Run Sonde Workflows End to End

Purpose

Provide practical command sequences for common Sonde usage patterns when building and evaluating AI-friendly CLIs.

Inputs

• A target CLI binary in PATH.
• Working directory where artifacts can be written.

Outputs

• Generated manifest.
• Run result payload.
• Score payload.

Recommended deterministic fixtures

For reproducible local demos, use:

• cli-good: fully automation-friendly fixture.
• cli-ok: partially automation-friendly fixture.
• cli-bad: intentionally noisy/interactive fixture.

Why good vs bad

• cli-good is designed for deterministic automation: stable help, --json, and agent-safe flags.
• cli-bad is designed to fail automation checks: prompt markers ([y/n]), noisy output, and weak machine-readable conventions.
• cli-ok sits in between and is useful as a regression middle-ground.

Example: baseline evaluation

sonde generate cli-good --json
sonde manifest --json
sonde run cli-good --json
sonde score cli-good --json

Example: protocol mode for tooling

sonde serve --json

Then write newline-delimited requests:

{"id":"1","method":"tools/list"}
{"id":"2","method":"tools/call","params":{"name":"deploy","input":{"target":"preview"}}}

Example: CI-safe invocation

sonde generate cli-ok --json
sonde manifest --json
sonde run cli-ok --json
sonde score cli-ok --json

Use JSON output in CI and parse ok + command result fields.

Edge cases

• If a target CLI lacks stable help output, regenerate manifest after CLI upgrades.
• If run output includes prompts, fix CLI non-interactive behavior before relying on score.
• sondage.manifest.json is Sonde's contract artifact used by run, score, and serve.

See also

• Quickstart
• CLI Reference
• Serve Protocol