[Bug]: claude-opus-4-7 fails with "thinking.type.enabled is not supported"

[centripetal-star]

Description

claude-opus-4-7 returns a 400 error because Hermes sends thinking: {type: "enabled", budget_tokens: N} which is not supported by this model. Opus 4-7 requires thinking: {type: "adaptive"} with output_config: {effort: ...} instead.

Error

BadRequestError [HTTP 400]
"thinking.type.enabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior.

Root Cause

The fix from #1126 / PR #1128 hardcoded version checks for "4-6" and "4.6" only in _supports_adaptive_thinking():

def _supports_adaptive_thinking(model):
    return any(v in model for v in ("4-6", "4.6"))

claude-opus-4-7 doesn't match, so it falls through to the old code path that sends thinking: {type: "enabled"}.

The _ANTHROPIC_OUTPUT_LIMITS dict also has no entry for claude-opus-4-7.

Suggested Fix

1. Update _supports_adaptive_thinking() to also match "4-7" / "4.7" (or ideally use a >= comparison to future-proof).
2. Add claude-opus-4-7 to _ANTHROPIC_OUTPUT_LIMITS.
3. Note: opus 4-7 also removes support for temperature, top_p, and top_k parameters entirely (400 errors). These may need to be stripped as well.

Additional Opus 4-7 API Changes

• thinking.type must be "adaptive" (both "enabled" and "disabled" return 400)
• budget_tokens replaced by output_config.effort (low / medium / high / xhigh / max)
• "xhigh" effort level is new and exclusive to opus 4-7
• max_tokens now covers both thinking + response output
• New tokenizer uses 1x-1.35x more tokens

Environment

• Hermes Agent (latest from main)
• Provider: Anthropic (direct API)
• Model: claude-opus-4-7

[vominh1919]

Opened PR #11152 to fix this: #11152

[trevthefoolish]

Opened #11157 as a comprehensive fix for this. Beyond the adaptive-thinking gate the earlier attempts (#11107, #11145, #11152 — all closed as "doesn't exist yet" before the announcement went live), it also covers the items you flagged in the original report:

• temperature / top_p / top_k stripping (400 on 4.7)
• output_config.effort with xhigh preserved as a distinct level (not aliased to max)
• New thinking.display = "summarized" so reasoning stays visible in streams
• 1M context and claude-opus-4-7 added to catalogs
• New refusal / model_context_window_exceeded stop_reason mapping
• 4 new tests

Each change traces to a specific line in the official migration guide at https://platform.claude.com/docs/en/about-claude/models/migration-guide — the model is public as of today: https://www.anthropic.com/news/claude-opus-4-7