开源贡献日报 · 2026-09-12

记录可验证的修复进展，也记录仍有空间参与的架构与规范讨论。

数据截止：2026-09-12T23:59:59+08:00（Asia/Singapore）。以下是当日截至该时间的记录。

公开资料覆盖：完整；All 13 enabled repositories received a complete incremental pass; six model evaluations each were completed for Lody, pi, OpenCode and LangGraph. Deeper design-discussion coverage remains partial.。

1 今天跟进与完成的结果

• Windows taskbar icon (apache/maka#5199，pr) — current-head CI passed；The rollback-verifier repair is now green and the contribution is waiting for review.
• Agent token-limit clamping (CherryHQ/cherry-studio#20240，pr) — collaborator approved；The shared clamp/reservation design and wire regression received approval with no blocking findings.
• Windows baseline Bun sidecar (cline/cline#14066，pr) — submitted; external allowlist gate；The validated patch remains open while the repository handles its author allowlist.
• Persist valid task progress in ACP history (LodyAI/Lody#445，pr) — submitted as draft fallback after Ready permission denial；Validated task metadata now reaches history, merges by taskId, preserves identity and purpose, and resumes settled tasks only on allowlisted lifecycle events.
• HTML artifact Finder fallback (apache/maka#5215，pr) — submitted; CI/review pending；A validated failure-only Finder recovery action now reuses the existing safe artifact IPC path, with successful previews and path guards unchanged.
• Grok 4.6 OAuth tool-shape rejection (makecindy/cindy#4336，issue) — current source already contains the proposed sanitizer; runtime evidence requested；Current main and the reported release tag were traced and focused request-body tests passed, narrowing the remaining question to packaged routing or later field injection.
• HTML Artifact preview handoff (apache/maka#5213，issue) — Reopened with maintainer implementation in progress；The corrected Artifact scope is already being implemented in maintainer PR #5216, so no duplicate contribution was started.
• Native Windows daemon discovery coverage (multica-ai/multica#8336，pr) — submitted; native Windows CI/review pending；The test harness now isolates USERPROFILE and LOCALAPPDATA as well as HOME, uses the platform-native Hermes root, and runs the two affected discovery regressions in Windows CI without changing production behavior.
• Preserve snoozed scheduled-task occurrence across edits and pause/resume (apache/maka#5221，pr) — PR #5226 merged; issue #5221 closed.；The lifecycle repair preserves a future snoozed occurrence through edit, pause and resume while recomputing elapsed occurrences.
• Opt-in OAuth callback page renderer (earendil-works/pi#5372，issue) — Validation passed and proposal published; waits valid maintainer lgtm before PR.；A fixed fork compare adds safe opt-in rendering across four OAuth browser providers and the legacy adapter.
• Orca recovery notice capture boundary (makecindy/cindy#4349，issue) — Exited after another contributor took the same fix；The local investigation was preserved, while duplicate implementation stopped when another contributor published the same bounded plan.
• Mobile device roster freshness (makecindy/cindy#4358，pr) — Independent acceptance passed; review pending；The request-generation fence passed focused acceptance; the remaining Windows failure was isolated to an unchanged desktop timeout.
• Long-running Cloud subagent duplicate execution (langchain-ai/langgraph#7417，issue) — OSS boundary isolated; no justified patch；A current-main 185-second probe ran once and located the unresolved mechanism outside the OSS repository in Cloud sweep/lease behavior.
• Repository-size dependent translation delay (CherryHQ/cherry-studio#20271，pr) — open; maintainer follow-up handed to user；The contribution remains green on major checks, but its waiting episode passed the maintainer-response window without substantive feedback.
• Long-conversation renderer performance (CherryHQ/cherry-studio#18356，issue) — Measured investigation completed; patch deferred；A 20k-token component fixture quantified expensive mixed-Markdown remounts, but a native trace is still needed to identify the live scroll bottleneck.
• Context-limit compaction regression (CherryHQ/cherry-studio#5996，issue) — Existing implementation found；The reopened failure maps to an active tested compaction fix in PR #18947, so duplicate code was avoided.
• Launch-to-tray startup regression (CherryHQ/cherry-studio#5256，issue) — Current paths checked; no safe patch yet；Both startup suppression layers remain in current code, so the new Windows regression needs a concrete bypass trace before modification.
• Bound mixed multiline math fences (CherryHQ/cherry-studio#20469，pr) — Ready pull request submitted; current CI passed；The parser now closes the mixed multiline display block at the matching physical-line fence instead of swallowing the following heading, links, table or HTML.
• Make provider probing reliable immediately after creation (LodyAI/Lody#652，issue) — Validated reference fix and proposal published；Empty provider keys are rejected before probing, while the daemon gets one bounded sync-and-reread opportunity for a just-saved Machine Flock configuration.

• Agent output-limit review repair (CherryHQ/cherry-studio#20353，pr) — Current review fix revalidated；Vertex Anthropic option resolution was checked against actual SDK wire semantics and the focused 93-test matrix without adding a duplicate progress comment.

