[danieldutu]

agents workflow diagram <config-path> prints the workflow of a local agent config as a Mermaid flowchart.

The config stores workflow.nodes and workflow.edges as raw JSON, which is hard to review in a diff. Mermaid renders natively on GitHub, so the output can be pasted into a mermaid code fence in a PR or doc. The command reads the file only: no API key, no network.

It reads the file as untyped JSON, so node types the CLI does not know yet still render.

Output rules:

• Nodes and edges are sorted by id, so the output diffs cleanly.
• Each node type (start, end, say, tool, update_state, override_agent, standalone_agent, phone_number) has its own classDef. The emitted class names carry a wf_ prefix (wf_start, wf_end, …): Mermaid treats end as a flowchart keyword, so a class named end is unsafe across Mermaid versions. Node ids get the same treatment through an n_ prefix.
• A type the CLI does not know renders with the wf_other class.
• An edge label is the forward condition's label, or its condition text, cut at 40 characters. Unconditional edges have no label.
• A config without a workflow key exits with a validation error.

I edited cli/elevenlabs/workflow/diagram.rs and cli/elevenlabs/workflow/mod.rs by hand; both are in .fernignore, so regeneration leaves them alone. The README section is also hand-written.

Checks: cargo test diagram (3 passed), rustfmt --check on the touched file, and cargo clippy with no findings there.

Not in this PR: the GitHub Action, backward_condition arrows and an agent-name argument from the #170 proposal. I can add them in a follow-up. If you prefer another command name than agents workflow diagram, say so and I rename it.

Part of #170.