claw-compactor

What to consider before installing: - Coherence: The code implements what the skill claims (workspace compression, session transcript observation extraction). That part is internally consistent. - Sensitive file access: The skill will read session transcripts from ~/.openclaw/sessions and will write memory/.observed-sessions.json and memory/observations/ inside the target workspace. Those session files can contain private data — back up and inspect them first. - Prompt-injection risk: The compressed_context module and SKILL.md suggest prepending decompression instructions to model/system prompts. That is functionally necessary to make compressed notation readable by an LLM, but it is also a high-risk operation because it alters model steering. Do NOT automatically modify your agent’s system prompt or global model configuration without manual review. If you must use decompression instructions, keep them minimal and review their exact text. - LLM use paths: The code contains LLM prompt templates and both rule-based and LLM-driven extraction paths. By default many commands use rule-based functions, but some utilities can generate prompts for LLMs. Before enabling any LLM-based compression, verify there are no unintended network endpoints and confirm which prompts/data would be sent out. - Run safely first: Use the --dry-run/benchmark modes and run on a copy of your workspace. Inspect generated artifacts (memory/.observed-sessions.json, memory/.codebook.json, memory/observations/) and the code paths that would call an LLM (search for code that invokes network libraries or external APIs in the omitted files). - Check for hidden characters: Scan README/SKILL.md and compressed text for zero-width or control characters that could hide instructions (tools like `cat -v`, hexdump, or editors that show invisibles can help). - Review compressed_context behavior: aggressive 'ultra' compression removes many function words and applies abbreviations; ensure the resulting compressed text + decompression instructions preserve all facts you care about and that you trust the decompression guidance. If you are not comfortable auditing the code paths that might call an external LLM or that will modify model prompts, consider running the tool in an isolated environment or container and only using the purely rule-based (no-LLM) modes.