The Agent Skills Directory

[DATA_EXFILTRATION]: Accesses sensitive session transcripts stored in ~/.claude/projects/. These files contain a full history of user interactions, including tool calls and outputs from past sessions. While the skill description mentions redacting credentials (API keys, tokens, PEM blocks, etc.), the programmatic reading and storage of session turn data constitutes a high-risk data exposure.
[COMMAND_EXECUTION]: Frequently invokes a local CLI utility mnemos and a shell script mnemos-post-compact-inject.sh via lifecycle hooks (PreToolUse, SessionStart, Stop). These executions are used to manage a local SQLite database and automate context injection tasks.
[PROMPT_INJECTION]: Provides specific instructions to influence the agent's context compaction and summarization. It aims to ensure that its 'checkpointed' data is manually preserved by the agent, effectively overriding the default summarization behavior.
[PROMPT_INJECTION]: Implements an indirect prompt injection surface by ingesting historical session transcripts and re-injecting them into the active prompt.
Ingestion points: Historical transcript JSONL files in ~/.claude/projects/ (referenced in SKILL.md).
Boundary markers: Uses a 'CONTEXT RESTORED AFTER COMPACTION' block as a delimiter (referenced in SKILL.md).
Capability inventory: Subprocess calls to mnemos CLI and mnemos-post-compact-inject.sh (referenced in SKILL.md).
Sanitization: States that common secret patterns (API keys, JWTs, etc.) are redacted from transcripts prior to storage.

mnemos