Files

Selig 4c966a3ad2 Initial commit: OpenClaw Skill Collection

6 custom skills (assign-task, dispatch-webhook, daily-briefing,
task-capture, qmd-brain, tts-voice) with technical documentation.
Compatible with Claude Code, OpenClaw, Codex CLI, and OpenCode.

2026-03-13 10:58:30 +08:00

1.7 KiB

Raw Blame History

Context

OpenClaw's "Context" represents everything the model receives for a run, constrained by the model's token limit. It encompasses the system prompt, conversation history, tool calls, and attachments.

Key Components

The system breaks down into several parts:

System prompt (built by OpenClaw): includes rules, tools, skills, time/runtime data, and workspace files
Conversation history: user and assistant messages within the session
Tool results and attachments: command outputs, file reads, media

Context Inspection Commands

Users can monitor context usage via:

Command	Description
`/status`	Shows window fullness and session settings
`/context list`	Displays injected files with approximate token counts
`/context detail`	Provides granular breakdown by file and tool schemas
`/usage tokens`	Appends token usage to replies
`/compact`	Summarizes older messages to free space

What Counts Toward the Window

Everything sent to the model consumes tokens:

System prompt sections
Conversation history
Tool calls and results
Attachments and transcripts
Compaction summaries
Provider wrappers

Workspace File Injection

OpenClaw automatically injects these files (if present):

AGENTS.md
SOUL.md
TOOLS.md
IDENTITY.md
USER.md
HEARTBEAT.md
BOOTSTRAP.md

Files exceeding bootstrapMaxChars (default 20,000) are truncated, with truncation status indicated in context reports.

Skills and Tools

Skills include metadata in the system prompt but load instruction details only when the model calls /read on the skill file.

Tools incur dual costs:

Text descriptions in the system prompt
JSON schemas that count toward context separately

1.7 KiB Raw Blame History