1. Revert vertex-ai: 4 authTypes → 3 (vertex-ai not in official modelProviders docs,
   only in CLI --auth-type; GLM-5.1 self-corrected their R2 suggestion)
2. Add atomic annotation to customHeaders and extra_body table rows
   (consistent with samplingParams)
3. Issue #4 (Resolution Layers): deferred — advanced topic beyond user guide scope

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

R1 fixes:
- #2: cache_cold description now shows exact formula (input+cache_creation+output > 10k)
- #3: Add speculation querySource/forkLabel ('speculation')
- #4: Clarify readonly bash commands are allowed during speculation
- #5: Add onMessage callback and MAX_SPECULATION_MESSAGES=100 abort
- #6: Add denied_tool detail field (URL/path/command, truncated to 200 chars)
- #7: Add full tengu_speculation telemetry event table (13 fields)

R2 fixes:
- R2-1: Add speculation feedback message ([ANT-ONLY] ... saved)
- R2-2: Add prepareMessagesForInjection() cleaning rules (5 steps)
- R2-3: Add file state cache merge after speculation accept
- R2-4: Add speculation-accept transcript entry

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Add Claude Code Prompt Suggestions deep-dive as standalone doc (+290 lines)

New file docs/tools/claude-code/08-prompt-suggestions.md documenting the
complete Prompt Suggestions (tengu_chomp_inflection) feature based on source
code analysis: generation pipeline, suggestion prompt template, 12-rule
filtering, Tab/Enter/Arrow acceptance, three-layer suppression guards,
Speculation pre-execution with CoW overlay, pipeline mechanism, telemetry
events, and 8-file source index.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Rename to 10-prompt-suggestions.md (avoid numbering conflict with 08/09)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Add 10-prompt-suggestions to README.md index

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Address review feedback: add source disclaimer, fix --bare vs -p, fix PR ref

- Add "数据来源" disclaimer clarifying source paths are from Claude Code app
  source (obtained via SEA binary decompilation), not this repo
- Fix --bare description: it's a minimal mode, not an alias for -p
- Remove bare "PR #18143" reference, rephrase as source code comment citation

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Address review: clarify env var three-state logic and timestamp default values

- Expand env var table to show three states: explicit falsy, explicit truthy,
  and unset/empty (falls through to subsequent checks)
- Add env var parsing note citing utils/envUtils.ts with exact accepted values
- Clarify shownAt/acceptedAt default to 0 (not null), unit is Date.now() ms,
  and acceptance check uses acceptedAt > shownAt

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Address qwen3.6-plus-preview R1+R2 review: 10 fixes

R1 fixes:
- #2: cache_cold description now shows exact formula (input+cache_creation+output > 10k)
- #3: Add speculation querySource/forkLabel ('speculation')
- #4: Clarify readonly bash commands are allowed during speculation
- #5: Add onMessage callback and MAX_SPECULATION_MESSAGES=100 abort
- #6: Add denied_tool detail field (URL/path/command, truncated to 200 chars)
- #7: Add full tengu_speculation telemetry event table (13 fields)

R2 fixes:
- R2-1: Add speculation feedback message ([ANT-ONLY] ... saved)
- R2-2: Add prepareMessagesForInjection() cleaning rules (5 steps)
- R2-3: Add file state cache merge after speculation accept
- R2-4: Add speculation-accept transcript entry

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Address qwen3.6-plus-preview R3 review: 7 fixes

- R3-1: tools_executed counts tool_result && !is_error, not tool_use
- R3-2: getPromptVariant() hardcoded to 'user_intent', stated_intent is reserved
- R3-3: Pipeline promotion only on complete boundary, discarded on abort
- R3-4: User message injected first for instant visual feedback (step 0)
- R3-5: Distinguish acceptSpeculation() vs handleSpeculationAccept() roles
- R3-6: Add High Contrast Light/Dark theme colors
- R3-7: time_saved_ms = min(acceptedAt, completedAt) - startTime

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- #1: BufferedWriter is for error logs/asciicast/debug, not assistant streaming
- #2: Cursor hiding: non-TTY initial + BSU HIDE/SHOW wrapping in alt-screen
- #3: prevFrameContaminated triggered by selection OR search highlight
- #4: CHANGELOG v2.1.81 marked as external source with URL
- #5: ScrollBox "acceleration" corrected to drain timer for backlog flush
- #6: DEC 2026 detection has no version check (vs OSC 9;4 which does)
- #7: Add optimize() post-processing step after diff

