Skip to content

fix(thinking): stop injecting fabricated redacted_thinking block on plain replay (#5945) #5953

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
diegosouzapw merged 1 commit into from
Jul 2, 2026
Merged

fix(thinking): stop injecting fabricated redacted_thinking block on plain replay (#5945) #5953

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
233 changes: 134 additions & 99 deletions open-sse/translator/request/openai-to-claude.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,6 +165,97 @@ export function openaiToClaudeRequest(model, body, stream) {
result.stop_sequences = Array.isArray(body.stop) ? body.stop : [body.stop];
}

// Thinking configuration
// NOTE: computed BEFORE message-block conversion (below) so that
// `getContentBlocksFromMessage` knows whether the outbound request actually has
// extended thinking enabled — required to correctly gate the `redacted_thinking`
// replay-placeholder injection (#5945). This block has no dependency on
// `result.messages`/`toolNameMap`, so moving it earlier is safe.
if (body.thinking) {
result.thinking = {
type: body.thinking.type || "enabled",
...(body.thinking.budget_tokens && { budget_tokens: body.thinking.budget_tokens }),
...(body.thinking.max_tokens && { max_tokens: body.thinking.max_tokens }),
};
} else if (body.reasoning_effort) {
// Convert OpenAI reasoning_effort to Claude thinking format (#627)
// Clients like OpenCode send reasoning_effort via @ai-sdk/openai-compatible
const requestedEffort = String(body.reasoning_effort).toLowerCase();
const normalizedEffort =
requestedEffort === "max" && !supportsClaudeMaxEffort(model)
? "high"
: requestedEffort === "xhigh" && !supportsXHighEffort("claude", model)
? "high"
: requestedEffort;
if (isAdaptiveThinkingOnly(model)) {
// Opus 4.7+/Fable 5 removed manual extended thinking: a fixed `budget_tokens`
// (or `type:"enabled"`) is a hard 400. Steer EVERY level via adaptive +
// output_config.effort instead of the budget buckets below. Unrecognized levels
// leave thinking unset so the model keeps its adaptive default rather than 400ing
// on an invalid effort value.
if (ADAPTIVE_EFFORT_LEVELS.has(normalizedEffort)) {
result.thinking = {
type: "adaptive",
};
result.output_config = {
...(result.output_config || {}),
effort: normalizedEffort,
};
}
} else if (normalizedEffort === "max" || normalizedEffort === "xhigh") {
result.thinking = {
type: "adaptive",
};
result.output_config = {
...(result.output_config || {}),
effort: normalizedEffort,
};
} else {
const effortBudgetMap: Record<string, number> = {
low: 1024,
medium: 10240,
high: 131072,
max: 131072,
};
const budget = effortBudgetMap[normalizedEffort];
if (budget !== undefined && budget > 0) {
result.thinking = {
type: "enabled",
budget_tokens: budget,
};
}
}
}

// Fit thinking budget within the model's output cap and ensure
// max_tokens > budget_tokens for all thinking configurations (#627).
// Replaces the previous unconditional `budget + 8192` inflation, which
// could exceed model caps (e.g. Opus 4.7's 128000 ceiling) and trigger
// HTTP 400 from Anthropic.
const fitted = fitThinkingToMaxTokens(model, Number(result.max_tokens) || 0, result.thinking);
result.max_tokens = fitted.maxTokens;
if (fitted.thinking === undefined) {
delete result.thinking;
} else {
result.thinking = applyCopilotSummarizedThinkingDisplay(fitted.thinking, body);
}

delete result[COPILOT_REASONING_SUMMARY_MARKER];

// Final guard: Claude rejects `temperature` whenever extended thinking is
// enabled. If `result.thinking` was set above from `body.thinking` or
// `body.reasoning_effort` (manual budget or adaptive effort), drop temperature
// defensively. The model-name strip earlier already covers Claude OAuth's
// forced-thinking case (claude-opus-4.x / claude-sonnet-4.x).
if (result.thinking && result.temperature !== undefined) {
delete result.temperature;
}

