Conversation
The combined model x reasoning menu carried a native range slider whose thumb travel did not match the tick label grid: the thumb centers run 7 + i*(W-14)/(n-1) while the labels sat at flex-column centers, so the end stops were off by roughly a character width in each direction and the middle stops drifted the opposite way (vastsa#417 follow-up). Make the ticks row an n-column grid and inset the range input on both sides by half a column minus the thumb radius, so the native thumb center lands on the matching column center for every stop count. Each label now centers in its column and carries a tick dot; the selected stop uses the accent token and the rest a muted token, per the D458 contract. Long ladders cannot fit every canonical word, so when more than four stops are listed only the selected stop and its immediate neighbours keep visible text; the rest collapse to their tick dot but stay clickable and keep their tooltip. Four or fewer stops show every label. Renderer only: no protocol, storage, host, or migration change. The range input remains the accessible control; tick labels are still not tab stops. Updates the source-contract test, D458 in both decisions logs, and E2E-050.
The previous stop alignment kept the tick dots in the label row below the track and collapsed labels on long ladders, neither of which matched the intended design. The dots belong on the rail itself, one per stop with the thumb gliding over them, and every label should stay visible. Restructure the slider into a rail (dots row plus the overlaid range input) and a separate labels row. Both are full-width n-column grids keyed to --stop-count, so dot i, label i and the thumb at index i share one column center W/(2n) + i·W/n. The input is inset by half a column minus the thumb radius to move its native stops onto those centers; its own track is transparent so the rail line and dots show through. The rail line spans the stop-to-stop distance and the fill uses i/(n-1), so it ends exactly at the active stop. Track dots use the muted token with the accent token for the selected stop (the thumb covers it); every label stays visible and ellipsizes in its column. Renderer only: the range input remains the accessible control and labels are still not tab stops. Updates the source-contract test, D458 in both decisions logs, and E2E-050.
The slider now gives feedback when a level change lands without giving up direct manipulation. While a drag is in flight the thumb, track dots and accent fill track the pointer with no transition so the control stays responsive; once the store confirms and the drag lead clears, the thumb plays a single overshoot pop and the fill slides to the stop. A settling flag marks the brief window after a drag commit lands, so the thumb pop only runs then — never mid-drag and never on open. The accent fill is split into its own ::after layer whose width transitions between stops, and the track dots cross-fade between muted and accent. The fill percentage is i/n so it reaches exactly the active stop's column center under the new layering. All of it is disabled under prefers-reduced-motion. Renderer only: no change to the accessible control, commit path, or persistence.
Keep the popover anchored when the trigger is pressed and move the visual thumb smoothly between selected levels without delaying native dragging. Preserve optimistic selection while writes settle and link label hover with unfilled stops without disturbing the filled track. Cover geometry, repeated menu opening, motion, hover, keyboard input, dragging, and reduced motion in an isolated Electron fixture.
Let long-running sessions and subagents recover from independent provider outages without spending retries from earlier successful responses. Keep continuous failures bounded and preserve cancellation and tool execution. Report terminal retry counts from the relevant budget and cover recovery through real agent loops and isolated desktop fault injection. Refs vastsa#699
pi-ai only clamps max_tokens on the streamSimple path; the low-level stream used for thinkingLevel "omit" sends options.maxTokens untouched, and pi-ai's chars/4 estimate under-counts CJK text. Both routes can push input + output past the model window and hit a maximum-context-length 400/413/503. Add a shared request-side clamp (output-cap.ts) applied in the parent runtime and subagent stream fns: CJK-aware input estimate plus a window-scaled reserve, so the clamp holds for streamSimple and the low-level stream alike. Unit tests are model-agnostic.
Use the desktop runtime's OS trust support so locally trusted inspection roots work without bypassing TLS validation. Preserve certificate causes through session and delegate recovery to stop futile retries and explain the failure in the transcript. Cover the real sidecar launch and local HTTPS path, synchronize the runtime specifications, and document the trust decision. fixes vastsa#714
fix(agent-runtime): cap output tokens to the model window
Plugin utility processes and plugin-declared stdio MCP servers inherited a closed environment that omitted HOME, USER, and USERPROFILE. A spawned binary that resolves ~ through $HOME then looked for its state under TMPDIR and exited 1. Both spawn sites now share one allowlist so the identity variables cannot drift out of only one of them. Provider keys still do not cross. fixes vastsa#717
fix(plugins): pass HOME/USER into plugin child processes
Merge the reviewed provider certificate trust fix.\n\nThe Vercel preview check was not available because deployment authorization is required.
