待决与问题看板

CAND-0007 · P5 RES-1｜史实域缺 skill 层受控入口：P-3 probe 零注入下 worker 自发够到 war-history EVIDENCE 方法并有效应用（applied_effective 反转预期），但依赖探索能力+运气非受控确定命中。
target=NEW · created_at=2026-07-20T12:20:31+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0007
CAND-0008 · P5 RES-2｜回路规范存量（error_ledger）对独立 session 完全不可达：P-5 probe（日常语言故障排查）零注入下 worker 够到 OBSERVATIONS 观测流但全程未触 error_ledger，ERR-000150 既往解决办法对新 session 归零。
target=rules/skills/_DAO_ROUTING.md:17 · created_at=2026-07-20T12:20:31+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0008
CAND-0009 · P5 RES-3｜open-ended 日常/复杂语言 × headless 超时：P-1（段A）与 P-4/P-6（段B，ADJ-1 预警重演）三点复现 open-ended prompt × 无项目 CLAUDE.
target=rules/skills/drafts/workflow_buildout_real_session_eval.md · created_at=2026-07-20T12:20:31+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0009
CAND-0010 · P5 RES-4｜receipts path 级归因只覆盖 rules/skills/：S1 收集期发现的测量不对称经 P5 probe（P-2/P-4）复证——methodology 正文/staging/career 本体的消费只能靠 Face-B transcript 抽取，receipts 层测不到。
target=NEW · created_at=2026-07-20T12:20:31+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0010
CAND-0011 · S2-W2 taxonomy 含混①｜处置动作词表中英并列无机读声明：§4 用中文别名（继续取证/降级存疑/关闭不可证），run 侧用英文枚举（continue_evidence/downgrade_to_doubtful/close_unprovable），一一对应但词表真源无『机读枚举以英文值为准』声明。
target=adhoc_jobs/chinese_internet_pseudohistory_research_20260716/research/theory_taxonomy.md §4 · created_at=2026-07-20T12:20:31+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0011
CAND-0012 · S2-W2 taxonomy 含混②｜pending 补齐完成标记未定义：§4 头注定义『缺声明的 pending 行下轮 run 收口时补齐』义务，但未定义补齐后如何标记已完成。
target=adhoc_jobs/chinese_internet_pseudohistory_research_20260716/research/theory_taxonomy.md §4 · created_at=2026-07-20T12:20:31+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0012
CAND-0013 · S2-W5 B7 方法边界①｜静态 grep 对运行时渲染站点系统性低估图数：站A（history.
target=adhoc_jobs/context_infra_base_tooling_buildout_20260615/reflection_decisions_visibility_20260719/design/VISIBILITY_DEFINITION_20260719.md §二 · created_at=2026-07-20T12:20:31+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0013
CAND-0014 · S2-W5 B7 方法边界②｜静态占位数字误判手维护漂移：站A hero 静态占位数字 17/9/7 会被静态检查判为与台账（17/13/10）漂移，实为 app.
target=adhoc_jobs/context_infra_base_tooling_buildout_20260615/reflection_decisions_visibility_20260719/design/VISIBILITY_DEFINITION_20260719.md §二 · created_at=2026-07-20T12:20:31+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0014
CAND-0015 · S4 方法论验证流回填①｜聚合噪声地板与 case 级翻转率是两个量纲，现行 E23（career/methodology/METHODOLOGY_OPTIMIZATION_PHASE_v1.
target=career/methodology/METHODOLOGY_OPTIMIZATION_PHASE_v1.md:197 · created_at=2026-07-20T14:20:16+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0015
CAND-0016 · S4 方法论验证流回填②｜分域可证伪预测的连续证伪＝层归因误判的暴露信号，现行 E24（career/methodology/METHODOLOGY_OPTIMIZATION_PHASE_v1.
target=career/methodology/METHODOLOGY_OPTIMIZATION_PHASE_v1.md:198 · created_at=2026-07-20T14:20:16+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0016
CAND-0017 · S4 方法论验证流回填③｜消费单调性原则（agent 工程通用，非竞赛专属）：概率组件（LLM）接入确定性流程时，其输出只得增加信号、不得关闭基线成功路径；违反即出现「信息更多分数更低」反直觉回归。
target=NEW · created_at=2026-07-20T14:20:16+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0017
CAND-0018 · AFAC 07-20 优化 campaign 回填①｜GATE-1 成对门禁：只含改善项的优化门禁会系统性选出过拟合配置——AFAC P1 门禁只量目标子集改善（keys_in/purity24），同一 grid.
target=NEW · created_at=2026-07-21T00:56:17+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0018
CAND-0019 · AFAC 07-20 优化 campaign 回填②｜判定层证据面上限：一切 LLM 判定（裁决器/反驳者/外部模型共识/近似 GT）的可靠性上限=它看到的证据面，机械核验是唯一不受此限的仲裁者。
target=NEW · created_at=2026-07-21T00:56:17+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0019
CAND-0020 · AFAC 07-20 优化 campaign 回填③｜审查指令压力扭曲：「必须找出问题」型指令使审查者制造异议——N12 双向验证复核轮 10 次改判仅 3 朝近似 GT、7 背离，且坐实虚构 1 例（fin_a_011 把比亚迪数字张冠李戴成美的，事实表机械核验不在表内）；背离共同签名=反方指令逼模型把「限定语/前提省略」抬成实质扭曲；N15 计量反驳轮在有锚题上 5/7 REFUTED 被语料…
target=NEW · created_at=2026-07-21T00:56:17+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0020
CAND-0021 · AFAC 07-20 优化 campaign 回填④｜INFRA-1 通知不作证据：harness 后台任务通知通道实测送达 2 例伪造完成事件（未来时间戳+虚构 COVERAGE 行+错误 token 数），与磁盘直读矛盾；受害 lane 弃用通知改磁盘轮询后判读未受影响。
target=rules/skills/workflow_parallel_subagents.md:98 · created_at=2026-07-21T00:56:17+02:00 · contexts/methodology/candidates/queue.jsonl#CAND-0021

