Emit SSE error frame for unsupported Cursor AgentService IDE tools
instead of assistant content, and drop frames after the turn finishes
to avoid double-closing the stream controller.
Update model lists and context lengths across free/apikey providers,
move bazaarlink, kilo-gateway, and kimchi into freeTier, demote llm7 to
apikey, and hide bluesminds, sambanova, zed, and mimo-free.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Default to the full agent with built-in tools, expose client function
tools as an MCP server, surface tool calls as OpenAI tool_use, resolve
workspace cwd from the request, and bump context windows.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Exchange Personal Access Tokens for short-lived job tokens, close the
SSE stream on terminal frames so non-streaming clients do not hang,
re-enable OAuth plus API-key auth modes, and replace the model catalog
with the current Qoder aliases.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Decode GetGrokCreditsConfig frames when REST billing returns empty
caps, so SuperGrok weekly quota shows in the usage dashboard.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Wire /v1/usages for Kimi (OAuth + API key) and balance API for DeepSeek,
flag both providers with usage/usageApikey, and normalize their quotas
in the dashboard ProviderLimits parser.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Wire Devin CLI as a routed provider that spawns the local `devin acp`
binary. Add the DevinCliExecutor, register it in the executor map, expose
its status through the cli-tools batch endpoint and devin-settings route,
and document setup in cliTools constants.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Register 9 new upstream providers with logos and update the auto-generated
registry index. Refresh providers/alias baselines and extend the alias
token allowlist so verify-alias stays green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rotating-RT providers (xAI/grok-cli) issue a new refresh_token on every
refresh; mutate credentials in-place so refreshWithRetry reuses the fresh
RT instead of the already-consumed one.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add gemini-3.6-flash tiered (high/medium/low) for Antigravity routing
via upstreamModelId "gemini-3.6-flash-tiered(level)" + thinkingLevel,
plus gemini-3.6-flash and gemini-3.5-flash-lite direct API models.
- getModelUpstreamId: split (level) suffix before lookup, re-append after
- Antigravity executor: preserve transformed body.model
- MITM extractModel: parse thinkingLevel for tiered model (default medium)
- Isolate Cloud Code endpoints: discovery (loadCodeAssist/onboardUser/
quota) on PROD cloudcode-pa, chat transport on daily-cloudcode-pa
to bypass prod 429
Clear stale provider error code and account lock after a successful web
fetch (the core fetch handler never consumed the onRequestSuccess
callback), switch Jina Reader to its documented JSON POST request, and
parse the Title: metadata line before falling back to a Markdown heading.
Strip the generic dashboard model(level) suffix before resolving Kiro
synthetic -thinking/-agentic variants so the upstream request no longer
carries an invalid parenthesized model id. Map explicit levels to native
Kiro effort fields only for supported Claude/GPT model families, and stop
advertising native levels for unsupported legacy Kiro models. Applies to
both OpenAI→Kiro and direct Claude→Kiro routes.
Real Cursor IDE now uses AgentService at agent.api5.cursor.sh (HTTP/2-only)
while 9router still spoke the retired ChatService at api2.cursor.sh with
outdated headers, producing HTTP 429 "Update Required". Add an executeAgent
path that builds an agent.v1.RunRequest Connect RPC over a raw http2 stream
and fetches the account-specific usable model catalog via GetUsableModels.
Also implement MCP tool calling over AgentService: encode OpenAI tools as
AgentRunRequest.mcp_tools (McpToolDefinition with google.protobuf.Value
input_schema), decode McpArgs tool calls, and forward them to the client as
OpenAI tool_calls so the client runs the tool and resumes in the next turn.
Reply to request_context_args with a non-empty RequestContext, to server
heartbeats with client_heartbeat, and to KV blob get/set with empty results,
so action queries no longer stall the stream. Fold the client system prompt
into the user message (custom_system_prompt makes the server return an empty
turn). Bump clientVersion to 3.12.17 and add the x-cursor-client-commit
header so the gateway identifies as a current Cursor IDE release.
Forward the service_tier field from OpenAI requests into the Responses API payload so clients can select priority/default/flex tiers instead of the field being silently dropped.
