Document and guard against empty optional fields in Codex spawn_agent calls

从 plugins/omo/skills/visual-qa/SKILL.md、plugins/omo/skills/ulw-loop/SKILL.md 和 plugins/omo/components/rules/bundled-rules/hephaestus/gpt-5.5.md 开始，然后检查 skill-sync 或 rule-generation 测试。运行 sync-skills、ulw-loop、visual-qa 和 Hephaestus 规则测试；当生成的 Codex 指南始终说明应省略未设置的可选 spawn_agent 键，并验证有效的 V1 payloads 时，即表示完成。

Summary

LazyCodex's Codex-facing skill/rule guidance depends heavily on native spawn_agent, but current Codex tool parsing rejects common empty optional-field shapes. Until upstream normalizes those fields, LazyCodex should make its generated guidance and any spawn guardrails explicitly tell agents to omit unused optional keys, especially items: [], message: \"\", model: \"\", reasoning_effort: \"\", and service_tier: \"\".

Upstream issue filed: https://github.com/openai/codex/issues/37037

Environment

• Codex CLI on PATH: codex-cli 0.146.0
• Runtime: Codex Desktop / ChatGPT-auth session exposing multi_agent_v1.spawn_agent
• OS: macOS 26.5.2, Darwin 25.5.0 arm64
• Node: v26.6.0
• LazyCodex source checked: code-yeongyu/lazycodex main at fb48ddc4bc8be02a0cfe0a509a30cf3543edf72a
• Upstream Codex source checked: openai/codex main at 1fe6be9719ac4a18ad08f8341b89f9a0f386105e

Repository Decision

• Target repository: code-yeongyu/lazycodex
• Why this belongs here: the root parser bug is upstream Codex, but LazyCodex ships the skills/rules that repeatedly instruct agents to dispatch subagents through multi_agent_v1.spawn_agent / flat spawn_agent. A downstream compatibility issue is useful so LazyCodex can harden its generated instructions and guardrails while waiting for upstream.
• LazyCodex evidence:
  • plugins/omo/skills/visual-qa/SKILL.md:6-22 maps OpenCode examples to Codex multi_agent_v1.spawn_agent calls.
  • plugins/omo/components/rules/bundled-rules/hephaestus/gpt-5.5.md:42-44 tells agents to spawn subagents via multi_agent_v1.spawn_agent({"message":"...","fork_context":false}).
  • plugins/omo/skills/ulw-loop/SKILL.md:56-66 already gives a good V2 warning not to set agent_type / model / reasoning_effort; V1 guidance could use similarly explicit "omit empty optional keys" wording.
• Upstream Codex evidence:
  • codex-rs/core/src/tools/handlers/multi_agents_common.rs:137-144 rejects Some(message), Some(items) without treating empty unused values as absent.
  • codex-rs/protocol/src/openai_models.rs:119-133 rejects empty reasoning_effort during deserialization.

Reproduction

In a Codex session exposing multi_agent_v1.spawn_agent, attempts with serialized empty optional fields failed as follows:

1. message: \"...task...\" plus items: [] returned:

Provide either message or items, but not both

2. items: [{ type: \"text\", text: \"...task...\" }] plus message: \"\" returned:

Provide either message or items, but not both

3. reasoning_effort: \"\" returned:

failed to parse function arguments: reasoning_effort must not be empty at line 1 column 219

Expected Behavior

LazyCodex-generated instructions should make the valid payload shape unambiguous for both tool surfaces:

• For V1: pass only message, agent_type when needed, and fork_context; omit items, model, reasoning_effort, and service_tier unless they are intentionally non-empty.
• For V2: pass only task_name, message, and fork_turns unless the active schema explicitly exposes additional non-empty fields.
• Never emit empty placeholder values for optional spawn fields.

Actual Behavior

The current shipped guidance mostly shows minimal calls, but it does not consistently warn that empty optional fields are actively dangerous on the native tool surface. In a real visual QA workflow, this caused subagent dispatch to fail and forced the parent agent to continue single-threaded.

Evidence

• Live Codex Desktop tool-call failures on 2026-08-05.
• Source-backed upstream root cause captured in https://github.com/openai/codex/issues/37037.
• LazyCodex source references above show the downstream surface area likely affected by this parser sharp edge.

Root Cause

Upstream Codex currently treats empty optional values as supplied, while LazyCodex guidance assumes agents can safely follow the model-visible shape and omit optional fields. When a caller/tool wrapper serializes empty defaults, the upstream parser rejects the call before a subagent is spawned.

Proposed Fix

• Update Codex subagent compatibility sections in shared LazyCodex skill/rule templates to explicitly say: omit optional keys entirely when unset; do not pass empty items, empty message, or empty override strings.
• Add a small regression/fixture in the skill sync or rule generation tests to preserve this warning in generated skills that mention spawn_agent.
• If LazyCodex has any hook or guard that inspects tool-call payloads before execution, consider warning when spawn_agent includes items: [], message: \"\", or empty override strings.

Verification Plan

• Regenerate/sync skills and confirm all shared Codex harness compatibility blocks contain the warning.
• Run LazyCodex skill/rule tests touching sync-skills, ulw-loop, visual-qa, and Hephaestus bundled rules.
• Manually validate a V1 spawn call with only non-empty keys in a Codex session.

---

This issue or PR was generated by LazyCodex.
Tag: lazycodex-generated