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Add Claude Code terminal rendering deep-dive (11-terminal-rendering.md, +342 lines)

Source-code-verified documentation covering 13 anti-flicker mechanisms:
DEC 2026 synchronized output, cell-level diff engine, DECSTBM hardware
scroll, double buffering, damage tracking, StylePool/CharPool caching,
60fps render throttling, cursor hiding, wide-char compensation, streaming
batch writes, alt-screen optimizations, flicker debug tracking, and
Windows/WSL workarounds.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Address qwen3.6-plus-preview R1 review: 7 fixes

- #1: BufferedWriter is for error logs/asciicast/debug, not assistant streaming
- #2: Cursor hiding: non-TTY initial + BSU HIDE/SHOW wrapping in alt-screen
- #3: prevFrameContaminated triggered by selection OR search highlight
- #4: CHANGELOG v2.1.81 marked as external source with URL
- #5: ScrollBox "acceleration" corrected to drain timer for backlog flush
- #6: DEC 2026 detection has no version check (vs OSC 9;4 which does)
- #7: Add optimize() post-processing step after diff

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Address Copilot + qwen R2 review: 7 fixes

Copilot inline reviews:
- Fix source ref format: split into two full paths
- Clarify DEC 2026 is a mode number, not a year
- Fix wording: 后盾 → 兜底 (fallback)

qwen R2 fixes:
- R2-1: VTE does have version check (>= 6800), clarify "except VTE"
- R2-2: Add ink/optimizer.ts (93 LOC) to source file index
- R2-3/R2-4: Low priority, deferred

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Replace generic one-liners with specific user scenarios and impacts:
- #1: npm postinstall reading ~/.ssh/ and ~/.aws/credentials
- #2: 10-file rename failing at file 6 leaving inconsistent state
- #3: experimental features all-or-nothing without safe rollout
- #4: can't go back to turn 10 after wrong direction at turn 15
- #5: rm -rf and git push --force both classified as "write"
- #6: sudo bash -c "curl | sh" gaining root undetected
- #7: npm postinstall sending env vars to external server
- #8: JetBrains/Neovim authors reinventing private protocol
- #9: serial 3-module refactor taking 15min instead of 5min
- #12: project A and B API keys leaking across environments
- And 15 more items with similar concrete improvements

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

整合 Claude Code 在 3 个源文件中的 9 项 Skill 装载优化，作为统一
追踪单元。

## 9 项优化（按 Tier）

Tier 1 冷启动（P0 子项）：
1. 外层 Promise.all 并行 5 路目录来源（5 行代码，~5×）
2. 内层 entries.map(async) 并行单目录（10 行，每 dir N×）
3. 顶层 memoize() 按 cwd 缓存（P3，10 行）

Tier 2 运行时 token 节省（P1 子项）：
4. sentSkillNames per-agent 去重 — 每轮省 600-1500 token（~50 行）
5. suppressNext on --resume — resume 省一次注入（~20 行）
6. Conditional skills（paths: frontmatter） — 按需激活（~30 行，
   可复用 Qwen 已有的 ConditionalRulesRegistry）

Tier 3 工程正确性（P2-P3 子项）：
7. 300ms reload debounce + 1s file stability — git checkout 不卡顿
8. Bun usePolling 规避 PathWatcherManager 死锁
9. realpath 并行去重 symlink — 解决嵌套/NFS/ExFAT inode 不可靠

## Qwen Code 现状

- skill-manager.ts:265-285 refreshCache() 4 层串行 for
- skill-manager.ts:695/723 provider + skill dir 串行
- 每 turn 注入完整 skill 列表（无 sentSkillNames）
- /resume 重复注入
- 未用 ConditionalRulesRegistry 给 skill lazy 激活

## 实施路线图（~190 行总改动）

P0: #1+#2 并行化（15 行）→ 冷启动 ~5×
P1: #4 sentSkillNames（50 行）→ 每轮省 token
P1: #6 conditional skills（30 行）→ 大 monorepo 救星
P2-P3: 剩余 6 项按需

