[tt-a1i]

节点预算修复和磁盘回读路径本身是对的，但新增的省略证据还没有穿过 settled/completion 内存投影。请补齐这个降级路径及回归测试后再合并。

[onychen]

@tt-a1i 已补齐降级路径,见 b63a3a7:

• retention.ts:transcriptsOmitted 现在穿过 makeProjection 的 candidate(与 logsDropped 一致),并加入字节压力下的可丢弃白名单,保证极小预算仍能收敛。这也让 completionEnvelope() 经由 projectWorkflowDetails 构造的信封保留该证据,重启恢复/驱逐后不再丢失。
• completion-projection.ts:在展开的 operator report 里紧挨 transcripts artifact 补了一行(折叠态 alert 上一版已加)。
• 回归测试:断言该字段能穿过 projectWorkflowDetails,并同时出现在 buildWorkflowCompletionDisplay 的 alerts 与 expanded 证据中。

全部 349 个 workflow 测试通过(2 个平台跳过),format/lint/typecheck 干净。

[onychen]

@tt-a1i 刚看到 Background terminals (Windows) check 失败了，排查了一下：

• 失败的是 tests/extensions/ai-providers/antigravity.test.ts:1183 的 streamAntigravity bounds the full SSE lifetime after the first event，报的是 Timed out waiting for the first SSE event（预期匹配 /next SSE event/）。
• 这个用例用真实定时器把 timeoutMs 压到 5ms 去竞速一个 mock 的 ReadableStream（同文件里的模式），本身就对调度抖动很敏感；Windows runner 在这次跑动里显然被挤到超过 5ms 才推进到第一个事件。
• 这个文件和 extensions/ai-providers/** 都不在本 PR 的 diff 里（git diff main --stat 只涉及 extensions/workflows/* 和新增测试），本地 checkout 对照 upstream/main 复现同样的用例逻辑，和这次改动无关，是环境性的偶发超时，不是本 PR 引入的回归。
• 我没有仓库的 admin 权限，gh run rerun --failed 提示需要仓库管理权限，麻烦您帮忙重跑一下这个失败的 job？其余 5 个 check（Node 22/24/26、Web E2E、Apply area labels）都是绿的。

[tt-a1i]

审查提交：b63a3a7629b6a6fedfd035729ddeb86ee42d442e。

本轮未发现新的阻断问题。以下结论仅覆盖所列验证范围，不是合并批准。

验证范围

15 artifact-node-budget/retention/completion tests pass. Latest head copies transcriptsOmitted into settled projection and covers completion display; prior retention finding addressed.

限制与后续

No full-suite run. Existing safeStringify depth/string truncation is unchanged and not claimed fixed.

已核对的 CI 失败

Run 35496500076：Windows：Antigravity SSE lifetime 测试期望 next SSE event，却在 first SSE event 超时；失败位于本 PR 未修改的 provider 用例，尚未证明是偶发还是底层回归。

[onychen]

补充一个自查时发现的相关问题，已一并修复（commit 6e17db9）。这个不在之前的 review 里，是我复查这次改动的 UI 影响面时找到的。

问题：transcriptsOmitted 之前只到达了三个地方——磁盘 workflow.json、completion alert、以及用户按 s 保存的 report.md（buildWorkflowReport）。而 /workflows 交互式详情页走的是 renderDetail，它根本没读这个字段。所以 transcripts.json 超过 2 MiB 后，用户进详情页、打开被丢弃的 agent，仍然看到空 transcript，而且 transcript 视图的默认文案是 "this run predates transcript capture"——正是 #558 想消除的误导，只不过之前变成了"磁盘/alert 有证据、主界面没有证据、而且主界面还在说错话"。

修复（extensions/workflows/dashboard.ts）：

1. renderDetail 现在在 run 级别显示省略提示（N of M agent transcript(s) omitted from transcripts.json (byte budget); full data on disk.），并为这一行预留高度，保持原有的精确行高布局不变。
2. transcript 视图的 emptyText 改为 run-aware：当 details.transcriptsOmitted 存在时，被打开的空 transcript 显示 "omitted ... to stay within its byte budget"，不再是错误的 "predates transcript capture"。

回归测试（tests/extensions/workflows/dashboard.test.ts）：新增 WorkflowDashboard.render() 用例，直接驱动详情页与 transcript 两个界面——断言详情页出现省略提示、行高仍精确、transcript 空态文案改为字节预算措辞且不含 "predates"。

验证：dashboard 套件 40/40 通过；全部 workflow 测试 0 失败（execute.e2e 的 2 个差值是既有的平台 skip）；format / lint 干净；typecheck 无新增错误（TuiMouseEvent 那条在干净 main 上就存在，是 pi-tui 版本漂移，与本改动无关）。

[tt-a1i]

复审提交 e5373b988cb26f5c1f628d8106f376c17cc6da7b。

未发现新的阻断问题。新增投影将 transcriptsOmitted 带到运行详情，详情页给出 run 级别的省略数量，进入空 transcript 时也不再错误显示为“旧运行尚未捕获”；高度预算同步扣除提示行，测试覆盖精确布局和文案。当前所有远端检查通过。

验证边界：分支目前落后于 main，合并前仍需更新并确认最终 head；本结论不扩大到既有 safeStringify 深度或字符串截断语义。