• 今日合并：2。
  • fix(scheduled-task): preserve snoozed occurrence on edits
• 今日关闭（未合并）：0。

今天合并的旧 PR

旧 PR 指今天之前创建、今天合并的 PR；已包含在今日合并总数中。

时间均为 Asia/Singapore。

2 今天吸取的经验

• Repository bots may report contribution-policy gates rather than code defects; classify the signal before changing a validated patch.
  • 证据：GitHub
• A shared resolver made output-token clamping consistent across wire requests and context reservations, which the collaborator review explicitly validated.
  • 证据：GitHub
• A fixed-head comparison is more useful than PR-title matching: the competing HTML artifact change expanded successful-path behavior and failed TypeScript, while the narrower failure-only route could be validated against the issue contract.
  • 证据：GitHub
• Portable tests must isolate platform-native profile variables, not only POSIX HOME; native CI should execute the exact path-sensitive regression.
  • 证据：GitHub
• Testing one schedule subtype was insufficient: moving the unchanged-schedule invariant ahead of type-specific resolution and covering both daily and interval paths prevented a snoozed occurrence from being lost on metadata-only edits.
  • 证据：GitHub
• A managed-environment browser launch denial is not an application regression: retain the successful build evidence, mark the browser smoke unperformed, and seek a privileged rerun before changing timeouts.
  • 证据：GitHub
• Scheduled-time correctness lives across transitions: preserving the recurrence anchor during edit is insufficient if pause clears the pending occurrence. Test edit, pause, persistence and resume as one lifecycle, and distinguish future pending times from elapsed ones.
  • 证据：GitHub
• Keep authentication customization opt-in and make renderer failure fall back to the escaped default page.
  • 证据：GitHub
• Durable user-visible recovery notices should not silently become captured worker output or progress signals.
  • 证据：GitHub
• When a caller establishes a reconciliation fence after a shared request began, request deduplication can reintroduce stale snapshots; start a new snapshot after the fence and keep epoch checks on the response.
  • 证据：GitHub
• A long-duration local probe can distinguish framework execution from managed scheduler replay; do not patch checkpoint semantics when the observed sweeper and lease code lives outside the repository.
  • 证据：GitHub
• Gateway adapters need to resolve configuration by the namespace that owns the option, not only by the runtime provider key; a Vertex request exposed why wire-body regressions matter.
  • 证据：未知（未提供公开证据）
• A fresh request must supersede older consumers as well as avoid promise coalescing; otherwise stale offline evidence can still win before the fresh online response arrives.
  • 证据：未知（未提供公开证据）
• Markdown fence fixes should be proven with the exact text that follows the closing delimiter, because the dangerous failure is often not the math node itself but everything the parser silently absorbs afterward.
  • 证据：未知（未提供公开证据）
• For renderer-to-daemon configuration visibility, pair source-side validation with one bounded synchronization attempt; this closes the real race without turning a missing configuration into an indefinite retry policy.
  • 证据：未知（未提供公开证据）

3 有意思的模块与行业动态

• Atomic idempotency for replayed long-running tool calls（langchain-ai/langgraph）
  • 讨论问题：Cloud replay can duplicate side effects; the open boundary is whether authority is checked before dispatch or atomically compared and committed at persistence time.
  • 行业动态：Exactly-once-visible effects need commit-time fencing; a check-then-write authority test retains a time-of-check/time-of-use window.
  • 总结点：Separate scheduler responsibility from saver/store commit contracts.；Use concurrent stale-writer tests to validate atomic compare-and-commit semantics.
• Tool-call argument lazy parsing and snapshot API boundary（earendil-works/pi）
  • 讨论问题：Large streaming tool arguments can starve a multi-session loop, but an incremental parser must preserve structured partial previews and stable snapshots.
  • 行业动态：Streaming performance work is an API-contract problem as much as a parser problem: consumers depend on malformed-preview repair and snapshot identity.
  • 总结点：Benchmark actual incremental strings and clone costs.；Keep malformed-preview and compatibility fixtures beside throughput tests.
• Define MCP disable boundaries without losing core file and image capabilities（LodyAI/Lody）
  • 讨论问题：Users need global and provider-level MCP disable controls for strict gateways, while built-in file and image channels must remain available.
  • 行业动态：Compatibility switches need capability-level contracts rather than a single global off bit.
  • 总结点：Separate external optional MCP from core file and image channels.；Specify precedence, session snapshot timing and runtime feedback before implementation.

4 今天在哪些项目做了贡献

• CherryHQ/cherry-studio：Submitted and validated PR #20469 for mixed multiline display-math boundaries, and revalidated the current #20353 Vertex Anthropic review repair.（来源 1、来源 2）
• cline/cline：Windows x64 Bun baseline fix was submitted; the remaining Greptile signal is an author allowlist gate.（来源 1）
• LodyAI/Lody：Published a validated fixed commit and proposal for provider creation/probe consistency in #652; upstream PR creation awaits the repository template's publishable original-prompt field.（来源 1）
• makecindy/cindy：Submitted PR #4366 to prevent stale mobile device rosters from overriding presence; review then added a request-generation fence so older offline callbacks cannot outrank a fresh online snapshot.（来源 1）
• multica-ai/multica：Submitted PR #8338 to make daemon local-skill fixtures safe on native Windows and add the missing Windows CI coverage; local focused tests and cross-compilation passed.（来源 1）
• apache/maka：Delivered the HTML artifact Finder fallback in PR #5218 and completed the scheduled-task snooze lifecycle repair: PR #5226 was merged and issue #5221 closed.（来源 1、来源 2）
• earendil-works/pi：Built and validated an opt-in OAuth callback renderer with safe default fallback, then published a fixed compare and the repository-required LGTM proposal for #5372.（来源 1）
• langchain-ai/langgraph：Isolated #7417 to the managed Cloud sweep/lease boundary with a 185-second current-main probe, avoiding a speculative OSS checkpoint patch.（来源 1）