メニュー
Monitors memory compaction for failures and enforces a three-level fallback chain — normal, aggressive, deterministic truncation — ensuring compaction always makes forward progress.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | compaction-resilience-guard |
| version | 1.0 |
| category | openclaw-native |
| description | Monitors memory compaction for failures and enforces a three-level fallback chain — normal, aggressive, deterministic truncation — ensuring compaction always makes forward progress. |
| stateful | true |
Memory compaction can fail silently: the LLM produces empty output, summaries that are larger than their input, or garbled text. When this happens, compaction stalls and context overflows.
Compaction Resilience Guard enforces a three-level escalation chain inspired by lossless-claw:
| Level | Strategy | When used |
|---|---|---|
| L1 — Normal | Standard summarization prompt | First attempt |
| L2 — Aggressive | Low temperature, reduced reasoning, shorter output target | After L1 failure |
| L3 — Deterministic | Pure truncation: keep first N + last N lines, drop middle | After L2 failure |
This ensures compaction always makes progress — even if the LLM is broken.
python3 guard.py --check # Validate recent compaction outputs
python3 guard.py --check --file <summary.yaml> # Check a specific summary file
python3 guard.py --simulate <text> # Run the 3-level chain on sample text
python3 guard.py --report # Show failure/escalation history
python3 guard.py --status # Last check summary
python3 guard.py --format json # Machine-readable output
The guard detects these compaction failures:
| Failure | How detected | Action |
|---|---|---|
| Empty output | Summary length < 10 chars | Escalate to next level |
| Inflation | Summary tokens > input tokens | Escalate to next level |
| Garbled text | Entropy score > 5.0 (random chars) | Escalate to next level |
| Repetition | Same 20+ char phrase repeated 3+ times | Escalate to next level |
| Truncation marker | Contains [FALLBACK] or [TRUNCATED] | Record as L3 usage |
| Stale | Summary unchanged from previous run | Flag for review |
Step 1 — Check recent compaction outputs
python3 guard.py --check
Validates all summary nodes in memory-dag-compactor state. Reports failures by level and whether escalation was needed.
Step 2 — Simulate the fallback chain
python3 guard.py --simulate "$(cat long-text.txt)"
Runs the 3-level chain on sample text to test that each level produces valid output.
Step 3 — Review escalation history
python3 guard.py --report
Shows how often each level was used. High L2/L3 usage indicates the primary summarization prompt needs improvement.
Failure counts, escalation history, and per-summary validation results stored in ~/.openclaw/skill-state/compaction-resilience-guard/state.yaml.
Fields: last_check_at, level_usage, failures, check_history.
Handles PR review feedback by fetching comments, grouping issues, fixing one group at a time, and verifying before replies.
Reviews whether a skill will trigger reliably, guide useful behavior, avoid overlap, and produce testable outcomes.
Diagnoses OpenClaw provider, fallback, channel, MCP, and gateway config issues with read-only scans and stateful summaries.
Tracks required validation gates, records pass/fail/waived results, and reports readiness before task completion.
YAML-based delegation grant ledger — issues, validates, and tracks scoped permission grants for sub-agent expansions with token budgets and auto-expiry.
Walks the memory DAG to recall detailed context on demand — query, expand, and assemble cited answers from hierarchical summaries without re-reading raw transcripts.