obsidian-vault/sources/bridge-workspace-4c6d747f-memory-2026-10-02-a398fd1f.md
2026-10-15 02:00:01 +00:00

202 lines
23 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

---
pageType: source
id: source.bridge.workspace-4c6d747f.memory-2026-10-02-a398fd1f
title: "Memory Bridge (main, coder-agent, evaluator-agent, humanizer-agent,
setup-agent): 2026-10-02"
sourceType: memory-bridge
sourcePath: /home/owner/.openclaw/workspace/memory/2026-10-02.md
bridgeRelativePath: memory/2026-10-02.md
bridgeWorkspaceDir: /home/owner/.openclaw/workspace
bridgeAgentIds:
- main
- coder-agent
- evaluator-agent
- humanizer-agent
- setup-agent
status: active
updatedAt: 2026-10-02T18:28:42.071Z
---
# Memory Bridge (main, coder-agent, evaluator-agent, humanizer-agent, setup-agent): 2026-10-02
## Bridge Source
- Workspace: `/home/owner/.openclaw/workspace`
- Relative path: `memory/2026-10-02.md`
- Kind: `markdown`
- Agents: main, coder-agent, evaluator-agent, humanizer-agent, setup-agent
- Updated: 2026-10-02T18:28:42.071Z
## Content
```markdown
# 2026-10-02
## LCM 1.1.1 upgrade — verification results (08:08–08:30 MSK)
User shared an upgrade report (1.1.0→1.1.1 done yesterday 10-01 ~20:48 MSK via in-process replacement); I verified it live (read-only). Confirmed:
- lossless-claw 1.1.1 enabled, `plugins.slots.contextEngine: lossless-claw`, `plugins.entries.lossless-claw.hooks.allowConversationAccess: true` — config OK.
- Engine initialized today 07:17 and 07:18 MSK (`lossless-claw-2026-10-02.log`); compaction ran cleanly at 08:08 MSK (conversation 17 = dashboard session). No read-only errors today.
- lcm.db: 30 summaries. Zombie pending_compaction_batches: 2 (conv 3, conv 14, status `planning`) — harmless, matches report.
**Corrections to the report (important for future INFRASTRUCTURE.md update):**
1. **Summarization model = primary local model, not glm-5.2.** Plugin log today: `[lcm] Compaction summarization model: ollama-local/qwen3.8-27b-q5xl (default)`. The `glm-5.2` rows in `conversation_compaction_telemetry` (2 rows) are the *conversation* model of conv 26 from 09-29, not LCM's summarization model. `summaryModel: ""` → inherits primary model; to change: explicit `plugins.entries.lossless-claw.config.summaryModel`.
2. **Trust flag still `provenance-invalid`** in live `openclaw plugins list --json`, BUT it does not gate runtime registration for non-channel plugins. Yesterday's "read-only plugin registration" errors (19:12 & 20:48 MSK) occurred inside the 1.1.0→1.1.1 replacement window; after `Plugin replacement applied` registration is full runtime, today's init+compaction clean. **My earlier verdict "catalog patch (variant A) required" is RETRACTED — no dist patch needed** (dist catalog files unpatched, mtime 10-01 17:11, 0 lossless mentions).
3. **TODO "5 plaintext secrets in openclaw.json" is stale** — scan 10-02: 0 plaintext token/key/secret/password values; all ENVREF (gateway.auth.token, channels.telegram.accounts.862485388.botToken, channels.discord.token, GITEA_ACCESS_TOKEN). Secrets live in `/home/owner/.openclaw/.env` (11 vars; TELEGRAM_BOT_TOKEN len 46).
**Open items:**
- User asked for a recommendation; I proposed **D** = fix up INFRASTRUCTURE.md with the corrections above (A dropped, B/C deferred). **Awaiting explicit user approval** (AGENTS.md §1: file modification needs "yes").
- Cleanup of `.disabled-martian-engineering-lossless-claw-fde018f0ba` (old 1.1.0 install dir, renamed 10-01 22:25 MSK): after 24h = after 22:25 MSK today.
- Telegram channel is account `862485388` (enabled, botToken via .env ENVREF) — note for future channel work.
