| pageType |
id |
title |
sourceType |
sourcePath |
bridgeRelativePath |
bridgeWorkspaceDir |
bridgeAgentIds |
status |
updatedAt |
| source |
source.bridge.workspace-4c6d747f.memory-2026-08-28-d96b9d18 |
Memory Bridge (main): 2026-08-28 |
memory-bridge |
/home/owner/.openclaw/workspace/memory/2026-08-28.md |
memory/2026-08-28.md |
/home/owner/.openclaw/workspace |
|
active |
2026-08-28T02:53:58.649Z |
Memory Bridge (main): 2026-08-28
Bridge Source
- Workspace:
/home/owner/.openclaw/workspace
- Relative path:
memory/2026-08-28.md
- Kind:
markdown
- Agents: main
- Updated: 2026-08-28T02:53:58.649Z
Content
# 2026-08-28 — durable memory notes
## Forensic analysis: OpenClaw/Ollama context growth & dreaming-sweep timeouts (read-only, 2026-08-28 05:17–05:49 MSK)
### Task
Sergey requested strictly read-only forensic analysis of context growth and model-call performance to explain dreaming-sweep timeouts. No changes allowed. Deliverables: static/dynamic context decomposition, per-call metrics, context_tokens→prefill correlation, WooCommerce MCP per-call inclusion check, final run table, root-cause A/B/C/D.
### Key measured facts
- **Ollama /api/ps:** qwen3.8-27b-q5xl:latest loaded, ~28GB VRAM, context_length=131072, expiry +12h (keep_alive confirmed warm). Q4_K_S, 27.3B, family qwen35.
- **Dell 192.168.1.200 SSH unreachable** (port 22 closed) → no direct Ollama server logs; prompt_eval_duration NOT persisted anywhere locally. OpenClaw maps prompt_eval_count→usage.input, eval_count→usage.output only.
- **Prompt eval time unavailable** in local evidence; used "model wall time" proxy (gap between last tool result and next assistant message).
### Crashed sweep 28.08 (session 694c06b0-c6d3-467a-ab55-bd44cb329531, cron timeout 900481ms, "last phase: model-call-started")
- context.compiled: 49 tools, WooCommerce MCP = 0 tools, Context7 = 6 tools (4917 B), core/other 43 (51667 B); systemPrompt 25674 chars, prompt 446 chars, messages 0.
- Per-call: #1 in=25696/out=1943/51.9s; #2 30828/1761/34.4s; #3 39045/1284/27.7s; #4 53575/9287/184.5s; #5 64657/9104/184.9s; #6 79139/5482/120.3s; #7 83036/7264/150.8s; #8 84025/2166/124.4s; #9 abort stub (0/0).
- Totals: input=460001, output=38291. Cumulative model wall ≈ 879s + tool time → hit 900s cron timeout DURING a model call. NOT a stuck tool (timedOutDuringToolExecution=false).
### Aborted sweep 27.08 (session fbb58067-f86f-40a2-b2be-e758ebddbad7, AbortError ~03:13:57, duration 839392ms)
- Per-call: #1 in=25686/out=2511/60.2s; #2 29371/1740/33.5s; #3 38442/898/21.0s; #4 44644/8096/155.7s; #5 53897/9562/189.2s.
- After call #5 (03:07:41) — 3 toolCalls made, NO assistant response until abort. Gateway showed "stalled session", "blocked_tool_call", lastProgress=tool:exec:ended aging 135s→375s. → tool/exec hang (C) in this run. Found stuck wiki_search as candidate.
### check-dreaming 27.08 21:50 (session 5e3ec90f, ok, 63s)
- call#1 in=49763/out=1037/46.3s; call#2 in=53058/out=650/15.4s.
### Current main interactive session (2cbea839)
- First call input=58643 (baseline ~57–58K confirmed), second call input=65638 (+6995).
- Trajectory 0d7bf512: tools=65 — WooCommerce MCP tools PRESENT in main interactive session; ABSENT in cron isolated sweep sessions.
### WooCommerce MCP inclusion verdict
- NOT universal: present in main interactive session (65 tools incl. WooCommerce), absent from cron isolated sweep sessions (49 tools, 0 WooCommerce). Cron isolated sessions get a reduced tool set.
- Within a session, tool schemas are re-sent on every model call (standard OpenClaw request build; context.compiled once but request includes tools each call — confirmed via tool list in compiled context and stable per-call input baseline).
### Root-cause classification (preliminary)
- 28.08 crash = **B (large context/prefill + decode)**: cumulative 900s+ of model time at 25K→84K growing context, warm model, no tool hang.
- 27.08 abort = **C (stuck tool — exec/wiki_search hang)** after 5 healthy calls.
- Overall: **D (combination)** across the failure series; no evidence of A (model load/eviction) — keep_alive 12h keeps 27B resident, first-call latencies (51.9s/60.2s) consistent with warm prefill of ~25K, not load.
### Open items (not finished before compaction)
- Older successful sweeps 17–22.08 (candidate files found by mtime ~00:00 UTC, e.g. 9f7e4cbc, 3ab0402f, 454857dd 17.08) — per-call extraction pending.
- coder/evaluator context.compiled inspection pending.
- Final table assembly pending (draft rows exist for all key runs above).
Notes
Related