OAuth Anthropic provider: clear_thinking_20251015 strategy forwarded without thinking block → 400 from upstream

[ojsol]

Summary

When Claude Code is configured with an OAuth-authenticated Anthropic provider in Tingly Box and the orchestrator model has extended thinking enabled (default for Opus 4.7), every request to POST /tingly/claude_code/v1/messages?beta=true fails with a 400 from upstream:

{"type":"error","error":{"type":"invalid_request_error","message":"`clear_thinking_20251015` strategy requires `thinking` to be enabled or adaptive"},"request_id":"req_011CasKQviFdqAv4S7D1CcqU"}

Anthropic is rejecting the upstream request because the body contains context_management.strategy: clear_thinking_20251015 but no matching thinking block. Hypothesis: Tingly's protocol-translation pipeline preserves the context_management field while dropping the thinking field somewhere on the OAuth Anthropic path. The two must travel together.

The error is non-recoverable from the client side — Claude Code retries up to 10 times with exponential backoff, each retry burning quota against the user's Anthropic 7-day window before the session times out.

Environment

• Tingly Box: v0.260507.1 (binary at ~/.cache/tingly-box/v0.260507.1/bin/tingly-box)
• Claude Code: 2.1.138
• OS: Linux x86_64
• Provider auth type: OAuth (Claude Code) — Claude.ai Pro/Max subscription
• Provider name (auto-generated): claude_code-YYYYMMDD-HHMM
• Upstream model selected by routing: claude-opus-4-7
• Endpoint receiving traffic: POST /tingly/claude_code/v1/messages?beta=true

Steps to reproduce

1. Add an OAuth Anthropic provider via the UI (OAuth → Add Auth Provider → Claude Code).
2. Run tingly-box apply cc --provider <oauth-provider-name> --model claude-opus-4-7 (or wire ~/.claude/settings.json manually) so Claude Code routes through http://127.0.0.1:12580/tingly/claude_code with ANTHROPIC_DEFAULT_OPUS_MODEL=tingly/cc-opus.
3. Configure the Tingly rule for tingly/cc-opus to point at the OAuth provider with model claude-opus-4-7.
4. Launch Claude Code with "model": "opus" in ~/.claude/settings.json and the default thinking effort (High). Do not set MAX_THINKING_TOKENS=0.
5. Send any prompt.

Expected: a normal response from Anthropic Opus 4.7.

Actual: tight retry loop with 500s from Tingly wrapping the 400 above.

Relevant log lines

From ~/.tingly-box/log/tingly-box.jsonl:

[error] POST /tingly/claude_code/v1/messages?beta=true 500 — error: "Failed to forward request: POST \"https://api.anthropic.com/v1/messages?beta=true\": 400 Bad Request ... `clear_thinking_20251015` strategy requires `thinking` to be enabled or adaptive"

Earlier in the same log (after rule reconfiguration but before the strategy error surfaced), a related 500 also appeared with CLAUDE_CODE_MAX_OUTPUT_TOKENS=32000:

[error] POST /tingly/claude_code/v1/messages?beta=true 500 — error: "streaming is required for operations that may take longer than 10 minutes"

That one was sidestepped by lowering CLAUDE_CODE_MAX_OUTPUT_TOKENS from the apply-default of 32000 to 8192. Worth noting because the apply CLI ships a default that Anthropic's policy now rejects without streaming — possibly a separate small bug.

Workaround

Setting MAX_THINKING_TOKENS=0 in ~/.claude/settings.json env block disables Claude Code's extended thinking, which in turn stops Claude Code from attaching the clear_thinking_20251015 strategy. Requests then succeed.

This is acceptable as a temporary workaround but disables Opus 4.7's reasoning quality on the orchestrator — the very feature most users pay for the subscription to get.

Suggested fix

In the protocol-transform pipeline that handles the OAuth Anthropic path (likely internal/protocol/transform/... and the assembler/dispatch layer touched by recent PRs like #865, #875), preserve the thinking field alongside context_management when forwarding to upstream Anthropic. The two are coupled — if context_management.strategy is one of the clear_thinking_* family, thinking must be present (or removed together).

Either option resolves it:

• (a) Forward thinking alongside context_management whenever both are present in the incoming request.
• (b) Strip context_management.strategy matching ^clear_thinking_ whenever thinking is being stripped.

Option (a) preserves Claude Code behavior; (b) is safer if there's a reason thinking must be removed for OAuth.

Related

• The streaming is required error suggests the apply CLI's default CLAUDE_CODE_MAX_OUTPUT_TOKENS=32000 value is now too high for non-streaming paths. Consider lowering the default to 8192 in internal/agent/apply.go::GenerateClaudeCodeEnv.
• This blocks the canonical "Opus orchestrator + DeepSeek subagents" topology Tingly Box markets toward Claude Code users with a Pro/Max subscription.