## 主矩阵 / p2-perf / README 联动

- 主矩阵：新增 P1 item-28 行
- p2-perf item-2（并行加载 P2）顶部加整合提示 —— 已合并到 item-28
  子项 #1+#2+#3，保留作为分解参考
- sub-report 计数：p0-p1-engine 27 → 28
- 总项数：274 → 275
- README 两处数字同步

## 为什么定位为 P1 而不是 P2

单子项单独看是 P2，但综合收益对重度 skill 用户（Qwen 的明确定位，
skills 已反超 OpenCode 6.8×）是 P1 级别：
- #4 每轮 600-1500 token × N turn = 日累计省 50K+ token
- #6 对 50+ 条件 skill 的大 monorepo 不可或缺
- #1+#2 是用户第一印象的冷启动优化

…04 详情

扫描窗口：2026-04-26 11:40 UTC（上次扫描 1cf3196）→ 2026-04-27 02:30 UTC。

## 🟢 OPEN→MERGED 转换（4 项 + 2 项 carryover）

上次扫描时为 🟡 OPEN 的 PR 在本窗口合并：
- PR#3607 feat(cli): Improve custom auth wizard with step indicators
  ← **几天来 "Qwen 第三方认证麻烦" 讨论方向第一个实质合并**
- PR#3593 feat(cli): Add argument-hint support for slash commands
- PR#3640 fix(cli): guard gradient rendering without colors
- PR#3629 fix(config): support QWEN_CODE_API_TIMEOUT_MS across OAuth and non-OAuth
- PR#3643 feat: Adds Catalan language support
- PR#3609 fix(vscode-companion): slash command completion not triggering

## 🟢 本窗口新合并（1 项）

- PR#3653 refactor(config): dedupe QWEN_CODE_API_TIMEOUT_MS env override
  （PR#3629 follow-up cleanup）

## 🟡 新 OPEN（6 项）

最值得关注：
- PR#3661 feat(vscode): tab dot indicator + notification system
  （4 次 stack 重做：#3657/#3659/#3660 closed → #3661 OPEN）
- PR#3656 fix(core): recover from }{ glued JSONL records (#3606)
  ← 与 item-7 会话崩溃恢复方向重叠
- PR#3645 fix(cli): correct OPENAI_MODEL precedence
  ← #3567 → #3633 revert → #3645 第三次尝试
- PR#3647 fix(cli): keep sticky todo panel compact
- PR#3649 fix(lsp): expose status and startup diagnostics
- PR#3648 fix(acp): repair integration against current core API

## 📝 item-28 详细更新（PR#3604 OPEN）

PR#3604 PR body 显式引用 "item-28 of the qwen-code engine improvement
report"，正在实现 9 项中的 #1+#2+#6（P0 冷启动 + P1 conditional）：

- 子项 #1 外层 Promise.all — refreshCache 4 层串行 → 并行
- 子项 #2 内层 Promise.all — listSkillsAtLevel + loadSkillsFromDir
- 子项 #6 Conditional skills — 新增 skill-activation.ts (118 行
  picomatch registry，project-root scoped) + coreToolScheduler.ts
  文件路径触发 hook + <system-reminder> 通知

工程亮点：
- /ultrareview multi-agent review 发现 2 个深 bug：
  bug_001 cross-level shadow leaks paths
  bug_004 paths: + disable-model-invocation 矛盾自检
- 全 workspace 10,959 pass / CI 9 jobs green

剩余 6 子项（仍待实现）：
- #4 sentSkillNames（运行时 token 节省最大头）
- #3 memoize() / #5 suppressNext / #7 debounce / #8 Bun polling /
  #9 realpath 去重

## README 同步

已合并 PR 计数：94 → 101。

reviewer 反馈 §05 §一拓扑图后的进一步 audit 发现，多个章节仍按老 multi-
session 模型描述当前设计：

§03 §3 决策 "MCP server 生命周期" 大改：
- 决策头：per-workspace MCP state → per-daemon MCP state（在 1 daemon =
  1 session 模型下，per-daemon === per-workspace === per-session 同一
  概念，三者合并）
- 共享语义图：删除 Workspace A / Workspace B 多 workspace 拓扑，改为单
  daemon-global McpState；加跨 daemon 实例的资源池化指针（§21 路径 A）
- 决策依据 #3-#4：reframe 为"OpenCode multi-session 工程实践仍可借鉴
  +  与决策 §2 协调"（不再是"§1 协调"）
- 重复 spawn 代价表：维度从"active workspace 数"改为"active daemon 数"
  + 区分 N<50 个人/团队 vs N≥50 大规模 SaaS（链接 §21 路径 A）
- 实现要点：Workspace 类 + workspaceMap → DaemonInstance 类 daemon-global
  singleton

