openclaw - 💡(How to fix) Fix [Bug]: subagent spawn validation rejects every non-off thinking level on all canonical openai/* models — error cites canonical alias even when openai-codex/* is requested

Summary

On v2026.5.19, sessions_spawn (and any subagent-creation path) rejects every non-off thinking level for the canonical openai/* model namespace. The error consistently cites openai/gpt-5.5 even when the caller specified openai-codex/gpt-5.5 — subagent validation appears to fold openai-codex/* through openai/* alias routing before consulting the catalog, then the canonical catalog's missing reasoning metadata causes the rejection.

This is the same root cause as the picker / /think directive surface tracked in #84646, but a distinct surface and worth its own narrow report because it breaks the subagent API contract and is environment-affecting (not just UX).

Environment

• OpenClaw v2026.5.19 stable
• Channel: Telegram (also reproduces via main agent calling the sessions_spawn tool)
• Active agent: main, runtime pi
• Auth: auth.order.openai = ["openai-codex:[email protected]"] (OAuth, no api-key fallback)
• /status confirms runtime: OpenAI Codex

Reproduction

A subagent spawn request:

sessions_spawn({
  model: "openai-codex/gpt-5.5",
  thinking: "xhigh",
  // ...prompt etc.
})

returns:

Thinking level xhigh is not supported for openai/gpt-5.5.
Use one of: off.

Notes from this run:

• The caller passed openai-codex/gpt-5.5, but the validation error names openai/gpt-5.5 — alias normalization happens before catalog lookup.
• Same error shape with high, medium, low, minimal — only off is accepted.
• Affects every canonical openai/* model in our setup (gpt-5.5, gpt-5.4, gpt-5.4-mini) — they all share the same catalog metadata gap.

Asymmetry vs. main-session thinking

Main-session thinking via the config-time path still applies correctly:

"agents": { "defaults": { "thinkingDefault": "xhigh" } }

/think reports Current thinking level: xhigh for the main session, and turns appear to honor it. So the routed gpt-5.5 runtime (OpenAI Codex OAuth) is fully capable of running non-off thinking — only the OpenClaw catalog metadata for openai/* claims off-only.

Regression vs. v2026.5.18

This was working on v2026.5.18 a few hours before upgrading to v5.19. I previously verified in #84489 that on v5.18:

• sessions_spawn(model: "openai/gpt-5.4", thinking: "high") → ✅ completed successfully
• sessions_spawn(model: "openai-codex/gpt-5.5", thinking: "high") → ✅ completed successfully

After upgrading to v2026.5.19 these exact spawn calls now fail with the Use one of: off rejection above. The catalog truth-source enforcement appears to have been tightened in v5.19 in a way that exposes the underlying canonical openai/* catalog reasoning-metadata gap on every code path, including subagent spawn, that previously bypassed the canonical lookup.

The /think picker on the same model also regressed in v5.19 (now shows only default, off, was showing the full range on v5.18 because #82975's picker fix was effectively a UI-side hack rather than a catalog correction) — that surface is tracked in #84646.

Expected

sessions_spawn with thinking: "high" (or any other non-off level supported by the underlying model) on any openai/* or openai-codex/* canonical OpenAI gpt-5 model should be accepted. The fix is to populate the reasoning-metadata for openai/gpt-5.5, openai/gpt-5.4, and openai/gpt-5.4-mini in the canonical catalog so that they advertise the full thinking-level capability the underlying API supports.

Workaround

• For main session: set agents.defaults.thinkingDefault: "<level>" at config time — this bypasses runtime validation and the level is honored for normal turns.
• For subagents: no working workaround in v5.19. Forcing thinking: "off" on every subagent spawn loses the capability entirely. Any subagent that needs reasoning is currently blocked.

Related

• #84646 — /think directive / picker surface; this is the subagent-spawn surface peer report for the same root cause
• #82975 — original "picker only shows default/off" (closed as completed; picker UI was patched in a way that did not fix the underlying catalog metadata, and v5.19 reverted that patch)
• #84489 — broader subagent / orchestration umbrella where the original v5.18 verify was recorded