fix(anthropic-oauth): opus-4-7 body-shape parity with Claude Code 2.1.112

[mssteuer]

What

Claude Code 2.1.112 was observed (capture 2026-04-17) sending this shape on opus-4-7 /v1/messages requests:

thinking:           {"type":"adaptive"}
output_config:      {"effort":"xhigh"}
context_management: {"edits":[{"type":"clear_thinking_20251015","keep":"all"}]}
temperature:        (absent)
tool_choice:        (absent when caller doesn't request one)

Hermes was sending:

thinking:    {"type":"enabled","budget_tokens":4000}
temperature: 1
tool_choice: {"type":"auto"}
(no output_config, no context_management)

Why now

PR #3 (fingerprint refresh) and PR #4 (Stainless/JS SDK spoofing) got the headers + client identity right but Extra Usage kept climbing. Michael asked me to check whether the system prompt / body shape was the remaining drift before chasing TLS/JA3 (the nuclear option). A mitmproxy capture of a real claude --model claude-opus-4-7 --print "..." session against our own OAuth token revealed body-shape drift, not TLS. Sent a hand-crafted pure-CC-shape request through our post-PR-#4 stack: Anthropic returned 200 OK, no issues. Body shape is the remaining drift layer.

Root cause

_supports_adaptive_thinking() only matched 4-6 / 4.6 substrings. Opus 4.7 (released after the adapter was written) fell through to the legacy manual-thinking branch written for 4.5 and earlier, which sets thinking:{enabled,budget} and temperature:1. Opus 4.7 rejects temperature with a 400 temperature is deprecated for this model — we saw this in journalctl right after the PR #3 deploy but it wasn't blocking because the errors were retried. The classifier reads the broken shape as drift.

Changes

Single commit, 2 files, +240/-11.

agent/anthropic_adapter.py (+72/-11)

• _supports_adaptive_thinking: match 4-7 / 4.7 in addition to 4.6. Per Michael: just 4.7, not broader.
• ADAPTIVE_EFFORT_MAP: xhigh → xhigh (identity). CC sends xhigh literally; the old xhigh → max mapping was a fingerprint-drift signal.
• _ANTHROPIC_OUTPUT_LIMITS: explicit claude-opus-4-7 → 128_000 entry so the output ceiling doesn't fall through the substring table to claude-opus-4 → 32_000. Kept 128K per Michael — CC clips to 64K but the model supports the full 128K.
• Adaptive branch now emits context_management with the clear_thinking_20251015 cleanup edit. Routed via extra_body because anthropic Python SDK 0.92.0 doesn't accept it as a native kwarg yet.
• Adaptive branch no longer sets temperature. Legacy branch (4.5, 3.7, etc.) still sets temperature=1 for manual-thinking models.
• tool_choice: emit only when caller explicitly passed a value. Previously emitted {"type":"auto"} unconditionally. CC never sends tool_choice for default (auto) behavior.

tests/agent/test_anthropic_adapter.py (+179)

• New TestOpus47BodyShapeParity class, 14 tests:
  • 3 tests for the three adaptive-thinking fields (thinking type, output_config.effort, context_management shape).
  • 4 tests for tool_choice behavior (omit when None, auto when asked, any for required, specific tool name).
  • 1 test asserting no temperature on adaptive-thinking models.
  • 1 test asserting max_tokens == 128_000 for opus-4-7.
  • 2 tests for effort passthrough (xhigh and medium).
  • 1 test for _supports_adaptive_thinking matching 4-7 variants + regression guards.
  • 1 test confirming legacy (sonnet-4-5) branch still uses manual-thinking shape.
  • 1 test confirming _get_anthropic_max_output('claude-opus-4-7') == 128_000.
• Existing test_reasoning_config_maps_xhigh_to_max_effort_for_4_6_models renamed + rewritten to assert the corrected xhigh → xhigh identity mapping.

Tests

172 adapter tests pass (was 158; +14 new, 0 regressions).

Verified end-to-end

Built a real OAuth Anthropic client with the patched code, routed through mitmproxy, fired a real /v1/messages call on claude-opus-4-7. Anthropic returned 200 OK; captured request body matches CC byte-for-byte on all 5 parity deltas. Side-by-side diff shows the only remaining difference is stream (present on CC; absent in the non-streaming smoke test — gateway streams in production).

Verification next steps for reviewer

1. Merge / deploy.
2. Let OAuth traffic run for 30-60 min at normal load.
3. Diff Extra Usage growth on Claude.ai dashboard before/after this PR.
4. If Extra Usage flattens → we've found the last drift layer, done with fingerprint work.
5. If it keeps climbing → the last remaining hypothesis is TLS/JA3 (Python httpx/OpenSSL vs Node undici handshake). Meaningfully harder to spoof; only worth chasing if this PR doesn't move the needle.

Note

No changes to the non-OAuth (regular API-key) code paths. Legacy manual-thinking for sonnet-4-5 and earlier is preserved unchanged.