Route GPT-5.6 reasoning effort through Kiro's native reasoning.effort field instead of the legacy Claude output_config.effort path. GPT-5.6 models now emit reasoning.effort for low/medium/high/xhigh, with max mapped to the xhigh wire value.
Preserve the Responses API reasoning.effort through the OpenAI intermediate by copying it to reasoning_effort before the field is dropped. Skip legacy thinking_mode prompt tags when a supported native GPT effort is emitted, while keeping the legacy fallback for unsupported values (auto/minimal/ultra) and explicit disable semantics (none/off/disabled). Claude adaptive effort continues to use thinking plus output_config.effort.
Validate AWS EventStream framing, header bounds, CRCs, error frames,
and terminal stop metadata before exposing Kiro output. Classify stop
reasons into dispositions (complete / retryable / terminal_incomplete /
refusal) and retry once when the stream ends with a malformed tool call,
ellipsis-only output, or a short future-action sentence.
Fail closed: propagate streaming failures as error SSE (502) instead of
collapsing them into a successful stop, so incomplete responses no longer
leak as final answers.
Detect the observed evidence-prefixed trailing progress final without
broadening the Chinese heuristic to completed findings.
8b9cac1 swapped alicode-intl to the DashScope compatible-mode endpoint to
fix#2591 for standard DashScope keys, but that broke Coding Plan keys
(sk-sp-...) which only work on coding-intl.dashscope.aliyuncs.com. The two
key types use two different hosts and are not interchangeable.
- alicode-intl: revert to coding-intl endpoint (Coding Plan keys)
- alims-intl: new provider for dashscope-intl/compatible-mode (standard keys)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Switch baseUrl from coding-intl.dashscope.aliyuncs.com (Coding Plan keys
only) to dashscope-intl.aliyuncs.com/compatible-mode so ordinary DashScope
API keys authenticate. Path /v1/chat/completions and preserveCacheControl
quirk unchanged.
Fixes#2591
Reshape Kiro direct requests so resumed client sessions reuse Kiro's
cache-affinity fields instead of starting unrelated CodeWhisperer
conversations.
- keep conversationState.conversationId stable when the client sends an
explicit session id (x-session-id, session_id, conversation_id, Claude
Code session metadata)
- add a stable conversationState.agentContinuationId per Kiro session
- send conversationState.agentTaskType: "vibe" and agentMode: "vibe",
matching the normal Kiro CLI/KAS chat path
- move Kiro thinking instructions into Kiro-compatible systemPrompt /
additionalModelRequestFields instead of generic top-level thinking
- keep volatile timestamp context out of the top-level systemPrompt; it
remains only in user content fallback
- suppress additionalModelRequestFields for legacy 4.5-era Claude/Kiro
models that reject it, while defaulting future Claude/Kiro model ids
to supported
- preserve Kiro meteringEvent credit usage internally for accounting
without leaking provider-specific fields into OpenAI-compatible usage
- prevent unrelated headerless Kiro requests from sharing one
connection-wide continuation
- cap/evict continuation sessions so long-running processes do not grow
the continuation map unbounded
- treat generated headerless Kiro sessions as one-shot so they do not
evict real explicit-session continuations
- keep credit-only Kiro metering valid for internal persistence when
token metrics are unavailable
Add GPT-5.6 Sol/Terra/Luna and their synthetic thinking/agentic/
thinking-agentic variants to the Kiro static catalog with the observed
272k context window and credit multipliers (2.4/1.2/0.6), register MITM
mapping slots for the new base ids, and override runtime capabilities so
the GPT-5.6 family reports the 272k window instead of the generic GPT-5
profile.
client_metadata is an OpenAI Responses API-specific field. When translating
openai-responses requests to openai (Chat Completions), it leaked through
to providers like NVIDIA, which rejected it with a 400 "Unsupported
parameter". Strip it in the Responses-specific cleanup block alongside
input, instructions, store, and reasoning.
GitHub Copilot's /chat/completions and /responses endpoints never surface
prompt-cache token counts for Claude models. Route Claude models (detected
by name pattern) to Copilot's Anthropic-native /v1/messages shim via a new
executeWithMessagesEndpoint(), translating OpenAI-shape requests to Claude
natively so cache_control gets injected and cached_tokens surface.
