GeezerVison

Opus 5.5 and Fable 5.1 bind signed thinking blocks to the exact prefix.
chatd rebuilds the system prompt and tools before each call, so a
mid-turn state change makes Anthropic reject replayed thinking with a
400. On that error, retry once with the thinking-binding-controls beta
and prefix_mismatch_behavior "drop_block", then keep sending both for
the chat and model config.

aibridge InvokeModel now forwards the beta and block_binding together,
or strips the field when the beta is absent.

aibridge removed thinking-binding-controls-2026-08-01 for every Bedrock
model except Opus 5.5 and Fable 5.1, so each new model that enforces
thinking block binding needed a code change. Clients send the beta only
to models that need it, so forward it for any model.

Bedrock names the field thinking.block_binding.mismatch_behavior. Some
regions reject the Anthropic name, prefix_mismatch_behavior, so
InvokeModel now renames it.

chatd now tracks drop_block per chat and model config. A binding error
in compaction with an override model turns on drop_block for the
override config only, and the main model config no longer turns it on
for the override.

chatgpt-codex-connector Bot left a comment

Move the thinking binding retry into a model decorator applied in
resolveModelCall, so turns, computer use, compaction, and the advisor
all get it. The decorator records a provider and model only after a
drop_block retry succeeds, which bounds the in-memory set to configured
models instead of growing with every chat.

chatgpt-codex-connector Bot left a comment

The cached provider and model entry only picks which request goes
first. With the hint set, drop_block is sent first and any 400 before
output retries once without it, clearing the hint if that succeeds.
BYOK keys, key rotation, or a gateway that rejects the beta now cost
one extra 400 instead of failing every request until restart.

chatgpt-codex-connector Bot left a comment

johnstcn left a comment