zhanghanduo · Oct 1, 2026 · Oct 1, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -7,6 +7,17 @@ the GitHub Release body, so a release with no entry here fails.
 
 Versioning follows [docs/versioning.md](docs/versioning.md).
 
+## [0.12.3] - 2026-10-01
+
+### Fixed
+
+- Anthropic clients omit sampling parameters from requests in both thinking and non-thinking modes, avoiding `TypeError` with Anthropic Python SDK v1 and HTTP 400 responses from newer models. Constructor, `chat`, and `stream` temperature arguments remain accepted for compatibility but are no longer sent; thinking and effort settings are unchanged.
+- Anthropic requests fold the trailing per-call runtime addendum into the preceding user message instead of sending a second consecutive user message; translating gateways (llm-hub before deepseek-flash) intermittently rejected the split form. The prompt-cache breakpoint stays on the last persistent block.
+- The Anthropic client now sends the tool results of one turn's parallel tool calls in a single user message, as Anthropic documents, instead of one user message per result. Translating gateways (llm-hub in front of a non-Claude model such as deepseek-flash) rejected the split form with a 400.
+- Failed Responses results now raise instead of returning their partial output, so tool calls from a response the server reported as `failed` are never executed. The raised `LLMError` keeps the provider error code and message with an equivalent HTTP status: rate limits use 429, server errors use 500, vector-store timeouts use 504, and other known SDK codes use 400. Unknown or missing codes retain the existing generic retry behavior, while existing text-based fallback and transient-network rules still take precedence. Streaming `response.failed` events now raise the same error instead of being ignored.
+- The runaway-retry reminder is now marked per-call (`transient`) in every recovery path, including early-stopped streams and retries after expanded-context overflow, like the runtime addendum it follows. Unmarked, it ended the Anthropic fold of that addendum and a retried request went out as three consecutive user messages. Cache breakpoints stay on persistent content and reminders never accumulate in history.
+- Summary LLM failures now log the provider HTTP response body, making rejected requests easier to diagnose.
+
 ## [0.12.2] - 2026-09-27
 
 ### Fixed

diff --git a/changes/a4-anthropic-omit-sampling.fix.md b/changes/a4-anthropic-omit-sampling.fix.md
diff --git a/changes/anthropic-fold-transient-addendum.fix.md b/changes/anthropic-fold-transient-addendum.fix.md
diff --git a/changes/anthropic-merge-tool-results.fix.md b/changes/anthropic-merge-tool-results.fix.md
diff --git a/changes/responses-failed-error-routing.fix.md b/changes/responses-failed-error-routing.fix.md
diff --git a/changes/runaway-reminder-transient.fix.md b/changes/runaway-reminder-transient.fix.md
diff --git a/changes/summary-log-error-body.fix.md b/changes/summary-log-error-body.fix.md
diff --git a/pyproject.toml b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "apodex-agent-core"
-version = "0.12.2"
+version = "0.12.3"
 description = "Shared, product-neutral runtime primitives for Apodex agents"
 readme = "README.md"
 license = "Apache-2.0"

diff --git a/uv.lock b/uv.lock