docs: add cookbook for tool calling and LLM patterns

[obrutjack]

Closes #26

Adds a Cookbook section to the README with common workflow patterns:

• Which command to use — table clarifying llm.invoke vs llm_task.invoke vs openclaw.invoke
• Fetch → LLM summarize — minimal two-step example
• Three-step chain — fetch → LLM → save result
• Approval gate — human confirmation before LLM call
• Shell-safe args — using env vars instead of raw ${arg} substitution

This directly addresses the confusion in #26 where the user couldn't figure out the correct syntax for calling LLMs mid-workflow.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3396dca6a0

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Build args JSON safely from env variables

This example is labeled as shell-safe for quotes/special characters, but --args-json "{\"prompt\":\"$Q\"}" still inlines $Q into a JSON string without escaping. If question contains a double quote (e.g. He said "hi"), the resulting JSON becomes invalid and openclaw.invoke fails with --args-json must be valid JSON, so the documented “safe” pattern breaks on common inputs. Use a JSON encoder (for example via jq --arg) instead of manual string interpolation.

Useful? React with 👍 / 👎.