错误台账 · 未彻底解决（105 条 / 6 类）

口径：resolution_status ∈ unresolved / workaround / workaround_only（已处置 resolved、已搁置 deferred 不计入本页）。

orchestration_failure · 59 条 · 2026-07-04T21:16:04.675911+02:00 ~ 2026-07-19T23:55:28.734455+02:00

ERR-000151 unresolved 复发×32（auto_regulator_runner:missing_task_id:no_tier:missing_task_id）
regulator_runner missing_task_id: regulator_runner requires TODO_ID

现有处置：Inspect regulator attempt receipts and rerun from the main environment if needed.

contexts/runtime/error_ledger/errors.jsonl#ERR-000151 · 2026-07-19T23:55:28.734455+02:00
ERR-000146 unresolved 复发×32（auto_regulator_runner:missing_task_id:no_tier:missing_task_id）
regulator_runner missing_task_id: regulator_runner requires TODO_ID

现有处置：Inspect regulator attempt receipts and rerun from the main environment if needed.

contexts/runtime/error_ledger/errors.jsonl#ERR-000146 · 2026-07-19T02:19:43.373390+02:00
ERR-000145 unresolved 复发×10（auto_regulator_runner:codex_shell_requires_explicit_allow:no_tier:codex_shell_requires_explicit_allow）
regulator_runner codex_shell_requires_explicit_allow: regulator default dispatch from Codex shell requires explicit REGULATOR_ALLOW_CODEX_SHELL=1 plus TODO_ID and CODEX_THREAD_ID

现有处置：Inspect regulator attempt receipts and rerun from the main environment if needed.

contexts/runtime/error_ledger/errors.jsonl#ERR-000145 · 2026-07-18T23:24:07.777351+02:00
ERR-000141 workaround 复发×4（stop_hook_dpath_stale_baseline_mass_delete_rc126_orphan）
resolution_status 更正(ERR-000139 修正): 团队复核后确认 rc=126 的精确崩溃机理未坐实——已排除 gateway serialize import 链/die()全部路径/pre-commit hook 直接导致；时序线索指向 orphan-guard 预算撞线(stop_hook.log 23:07:19 orphan guard armed budget 42s of 45s elapsed=3…