// Whether the OUTBOUND request actually has extended thinking enabled. Anthropic's
// schema only requires a precursor thinking/redacted_thinking block before a tool_use
// block when thinking mode is active for THIS request — never unconditionally (#5945).
const thinkingEnabledForRequest = Boolean(result.thinking) && result.thinking.type !== "disabled";

// Messages
const systemParts = [];

Expand Down Expand Up @@ -199,7 +290,12 @@ export function openaiToClaudeRequest(model, body, stream) {

for (const msg of nonSystemMessages) {
const newRole = msg.role === "user" || msg.role === "tool" ? "user" : "assistant";
const blocks = getContentBlocksFromMessage(msg, toolNameMap, disableToolPrefix);
const blocks = getContentBlocksFromMessage(
msg,
toolNameMap,
disableToolPrefix,
thinkingEnabledForRequest
);
const hasToolUse = blocks.some((b) => b.type === "tool_use");
const hasToolResult = blocks.some((b) => b.type === "tool_result");

Expand Down Expand Up @@ -387,87 +483,6 @@ export function openaiToClaudeRequest(model, body, stream) {
: [{ type: "text", text: String(body.system) }];
}

// Thinking configuration
if (body.thinking) {
result.thinking = {
type: body.thinking.type || "enabled",
...(body.thinking.budget_tokens && { budget_tokens: body.thinking.budget_tokens }),
...(body.thinking.max_tokens && { max_tokens: body.thinking.max_tokens }),
};
} else if (body.reasoning_effort) {
// Convert OpenAI reasoning_effort to Claude thinking format (#627)
// Clients like OpenCode send reasoning_effort via @ai-sdk/openai-compatible
const requestedEffort = String(body.reasoning_effort).toLowerCase();
const normalizedEffort =
requestedEffort === "max" && !supportsClaudeMaxEffort(model)
? "high"
: requestedEffort === "xhigh" && !supportsXHighEffort("claude", model)
? "high"
: requestedEffort;
if (isAdaptiveThinkingOnly(model)) {
// Opus 4.7+/Fable 5 removed manual extended thinking: a fixed `budget_tokens`
// (or `type:"enabled"`) is a hard 400. Steer EVERY level via adaptive +
// output_config.effort instead of the budget buckets below. Unrecognized levels
// leave thinking unset so the model keeps its adaptive default rather than 400ing
// on an invalid effort value.
if (ADAPTIVE_EFFORT_LEVELS.has(normalizedEffort)) {
result.thinking = {
type: "adaptive",
};
result.output_config = {
...(result.output_config || {}),
effort: normalizedEffort,
};
}
} else if (normalizedEffort === "max" || normalizedEffort === "xhigh") {
result.thinking = {
type: "adaptive",
};
result.output_config = {
...(result.output_config || {}),
effort: normalizedEffort,
};
} else {
const effortBudgetMap: Record<string, number> = {
low: 1024,
medium: 10240,
high: 131072,
max: 131072,
};
const budget = effortBudgetMap[normalizedEffort];
if (budget !== undefined && budget > 0) {
result.thinking = {
type: "enabled",
budget_tokens: budget,
};
}
}
}

// Fit thinking budget within the model's output cap and ensure
// max_tokens > budget_tokens for all thinking configurations (#627).
// Replaces the previous unconditional `budget + 8192` inflation, which
// could exceed model caps (e.g. Opus 4.7's 128000 ceiling) and trigger
// HTTP 400 from Anthropic.
const fitted = fitThinkingToMaxTokens(model, Number(result.max_tokens) || 0, result.thinking);
result.max_tokens = fitted.maxTokens;
if (fitted.thinking === undefined) {
delete result.thinking;
} else {
result.thinking = applyCopilotSummarizedThinkingDisplay(fitted.thinking, body);
}

delete result[COPILOT_REASONING_SUMMARY_MARKER];

// Final guard: Claude rejects `temperature` whenever extended thinking is
// enabled. If `result.thinking` was set above from `body.thinking` or
// `body.reasoning_effort` (manual budget or adaptive effort), drop temperature
// defensively. The model-name strip earlier already covers Claude OAuth's
// forced-thinking case (claude-opus-4.x / claude-sonnet-4.x).
if (result.thinking && result.temperature !== undefined) {
delete result.temperature;
}

