feat(sandbox): make codex reasoning summaries configurable via env

[tarrencev]

Why

The codex CLI bump to 0.139 (#…recent releases) changed behavior: codex no longer emits reasoning summaries unless model_reasoning_summary is set in its config. Since harness/codex/config.toml doesn't set it, the app-server now emits reasoning items with empty summary/content, and downstream renderers (e.g. the Slack adapter, which reads the reasoning item text) silently lost the embedded thinking trace.

Reproduce on codex 0.139.0:

codex exec --json --skip-git-repo-check -c model_reasoning_effort='"low"' "What is 17 * 23 - 19?"
# -> no reasoning item, but usage shows reasoning_output_tokens > 0

codex exec --json --skip-git-repo-check -c model_reasoning_effort='"low"' -c model_reasoning_summary='"detailed"' "What is 17 * 23 - 19?"
# -> item.completed {type: reasoning, text: "**Calculating...**"}

What

Make it configurable per deployment instead of baking a default into the image: the sandbox entrypoint merges CODEX_MODEL_REASONING_SUMMARY (auto | concise | detailed | none) into the generated ~/.codex/config.toml — replacing an existing top-level model_reasoning_summary or inserting one before the first table. Invalid values warn to stderr and are ignored so a typo can't break every sandbox boot.

Deployments opt in through the existing sandbox.extraEnv chart value; no chart changes needed.

Tested the merge against harness/codex/config.toml for all four cases (unset, insert, replace-existing, invalid) — output stays valid TOML and the existing [features] rewrite is unaffected.

🤖 Generated with Claude Code