## LCM config restored to 09-27 baseline (15:20 MSK) — supersedes 08:08 correction #1
Per Sergey's decision, the 09-29 LCM tuning (0.75→0.40, 64→32, 24000→12000) was REVERTED; original 09-27 install restored from Config Git `openclaw.json @ ddba328`: `contextThreshold=0.75`, `freshTailCount=64`, `freshTailMaxTokens=24000`, `summaryModel=largeFileSummaryModel=expansionModel=zai/glm-5.2`, `enableSummaryThinking=true`, plus session filters and `stripInjectedContextTags`.
**This supersedes the 08:08 MSK "correction #1"** (which said summarization model = primary local model with glm-5.2 rows being a misread): the live config now carries an EXPLICIT `summaryModel=zai/glm-5.2` override. Verified after gateway restart: plugin log `Compaction summarization model: zai/glm-5.2 (override)`, DB compaction rows model=glm-5.2. Hot-apply initially failed (retained work would not settle behind this session) — the restored config took effect at the user's manual `openclaw gateway restart` (~15:58/18:44 MSK window depending on log line). num_ctx=131072 unchanged.
## Delegation bug FIXED + verified (18:30–19:34 MSK)
Root cause of "No API provider registered for api: ollama" (inference probe on subagent spawn, broke after 2026.9.7 update): the `api: "ollama"` adapter did not register in CLI/probe contexts. Fix applied 18:30 MSK: `models.providers.ollama-local` → `api: "openai-completions"`, `baseUrl: http://192.168.1.200:11434/v1`, keep `injectNumCtxForOpenAICompat: true`; duplicate `ollama` provider deleted (null). Backup: `~/.openclaw/openclaw.json.bak.20261002-pre-ollama-api-switch`. Gateway restarted by user (~18:41 MSK "да").
Verification (3 spawns, all `resolvedProvider=ollama-local`, `modelApplied=true`):
1. setup-agent probe → "OK". 2. coder-agent recheck → transport fine (25 tool-calls executed+delivered) but its `tools.profile="minimal"` has NO exec by design. 3. setup-agent recheck-2 (profile "full") → exec worked: `DELEG-OK`, os-release shows **Debian GNU/Linux 13 (trixie)** on clawpi (Pi 5).
**Subagent tool profiles (from live openclaw.json, 19:3x):** coder-agent=`minimal` (no exec), evaluator-agent=`coding`, setup-agent=`full`, humanizer-agent=`messaging`; `agents.defaults.tools` empty.
**OPEN ASK to Sergey (unanswered at flush time):** whether to change coder-agent profile `minimal` → `coding` so it gets exec for shell-heavy coding tasks. Do not change without explicit yes.
Docs: MEMORY.md final-state line updated ("подтверждено 19:34 MSK"); CHANGELOG [VERIFY] 19:34 MSK + [CONFIG] 18:30 MSK; INFRASTRUCTURE.md §49.7.
## Wiki-ingest event 43d023c919493bdcba10c8df: timeout + recovery (19:36–20:01 MSK)
Triggered by MY MEMORY.md edit. First setup-agent spawn FAILED: I passed `runTimeoutSeconds: 900`, which cut the run at ~1039s before complete. **Lesson: for local-27B subagent runs, never override runTimeoutSeconds below the configured default (3600s) — `openclaw wiki ingest` CLI alone takes 1–2 min, plus slow tool-call latency.** Recovery per wiki-ingest-event skill: `fail <event> <claim_token>` (released lease, dead-letter) → `rearm` → fresh `claim` (token 9ea0c0aa…, lease 7200s) → retry spawn WITHOUT runTimeoutSeconds override → completed in 449s, `complete` ok, `fire:false up_to_date`. Result: pageId `source.memory`, 13865 bytes, ingestedAt 2026-10-02T16:55:26.105Z, Content-block MATCH, stale-fact sweep clean (no live claims asserting "delegation broken"), exactly one CHANGELOG [WIKI_INGEST] 20:00 MSK entry.
## lossless-claw full re-verification (20:06–20:08 MSK) — all green
- Plugin: enabled, `threshold=0.75`, deferred, summarization `zai/glm-5.2 (override)`; last loads 12:48Z & 15:44Z.
- Live compaction working: 17:06Z (20:06 MSK) force-compact 46533→29166 tokens, sum_b5e5f857dc1f34c9, 9.6s.