现有处置：结构性修复已落地(解释器改绑+D-path护栏+rc=126/127诊断埋点)，但rc=126本身的确切触发机理待下次真实复现坐实；观察窗判据=埋点上线后30天无新rc=126或出现一次即按诊断快照定谳

contexts/runtime/error_ledger/errors.jsonl#ERR-000141 · 2026-07-18T13:51:38.238730+02:00
ERR-000135 unresolved 复发×10（auto_regulator_runner:codex_shell_requires_explicit_allow:no_tier:codex_shell_requires_explicit_allow）
regulator_runner codex_shell_requires_explicit_allow: regulator default dispatch from Codex shell requires explicit REGULATOR_ALLOW_CODEX_SHELL=1 plus TODO_ID and CODEX_THREAD_ID

现有处置：Inspect regulator attempt receipts and rerun from the main environment if needed.

contexts/runtime/error_ledger/errors.jsonl#ERR-000135 · 2026-07-16T11:29:58.197021+02:00
ERR-000132 unresolved 复发×32（auto_regulator_runner:missing_task_id:no_tier:missing_task_id）
regulator_runner missing_task_id: regulator_runner requires TODO_ID

现有处置：Inspect regulator attempt receipts and rerun from the main environment if needed.

contexts/runtime/error_ledger/errors.jsonl#ERR-000132 · 2026-07-12T15:12:01.857163+02:00

另有 53 条同类未展开，真源：contexts/runtime/error_ledger/errors.jsonl（category=orchestration_failure）

tool_failure · 27 条 · 2026-07-03T19:52:21.341426+02:00 ~ 2026-07-21T02:31:44.767910+02:00

ERR-000159 workaround
wrangler pages deploy 输出吞没复发（r002 已知坑加剧版）：不仅管道下，文件重定向下最终 deployment URL 行也会丢（本次三跑：一次假成功未产 deployment、一次只出 banner、一次 exit=0 停在 Deploying... 但服务端实际成功）。输出解析不可作为部署成败依据

现有处置：已定加固验证法：部署成败一律以 wrangler pages deployment list 的新 Production 行（source commit 匹配）+ 部署专属 URL 内容 grep（记得 -L 跟 308）为准，永不解析 deploy 命令输出

contexts/runtime/error_ledger/errors.jsonl#ERR-000159 · 2026-07-21T02:31:44.767910+02:00
ERR-000157 unresolved
infra 站共享消毒器泄漏 tilde 路径：build.mjs FORBIDDEN 黑名单与 decisions_data/dashboard_data 的 clean() 只脱敏 /Users/<user> 绝对路径和一个字面 [本地路径已脱敏] 前缀，通用 ~/ 路径全漏。线上实测 dist/decisions.html 33 处未脱敏（[本地路径已脱敏]、[本地路径已脱敏]

现有处置：建议：把 build_pending_board.mjs 的宽 tilde 正则移植进共享 clean()/FORBIDDEN 并全站回归（26 页隐私断言重跑）；作为独立小改动走，勿与功能变更混批

contexts/runtime/error_ledger/errors.jsonl#ERR-000157 · 2026-07-21T02:24:08.853768+02:00
ERR-000150 unresolved 复发×1（reflector_preflight_opencode_recovery_exit127）
reflector cron (Sun 14:00, periodic_jobs/ai_heartbeat/src/v0/reflector.py) has refused OBSERVATIONS.md -> memory promotion for 7+ days straight (stalled since 2026-07-12, last successful promotion date:2026-07-11 ran-at:…

现有处置：Not yet fixed. Needs: (1) root-cause the isolated OpenCode recovery session exit 127 (likely PATH/interpreter issue in the isolated recovery subprocess, same family as other exit-127 interpreter-wall incidents this repo has hit before); (2) manually quarantine…

contexts/runtime/error_ledger/errors.jsonl#ERR-000150 · 2026-07-19T22:46:48.776383+02:00
ERR-000149 unresolved 复发×1（stop_hook:synthetic_add:settings_json）
stop hook synthetic staging repeatedly fails to git add .claude/settings.json; partial capture each turn, only [本地路径已脱敏] text log records it

