OpenClaw Failure Layer Map

過去レポートのView/ソース規律バックフィルで生成したView。

Generated: 2026-05-20T07:55:00+09:00

Updated: 2026-05-20T12:48:00+09:00

健人くんの「すすめよ」を受けて、2026-05-20 の dev intel「AGENT = MODEL + HARNESS」を実運用に落とした。狙いは、同じ失敗を見つけた時に「注意します」で止めず、どの層のどの部品を直すかへ即進むこと。

Layer Taxonomy

| layer | scope | typical fix |

| --- | --- | --- |

| L1 UI/REPL | Telegram文面、承認UI、Viewの見え方、スマホ上の体験 | copy rule, screenshot check, delivery policy |

| L2 Orchestration/context | memory、prompt、mode、compaction、artifact shape | context sheet, memory rule, report schema |

| L3 Agent loop/gates | retry、timeout、human gate、verification、done条件 | guard, lint, timeout contract, task ledger gate |

| L4 Tools/skills/connectors | scripts、MCP/apps、browser、GOG、View renderer、plugin/skill | tests, connector evidence, source scan, wrapper fix |

| L5 LLM service | model/provider/runtimeの外部挙動 | provider fallback, model switch, isolate from harness failure |

First Mapped Failures

| failure_signature | layer | runtime_contract | fix_surface | regression_gate | notify_policy |

| --- | --- | --- | --- | --- | --- |

| x_original_text_confusion | L2 + L3 + L4 | observability_event | X/browser rule, source-observed flags, report schema | block quoted X text unless original is verified | verified original only; otherwise label 原文未確認 |

| chart_sizing_oscillation | L4 + L1 | observability_event | meal-tracker View renderer and mobile chart constraints | screenshot/DOM-size test for compact + readable chart | notify after fixed artifact, not explanation-only |

| thin_report_accumulation | L2 + L3 | observability_event | report quality classifier and task ledger done gates | report-quality lint requiring conclusion/evidence/sources/next action | silent internal cleanup unless finished View is useful |

| gog_calendar_degraded_output | L4 + L3 | settings_inheritance + observability_event | GOG calendar connector wrapper and daily-news wording | connector failure fixture must not render as 予定なし | ask/notify only for reauth or day-affecting events |

| unknown_skill_plugin_install | L4 + L3 | settings_inheritance + approval_queue | manifest/source/package/hooks/MCP/env scan, trigger-collision notes, and approval packet | block installs without inspected manifest/scripts/secrets/triggers summary | always ask before external/unknown plugin install |

| resident_agent_contract_gap | L2 + L3 | settings_inheritance + session_persistence + scheduler_policy + approval_queue + observability_event | runtime-contract ledger and resident event schema | audit at least one durable event for each contract category | notify only when the gap creates owner-visible risk or a useful finished report |

Changed

• Replaced the old memory/heartbeat-known-failures.md table with failure_signature / layer / evidence / fix_surface / regression_gate / notify_policy.
• Added two missing but important signatures: gog_calendar_degraded_output and unknown_skill_plugin_install.
• Kept existing evidence and current controls where they already existed, but converted each row into a routing rule.
• Added runtime_contract as a second routing axis from the resident-agent intel: settings_inheritance / session_persistence / scheduler_policy / approval_queue / observability_event.
• Added resident_agent_contract_gap so future OpenClaw/Claude Code/afk-code daemon ideas become concrete event/guard work instead of a vague design note.

Sources

• Local source-backed intel: projects/heartbeat-editorial-room/20260520-0702-dev-intel-harness-layer-failure-routing.md
• Zenn: https://zenn.dev/luoxi/articles/harness-engineering-everyday
• Zenn: https://zenn.dev/r_kaga/articles/89e4659ca691bc?locale=en
• GitHub: https://github.com/VoltAgent/awesome-agent-skills
• Resident-agent intel: projects/heartbeat-editorial-room/20260520-0932-dev-intel-resident-agent-forks.md

Verification

Run:

rg -n "runtime_contract|resident_agent_contract_gap|settings_inheritance|approval_queue|observability_event" memory/heartbeat-known-failures.md projects/openclaw-operations-improvement/openclaw-failure-layer-map.md