- Recall: `lcm_grep` full_text "openai-completions" → 3 valid hits (conv 32). DB: 124 summaries, 3917 messages.
- Errors: all 98 `level=error` lines are ≤13:01Z (old generation, pre-fix ollama provider bug + spend guard). **Post-restart: 0 errors.**
- Known non-failure: conv 32 (skill-workshop) hit `summary spend guard` (24/24 zai calls) + 120s wall-clock sweep deadline — protective backoff, compaction partially done (−400k tokens) and resumes after backoff.
## Other durable notes (2026-10-02)
- **tool_call arg degradation recurred today** (openclaw-local qwen3.8-27b-q5xl): empty/missing args for `sessions_spawn` and `lcm_grep` (`query` vs required `pattern`), XML fragment errors earlier. Reliable recovery: `catalog.describe` → exact schema → retry with explicit full args. Matches openclaw-tool-call-failure-triage skill. Happened ~8x in one turn before succeeding — do not treat early failures as tool absence.
- Open items carried forward: (a) Sergey's pending "yes/no" on coder-agent profile change; (b) cleanup of `.disabled-martian-engineering-lossless-claw-fde018f0ba` (24h since 10-01 22:25 MSK expiry = tonight); (c) the 08:08 "option D" INFRASTRUCTURE.md corrections (summarization-model note, etc.) — premise of #1 invalidated by the 15:20 restore; if revisiting INFRASTRUCTURE.md, only corrections #2/#3 + §49.7 delegation entry matter.
# 2026-10-02
## LCM 1.1.1 upgrade — verification results (08:08–08:30 MSK)
User shared an upgrade report (1.1.0→1.1.1 done yesterday 10-01 ~20:48 MSK via in-process replacement); I verified it live (read-only). Confirmed:
- lossless-claw 1.1.1 enabled, `plugins.slots.contextEngine: lossless-claw`, `plugins.entries.lossless-claw.hooks.allowConversationAccess: true` — config OK.
- Engine initialized today 07:17 and 07:18 MSK (`lossless-claw-2026-10-02.log`); compaction ran cleanly at 08:08 MSK (conversation 17 = dashboard session). No read-only errors today.
- lcm.db: 30 summaries. Zombie pending_compaction_batches: 2 (conv 3, conv 14, status `planning`) — harmless, matches report.
**Corrections to the report (important for future INFRASTRUCTURE.md update):**
1. **Summarization model = primary local model, not glm-5.2.** Plugin log today: `[lcm] Compaction summarization model: ollama-local/qwen3.8-27b-q5xl (default)`. The `glm-5.2` rows in `conversation_compaction_telemetry` (2 rows) are the *conversation* model of conv 26 from 09-29, not LCM's summarization model. `summaryModel: ""` → inherits primary model; to change: explicit `plugins.entries.lossless-claw.config.summaryModel`.
2. **Trust flag still `provenance-invalid`** in live `openclaw plugins list --json`, BUT it does not gate runtime registration for non-channel plugins. Yesterday's "read-only plugin registration" errors (19:12 & 20:48 MSK) occurred inside the 1.1.0→1.1.1 replacement window; after `Plugin replacement applied` registration is full runtime, today's init+compaction clean. **My earlier verdict "catalog patch (variant A) required" is RETRACTED — no dist patch needed** (dist catalog files unpatched, mtime 10-01 17:11, 0 lossless mentions).
3. **TODO "5 plaintext secrets in openclaw.json" is stale** — scan 10-02: 0 plaintext token/key/secret/password values; all ENVREF (gateway.auth.token, channels.telegram.accounts.862485388.botToken, channels.discord.token, GITEA_ACCESS_TOKEN). Secrets live in `/home/owner/.openclaw/.env` (11 vars; TELEGRAM_BOT_TOKEN len 46).
**Open items:**
- User asked for a recommendation; I proposed **D** = fix up INFRASTRUCTURE.md with the corrections above (A dropped, B/C deferred). **Awaiting explicit user approval** (AGENTS.md §1: file modification needs "yes").
- Cleanup of `.disabled-martian-engineering-lossless-claw-fde018f0ba` (old 1.1.0 install dir, renamed 10-01 22:25 MSK): after 24h = after 22:25 MSK today.
- Telegram channel is account `862485388` (enabled, botToken via .env ENVREF) — note for future channel work.