// Attach toolNameMap to result for response translation
if (toolNameMap.size > 0) {
result._toolNameMap = toolNameMap;
Expand All @@ -488,7 +503,12 @@ export function openaiToClaudeRequest(model, body, stream) {
}

// Get content blocks from single message
function getContentBlocksFromMessage(msg, toolNameMap = new Map(), disableToolPrefix = false) {
function getContentBlocksFromMessage(
msg,
toolNameMap = new Map(),
disableToolPrefix = false,
thinkingEnabledForRequest = false
) {
const blocks = [];

if (msg.role === "tool") {
Expand Down Expand Up @@ -555,22 +575,6 @@ function getContentBlocksFromMessage(msg, toolNameMap = new Map(), disableToolPr
}
}
} else if (msg.role === "assistant") {
// Add reasoning_content as a replay placeholder (OpenAI extended thinking format).
// #5312 RC-D: reasoning_content carries NO real Claude signature. Emitting a
// `thinking` block with the fabricated DEFAULT signature makes Anthropic reject the
// replay with 400 "Invalid signature in thinking block" — and claudeHelper's
// latest-assistant guard (prepareClaudeRequest) preserves it verbatim, so the fake
// signature leaks upstream. Emit a signature-less redacted_thinking block instead
// (the same shape prepareClaudeRequest produces for Anthropic-native replay);
// Anthropic accepts it without signature validation and non-Anthropic Claude-shape
// upstreams re-hydrate the real text downstream from reasoningCache.
if (msg.reasoning_content) {
blocks.push({
type: "redacted_thinking",
data: DEFAULT_THINKING_CLAUDE_SIGNATURE,
});
}

if (Array.isArray(msg.content)) {
for (const part of msg.content) {
if (part.type === "text" && part.text) {
Expand Down Expand Up @@ -619,6 +623,37 @@ function getContentBlocksFromMessage(msg, toolNameMap = new Map(), disableToolPr
}
}
}

// Add reasoning_content as a replay placeholder (OpenAI extended thinking format) —
// ONLY when Anthropic's schema actually requires a precursor thinking block: the
// outbound request has extended thinking enabled AND this assistant turn contains a
// tool_use block (Anthropic rejects a tool_use turn without a preceding
// thinking/redacted_thinking block when thinking is active). #5312 RC-D:
// reasoning_content carries NO real Claude signature. Emitting a `thinking` block
// with the fabricated DEFAULT signature makes Anthropic reject the replay with 400
// "Invalid signature in thinking block" — and claudeHelper's latest-assistant guard
// (prepareClaudeRequest) preserves it verbatim, so the fake signature leaks
// upstream. Emit a signature-less redacted_thinking block instead (the same shape
// prepareClaudeRequest produces for Anthropic-native replay, gated the same way at
// claudeHelper.ts `thinkingEnabled && !hasThinking && hasToolUse`); Anthropic
// accepts it without signature validation and non-Anthropic Claude-shape upstreams
// re-hydrate the real text downstream from reasoningCache.
// #5945: injecting this unconditionally — for ANY assistant turn carrying
// reasoning_content, regardless of tool_use or thinking state — fabricates a content
// block the client never sent. Some upstream clients (reported: Claude Sonnet 5 via
// the "Pi" harness) detect the extra block and refuse the turn as prompt injection.
// Drop reasoning_content silently when it is not required by the schema, mirroring
// how other echo-only fields are dropped (see OPENAI_INCOMPATIBLE_ECHO_FIELDS).
const hasThinkingBlock = blocks.some(
(b) => b.type === "thinking" || b.type === "redacted_thinking"
);
const hasToolUseBlock = blocks.some((b) => b.type === "tool_use");
if (msg.reasoning_content && thinkingEnabledForRequest && hasToolUseBlock && !hasThinkingBlock) {
blocks.unshift({
type: "redacted_thinking",
data: DEFAULT_THINKING_CLAUDE_SIGNATURE,
});
}
}

return blocks;
Expand Down
Loading
Loading