Summary

Add support for invoking OpenClaw tools (exec, read, write, message, MCP skills, etc.) directly from a chat message or slash command, bypassing the LLM inference step entirely.

Problem to solve

Currently, every tool call in OpenClaw must go through an LLM turn: the user sends a message → the model decides to call a tool → the tool executes → the model summarizes the result. This round-trip is unnecessary for deterministic, well-defined operations where the intent is unambiguous.
Common pain points:

• Running a known shell command ("show me disk usage") burns tokens just to trigger a one-liner
• Calling a specific MCP skill tool with fixed inputs requires a full LLM turn
• Scheduled cron jobs using systemEvent can inject text, but can't directly fire a tool
• Power users who know exactly which tool they want are forced to pay latency + cost for an LLM that adds no value in these cases

Proposed solution

Introduce a direct tool invocation path — a way to call a tool with explicit inputs, skipping model inference entirely. Possible surfaces:

1. Slash command syntax:
   /tool exec command="df -h"
   /tool read path="/etc/hosts"
   /tool message target="alice" message="hello"
2. CLI flag:
   openclaw agent --tool exec --args '{"command": "df -h"}'
3. Cron payload kind:
   {
   "kind": "toolCall",
   "tool": "exec",
   "args": { "command": "df -h" }
   }
   The result could be
   returned directly to the session as a synthetic tool result (injected into conversation history), or delivered to the channel as a formatted reply.
   Use cases

• Run shell commands on a schedule without burning tokens on LLM inference
• Directly invoke a specific MCP skill tool (e.g., weather lookup, KM search) with known parameters
• Build lightweight automations where the tool call is deterministic and intent is explicit
• Reduce latency and API cost for high-frequency or time-sensitive operations

Alternatives considered

• Thin system prompt + forced tool call: Works but still requires a full LLM turn and wastes tokens on a trivial decision
• Cron agentTurn with explicit instructions: Closest current workaround, but spins up a full isolated session for what should be a single function call
• Direct shell / API calls outside OpenClaw: Bypasses the tool ecosystem entirely and loses observability, auth handling, and channel delivery

Impact

• Affected users: Power users, automation builders, cron-heavy deployments
• Frequency: Every time a user wants a deterministic action that doesn't need AI reasoning
• Consequence: Unnecessary token cost, added latency, and cognitive overhead of writing "please run X" instead of just running X

Evidence/examples

No response

Additional information

No response