Only clear relative file references when the workspace actually changes. Returning from Settings must retain the restored session draft. Cover remount preservation and workspace-change cleanup in the existing real React composer regression fixture.
The session runtime's budget math (hardLimit, requestHeadroom, keepRecentTokens) is about to be needed by subagents as well (ADR 0299, issue vastsa#708). Two drifting copies of the same safety boundary would let a delegate issue requests its parent already considers unsafe, so the formula moves to a dependency-light module both sides import. Pure extraction: contextBudgetFor() and retainedUserMessageBudget() reproduce the previous numbers exactly, which the unchanged runtime.test.ts budget assertions verify.
Record the decision to extend the session's compaction contract to delegates (issue vastsa#708): a delegate gets its own budget-derived turn boundary compaction, a degraded tail-only retry, and a dedicated SUBAGENT_CONTEXT_OVERFLOW terminal failure with actionable guidance instead of a raw provider error. The commit also declares the new error code in the shared registry and its localized strings ahead of the runtime implementation, which lands in the follow-up commit. Specs (agent runtime, error codes, E2E plan) are updated in both languages per the change checklist.
A delegate ran its Agent with none of the session's context protection: no budget, no turn-boundary compaction, no degradation, and a provider overflow ended the run as an unactionable CONTEXT_TOO_LARGE that the parent model could only read as a raw provider sentence. The feature whose purpose is keeping large reads out of the session window was the one path with no window protection at all (closes vastsa#708, ADR 0299). SubagentRun now wires prepareNextTurnWithContext and evaluates the shared budget formula against the run's resolved model. At or above the hard limit the run compacts synchronously through pi-agent-core's prepareCompaction/generateSummaryWithUsage, in memory only; if no fitting summary exists it degrades to the task brief plus the most recent messages and says so on its result; if even that does not fit the run fails with SUBAGENT_CONTEXT_OVERFLOW naming the three remedies the parent can act on. Fallback alternatives are re-evaluated against their own windows before switching, and resume seeding truncates the oldest tool call/result pairs first so a resumed chain starts below its limit instead of overflowing on arrival.
Keep the app default when it remains in the provider model list. Only fall back to the first binding when the selected model is removed.
Refresh the pinned pi runtime packages and rebase the Desktop-specific adapter patches onto the new upstream artifacts. Keep provider transcript compatibility and expose the upstream Meta/Muse OAuth catalog through the existing dynamic vendor flow.
Preserve captured certificate causes through one-shot completion so certificate errors remain terminal and diagnosable. Synchronize the Chinese runtime and E2E documentation with the pi 0.86.1 system-message contract.
fix(agent-runtime): give delegates the session's context budget
# Conflicts: # packages/agent-runtime/src/delegation-history.ts
Merge pi 0.86.1 sync, provider adapter fixes, and context-budget compatibility into main.
# Conflicts: # docs/spec/06-delivery/04-e2e-test-plan.md # packages/agent-runtime/src/runtime.ts
…ranscript-disclosure # Conflicts: # docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md
feat(transcript): land nested thinking-process disclosure
vastsa
reviewed
Sep 20, 2026
vastsa
left a comment
vastsa
left a comment
Owner
There was a problem hiding this comment.
感谢这个重试预算修复，代码逻辑和新增 agent-loop 覆盖看起来是对的。当前不能合并：PR 处于 DIRTY，且 Rust 检查失败，cargo test -p host-core --locked 为 558 passed / 1 failed，失败是 plugins::tests::an_install_reports_progress_and_honours_a_cancel，报 PLUGIN_INTEGRITY: package is 0 bytes, 6540 were announced。
请先基于当前 main 解决冲突，再重跑完整 Rust 检查；如果确认是基础设施竞态，也请提供可复现的分类和新的绿色结果。谢谢。
…lider-alignment fix(composer): stabilize reasoning menu and animate slider interactions
…omposer fix(composer): 遮住输入框下方的会话正文
fix(settings): preserve selected default model when saving providers
fix(composer): preserve existing attachments when switching chats during paste
Use the PluginManager data directory when the built-in marketplace catalog is materialized instead of consulting process-global test state. Add a regression test that keeps an unrelated data-directory override from redirecting package writes.
