Build a flight briefing agent with Codex

Traveler: “That forecast is for now. Can you check when my flight actually arrives?” The starter desk from your first call can check weather and exchange rates. Now add flight-status and departure lookups so your new desk can find the arrival airport and time before requesting a forecast.

Switch roles. You now maintain this desk. The change request from the opening gives you an acceptance check: when the same flight number has records for two dates, the desk should ask which date the traveler means before looking up weather. Keep that check beside your brief.

1. Inspect a Mantle project (10 minutes)

Open builds/starter/agent.yml, then skills/travel_briefing/skill.md and tools.py underneath that project. The first sets identity and rules. The skill tells Mantle when to route a question and which tools to call. The Python functions return provider data in a ToolResult. Open the shared readers and predict which field changes when an estimated arrival becomes available.

In the full desk, destination_weather requires flight_number and flight_date (YYYY-MM-DD, or empty if unknown); a unique matching record supplies its arrival forecast, ambiguity prompts a clarification, and airport_code/arrival_time provide a labelled independent forecast only if the board has no record or is unavailable.

Review two possible skill instructions

Choose wording you could test

B gives you an observable sequence: flight results, clarification, traveler’s choice, then a weather call for the chosen arrival. A can produce a plausible forecast for the wrong day. When Codex returns its candidate, find the sentence that controls this exact decision.

2. Give Codex the build brief (15 minutes)

Use a new workshop terminal. To authenticate Codex with the saved workshop project key, run bash desk login. This sends the key to the login command over standard input. Running Codex with this account can incur model charges. Then run:

bash desk propose codex "Build a travel desk for Oslo to Bergen. If two records match my flight number, ask which date I mean. After I choose, check that flight’s arrival before looking up its weather. Give me tonight’s other departures and convert my dollar budget into Norwegian kroner. Ask for missing airports and distinguish scheduled from estimated arrival."

Codex writes the agent specification in a fresh run directory. The specification contains the agent name, routing description, skill instructions and selected tools. The workshop compiler turns that specification into a runnable Mantle project using reviewed provider functions. Read the resulting instructions: is it clear what happens when two flights share a number or when a provider is unavailable?

Without Codex access: bash desk propose sample copies the authored example through the same build steps. Label the result as a sample. Later, rerun this section with Codex and compare the generated skill.

3. Review, build and try it (20 minutes)

Open the printed candidate path in the editor. Require these four capabilities: status, weather, fx, flights. Read the instructions as if you were handling an uncertain arrival. If they are unsatisfactory, keep the candidate for comparison and request a revision. Each new generation can be billed.

Copy the directory containing that candidate into the following command, replacing the placeholder. Approval compares the candidate with its receipt and builds a new project. It refuses to overwrite an existing output.

bash desk approve runs/YOUR_RUN_ID builds/codex-desk

Stop the starter Inspector with Ctrl+C in its own terminal. Run bash desk inspect builds/codex-desk from the workshop terminal. Open the private Inspector again, click Start voice conversation, and ask aloud “Which flights are scheduled from Oslo to Bergen tonight?” Compare its result with bash desk data flights OSL BGO from a second terminal. If the board is unavailable, the agent should say so.

Improve the actual Mantle skill with Codex

Once the generated desk works, ask a harder question: “I might fly into London or Paris. Which weather did you check?” Use Codex to change what the agent asks before calling a tool. Start from a copy of the working project.

cp -R builds/codex-desk builds/clarifying-desk
cd builds/clarifying-desk
"$HOME/.local/node_modules/.bin/codex"

Give Codex this instruction:

Edit only skills/travel_briefing/skill.md. For an ambiguous destination or a flight number shared by several records, ask one clarification question using the returned dates or airport codes. After the traveler chooses, use the selected record’s estimated or scheduled arrival for weather. Preserve all four tool references. Show me the diff and explain which question should trigger each branch.

Exit Codex with /quit. Return to the workshop and compare the two skill files:

cd ../..
diff -u builds/codex-desk/skills/travel_briefing/skill.md builds/clarifying-desk/skills/travel_briefing/skill.md

Read the diff before restarting. Differences are expected; check that they implement the clarification you requested and preserve all four tool references. Stop the old Inspector, then start the changed desk:

bash desk inspect builds/clarifying-desk

Try the same ambiguous question again. Success: a clarification precedes a destination-specific lookup; an unambiguous request still reaches the tool. If time is short, resume this iteration during home practice.

Checkpoint. Find all four tool references in the generated skill. In the flight result, identify the origin-local date, provider coverage and scheduled departure. The board can list schedules; it cannot promise a seat.

Break, minutes 75–85. Save the run ID and generated directory. You can reopen the same Codespace and train this project again without calling Codex again.

If you want to compare files before spending a model call, open the finished reference agent and its travel skill. Your generated project uses the same layout.