## LCM config restored to 09-27 baseline (15:20 MSK) — supersedes 08:08 correction #1
Per Sergey's decision, the 09-29 LCM tuning (0.75→0.40, 64→32, 24000→12000) was REVERTED; original 09-27 install restored from Config Git `openclaw.json @ ddba328`: `contextThreshold=0.75`, `freshTailCount=64`, `freshTailMaxTokens=24000`, `summaryModel=largeFileSummaryModel=expansionModel=zai/glm-5.2`, `enableSummaryThinking=true`, plus session filters and `stripInjectedContextTags`.
**This supersedes the 08:08 MSK "correction #1"** (which said summarization model = primary local model with glm-5.2 rows being a misread): the live config now carries an EXPLICIT `summaryModel=zai/glm-5.2` override. Verified after gateway restart: plugin log `Compaction summarization model: zai/glm-5.2 (override)`, DB compaction rows model=glm-5.2. Hot-apply initially failed (retained work would not settle behind this session) — the restored config took effect at the user's manual `openclaw gateway restart` (~15:58/18:44 MSK window depending on log line). num_ctx=131072 unchanged.
## Delegation bug FIXED + verified (18:30–19:34 MSK)
Root cause of "No API provider registered for api: ollama" (inference probe on subagent spawn, broke after 2026.9.7 update): the `api: "ollama"` adapter did not register in CLI/probe contexts. Fix applied 18:30 MSK: `models.providers.ollama-local` → `api: "openai-completions"`, `baseUrl: http://192.168.1.200:11434/v1`, keep `injectNumCtxForOpenAICompat: true`; duplicate `ollama` provider deleted (null). Backup: `~/.openclaw/openclaw.json.bak.20261002-pre-ollama-api-switch`. Gateway restarted by user (~18:41 MSK "да").
Verification (3 spawns, all `resolvedProvider=ollama-local`, `modelApplied=true`):
1. setup-agent probe → "OK". 2. coder-agent recheck → transport fine (25 tool-calls executed+delivered) but its `tools.profile="minimal"` has NO exec by design. 3. setup-agent recheck-2 (profile "full") → exec worked: `DELEG-OK`, os-release shows **Debian GNU/Linux 13 (trixie)** on clawpi (Pi 5).
**Subagent tool profiles (from live openclaw.json, 19:3x):** coder-agent=`minimal` (no exec), evaluator-agent=`coding`, setup-agent=`full`, humanizer-agent=`messaging`; `agents.defaults.tools` empty.
**OPEN ASK to Sergey (unanswered at flush time):** whether to change coder-agent profile `minimal` → `coding` so it gets exec for shell-heavy coding tasks. Do not change without explicit yes.
Docs: MEMORY.md final-state line updated ("подтверждено 19:34 MSK"); CHANGELOG [VERIFY] 19:34 MSK + [CONFIG] 18:30 MSK; INFRASTRUCTURE.md §49.7.
## Wiki-ingest event 43d023c919493bdcba10c8df: timeout + recovery (19:36–20:01 MSK)
Triggered by MY MEMORY.md edit. First setup-agent spawn FAILED: I passed `runTimeoutSeconds: 900`, which cut the run at ~1039s before complete. **Lesson: for local-27B subagent runs, never override runTimeoutSeconds below the configured default (3600s) — `openclaw wiki ingest` CLI alone takes 1–2 min, plus slow tool-call latency.** Recovery per wiki-ingest-event skill: `fail <event> <claim_token>` (released lease, dead-letter) → `rearm` → fresh `claim` (token 9ea0c0aa…, lease 7200s) → retry spawn WITHOUT runTimeoutSeconds override → completed in 449s, `complete` ok, `fire:false up_to_date`. Result: pageId `source.memory`, 13865 bytes, ingestedAt 2026-10-02T16:55:26.105Z, Content-block MATCH, stale-fact sweep clean (no live claims asserting "delegation broken"), exactly one CHANGELOG [WIKI_INGEST] 20:00 MSK entry.
## lossless-claw full re-verification (20:06–20:08 MSK) — all green
- Plugin: enabled, `threshold=0.75`, deferred, summarization `zai/glm-5.2 (override)`; last loads 12:48Z & 15:44Z.