Also fixes translateRequest()'s internal _toolNameMap being sent upstream,
which made Anthropic's strict schema reject tool-call requests with a 400 —
now stripped and threaded through response state. Removes the now-dead
response_format Claude JSON-mode workaround.
Broaden strip rule from /claude-opus-4/i to /claude/i so temperature is
removed for every Claude model, not just opus-4. Fixes Anthropic 400 on
OpenAI-compatible routes. #1748
Add pxpipe as an experimental fifth Token Saver: Claude-format request
bodies above a configurable size threshold are rendered as dense PNGs
via the pxpipe-proxy library API (transformAnthropicMessages) before
dispatch, cutting estimated input tokens by ~35-60% on token-dense
contexts. Integration follows the Headroom pattern: applied to the final
body in chatCore just before dispatch, fail-open on any error/timeout.
Managed npm install into DATA_DIR/pxpipe, dynamic loader with per-version
cache-bust, JSONL event log with rotation, /api/pxpipe/* endpoints, Token
Saver card (marked experimental) + /dashboard/pxpipe page, and per-request
Activated/Skipped annotation in Request Details. Disabled by default.
- Backup DB only on real SCHEMA_VERSION change, not every app version bump
- Kill idle MCP stdio bridge children to prevent orphan process leaks
- Add getDistinctProviders to avoid loading every row JSON blob (OOM fix)
- Update codex model list (gpt-5.6 sol/terra/luna, drop 5.3 codex variants)
- Reorder Claude default models (fable first)
Co-authored-by: Cursor <cursoragent@cursor.com>
Only update an existing Codex OAuth row when both rows share the same
chatgptAccountId, so a second Codex login no longer overwrites the first
account's rotated token pair. Also fall back to
workspaceId || chatgptAccountId || accountId for the chatgpt-account-id header.
Project conversationState history/currentMessage into OpenAI-style
messages for /v1/compress, then write compressed text back into the
original Kiro fields while preserving provider payload shape. Fail open
when the proxy returns malformed or reordered messages.
Co-authored-by: Cursor <cursoragent@cursor.com>
Gemini CLI requests with small max_tokens spend the whole output budget on
thoughts after reasoning_effort maps to thinkingConfig, returning blank
content or finish=length. Raise maxOutputTokens floors per thinking level/
budget (clamped to caps.maxOutput). Also emit toolConfig
functionCallingConfig.mode=VALIDATED for Gemini CLI tool requests to avoid
MALFORMED_FUNCTION_CALL.
Co-authored-by: Cursor <cursoragent@cursor.com>
Expose max in the Codex thinking dropdown for gpt-5.6-sol only (maps to
xhigh on wire; live probe rejected ultra). Include custom/kilo models
when computing provider thinking options so manually added gpt-5.6-sol
contributes its max level.
Co-authored-by: Cursor <cursoragent@cursor.com>
Translator only unwrapped tool.function when both tool.type==="function"
and tool.function were truthy. Loose/legacy OpenAI clients emit the bare
{ function: { name, parameters } } shape (no parent type), which fell
through and forwarded name: undefined upstream, rejected by strict
Anthropic-compatible gateways (MiniMax M3) as (2013) invalid tool type.
Unwrap tool.function whenever present. Built-in tools stay pass-through.
Adds regression coverage for the 4 tool shapes. See #2435.
Co-authored-by: Cursor <cursoragent@cursor.com>
Claude Code sends reasoning_effort "max" (its top level); OpenAI enum caps
at "xhigh" and rejects "max" with HTTP 400 "max effort not support".
applyFormat case "openai" now clamps "max"->"xhigh" before assigning
body.reasoning_effort; other levels pass through unchanged.
Add regression test covering client output_config.effort, direct
reasoning_effort, passthrough of xhigh/high, and budget_tokens capping.
Co-authored-by: Cursor <cursoragent@cursor.com>
isPathLike rejected any line with a colon, so Windows absolute paths
(C:\Users\me\a.js) were never recognized and find dumps went uncompacted.
find.js also split only on "/", mis-grouping backslash paths.
