Video Search And Summarization TOOLS.md

This TOOLS.md is the environment guide for an OpenClaw agent running inside a locked-down sandbox, driving a video search and summarisation stack that lives on the sandbox host. It is the most technical file in this collection and the best illustration of what a TOOLS.md is actually for: not a list of available commands, but the accumulated knowledge of how this environment misbehaves and what to do about it.

Roughly half the file is devoted to things that look like failures and are not. The sandbox presets an HTTP proxy that every backend call legitimately travels through, so the agent is told not to "fix" it by unsetting the variable. A permission-denied warning at shell startup is cosmetic and should be ignored. Several healthy backend endpoints return a 404 from their root path, so the file bans curl -f for reachability probes and explains the exit-code semantics that make a bare curl the right test. Each of these is a debugging session an agent would otherwise repeat every time it started work.

How the file is organised

• Sandbox host alias — why every runtime call must use the host alias rather than localhost or a literal IP, which file is the single source of truth for its value, why it cannot be persisted to a shell profile, and how to recognise a call the egress policy denied.
• Preset proxy and harmless warnings — the two non-problems described above.
• HTTP-response curl checks — when to use a strict failure flag and when not to, plus the orchestrator reachability probe and an explicit warning against adding a DNS precondition that would produce false negatives.
• Deployment — deployment is delegated to an orchestrator service; the agent is told not to run the deploy scripts itself or prompt for credentials the service already holds.
• Calling MCP tools — which orchestrator tools register natively and which must be invoked over HTTP, the three-message handshake with the reason each message is required, and the shape of a tool call.
• Mapping user intent to tool chains — a table pairing what the user said with the tool chain to run and, in its own column, what not to also run.
• Long-running deploys — the fire-and-return semantics of the deploy calls, how to poll for completion, the recommended cadence, and how to tell success from error from cancellation.
• Unknown operation ids and after a deploy completes — a two-strike rule for retrying a dead id, and which cheap checks to use once an operation is finished.
• Skills — discover and invoke skills through the CLI rather than reading files off disk.

What is worth borrowing

The intent-mapping table is the standout idea. Its third column — what not to also run — encodes the boundary an agent most often oversteps: asked to generate deployment artifacts, it helpfully deploys them too. The file adds an explicit rule to stop at the verb the user used, and to ask once when the phrasing is ambiguous, on the grounds that a clarifying question costs less than unwinding an unwanted deployment.

The polling guidance is the other lesson, and it is one almost no environment file thinks to include. It caps how many log lines each status poll may request, because the default would return tens of kilobytes per poll and exhaust the agent's context window before the deploy finished. Token budget is an operational constraint like disk or memory, and this file treats it as one.