openclaw doctor --fix silently strips agentRuntime: claude-cli runtime mapping in 2026.5.18

OpenClaw version: 2026.5.18 (50a2481) Severity: Critical — silently breaks Claude MAX OAuth routing, falls back to direct-API with the OAuth token (rejected by Anthropic with misleading "billing" error)

Symptoms

After running openclaw doctor --fix:

1. JARVIS (and other Claude MAX-OAuth-using agents) starts failing with what looks like a billing/quota error
2. Logs reveal Anthropic is rejecting calls because the request is going down the direct-API path with an OAuth token (which MAX rejects with "extra usage" / "billing" style errors)
3. The cause: doctor --fix removed agentRuntime: claude-cli from per-model runtime mappings in defaults.models
4. A persistent billing-cooldown in auth-state.json (5hr disable that survives restarts) compounds the problem until manually cleared

Root cause (per Claude Code's 2026-05-18 patch)

In 2026.5.18, the runtime resolver only reads per-model runtime mappings in defaults.models. Adding agentRuntime: claude-cli ONLY at the agent level is insufficient — doctor --fix strips it from the agent and the per-model mapping is required.

Workaround required to restore working state:

1. Re-add { "agentRuntime": { "id": "claude-cli" } } to each anthropic/* model entry in defaults.models
2. Clear the persistent billing cooldown in auth-state.json (manual edit)
3. Restart Gateway to flush in-memory provider cooldown

Why this needs upstream fix

• doctor --fix is presented as a safe normalization tool. Silently breaking MAX OAuth routing is the opposite of safe.
• The error surface (Anthropic "billing" error) hides the actual config issue — users will assume their MAX plan was wrong, not their config
• The 5hr persistent cooldown means the symptom outlives the user noticing and investigating

Suggested fixes

1. doctor --fix should NEVER strip runtime mappings — at minimum, preserve agentRuntime blocks if present
2. If the resolver has changed to require per-model runtime mappings, doctor should ADD them (not remove them) when it sees agent-level agentRuntime
3. Billing cooldown should be discoverable + clearable via a doctor command (e.g. openclaw doctor --clear-cooldown) so users don't need to hand-edit auth-state.json
4. The Anthropic "billing" error path should log diagnostic info about which path was taken (direct-API vs claude-cli subprocess) so users can self-diagnose

Related issue

The subagent silent crashes filed in #84053 may share root cause — if subagents are also affected by stripped runtime mappings, the streaming/socket behavior could be the surface symptom.

Reproducibility

Trivial — openclaw doctor --fix on a 2026.5.18 install with MAX OAuth configured will reproduce. The diagnosis + patch above was performed by Anthropic's Claude Code on 2026-05-18.