- autodetect: treat drive-letter prefix (X:\ or X:/) as path-like before
the general colon rejection.
- find.js: split on the last "/" or "\" separator and normalize emitted
directory labels to forward slashes.
Co-authored-by: Cursor <cursoragent@cursor.com>
- map service_tier=fast to upstream priority; drop unsupported tiers
- normalize reasoning effort max to xhigh (codex-only)
- convert 200-SSE model-capacity errors into 503 so account fallback rotates
- keep normal SSE output intact after peeking
Co-authored-by: Cursor <cursoragent@cursor.com>
VolcEngine Ark caps the Kimi family at max_tokens <= 32768, but the
model's advertised ceiling is far higher (Kimi-K2.7-Code resolves to
maxOutput 262144), so clampToModelMaxOutput alone leaves it uncapped and
the request 400s. Add a Kimi-scoped rule with an explicit maxOutputCap of
32768, combined with the model ceiling via min(). Covers max_tokens,
max_completion_tokens, max_output_tokens.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Match captured official Antigravity IDE traffic: cloudcode-pa host,
antigravity/ide/2.1.1 User-Agent, IDE-shaped agent requestId, and drop
router-only stream/usage headers plus the legacy double system prompt.
Co-authored-by: Cursor <cursoragent@cursor.com>
Ark rejects max_tokens above 128000 for GLM-5.2. Add a config-driven STRIP_RULES entry that clamps max_tokens, max_completion_tokens and max_output_tokens down to the model maxOutput before the upstream call.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add four shared Caveman prompt fragments (no invented abbreviations,
preserve user language, no self-reference, no decoration) across all six
levels, and remove ULTRA contradictions around abbreviations/arrow
shorthand. Adds regression tests for the prompt rules.
Co-authored-by: Cursor <cursoragent@cursor.com>
Map role="developer" messages to top-level instructions alongside
role="system" in openaiToOpenAIResponsesRequest. Previously developer
messages matched no branch and were silently dropped from the Responses
request, losing GPT-5/Codex system-level prompts.
Co-authored-by: Cursor <cursoragent@cursor.com>
On the translated OpenAI->Claude path, adjustMaxTokens capped max_tokens
before applyThinking set thinking.budget_tokens, so max-effort budget
(128000) could exceed a 64k-clamped max_tokens -> Anthropic 400.
prepareClaudeRequest now reconciles after the budget is known: prefer
raising max_tokens, only shrink budget when it meets/exceeds the ceiling.
Also lift the global 64000 cap: the ceiling is now the model's real
maxOutput, so high-output models (fable/mythos, opus-4.8/sonnet-4.6) get
their full budget. adjustMaxTokens gains an optional ceiling arg (default
unchanged, callers untouched); openai-to-claude passes the model maxOutput.
Native Claude Code passthrough is unaffected.
Co-Authored-By: Claude <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
- Pass system prompt via native systemInstruction field (+ <instructions> fallback)
so Claude models stop treating it as info-only <system-reminder>
- Add Opus 4.5/4.7/4.8 (base/thinking/agentic/thinking+agentic) to Kiro registry
- normalizeModelId(): dash->dot version separator, scoped to Kiro provider only
- Replace <system-reminder> with <instructions> in claude-to-openai/openai-to-kiro
Co-authored-by: Cursor <cursoragent@cursor.com>
Normalize every provider to one cache-inclusive convention via
canonicalizeUsage() before persist, and price cached + cache_creation as
subsets of prompt_tokens in calculateCostFromTokens() to stop
double-counting. usageRepo now delegates cost math to a single source.
Surface Cached tokens/cost across dashboard (overview, tokens, cost,
details). Merge Claude message_start cache with message_delta output so
cache counts survive. Compatible LLM nodes now allow multiple API-key
connections (key pool).
Co-authored-by: Cursor <cursoragent@cursor.com>
handleStreamingResponse and buildOnStreamComplete each generated their
own streamDetailId for what should be one logical record — the
placeholder row (0 tokens) and the final row (real usage) never shared
an id, so the DB's ON CONFLICT(id) upsert never merged them, leaving a
permanent 0-token stub for every streaming request.
