← Back to Skills Marketplace
mar-context-compression
by
MarjorieBroad
· GitHub ↗
· v1.0.0
· MIT-0
77
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install mar-context-compression
Description
This skill should be used when the user asks to "compress context", "summarize conversation history", "implement compaction", "reduce token usage", or mentio...
Usage Guidance
Do not install or run this skill until the author clarifies and documents external-network behavior and credentials. Specifically: (1) ask the author to explain why SKILLBOSS_API_KEY and calls to api.heybossai.com are required and what data is sent; (2) request that required env vars be declared in the skill metadata and that a privacy/security policy be provided; (3) review the compression_evaluator.py code yourself (or have a trusted engineer do so) to confirm no additional endpoints or hidden behaviors exist; (4) if you must test, run the skill in a sandboxed environment without any sensitive credentials and monitor outbound network traffic; and (5) consider removing or replacing the external-API calls if you prefer an offline/local evaluation option.
Capability Analysis
Type: OpenClaw Skill
Name: mar-context-compression
Version: 1.0.0
The skill bundle includes a Python script (`scripts/compression_evaluator.py`) that extracts conversation history, file paths, and decision logic to send to an external API endpoint (`api.heybossai.com`) for evaluation. While this behavior is consistent with the 'LLM-as-a-judge' framework described in the documentation (`references/evaluation-framework.md`), the automated transmission of potentially sensitive session data and the use of an environment variable (`SKILLBOSS_API_KEY`) for external communication represent significant data exfiltration risks that are not explicitly highlighted as a privacy concern in the main `SKILL.md` instructions.
Capability Tags
Capability Assessment
Purpose & Capability
The skill claims to be an instruction-only context-compression helper, but the bundle includes a Python evaluator that calls an external LLM judge endpoint and reads SKILLBOSS_API_KEY from the environment. The SKILL.md and metadata declare no required credentials or external API usage, so the presence of an external API client and hidden credential requirement is disproportionate to the stated purpose and not documented.
Instruction Scope
SKILL.md focuses on summarization strategies and probe evaluations but does not instruct the agent to call external services or require credentials. The Python script, however, will transmit probe data, compressed context, and responses to https://api.heybossai.com/v1/pilot using the SKILLBOSS_API_KEY. That means conversation history and artifact descriptions could leave the host if the script runs — this behavior is not described or justified in the SKILL.md.
Install Mechanism
There is no install spec (lower risk from arbitrary downloads), but the included script requires the 'requests' library and network access. Because no install/runtime environment is declared, running the script may fail or unexpectedly attempt network calls. Lack of a declared runtime or dependency list is sloppy and increases operational risk.
Credentials
The Python code reads os.environ['SKILLBOSS_API_KEY'] but the skill metadata lists no required env vars or primary credential. Requesting a bearer API key for a third-party service is not proportional to the documented scope unless the skill explicitly states that it will outsource evaluation to that service. Hidden credential requirements create risk of accidental credential reuse or exposure.
Persistence & Privilege
The skill is not always-enabled and does not request system-wide persistent privileges. It does not modify other skills' configs in the provided files. The main concern is outbound network activity rather than elevated platform privileges.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install mar-context-compression - After installation, invoke the skill by name or use
/mar-context-compression - Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
Initial release of context-compression skill.
- Provides guidance for compressing conversation history and context in long-running agent sessions or large codebases.
- Describes three main compression approaches: anchored iterative summarization, opaque compression, and regenerative full summaries.
- Emphasizes structuring summaries with explicit sections (session intent, files modified, decisions, next steps) to preserve critical information.
- Offers strategies for when and how to trigger compression, focusing on reducing tokens-per-task rather than tokens-per-request.
- Details evaluation methods and dimensions for measuring compression quality, especially artifact trail integrity and continuity.
- Includes practical workflows for research, planning, and implementation phases in large systems.
Metadata
Frequently Asked Questions
What is mar-context-compression?
This skill should be used when the user asks to "compress context", "summarize conversation history", "implement compaction", "reduce token usage", or mentio... It is an AI Agent Skill for Claude Code / OpenClaw, with 77 downloads so far.
How do I install mar-context-compression?
Run "/install mar-context-compression" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is mar-context-compression free?
Yes, mar-context-compression is completely free, licensed under MIT-0. You can download, install and use it at no cost.
Which platforms does mar-context-compression support?
mar-context-compression is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created mar-context-compression?
It is built and maintained by MarjorieBroad (@marjoriebroad); the current version is v1.0.0.
More Skills