sscoble/pi-mono
1
0
Fork 0
mirror of https://github.com/badlogic/pi-mono.git synced 2026-11-05 23:37:25 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

feat(ai): keep the prompt cache when changing thinking level on OpenAI GPT-6 (#10930)

Browse source 
With an API key, GPT-6 models on the openai provider now keep reasoning.effort
fixed and send the thinking level as configuration_update input items. Each
reply records its level, so earlier items are rebuilt unchanged on every request.
Controlled by compat.supportsMidConvoEffort; skipped for Sign in with ChatGPT
and thinking off.

closes #10853
This commit is contained in:
David Brailovsky 2026-10-14 10:56:20 +02:00 • committed by GitHub
parent 195120a915
commit 60ac3bcc3b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
8 changed files with 307 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

16
packages/ai/scripts/generate-models.ts
View file

@ -388,6 +388,9 @@ const OPENAI_CODEX_ADDITIONAL_TOOLS_MODEL_IDS = new Set([
"gpt-6-luna",
"gpt-6.1-sol",
]);
// OpenAI documents configuration_update for the GPT-6 family.
// https://developers.openai.com/api/docs/guides/reasoning
const OPENAI_MID_CONVO_EFFORT_MODEL_IDS = new Set(["gpt-6-astra", "gpt-6-sol", "gpt-6-luna", "gpt-6.1-sol"]);
const OPENAI_LONG_CONTEXT_INPUT_THRESHOLD = 272000;
const OPENAI_SHORT_CONTEXT_CAPPED_MODEL_IDS = new Set([
"gpt-5.4",
@ -976,6 +979,18 @@ function applyOpenAIExplicitPromptCacheMetadata(model: Model<Api>): void {
};
}
// GPT-6 models accept `configuration_update` input items that change the reasoning effort
// without changing the cached prompt prefix. Not enabled for openai-codex (Codex CLI keeps it
// behind a flag that is off by default) or proxies.
function applyOpenAIMidConvoEffortMetadata(model: Model<Api>): void {
if (model.provider !== "openai" || model.api !== "openai-responses") return;
if (!OPENAI_MID_CONVO_EFFORT_MODEL_IDS.has(model.id)) return;
model.compat = {
...(model.compat as OpenAIResponsesCompat | undefined),
supportsMidConvoEffort: true,
};
}
// Anthropic ephemeral entries have a hard five-minute lifetime; `ttl: "1h"`
// extends it to one hour. Only direct Anthropic is annotated so cache warming
// does not assume equivalent behavior through proxies.
@ -3348,6 +3363,7 @@ async function generateModels() {
applyOpenAICompletionsTranscriptMetadata(model);
applyOpenAIResponsesTranscriptMetadata(model);
applyOpenAIExplicitPromptCacheMetadata(model);
applyOpenAIMidConvoEffortMetadata(model);
applyPromptCacheMetadata(model);
applyImageInputMetadata(model);
}