Share the id from buildOnStreamComplete with handleStreamingResponse
so both writes hit the same row.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add read-only GET to inspect per-credit reset inventory (status, granted,
expiry, remaining) with a Quota Tracker modal. DRY the route via shared
connection/refresh helpers; keep existing consume POST unchanged.
Co-authored-by: Cursor <cursoragent@cursor.com>
Combo mixes models, so non-Claude thinking signatures leak into
conversation history. Native passthrough forwarded them verbatim and
Anthropic rejected the request. Validate signatures and drop invalid
thinking blocks, re-inserting a placeholder when tool_use requires one.
Co-authored-by: Cursor <cursoragent@cursor.com>
- Updated capabilities for NVIDIA models to enforce OpenAI-compatible reasoning formats.
- Added new models: MiniMax M3, GLM 5.2, DeepSeek V4 Pro, DeepSeek V4 Flash, Kimi K2.6, and Nemotron 3 Ultra to the NVIDIA registry.
This enhances the provider's functionality and aligns with OpenAI standards.
IAM Identity Center (authMethod=idc) tokens failed every request with 403
"bearer token invalid". Treat idc like api_key/external_idp:
- executors/kiro.js: route idc to *.amazonaws.com CodeWhisperer surface,
region-aware from credentials.region instead of hardcoded us-east-1.
- openai-to-kiro.js / claude-to-kiro.js: send resolved profileArn or empty
for idc/external_idp, never the shared builder-id placeholder ARN.
Co-authored-by: Cursor <cursoragent@cursor.com>
Filter empty text blocks from thoughtSignature-only parts, preserve
tool_calls when functionResponse and functionCall coexist in the same
content, and skip empty regular text parts before they reach Claude.
Co-authored-by: Cursor <cursoragent@cursor.com>
Clients echo full message history each turn including reasoning_content,
which the Kimchi OpenAI gateway counts as input tokens. Multi-turn convos
balloon to 100k+ tokens and the model returns empty content.
KimchiExecutor.transformRequest now strips reasoning_content from assistant
messages when it exceeds an 8-char threshold, preserving the 1-char
placeholder injectReasoningContent sets and keeping content intact.
Co-authored-by: Cursor <cursoragent@cursor.com>
CodeBuddy CN includes "tool_calls": [] in every SSE streaming delta.
@ai-sdk/openai-compatible checks delta.tool_calls != null — an empty
array passes ([] != null is true in JS), triggering premature
reasoning-end on every reasoning chunk (0/1ms durations in OpenCode).
Strip empty tool_calls arrays in passthrough before hasValuableContent.
Zero side-effect: real tool_calls always have at least one element.
Closes#2176
Co-authored-by: Cursor <cursoragent@cursor.com>
Gemini response translation wrote OpenAI-shaped bookkeeping into the
shared state.toolCalls map, which the downstream openai-to-claude
translator uses for Claude block metadata. That pre-population skipped
blockIndex creation, so Anthropic input_json_delta events lost index.
Track Gemini function calls via state.geminiToolCallCount instead,
leaving state.toolCalls clean for the Claude translator.
Closes#2218
Co-authored-by: Cursor <cursoragent@cursor.com>
Treat response.done as a terminal OpenAI Responses stream event so
passthrough streams ending with response.done are not flagged incomplete
and no synthetic response.failed is emitted. Restore the data: [DONE]
sentinel for same-format Responses passthrough streams.
Co-authored-by: Cursor <cursoragent@cursor.com>
Guard openai-responses compression: skip Headroom when body.input
contains non-message items (function_call, function_call_output,
reasoning) to preserve the Responses contract instead of collapsing
them into chat messages.
Co-authored-by: Cursor <cursoragent@cursor.com>
Claude Code chèn role:system cuối messages[], trước đây bị map thành assistant
khiến hội thoại không kết thúc bằng user → provider OpenAI-compat (LiteLLM)
dịch ngược Anthropic trả 400 "assistant message prefill". Map system -> user
và wrap <system-reminder> để giữ ngữ nghĩa instruction.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add qwen omni (audio/video input), qwen3.5/3.6/3.7 (native vision/video),
and mark qwen coder & max as text-only reasoning models.