§06 §1.1 决策回顾：per-workspace MCP state → per-daemon MCP state，加
"OpenCode 在 multi-session 模型下走 per-workspace；qwen-code 在单 session
模型下自然合并为 per-daemon" 注

README 核心设计哲学（line 125）大改：
- 删除"多 session 共享 daemon 进程；用 AsyncLocalStorage 做 cwd /
  context 隔离"——直接矛盾 1 daemon = 1 session
- 改为"1 Daemon Instance = 1 Session = 1 Workspace；daemon 进程级隔离，
  无 multi-session 路由层；多 session 由 External orchestrator spawn 多
  daemon 实现"
- 与 OpenCode 不同段加首句"进程模型分歧"
- 持久化分层段简化为"per-daemon transcript JSONL + 外部 orchestrator 可
  选 SQLite/Postgres"（删除 Stage 1-2 / Stage 3 / Stage 6 演进语，移到
  §15）

cross-references inline 修：
- §02 line 241：multi-workspace router → 1 daemon = 1 session 自动成立
- §10 §SDK 兼容性表：per-workspace MCP 共享 → per-daemon MCP
- §12 §六 sandbox 防御：per-workspace MCP → per-daemon MCP
- §16 §三 状态可恢复性矩阵：AsyncLocalStorage Instance ctx 行 →
  ~~删除~~ + 标"1 daemon = 1 session 后不需要"
- §20 §3.3 Tool 层：per-workspace MCP → per-daemon MCP
- §04 §一 路由表注释：多 workspace 路由 → daemon 绑定 1 workspace；
  zod schema 注释：多 workspace 路由 → daemon 绑定 workspace 校验
- §09 §二 Qwen Daemon 路由：多 workspace 路由 → daemon 绑定唯一 workspace
- §05 §三 AsyncLocalStorage 代码：加 inline 注 "当前不需要；multi-session
  扩展参考 §21 路径 C"

§21 / §22 中 per-workspace / Map<workspaceId> 等表述保留——这些章节本
就讨论 multi-session 路径作为对比/演进参考。

完整覆盖过去 7 天 28 项业务 PR（剩余 22 项为 CI/release/chore 不单独列）：

新增 MERGED 表行（22 项）：
- PR#3871 (+6253/-4423) feat(cli): core built-in i18n coverage — 本期最大 PR
- PR#3897 (+1327/-198) perf(core): session-list head/tail 64KB +buffer pool
- PR#3879 (+797/-12) feat(core): reactive compression initial impl（之前 §10 状
  态升级但 PR 未直接列出）
- PR#3880 (+1731/-109) feat(cli): searchable /resume picker
- PR#3902 (+494/-64) fix(core): throttle shell tool live text updates
- PR#3905 (+432/-4) fix(cli): unfreeze Ctrl+O on long conversations
- PR#3882 (+210/-0) fix(core): filter Mistral reasoning at request boundary
- PR#3963 (+467/-63) fix(cli): validate /model arguments
- PR#3894 (+935/-15) feat(core): promote integration PR-2/3 (#3831)
- PR#3873/3887/3892 Subagent isolation 系列 3 PR
- PR#3893 (+414/-52) feat(telemetry): sensitive span attr opt-in
- PR#3915 (+86/-383) fix(skills): symlinks outside skills dir
- PR#3908 (+262/-14) feat(web-templates): /export light theme
- PR#3872 (+616/-20) fix(core): shrink file diff session records
- PR#3903 (+76/-3) fix(cli): tmux-safe dots spinner
- PR#3921/3922 LiveAgentPanel follow-ups
- PR#3875 (+10/-0) temp dir before truncated shell output
- PR#3883 (+296/-5) warn on ignored provider generation config
- PR#3948 (+648/-17) fix(vscode): coder-model Discontinued

新增 2 个重点解析：
- 🎯 #3 Subagent isolation 5 PR 系列收官（PR#3735/3873/3887/3892/3707）
  跨多周工程序列 +1325 行隔离修复
- 🎯 #4 i18n PR#3871 大体量国际化收官 — Qwen Code 国际化路线关键里程碑