现有处置：de-ignore or special-case the path in synthetic staging filter; see audit memo trace_base_tooling_state_audit_20260719_manual.md; todo T3923

contexts/runtime/error_ledger/errors.jsonl#ERR-000149 · 2026-07-19T15:09:03.671226+02:00
ERR-000147 unresolved 复发×3（auto_unified_gate:missing_task_id:codex:high）
unified_gate regulator failure missing_task_id: missing_task_id

现有处置：Inspect regulator runner output and gate parser contract.

contexts/runtime/error_ledger/errors.jsonl#ERR-000147 · 2026-07-19T02:19:45.560222+02:00
ERR-000134 unresolved 复发×2（auto_regulator_runner:cannot_import_name_UnionType_from_types_Library_Developer_CommandLineTools_Libra:claude-kimi:medium:ImportError）
regulator_runner cannot_import_name_UnionType_from_types_Library_Developer_CommandLineTools_Libra: cannot import name 'UnionType' from 'types' (/Library/Developer/CommandLineTools/Library/Frameworks/Python3.framework/Ver…

现有处置：Inspect regulator attempt receipts and rerun from the main environment if needed.

contexts/runtime/error_ledger/errors.jsonl#ERR-000134 · 2026-07-13T00:55:17.588598+02:00

另有 21 条同类未展开，真源：contexts/runtime/error_ledger/errors.jsonl（category=tool_failure）

environment · 11 条 · 2026-07-04T08:57:06.807914+02:00 ~ 2026-07-21T02:31:43.038674+02:00

ERR-000158 unresolved
infra 站每日 publish cron（08:15）静默失败：publish.mjs 子进程 spawn npx/node 依赖 PATH，cron 环境无 nvm PATH，日志唯一输出=env: node: No such file or directory。crontab 行本身用了绝对 node 路径但子进程未继承等效 PATH。站点自动发布事实停摆（最新成功发布均为 session 手动）

现有处置：建议 publish.mjs 内 spawn 用 process.execPath 与其同目录 npx 的绝对路径（或 env 注入 PATH），修后看次日 site_publish.log 首行

contexts/runtime/error_ledger/errors.jsonl#ERR-000158 · 2026-07-21T02:31:43.038674+02:00
ERR-000096 unresolved 复发×2（cron_daemon_duplicate_lock）
crontab 写入持续挂死（读正常）：管道式与文件式 crontab - 均无限阻塞；ps 见 3 个 /usr/sbin/cron root 守护并存（Fri 常驻+今晨 8:00/8:30 各新增一个），疑 /var/at 锁争用。同 session 08:08 的 lock_watchdog 行安装成功，8:30 后全部写入挂死

现有处置：未解：杀 root cron 守护需 sudo 超自主档位；crontab 内容已核完好；待用户手动装一行或重启 cron 服务后重试

contexts/runtime/error_ledger/errors.jsonl#ERR-000096 · 2026-07-07T08:49:36.560640+02:00
ERR-000087 unresolved 复发×1（codex_workspace_write_git_index_denied）
commit_wrapper failed while staging r031 W5-G2 report: fatal: Unable to create '[本地路径已脱敏]': Operation not permitted; git add skipped path and reported nothing to commit

现有处置：Unresolved in this Codex sandbox because .git writes are denied; rerun the same commit_wrapper command from a main environment with .git write permission, staging only r031_w5g2_tsm_semantic/

contexts/runtime/error_ledger/errors.jsonl#ERR-000087 · 2026-07-06T10:31:23.396996+02:00
ERR-000065 workaround 复发×1（sandbox_ps_permission_denied）
T1172 pressure evidence attempted global ps process count, but current sandbox denied ps with operation not permitted; switched to launched-PID live count evidence.

