[A8Chann]

What this does

Adds CNB2API_ALLOW_THINKING=1 as an explicit opt-in that lets client thinking toggles reach the upstream thinking variant, instead of the current unconditional strip. Stripping stays the default — this only adds an escape hatch for deployments that supply a large max_tokens.

Why

331f5aa strips reasoning_effort / reasoning / enable_thinking / thinking at the single outbound choke point. The reasoning behind it is sound and I reproduced the failure it prevents against the real upstream with a hard prompt:

request	reasoning	content	finish
max_tokens=2048 + toggle	4,589 chars	0	length
max_tokens=8192 + toggle	20,305 chars	0	length
max_tokens=2048, no toggle	0	2,994 chars	stop

So the degenerate stream is real, and the guard is justified.

The problem is that the guard is absolute. It also excludes callers that do provide an adequate budget, and for them thinking works correctly — the pre-331f5aa traffic did exactly this at max_tokens=60000.

Concretely: this repo was synced into a private deployment on 2026-10-08. Reasoning output across that deployment's session logs went from 1,339,875 characters on 09-18 and 3,031,411 on 09-20 to exactly 0 on 10-08, 10-09 and 10-10 — three days of traffic where thinking silently disappeared. I also confirmed there is no client-side workaround against the live gateway: reasoningEffort (camelCase), enableThinking, chat_template_kwargs, extra_body, and thinking_budget all returned zero reasoning_content, because the upstream recognises only the four stripped spellings.

The change

Three lines of behaviour, plus an Anthropic path fix:

• src/config.mjs — allowThinking: process.env.CNB2API_ALLOW_THINKING === '1'
• src/proxy.mjs — stripReasoningTriggers() returns the payload untouched when the flag is set
• src/anthropic.mjs — fromAnthropicRequest() previously dropped body.thinking unconditionally, so the flag alone would not have helped /v1/messages clients (Claude Code). It now maps thinking:{type:"enabled"} to the upstream shape, plus reasoning_effort when the caller supplies one. budget_tokens is not mapped — there is no upstream field for it. The flag is passed in as an option rather than imported, so anthropic.mjs stays dependency-free.
• Docs: .env.example, README.md, README.zh-CN.md

Safety

• Default behaviour is unchanged; the existing strip test passes untouched.
• Strict opt-in: only the exact string 1 enables it (0 / true / yes / empty all keep stripping).
• Requests without a thinking toggle behave identically either way.

Tests

ℹ tests 73
ℹ pass 73
ℹ fail 0

New coverage: opt-in forwards reasoning_effort and thinking with other fields untouched; Anthropic adapter maps thinking when enabled, honours type:"disabled", and leaves plain requests alone.

One unrelated note

test/proxy.test.mjs derives the child cwd from new URL('..', import.meta.url).pathname, which yields /D:/... on Windows, so every spawn fails with port not up. This is pre-existing on main — I verified it on an unmodified checkout — and I left it alone to keep this diff focused. Happy to send it as a separate PR if useful (fileURLToPath(new URL('..', import.meta.url)) fixes it).