Keep beta and rc version surfaces aligned with the package version while checking their stable release documentation and changelog entry. Cover the preview resolution rules with focused Node tests so the release preflight matches the documented workflow.
Bump every package, Cargo, protocol, and bundled model-catalog surface for the next test release. Add stable 0.15.2 notes in every shipped locale and declare the 0.15.x release line in both READMEs.
All repository CI and docs gates passed. Vercel preview failed independently and is not a protected branch check.
Preserve both the workspace-remount regression and the pending-paste regression after upstream integration of vastsa#732. Keep both draft contracts in the shared Composer specification.
Cache the workspace associated with a composer draft so a remount can distinguish an unchanged workspace from a project switch while ChatSurface is unmounted. Extend the real composer fixture and cache tests to cover the transition.
…unt-726 fix(composer): preserve file references across workspace remounts
Preserve provider failure diagnostics while reporting the exhausted retry budget. Give the pi loop a message-array snapshot so stream recovery after tool calls does not duplicate assistant history. Incorporate the upstream marketplace data-path isolation fix and retain contributor history. Stabilize Windows E2E screenshots without weakening recovery assertions.
Contributor
Author
|
已按建议更新到当前 main，解决冲突并推送至 main 已包含插件市场数据目录隔离修复；本次重新运行的完整 Rust CI 已通过，原失败用例通过，JS 和文档检查也均通过。 集成新版 pi 时还发现消息数组共享导致流恢复失败，已修复并通过回归测试。隔离桌面 E2E 全部通过，包括连续故障耗尽、点击“继续”恢复，以及多次独立断网后的长任务。 当前 PR 已无冲突，仅 Vercel 预览仍需部署授权。麻烦再看一下，谢谢！ |
Owner
|
已核实该网络错误恢复问题确实存在，但修复内容已经随当前 main 的提交 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
8 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
问题与修复
长任务中，网络故障即使已经重试成功，次数仍按整个用户回合累计。上游
aad46adb的桌面模拟中，完成 10 次文件读取后，第 11 次独立断网直接终止任务。现在主会话和内置子代理在完整模型响应成功后重置网络错误及限流的重试预算，包括成功返回工具调用。下一次请求从第 1 次重试开始；只收到响应头、部分文字或切换故障阶段不会重置预算。
终止诊断改为读取对应预算的计数，避免临时重试状态清理后漏报
retryAttempt。范围与兼容性
验证
基线 main：
aad46adbc1aa8120659681125ae11a28014868bf；Windows x64、Electron 43.6.0、Node 24.2.0。使用重新构建的桌面及 runtime、隔离数据目录和本机故障模拟 API。node scripts/e2e-provider-recovery.mjs通过：请求断线、输出中断、Responses 恢复、连续故障耗尽及恢复后点击“继续”。持续断线仍在 10 次重试、约 63 秒后终止。pnpm lint、架构检查、pnpm docs:check、pnpm docs:build、agent-policy 和 diff 检查通过。验证边界
Runtime 全量测试 642 项通过、1 项失败：未改动的
native-pi-session.test.ts:759使用/拆分 Windows 路径，导致文件名比较失败。该测试及对应实现与上游一致，未纳入本次修复，不将全量结果标为通过。未验证原生 macOS／Linux、真实 ChatGPT OAuth 或原报告者的网络环境。本次解决已复现的预算累计问题，不能据此认定原报告的具体网络根因。与 #476 的可选无限重试功能范围不同。远程 CI 以 PR 发布后的实际运行结果为准。
Refs #699
CI 当前结果
本候选的 JS 构建／类型检查／lint／架构／测试及 Docs checks 已通过。Rust 的格式检查和 Clippy 通过，测试阶段 558 项通过、1 项失败：
plugins::tests::an_install_reports_progress_and_honours_a_cancel报告PLUGIN_INTEGRITY: package is 0 bytes, 6540 were announced（运行日志）。本 PR 未修改 Rust 源码或 Cargo 文件；基线
aad46adb的 Rust CI 通过。相同失败用例在本候选的 Windows 本地单独复测通过，尚不能确定 Linux 并行测试的完整根因。已尝试重跑失败任务，但 GitHub 返回需要仓库管理员权限。该 Rust 检查仍为失败，需项目方重跑或继续定位。Vercel 预览另需部署授权，不属于代码测试失败。