[business-logic] #1440-A's bgorphan terminal-status exit is dead code: it matches uppercase 'SUCCESS'/'FAILED'/'CRASHED'/'REMOVED' but the real CommandJobStatus values are lowercase 'completed'/'failed'/'cancelled'/'timed_out' (CRASHED/REMOVED don't exist at all) - completed jobs keep cycling false 'may have errored, hung' warnings until the 3-injection budget burns out, after which genuinely orphaned hung jobs stay silent forever (charter inverted); the test feeds a fictional 'Status: SUCCESS' sample so it only ever validated the dead branch

[topcheer]

文件行号

internal/agent/bgorphan_detect.go L173（terminal-status exit：if status == "SUCCESS" || status == "FAILED" || status == "CRASHED" || status == "REMOVED"——L174-176 delete 生产不可达）对照 internal/tool/command_jobs.go L38-42（值域全小写：running/completed/failed/cancelled/timed_out——"CRASHED"/"REMOVED" 在值域中根本不存在）+ L609（fmt.Sprintf("Status: %s\n", snapshot.Status) 渲染 Status: completed）+ 三读取通道 internal/tool/command_job_tools.go L183/276-287/457（read_command_output/wait_command/list 全走此渲染）+ bgorphan_detect_test.go L253（喂虚构 "Status: SUCCESS\n"——渲染层永不产生的值，测试通过但只验证死分支）+ premature_success.go L183（同库先例：同类 Status 行正是 ToLower 后比较）

问题

#1440-A 修复是死代码——大写字面量 vs 小写真实值域零匹配

• [business-logic] bgorphan has no completed-job exit - finished jobs loop-warn until the 3-injection budget is burned and real orphans go silent (result Status is in hand but recordOutputCheck takes only args; stop_command errors on completed jobs); strategist tail-truncation whitelist misses read_command_output/wait_command - the standard monitor channel for background tests gets head-only 500 chars cutting the verdict lines its own docblock says must be preserved; post-summary path has no message-count cap (fallback caps at 20) so long autopilot runs blow provider limits and hard-stop #1440 commit 自述目标 "completed jobs stop cycling warnings" 完全未达成：四个大写字面量与五个真实值交集为空（"SUCCESS"≠"completed"、"FAILED"≠"failed"）；全库无 ToUpper 归一层（grep 实证 internal/tool 仅 iterm2/scan_todos 有 ToUpper 与 Status 无关）
• 双向后果：completed job 留在 activeJobs → gap 达阈值 fire 假警告 "may have errored, hung, or produced important output"（误报——L181 每次 read 清 warned 下个 gap 再 fire）→ 直到烧尽 bgOrphanMaxInjections=3（L205）→ 此后真孤儿 hung job 永久静默（漏报——charter 倒置）
• 触发：任何 start_command 启动的 job 完成后 agent 再 read/wait 一次且 gap 达阈值即触发

修复建议

L173 改为小写值域匹配（completed/failed/cancelled/timed_out）或复用 ToLower+premature_success.go:183 模式；同步把测试样本改为真实小写格式防再犯（自证陷阱）。

严重程度

High（误报+漏报双向、#1440 名义已修实际未修、测试虚构样本掩盖）

---

独立复核确认：值域原文、渲染链 609、三通道、四 vs 零交集、测试虚构样本、假警告循环+预算烧尽链、先例对照。

#1440-B 与其余四件验证结论

• [business-logic] bgorphan has no completed-job exit - finished jobs loop-warn until the 3-injection budget is burned and real orphans go silent (result Status is in hand but recordOutputCheck takes only args; stop_command errors on completed jobs); strategist tail-truncation whitelist misses read_command_output/wait_command - the standard monitor channel for background tests gets head-only 500 chars cutting the verdict lines its own docblock says must be preserved; post-summary path has no message-count cap (fallback caps at 20) so long autopilot runs blow provider limits and hard-stop #1440-B 通过：strategist tail-whitelist（L290 六工具+head 300/tail 700）在位正确；<GOAL_ACHIEVED/> sentinel 防护+调用上限+fe920bf9 deadlock 修复在位；list_commands 不在 whitelist 为 Low 观察不立案。
• verify_lint：sync（agent.go:676）+async（108）双路对称/gofmt -l 按行解析/编辑钩子触发无误——# 前缀同族排查不适用（无命令字符串识别路径，[business-logic] the '# activity-label' prefix that run_command prompts mandate blinds three sibling detectors at once: isVerificationCommand / coverageIsVerifyCommand / eaIsVerifyTool all tokenize '#' as the first token, so verification debt never zeroes (false warnings), verify_coverage_gap is completely dead on conforming agents, and edit_abandon inherits the same blindness - stripLeadingShellComment exists but only 2 call sites got the #471/#950 fix; exit-127 isNonFailureExit swallows real go-test failures inside '||' chains as 'tool not available' green #1522 模式无载体）。无案。
• verify_regression：[business-logic] the empty-result fallback rules in tool_error_fallback/tool_fallback_hints are dead on arrival: they match 'no match'/'no results' strings on the error path, but grep/search_files/glob/code_search and four LSP tools all return empty results with IsError=false on the success channel, so the advertised 'when grep returns nothing, suggest alternatives, save 1-2 iterations' behavior never fires; tool_integration_monitor clears pending evidence BEFORE the empty-text early-return, so in the most common agentic pattern (pure tool-call turns with no assistant text) accumulated read/grep evidence evaporates unchecked and evidence-integration detection silently no-ops; resetToolFallback has zero production callers - the fired map leaks across runs and its per-tool slot is burned before rule evaluation, so even matching errors get no suggestion from run 2 onward #1508 系修复在位，全文深读无新案。
• verify_scope_narrow：收窄逻辑无误报路径/reset 在位（1225/1538）——同样无 # 载体。无案。