[Chee-0806]

Thanks for the clear analysis. You are right — the fix should follow one of the two approaches you outlined.

I implemented both (a) and (b) for defense in depth and verified locally (tests pass, build succeeds). You can find the changes on my fork at Chee-0806/tingly-box on the fix/thinking-claude-opus-4-7 branch.

Commit: Chee-0806@10a9448f

(a) — Don't strip thinking from claude-opus-4-7 (the real fix)

The root cause is that isThinkingSupportedModel in internal/protocol/transform/ops/request_anthropic_model.go only matched claude-opus-4-6 / claude-sonnet-4-6 by exact string. Since Claude Code routes to claude-opus-4-7, the match failed and thinking was stripped while context_management survived.

Fix: Changed match from suffix -4-6 to prefix -4-, so claude-opus-4-7 (and any future Gen 4.x model) is recognized as thinking-capable.

(b) — If thinking is stripped, strip clear_thinking_ too (safety net)

Added stripBetaThinkingContextManagement in applyAnthropicBetaThinkingFilter. If a future model slips through and thinking gets removed, the function also removes any clear_thinking_20251015 edits from context_management.

[Chee-0806]

Just to follow up — this fix is available on my fork at Chee-0806/tingly-box branch fix/thinking-claude-opus-4-7 (commit 10a9448f). Feel free to give it a try.

We're keeping this as a fork-only fix for now — no PR to main, as we want to keep PR #875 as the sole target for related work.

The fix itself is unchanged from the earlier analysis:

• (a) Updated isThinkingSupportedModel to match claude-opus-4-7 (and future 4.x models) so thinking isn't stripped while context_management survives.
• (b) Safety net: if thinking somehow gets stripped, the code also removes any clear_thinking_20251015 edits from context_management.

[0x0079]

Got this point, and glad to see the new task result (@Chee-0806).
I found claude code oauth support using a special http round trip, it works but hard to maintain.

I am going to provide an individual claude code client (just like anthropic client) like codex client.
We can keep going with round trip before the claude code client is confirmed.

[FFengIll]

Got this point, and glad to see the new task result (@Chee-0806). I found claude code oauth support using a special http round trip, it works but hard to maintain.

I am going to provide an individual claude code client (just like anthropic client) like codex client. We can keep going with round trip before the claude code client is confirmed.

Got it, and I push it with edits and validations here #897

[ojsol]

Awesome — thanks for the quick turnaround, guys.

[0x0079]

Just to follow up — this fix is available on my fork at Chee-0806/tingly-box branch fix/thinking-claude-opus-4-7 (commit 10a9448f). Feel free to give it a try.

We're keeping this as a fork-only fix for now — no PR to main, as we want to keep PR #875 as the sole target for related work.

The fix itself is unchanged from the earlier analysis:

• (a) Updated isThinkingSupportedModel to match claude-opus-4-7 (and future 4.x models) so thinking isn't stripped while context_management survives.
• (b) Safety net: if thinking somehow gets stripped, the code also removes any clear_thinking_20251015 edits from context_management.

Hi @Chee-0806 , would you like to submit this PR?

[Chee-0806]

@0x0079

would you like to submit this PR?

Already done — #992

A quick note on the implementation: I originally used a prefix-match approach (strings.Contains(model, "claude-opus-4-")) but after reviewing the full model list in the Anthropic SDK, I switched to a deny-list on Gen 3:

func isThinkingSupportedModel(model string) bool {
    if !strings.HasPrefix(strings.ToLower(model), "claude-") {
        return false
    }
    return !strings.Contains(strings.ToLower(model), "claude-3-")
}

Rationale:

• Gen 3 (claude-3-*) is a closed, non-expanding model family — no new Gen 3 models will be released. A deny-list is stable and won't need updates.
• All Gen 4+ models (claude-opus-4-*, claude-sonnet-4-*, claude-haiku-4-*, claude-mythos-preview, and future Gen 5+) automatically get thinking support without code changes.
• Non-Claude models and empty strings return false (safe conservative default for non-Anthropic routing paths).

The PR is marked as draft with two discussion points open:

1. CLAUDE_CODE_MAX_OUTPUT_TOKENS=32000 default causing "streaming required" errors — worth lowering to 8192?
2. Long-term direction for Claude Code OAuth — dedicated client vs. round-trip

Happy to iterate if you have any concerns.

[andreasfoo]

Adding the setting "alwaysThinkingEnabled": false fixes the issue:

400 Bad Request (Request-ID: req_011CbT8kZE6JCr1WwRFyQJtL) {"type":"error","error":{"type":"invalid_request_error","message":"\clear_thinking_20251015strategy requiresthinking to be enabled or adaptive"},"request_id":"req_011CbT8kZE6JCr1WwRFyQJtL"}

According to the settings reference at https://claudefa.st/blog/guide/settings-reference, “Enable extended thinking by default” is already false by default. So I honestly can’t explain why this works, but it does.