Codex: no built-in sub-agents; web search per setting (fixes chat-endpoint sessions)

[renmengye]

Codex 0.160 (pinned in #456) adds two tools to every request: a namespace tool, multi_agent_v1 (built-in sub-agents), and the provider-side web_search tool. Through a chat-only endpoint the kernel's Codex bridge rejects both tool types, so every such session fails at once with "unsupported bridge request" (zero turns), and a Codex author on a self-hosted chat endpoint churns: run, session error, new run.

Change

• Every kernel-launched Codex session passes -c features.multi_agent=false, after the operator's own args, so it cannot be re-enabled. Multi-agent work goes through the kernel's own channels (sibling view, messages, reports, lines), which contracts can switch per run; built-in sub-agents would bypass them.
• Codex's built-in web search follows a new operator setting, OUTERLOOP_CODEX_WEB_SEARCH: auto (default) keeps it on native provider sessions and drops it on endpoint sessions, where the provider-side tool cannot run anyway; on and off force it. Exported to ticks and jobs (allowlist and deploy script).

Verification

• Captured a real Codex 0.160 request: without the flags codex_bridge.supported_request returns False (tool types namespace, web_search); with them it returns True (function tools only).
• Launch test now asserts the sub-agent flag follows operator args in every containment/endpoint/resume combination, and web search is dropped exactly for endpoint sessions; unit test for the setting.
• Gate: ruff check, ruff format --check, mypy, pytest.

Written and reviewed by Claude.

[github-actions]

Round 1 — reviewed head 8bc1160f — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.

terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.

Verdict: 3 blocking, 1 advisory.

3 findings attached to the lines below.

The on setting cannot enable web search on chat-bridge sessions. [credentials+general+lifecycle] When OUTERLOOP_CODEX_WEB_SEARCH=on, harness.py omits its disable override, but every chat-bridge launch later appends web_search="disabled", so endpoint sessions still run with web search disabled. (src/outerloop/codex_bridge.py:362; high confidence)

Merged four findings. The chat-bridge forced-on failure is corroborated by credentials, general, and lifecycle; its blocking severity follows the general review. No findings rejected.

[renmengye]

Fixed in e18a4d7: the setting is now auto (default) or off. A forced on could not be honoured (the chat bridge disables the provider-side tool on endpoint sessions, and a Codex config can disable it on native ones), so it is removed rather than documented as something it is not. The override uses the quoted TOML form web_search="disabled" (as the bridge does), verified against the real 0.160 binary. Changelog entry shortened.

[github-actions]

Round 1 — reviewed head e18a4d76 — reviewer hermes/gpt-5.6-terra.

terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.

Verdict: no defects found.