现有处置：Use subprocess Popen handles for controlled bridge PID live-count pressure evidence; keep global ps count unavailable noted in REPORT.

contexts/runtime/error_ledger/errors.jsonl#ERR-000065 · 2026-07-06T01:41:41.533422+02:00
ERR-000064 unresolved 复发×4（codex_managed_profile_git_index_lock_denied）
commit_wrapper could not stage requested S3 gate paths because the managed Codex filesystem profile denied creation of .git/index.lock; HEAD commit could not be produced from this session.

现有处置：No workaround available inside current approval policy/sandbox; continue writing workspace artifacts and report commit blockage explicitly.

contexts/runtime/error_ledger/errors.jsonl#ERR-000064 · 2026-07-06T01:36:12.995680+02:00
ERR-000061 unresolved 复发×3（git_index_lock_eperm）
W5-B commit_wrapper could not stage or commit verified T1249/T1250/T1251/T1252/T1253/T1248 artifacts because current Codex sandbox cannot create .git/index.lock: Operation not permitted.

现有处置：Implementation, tests, todo completion events, and report remain in worktree; rerun the same scoped commit_wrapper command from an unrestricted main environment to land in HEAD.

contexts/runtime/error_ledger/errors.jsonl#ERR-000061 · 2026-07-06T01:24:41.495772+02:00

另有 5 条同类未展开，真源：contexts/runtime/error_ledger/errors.jsonl（category=environment）

provider_silent_failure · 6 条 · 2026-07-01T00:54:12.772870+02:00 ~ 2026-07-08T11:44:24.426354+02:00

ERR-000106 workaround 复发×1（provider_review_router_timeout_no_stdout）
r7 Test System provider review router attempts for codex:xhigh, claude:high, and claude-ollama-glm:high each timed out with rc=124 and no stdout/log evidence; direct codex exec succeeded separately for the primary GPT-5.…

现有处置：Use direct codex exec as primary bounded verdict; record Opus/GLM support as unavailable caveats and keep multi-provider solid decision incomplete.

contexts/runtime/error_ledger/errors.jsonl#ERR-000106 · 2026-07-08T11:44:24.426354+02:00
ERR-000098 unresolved 复发×1（auto_regulator_runner:subprocess_timeout:claude-kimi:medium:subprocess_timeout）
regulator_runner subprocess_timeout: subprocess_timeout

现有处置：Inspect regulator attempt receipts and rerun from the main environment if needed.

contexts/runtime/error_ledger/errors.jsonl#ERR-000098 · 2026-07-07T18:06:02.472844+02:00
ERR-000023 workaround
cli_agent router hangs with zero stdout/stderr for claude:low, claude-zai:low (subprocess_timeout, TTY-park class), and ALSO times out for opencode:low and codex:low canaries while a concurrent codex:high worker runs fin…

现有处置：S1 arms bypass router: direct codex exec with isolated workdir; router issue deferred to owner

contexts/runtime/error_ledger/errors.jsonl#ERR-000023 · 2026-07-04T21:24:40.191347+02:00
ERR-000003 workaround 复发×9（opencode_completion_detection）
根因定位: deepseek-v4-pro 是 reasoning 模型, 答案进 reasoning_content, 正式 content 返回空(finish_reason=length); opencode 通道只读 content -> 0-turn -> router 判 fake_success. 非 auth/模型名问题(API 直连验证: key 有效, /v1/models 含 v4-pro)

现有处置：三选一: (a)opencode deepseek 适配器读 reasoning_content+调大 max_tokens; (b)改用 deepseek-v4-flash(疑非reasoning); (c)第三评价者用 GLM claude-zai:high(已验证真跑 turns=11 cost0.64 modelUsage=glm-5.2[1m]) 顶替. 已采用(c)默认, (a)deepseek修复列backlog

contexts/runtime/error_ledger/errors.jsonl#ERR-000003 · 2026-07-01T01:47:41.973154+02:00
ERR-000002 unresolved 复发×9（opencode_completion_detection）
deepseek:high 重试仍 fake_success (turns=0, result_path=null). 确认非瞬时, deepseek(OpenCode 通道) eval 输出不可用

