Coordinate a small team without assuming a team mode

Understand the approach

A small agent team is useful when each member owns an independent deliverable and the parent retains integration. Codex's documented subagent workflow should not be presented as a one-to-one copy of Claude's experimental team controls. The capstone needs only two implementation packages and a parent integrator. More titles would add coordination without creating a new independent output.

The ledger assigns A to starter/normalize.cjs and B to starter/report.cjs after a shared interface is established. A handles padded case and invalid status values. B receives normalized records and computes completed-only counts and effort. The parent waits for both code outputs, reconciles their handoffs and checks the combined state. A and B may proceed independently against CONTRACT.md; final acceptance depends on both.

Deliver a candidate that produces total=4, done=2 and doneMinutes=55, passes the four learner grading tests and preserves source inputs. Distinguish the runner's five named integration predicates from the seven frozen-baseline/reference tests. The deterministic simulation supplies reference handoffs with actualAgentsSpawned:false; any real worker activity needs its own observation. A ledger marked complete is useful only when its files and actual checks support that status.

Use the exact two-owner prompt and inspection flow in native-agent-guide.md for actual Codex coordination. The offline route already supplies working solution/normalize.cjs and solution/report.cjs, while candidate.test.cjs grades the editable starter. The real-worktree program proves an A-only handoff is rejected and the destination must be checked after integration. Codex’s documented parent/subagent orchestration reaches this outcome without claiming Claude’s experimental mailbox, automatic task-dependency service or teammate-pane controls. Judge working files and actual worker evidence rather than a fictional team-mode switch.

Practice and acceptance

A team ledger is useful when it connects ownership, dependencies, artifacts and actual integration evidence.

Try it step by step

1. Run the broken starter and establish the final report contract.
2. Use work-packages.json to assign two disjoint outputs and one parent-owned integration decision.
3. Implement both modules with actual agents or sequentially, recording the method honestly.
4. Inspect both handoffs and run the combined candidate plus boundary cases before marking the ledger complete.
5. Compare with the reference solution only after attempting the task; report unverified Codex client behavior separately.

A useful starting example

Illustration · adapt before use.

| Task | Owner | Owns | Depends on | Done when |
| --- | --- | --- | --- | --- |
| Freeze interfaces | Parent | CONTRACT.md review | Baseline evidence | Inputs and acceptance are clear |
| Normalize statuses | A | starter/normalize.cjs | Frozen contract | Padded and invalid inputs meet contract |
| Aggregate report | B | starter/report.cjs | Frozen contract | Completed-only count/effort and validation work |
| Integrate candidate | Parent | Combined review | A + B | run.cjs starter and four candidate tests pass |

Return actual file/check evidence and explicitly identify whether real agents, sequential edits or deterministic simulation produced each artifact.

node -e "const {buildReport}=require('./labs/delegation-capstone/solution/report.cjs'); console.log(buildReport([{id:'open',status:'open',minutes:90},{id:'done',status:'done',minutes:5}]))"
# Reference: {total:2, done:1, doneMinutes:5}. The candidate must generalize too.

Check your result

• Both module contracts and all four edited-candidate grading tests pass.
• run.cjs produces the exact report and all five named integration predicates pass.
• The seven reference/baseline tests, four candidate tests and any actual agent observations are reported as different evidence.

Common mistakes

• Claiming Claude team flags, mailbox paths or split-pane controls are current Codex commands.
• Treating several agreeing agent summaries as a substitute for checking the product.