← View index

OpenClaw Ops Index — 2026-05-19

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: 2026-05-19T19:34:00+09:00

Lane: OpenClaw改善 / context_secretary

Why this is useful:

The stale docs/intel intake loop was blocked by synthesis debt, not by missing raw material. This index turns repeated Discord/OpenClaw patrol reports and recent agent-extension intelligence into one operating map: what matters, what to steal, and what the next local guard should be.

What I made/changed:

Consolidated existing local reports into a source-linked OpenClaw operations index and refreshed the task ledger so the P1 item no longer only says "make an index someday."

Sources/Evidence:

Prediction:

OpenClaw's next useful improvement is not more broad collection. It is an intake/review layer for operational knowledge: when a Discord/docs/X/Zenn signal repeats, it should land in one of four buckets and either change a rule/test/guard or be archived as weak.

Verify by:

Run python3 scripts/task_ledger.py lint and confirm openclaw-docs-intake-loop has an applied/verified record pointing to this index.

Observed:

The local evidence repeats the same pattern: OpenClaw is interesting because it can run as a personal ops/runtime, but the practical edge lives in implicit operating conventions, not only official docs. Recent agent tooling signals also point toward skills/plugins/AGENTS files becoming an extension supply chain that needs review before adoption.

Next safe action:

Create a local plugin_intake_check or manifest template that classifies unknown skills/plugins by instruction, worker, enforcement, and connection layers, then wires that manifest into future heartbeat/plugin-install decisions.

Notify: no — this is internal synthesis and ledger closure. It is useful groundwork, but not a fresh phone-worthy finding by itself.

Operating Buckets

1. Runtime and model routing

Signal:

Discord notes repeatedly compare Codex, Claude Code, Chrome/browser operation, and which agent should do which work.

Implication:

OpenClaw should preserve "which runtime handled this and why" in artifacts when the choice matters. Otherwise later patrols cannot distinguish model limits, browser-profile limits, and missing instruction.

Local rule to steal:

For nontrivial automation reports, add a short runtime note only when it changes the diagnosis: model, browser/profile, app connector, cron, or local script.

2. Skills and extension supply chain

Signal:

Recent Codex/Claude/OpenClaw material converges on the same layers: instructions, subagents/workers, deterministic guards/hooks, and external connections/MCP/apps.

Implication:

Unknown skills/plugins are not just prompt files. They can carry scripts, secrets access, browser profiles, hooks, MCP servers, and external writes.

Local rule to steal:

Before adding an unknown extension, classify it by:

3. Heartbeat as editor, not monitor

Signal:

The useful heartbeat pattern is not "alive" or raw patrol output. It is bounded editorial work: summarize, decide, patch a guard, or save a context artifact.

Implication:

Heartbeat lanes should continue suppressing internal hygiene unless the message itself helps 健人くん decide quickly.

Local rule to steal:

Every heartbeat artifact needs a prediction, observed evidence, next safe action, and notify yes/no. Weak findings should stay local.

4. Docs gap into View/report shape

Signal:

The recurring Discord complaint is less "docs do not exist" and more "the real operating pattern is scattered across docs, Discord, memory, local scripts, and examples."

Implication:

The right artifact is often an index or View that maps use cases to exact source links, local scripts, and failure modes.

Local rule to steal:

When a patrol sees the same docs/intel topic twice, choose one of:

MGR / EM Pass

Himeno self-check:

The task had plenty of evidence but no applied synthesis. Repeating "make an ops index" in the ledger was itself becoming the stale behavior.

MGR challenge:

Stop collecting. Produce the smallest index that makes the existing evidence usable and update the task ledger with applied/verified evidence.

EM final instruction:

Accept this pass only if the task ledger points to this index and a verification command passes. Do not notify owner unless the summary contains fresh, actionable source-backed value beyond internal cleanup.

質問したい箇所を選択
この箇所について質問
✓ 質問を送信しました