openclaw - 💡(How to fix) Fix [Bug]: thinkingDefault=xhigh silently remapped to high for anthropic/claude-opus-4-7 — provider-prefixed modelId bypasses isAnthropicOpus47Model matcher [1 participants]

Summary

thinkingDefault: xhigh set in openclaw.json at agents.defaults.thinkingDefault is silently remapped to high on new DM sessions for anthropic/claude-opus-4-7, with the notice:

Thinking level set to high (xhigh not supported for anthropic/claude-opus-4-7).

Opus 4.7 does support xhigh (per CHANGELOG: "Anthropic/models: add Claude Opus 4.7 xhigh reasoning effort support and keep it separate from adaptive thinking").

Root cause is a model-id format mismatch between the directive-handling layer (passes provider-prefixed id like anthropic/claude-opus-4-7) and the Anthropic runtime plugin's isAnthropicOpus47Model matcher (expects unprefixed id via startsWith("claude-opus-4-7")).

Environment

• OpenClaw 2026.4.22 (00bd2cf) (reproducible on stock install)
• macOS 25.3.0, Node v25.9.0
• Channel: Telegram DM (reproducible on any channel that spawns a fresh session after the default is set)
• Config:

  {
    "agents": {
      "defaults": {
        "model": { "primary": "anthropic/claude-opus-4-7" },
        "thinkingDefault": "xhigh"
      }
    }
  }

Repro

1. Clean install, set agents.defaults.thinkingDefault = "xhigh" via openclaw config set.
2. openclaw gateway restart.
3. Send first message to any channel DM where no prior session exists (e.g. Telegram DM).
4. Gateway replies with:

   Thinking level set to high (xhigh not supported for anthropic/claude-opus-4-7).

5. Inspect ~/.openclaw/agents/<agent>/sessions/sessions.json:

   "provider": "anthropic",
   "model": "claude-opus-4-7",
   "thinkingLevel": "high"

   Provider/model stored unprefixed; thinkingLevel is the remapped value.

Expected

New Opus 4.7 sessions should honor thinkingDefault: xhigh and persist thinkingLevel: "xhigh". Remap should not fire for a level the model actually supports.

Actual

Remap silently downgrades to high because the runtime thinks xhigh is unsupported.

Root cause (code trace)

The notice string is produced at dist/directive-handling.impl-*.js ~L681:

if (!directives.hasThinkDirective && shouldRemapUnsupportedThinkLevel && remappedUnsupportedThinkLevel)
  parts.push(`Thinking level set to ${remappedUnsupportedThinkLevel} (${nextThinkLevel} not supported for ${resolvedProvider}/${resolvedModel}).`);

shouldRemapUnsupportedThinkLevel is driven by isThinkingLevelSupported({ provider, model, level }) a few lines above — and resolvedModel at this point is the provider-prefixed id, as evidenced by the error message rendering anthropic/claude-opus-4-7 verbatim.

isThinkingLevelSupported → supportsThinkingLevel → resolveThinkingProfile (in dist/thinking-*.js):

const pluginProfile = resolveProviderThinkingProfile({
  provider: context.normalizedProvider,
  context: providerContext,   // providerContext.modelId == context.modelId, unnormalized
});
if (pluginProfile) {
  const normalized = normalizeThinkingProfile(pluginProfile);
  if (normalized.levels.length > 0) return normalized;   // EARLY RETURN — xhigh never re-appended
}

Note the early return: when the Anthropic plugin returns a profile, the fallback resolveProviderXHighThinking path below is skipped, so the profile is authoritative.

Anthropic plugin at dist/extensions/anthropic/register.runtime.js ~L380:

resolveThinkingProfile: ({ modelId }) => {
  const levels = [ {id:"off"}, {id:"minimal"}, {id:"low"}, {id:"medium"}, {id:"high"} ];
  if (isAnthropicOpus47Model(modelId)) levels.push({ id: "xhigh" }, { id: "adaptive" }, { id: "max" });
  ...
},

And the matcher (same file, ~L178):

const ANTHROPIC_OPUS_47_MODEL_ID = "claude-opus-4-7";
const ANTHROPIC_OPUS_47_DOT_MODEL_ID = "claude-opus-4.7";

function isAnthropicOpus47Model(modelId) {
  const lowerModelId = normalizeLowercaseStringOrEmpty(modelId);
  return lowerModelId.startsWith(ANTHROPIC_OPUS_47_MODEL_ID) || lowerModelId.startsWith(ANTHROPIC_OPUS_47_DOT_MODEL_ID);
}

When modelId === "anthropic/claude-opus-4-7" (provider-prefixed form used by the directive layer), both startsWith checks fail → xhigh is never pushed → runtime concludes it's unsupported → remap to high.

In the session record the provider/model are stored unprefixed ("anthropic" + "claude-opus-4-7"), so the directive layer and the storage layer disagree on the canonical model id format. Either side alone is fine; the mismatch breaks the matcher.

Suggested fix

One of:

A. Strip provider prefix before plugin dispatch (best — keeps plugin matchers simple):

In resolveThinkingPolicyContext (or at the directive-layer entry point), normalize model by stripping a leading <providerId>/ when it matches the resolved provider. Then providerContext.modelId passed into resolveProviderThinkingProfile is always plain claude-opus-4-7.

B. Make isAnthropicOpus47Model prefix-tolerant (localized fix):

function isAnthropicOpus47Model(modelId) {
  const raw = normalizeLowercaseStringOrEmpty(modelId);
  const id = raw.startsWith("anthropic/") ? raw.slice("anthropic/".length) : raw;
  return id.startsWith(ANTHROPIC_OPUS_47_MODEL_ID) || id.startsWith(ANTHROPIC_OPUS_47_DOT_MODEL_ID);
}

Same treatment would be needed for matchesAnthropicModernModel and any sibling matcher that uses startsWith on a bare model id.

A is the cleaner contract: the plugin is already addressed by provider, so it should never need to inspect its own provider prefix. Most callers already pass unprefixed ids (that's why plugin matchers look the way they do) — this one path is the outlier.

Impact

Users setting thinkingDefault: xhigh cannot get xhigh on Opus 4.7 sessions despite both config and model supporting it, and the notice blames the wrong party ("not supported for anthropic/claude-opus-4-7").

Workaround: none — /think xhigh directive hits the same isThinkingLevelSupported check and fails identically.

Related

• CHANGELOG 2026.x: "Anthropic/models: add Claude Opus 4.7 xhigh reasoning effort support and keep it separate from adaptive thinking."
• #67888 (Opus 4.7 forward-compat — supportsAdaptiveThinking() omits opus-4-7) — adjacent symptom class, different matcher, same root cause (missing/mismatched Opus 4.7 recognition across plugin matchers).