Co-authored-by: Cursor <cursoragent@cursor.com>
Merge adjacent same-role blocks and strip empty parts before sending to
Gemini, avoiding 400 INVALID_ARGUMENT on consecutive same-role messages.
Co-authored-by: Cursor <cursoragent@cursor.com>
Opt-in quirk preserveCacheControl keeps cache_control on content blocks
for alicode/alicode-intl, enabling DashScope prompt caching. signature
is always stripped; all other providers unchanged.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add Kimchi as a browser-token OAuth provider routed through its
OpenAI-compatible gateway. Discover live models for /v1/models and
provider models, normalize Claude-compatible requests, and wire up
provider connection tests.
Co-authored-by: Cursor <cursoragent@cursor.com>
Backfill DEFAULT_THINKING_AG_SIGNATURE onto functionCall parts missing it
(client history replay) and on Claude tool_use blocks, fixing 400
INVALID_ARGUMENT from Gemini-family APIs. Suppress the OpenAI-style
data: [DONE] sentinel for antigravity/gemini/vertex to avoid parser crashes.
Fixes#2193.
Co-authored-by: Cursor <cursoragent@cursor.com>
Gemini rejects the non-standard 'deprecated' keyword in nested tool
schemas with INVALID_ARGUMENT (400). Add it to UNSUPPORTED_SCHEMA_CONSTRAINTS
alongside 'optional' so it gets stripped during translation.
Co-authored-by: Cursor <cursoragent@cursor.com>
CodeBuddy CN bonus packs ("Bonus Pack N") are one-shot credits whose
CycleEndTime equals DeductionEndTime — they expire for good and never
replenish. The dashboard rendered their resetAt as "Reset in Xd",
implying a monthly refill.
Tag bonus packs recurring:false (refill packs recurring:true) in the
usage handler, forward the flag through parseQuotaData, and word the
quota table / progress bar as "Expires in" / "Expires at" for
one-shot packs.
CodeWhisperer leaks literal <thinking> blocks into assistantResponseEvent,
duplicating reasoning already routed via reasoningContentEvent. Track
inThinking state to strip these tags during SSE transform, handling split
chunks across tag boundaries.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add modelsFetcher + passthroughModels so the dynamic Kilo Gateway
catalog surfaces in the combo model picker, matching openrouter.js.
Co-authored-by: Cursor <cursoragent@cursor.com>
Gemini/Antigravity generateContent rejects the JSON Schema annotation
keywords deprecated, readOnly, writeOnly with a 400 INVALID_ARGUMENT.
MCP tool schemas (e.g. Claude Code) commonly set deprecated:true, making
every request with such a tool fail. Add them to
UNSUPPORTED_SCHEMA_CONSTRAINTS so cleanJSONSchemaForAntigravity removes
them recursively before the request is sent.
Co-authored-by: Cursor <cursoragent@cursor.com>
Cloudflare AI rule only sets flattenContent. Treat missing match as
provider-wide and missing drop as empty list to avoid crash. Fixes#1960.
Co-authored-by: Cursor <cursoragent@cursor.com>
Retry short-lived 5xx/capacity errors (500/502/503/504 + message
patterns) with bounded backoff capped at 15s; honor Retry-After/reset
hints and skip when wait is too long. Keep 400 non-retryable. Enable the
retry hook for 500 alongside existing 429/503.
Deduplicate sanitized Antigravity tool names before emitting the single
functionDeclarations group to avoid upstream "Tool names must be unique"
rejections.
Add Headroom size diagnostics and phantom-savings warning when reported
token delta does not shrink the outbound payload.
Co-authored-by: Cursor <cursoragent@cursor.com>
Carry Claude reasoning_effort/reasoning into OpenAI Chat, map into
OpenAI Responses reasoning.effort, and keep request-level effort
(incl. xhigh) across tool-result turns instead of collapsing to high.
Co-authored-by: Cursor <cursoragent@cursor.com>