- Live compaction working: 17:06Z (20:06 MSK) force-compact 46533→29166 tokens, sum_b5e5f857dc1f34c9, 9.6s.
- Recall: `lcm_grep` full_text "openai-completions" → 3 valid hits (conv 32). DB: 124 summaries, 3917 messages.
- Errors: all 98 `level=error` lines are ≤13:01Z (old generation, pre-fix ollama provider bug + spend guard). **Post-restart: 0 errors.**
- Known non-failure: conv 32 (skill-workshop) hit `summary spend guard` (24/24 zai calls) + 120s wall-clock sweep deadline — protective backoff, compaction partially done (−400k tokens) and resumes after backoff.
## Sergey question: does the "glm + qwen3.8" LCM setup work? (20:34–20:36 MSK)
Answered with live evidence — the current production pair is **ollama-local/qwen3.8-27b-q5xl (primary, free/local) + zai/glm-5.2 (LCM summarizer override, paid z.ai)**, and it IS working:
- `lcm.db` `summaries.model` breakdown: **76 summaries authored by glm-5.2, 49 by qwen3.8** (qwen3.8 rows = earlier periods when summarizer inherited the local primary). Per-summary model recorded in DB — proves glm-5.2 is actually writing summaries, not just sitting in config.
- `conversation_compaction_telemetry.model`: glm-5.2 (5) / qwen3.8 (19) / null (1) — recent conversations compacted with glm-5.2.
- Plugin log override line + 20:06 live compaction (sum_b5e5f857…, model glm-5.2) already logged above.
**Important disambiguation:** user wrote "glm 5.5" — **no glm-5.5 exists in the config** (only glm-5.2). Treated as likely typo for 5.2; if he really wants 5.5 that's a new config change (availability + allowModelOverride check first).
**Cost guardrail to remember:** zai spend guard = 24 summarization calls per window; big conversations (conv 32) hit it today. Cloud spend (z.ai) goes ONLY to LCM summaries; everything else runs local. If summaries get frequent, the guard will throttle more often — watch for `summary spend guard opened` warnings.
Also: user sent "ау" (ping) at 20:34 — no content, no action.
## Other durable notes (2026-10-02)
- **tool_call arg degradation recurred today** (openclaw-local qwen3.8-27b-q5xl): empty/missing args for `sessions_spawn` and `lcm_grep` (`query` vs required `pattern`), XML fragment errors earlier. Reliable recovery: `catalog.describe` → exact schema → retry with explicit full args. Matches openclaw-tool-call-failure-triage skill. Happened ~8x in one turn before succeeding — do not treat early failures as tool absence.
- Open items carried forward: (a) Sergey's pending "yes/no" on coder-agent profile change; (b) cleanup of `.disabled-martian-engineering-lossless-claw-fde018f0ba` (24h since 10-01 22:25 MSK expiry = tonight); (c) the 08:08 "option D" INFRASTRUCTURE.md corrections (summarization-model note, etc.) — premise of #1 invalidated by the 15:20 restore; if revisiting INFRASTRUCTURE.md, only corrections #2/#3 + §49.7 delegation entry matter.
# 2026-10-02 — долговременные записи (append в 21:26 MSK)
## Делегирование: баг починен и подтверждён (18:30–19:34 MSK)
- Корень: `ollama-local` был на `api: "ollama"` — adapter не регистрировался в CLI-контексте (inference probe для subagent spawn падал с `No API provider registered for api: ollama`).
- Фикс (setup-agent, config patch, hot-apply): `api` → `openai-completions`, `baseUrl` → `http://168.168.1.200:11434/v1` (точнее `http://192.168.1.200:11434/v1`), `injectNumCtxForOpenAICompat: true`; дубль-провайдер `ollama` удалён (null). Бэкап: `~/.openclaw/openclaw.json.bak.20261002-pre-ollama-api-switch`.
- Подтверждено: 3 spawn через новый route — probe (setup-agent, OK), coder-agent (25 tool-calls, транспорт ok), setup-agent exec-тест (`DELEG-OK 19:34`, Debian trixie). CHANGELOG [CONFIG] 18:30 + [VERIFY] 19:34. Баг §48.3 закрыт.