更新累计计数 176 → 193（+28），标题描述从"11 项合并"→"28 项合并"，加
本周 Top 5 体量 PR 排名（PR#3871 / 3880 / 3933 / 3909 / 3897）+ 扫描盲
区检查备忘（22 项 CI/release/chore 不单独列）

PR#3889 qwen serve daemon Stage 1 已合并，merge commit 870bdf2a，最终 +12993/-194 / 84 commits（之前文档记录 +12393/-194 / 78 commits 是合并前 2026-05-12 状态）。

更新内容：
- §02 §2 实现现状表 + 决策依据 #4 + 决策矩阵 row 7：OPEN → ✅ MERGED 2026-05-13
- §03 §八 API 总览：OPEN → MERGED
- §06 总览 / Stage 1 标题 / audit 头部 / 体量对比表 / 综合结论 / 时间线 / 里程碑：全部 OPEN/CHANGES_REQUESTED → MERGED；数字 78 → 84 commits / +12393/-194 → +12993/-194
- §07 迁移路径 bash 注释：GA-ready → ✅ MERGED 2026-05-13
- §09 §〇·五 决策建议：PR#3889 OPEN → ✅ MERGED 2026-05-13
- §13 TL;DR + §3.6：OPEN/CHANGES_REQUESTED → MERGED + 数字更新
- README 顶部状态块：完整重写为已合并版本（merge commit + 84 commits + +12993/-194 + 完整路径）
- improvement-report 2026-05-13 entry：加 PR#3889 merge 作为关键事件（系列追踪以来最大单 PR，~25 轮 audit 收敛）+ 9 STAGE1_FEATURES + 内存对比 + Stage 1.5 deferred 清单 + 经验沉淀链接

… PR#4269 W4 PR 19 OPEN

PR#4249 workspace memory + agents CRUD MERGED 2026-05-18 06:27:00
doudouOUC, 15h20m open→merge (破 PR#4250 13h55m 新 daemon 项目最长 lifetime 纪录)
11 rounds review + 6 LLM model 跑 review pipeline:
- Copilot
- gpt-5.5
- DeepSeek-v4-pro
- glm-5.1
- qwen-latest
- github-advanced-security
- wenshao R1-R7 (人工)

最终 +5318/-8 23 文件 (review 翻 +2364 LOC, 初版 +2954/-3)
wenshao R11 quote: "fold-in cascade 2a → 2j has addressed every critical
and high-impact finding"
= 10 sub-fix iteration, daemon 项目最系统化修复 trail

PR#4269 safe workspace file read routes OPEN 2026-05-18 06:23
doudouOUC, +1177/-3 10 文件
PR#4249 merge 前 4 分钟开 - 零空窗接力

4 read routes 通过 PR 18 WorkspaceFileSystem boundary:
- GET /file?path=... UTF-8 text + meta(encoding/BOM/lineEnding)
  + ?maxBytes soft-truncate + ?line+?limit window
- GET /list?path=... {name, kind, ignored} 2000-entry cap
  + ?includeIgnored=1 opt-in
- GET /glob?pattern=... workspace-relative + ?cwd + ?maxResults
- GET /stat?path=...

单 capability tag workspace_file_read 覆盖 4 个 read route

wire runQwenServe 构 + inject fsFactory
= PR 18 立的 trusted: false default invariant 在 PR 19 验证
(第一个真 fsFactory consumer)

关闭 PR 18 follow-up:
- #2 fsFactory injection contract test
- #4 glob audit hashes boundWorkspace 但 emit 文字 pattern field on
  成功 + parse-rejection 两 path

trust snapshot 默认 true (operator-chosen workspace)

Wave 状态:
- Wave 4: ✅ PR 15 + PR 16 + PR 18; 🔧 PR 19 (#4269) + PR 21 (#4255) OPEN; PR 17/20 ⏳
- 24 MERGED + 8 OPEN/draft, Wave plan 17/31 ≈ 55%

doudouOUC 节奏: PR 16 merge 前 4 分钟开 PR 19, 零空窗
- 06:23 doudouOUC 开 PR#4269
- 06:27 PR#4249 MERGED
- review 期间 doudouOUC 必然已经在写 PR 19, R11 approving 后立即推送