现有处置：确认 deepseek eval 通道 fake_success 复现. 升级用户三选一: (a) 修 opencode completion detection 让 deepseek 真产出; (b) 换第三评价者(claude-zai/kimi); (c) 暂以 Codex+Opus 双评价者. S0 eval 最终 = 2/3 (Codex+Opus 双 PASS)

contexts/runtime/error_ledger/errors.jsonl#ERR-000002 · 2026-07-01T00:59:11.397076+02:00
ERR-000001 workaround 复发×9（opencode_completion_detection）
deepseek:high dispatch 返回 fake_success: turns=0 cost=0 tokens=0 duration=142s, result_path=null, 未产出 eval verdict. router 检测到 fake_success 但仍标 dispatch OK 且未 fallback

现有处置：先同通道重试一次验证是否瞬时; 若复现则 deepseek(OpenCode 通道) eval 暂不可用, 升级用户(需修 opencode completion detection); S0 eval 以 2/3 (Codex+Opus 双 PASS) 继续, 不让 fallback 冒充 deepseek

contexts/runtime/error_ledger/errors.jsonl#ERR-000001 · 2026-07-01T00:54:12.772870+02:00

provider_auth · 1 条 · 2026-07-04T20:38:18.015283+02:00 ~ 2026-07-04T20:38:18.015283+02:00

ERR-000016 workaround 复发×1（claude_native_login_required）
Native 'claude -p' probe exited 1 with 'Not logged in · Please run /login'; project UserPromptSubmit hook could not be exercised through native Anthropic subscription.

现有处置：Switched verification to claude-kimi wrapper (same CC binary, Kimi Coding Plan subscription); UserPromptSubmit hook fired and capture/kill-switch both verified.; receipt_id backfilled by tools/error_ledger/triage.py: TR-124944 score=80 reason=task_id+chain_ste…

contexts/runtime/error_ledger/errors.jsonl#ERR-000016 · 2026-07-04T20:38:18.015283+02:00

provider_behavior · 1 条 · 2026-07-06T02:02:49.526915+02:00 ~ 2026-07-06T02:02:49.526915+02:00

ERR-000071 unresolved
T1153 canary from MAIN env: codex-oc:high returns degenerate_success (empty/degenerate reply via OpenCode gpt-5.5); sandbox network was not the only blocker. Channel currently not viable as true-hit tier

现有处置：T1153 verify-failed back to in_progress; unlock predicate=non-degenerate reply recorded to coverage ledger; feeds T1273 tier-viability config

contexts/runtime/error_ledger/errors.jsonl#ERR-000071 · 2026-07-06T02:02:49.526915+02:00

provenance：contexts/runtime/error_ledger/errors.jsonl contexts/methodology/candidates/queue.jsonl adhoc_jobs/context_infra_migration_20260709/design/OPEN_PROBLEMS.md adhoc_jobs/context_infra_base_tooling_buildout_20260615/reflection_decisions_visibility_20260719/site_candidates/kimi/data/pending_board_seeds.json

投影生成于 2026-07-21T04:36:44.982Z。

dashboard data: generated 2026-07-21T04:36:44.946Z from 5 SSOT files

待决与问题看板

待用户拍板（1）

GATE-C-R3 · 竞赛方法论织入授权

问题域登记簿 · HL/GIT/VIS 全族（18 条 / 3 族）

HL 族 · hook × 机器负载 （6 条 · open 0 / partial 4 / closed 2）

GIT 族 · 解释器墙致守卫失效 + 共享 checkout 安全网缺口 （4 条 · open 2 / partial 2 / closed 0）

VIS 族 · 可见性与状态呈现 （8 条 · open 7 / partial 0 / closed 1）

方法论候选队列 · proposed（15）

错误台账 · 未彻底解决（105 条 / 6 类）

HL 族 · hook × 机器负载（6 条 · open 0 / partial 4 / closed 2）

GIT 族 · 解释器墙致守卫失效 + 共享 checkout 安全网缺口（4 条 · open 2 / partial 2 / closed 0）

VIS 族 · 可见性与状态呈现（8 条 · open 7 / partial 0 / closed 1）