## Tool-профили субагентов (openclaw.json, проверено 2026-10-02)
- coder-agent: `tools.profile: "minimal"` — **нет exec по дизайну** (исправлено: не баг)
- evaluator-agent: `coding`; setup-agent: `full` (exec работает); humanizer-agent: `messaging`
- `agents.defaults.tools: {}`
- **Открытый вопрос к пользователю (без ответа, 19:40):** сменить ли coder-agent на `coding` для shell-задач. Пока нет — кодозадачи с exec делегировать setup-agent.
## LCM (lossless-claw) — состояние на 21:26 MSK
- Конфиг = baseline 09-27: `contextThreshold=0.75`, `freshTailCount=64`, `freshTailMaxTokens=24000`, `proactiveThresholdCompactionMode=deferred`, summaryModel/largeFileSummaryModel/expansionModel = `zai/glm-5.2` (override), enableSummaryThinking=true.
- Здоровье (20:06 MSK, проверено live): plugin loaded 15:44:48Z, live compaction 46533→29166 tok (sum_b5e5f857dc1f34c9, model glm-5.2, 9.6s), `lcm_grep` recall работает, DB 124 summaries (76 glm-5.2 / 49 qwen3.8) + 3917 messages.
- 0 ошибок в плагин-логе после перезагрузки 15:44:48Z (все 98 ошибок дня — до 10:0:41Z, старый generation, ollama-баг).
- Watch: conv 32 (skill-workshop-review) — zai spend guard сработал (24/24 calls, 16:37Z, backoff до 17:07Z), компакция остановлена частично (1.75M→1.35M tok), deferred-продолжение.
- **Решение Сергея (21:23): «остаёмся пока модели 5.2»** — платный zai/glm-5.2 для LCM оставляем. Локализация LCM (qwen3.8+qwen3.8) отложена.
## Отложенный план: LCM → локальный qwen3.8 (не выполнен, готовый промпт передан пользователю в чате 21:23)
6 правок в `~/.openclaw/openclaw.json` (только `openclaw config patch`, бэкап до правок):
1. `plugins.entries.lossless-claw.config.summaryModel` = `ollama-local/qwen3.8-27b-q5xl`
2. `...largeFileSummaryModel` — то же
3. `...expansionModel` — то же
4. `...llm.allowedModels` = `["ollama-local/qwen3.8-27b-q5xl","zai/glm-5.2"]`
5. `...subagent.allowedModels` — аналогично
6. Остальной блок LCM не трогать.
Не удалять: провайдер zai + models-запись `zai/glm-5.2` (alias GLM) — рычаг отката одной правкой. Модель уже в VRAM (KEEP_ALIVE=12h) — свопова нет; суммаризация будет медленнее, deferred-режим сглаживает. Проверка: строка override в plugin log + новые summaries в lcm.db с model=qwen3.8-27b-q5xl + lcm_grep.
## Wiki-ingest: event 43d023c919493bdcba10c8df закрыт (20:01 MSK)
- Дельта: строка OpenClaw version в MEMORY.md — статус делегирования (баг закрыт, подтверждено 19:34 MSK).
- Первый spawn setup-agent **таймаут на 900s** (я передал runTimeoutSeconds: 900 — ошибка спавнера; defaulgt=3600). Recovery по wiki-ingest-event skill: `fail <evt> <token> "reason"` → `rearm` → fresh `--trigger`+`claim` → повторный spawn **без** override таймаута → PASS за 449s.
- Результат: `openclaw wiki ingest MEMORY.md --agent main` → pageId source.memory, 13865 bytes, ingestedAt 2026-10-02T16:55:26Z; Content-block MATCH; stale-fact sweep — 0 живых устаревших claims («delegate routes dead», «требует тестового spawn» — только в snapshot-страницах, не трогать); 1 CHANGELOG entry [WIKI_INGEST] 20:00 MSK; `complete` ok, watermark 2026-10-02T20:01:21. Finaal state: `--trigger` → `fire:false, reason:up_to_date`.
- **Урок:** spawn на wiki-ingest не переопределять runTimeoutSeconds (CLI ingest 1–2 мин + sweep + renew) — давать стандартные 3600s.
```
## Notes
<!-- openclaw:human:start -->
<!-- openclaw:human:end -->
## Related
<!-- openclaw:wiki:related:start -->
- No related pages yet.
<!-- openclaw:wiki:related:end -->