gpt-5.6-sol `ultra` effort spawns multi-agent swarms despite `features.multi_agent=false`, consuming 250–600 credits per read-only task without batching guidance

Summary

On gpt-5.6-sol at model_reasoning_effort="ultra", codex exec spawns
20+ sub-agent sessions per task even when the config explicitly sets
features.multi_agent=false. Combined with the Code Mode serialization
described in #32503 / #35050, a single bounded read-only investigation task
consumed 7.5–17.6M raw tokens (267–605 credits at the published rate card) —
about 1% of a Pro weekly allowance per task. The identical task at xhigh
with batching guidance: ~150k raw tokens / ~10 credits, with an
identically-graded answer.

Environment

• Codex CLI 0.145.0 (linux x86_64, WSL2), ChatGPT Pro auth
• gpt-5.6-sol, sandbox read-only, approval_policy="never"
• Overrides per run: mcp_servers={}, plugins={},

features.multi_agent=false, features.memories=false

Reproduction

Fixed synthetic repo (48 small modules + config + manifest; generator in
https://github.com/ForrestGrump/codex-batch-bench), prompt = per-module inventory table (all reads independent
after reading the manifest). 2 trials per arm, arms differ only in a
batching instruction vs length-matched placebo via developer_instructions.

Observed (per trial)

| arm | sessions spawned | model cycles | raw tokens | credits | Promise.all* cells |
|---|---|---|---|---|---|
| no batching guidance | 23, 25 | 305, 652 | 7.5M, 17.6M | 267, 605 | 0, 0 |
| batching guidance | 8, 6 | 120, 78 | 3.4M, 2.3M | 138, 101 | 18, 19 |

Both arms produced correct, equivalent answers (48/48 modules, 40/40 config
keys). Sub-sessions were identified by session_meta.cwd in
~/.codex/sessions rollout logs; counts are a lower bound.

Expected

1. features.multi_agent=false should gate all sub-agent spawning,

including ultra's internal orchestration — or the docs should state that
ultra ignores it.

1. Ultra's orchestrator and sub-agents should batch independent Code Mode

reads (see #35050) — at minimum, effort-conditional guidance, since the
serialization also appears at xhigh/max but not at high and below.

Notes

• Sub-agents inherit the parent process's resolved config, including

-c developer_instructions overrides (verified: 0/48 placebo-arm
sub-sessions contained the globally-configured instruction text when the
CLI override replaced it).

• Full effort curve and real-repo replication data in my comment on #35050:

#35050