@fractaal/pi-ai 0.86.9 → 1.6.0

package/dist/api/openai-decisions.js

@@ -0,0 +1,162 @@
+import { formatProviderError, normalizeProviderError } from "../utils/error-body.js";
+import { isRecord, parseClassifierUsage, postClassifierRequest, requiredNumber, } from "./classifier-shared.js";
+/**
+ * OpenAI's Decisions API: `POST /v1/decisions` with `{ model, input, questions }`.
+ * https://developers.openai.com/api/docs/guides/decisions
+ *
+ * The state is sent as JSON text. With images, the input becomes one user message with the state
+ * as `input_text` followed by `input_image` data URLs. Questions map to Decisions types:
+ * `choice` to `choice`, `score` to `score`, and `bool` to `predicate`. Predicates have no
+ * criteria field, so the meanings of true and false are appended to the instructions.
+ *
+ * Only OpenAI API keys work: Sign in with ChatGPT tokens are rejected on this route.
+ */
+const LABEL = "OpenAI Decisions";
+/** The endpoint accepts at most this many image parts per request. */
+const MAX_IMAGES = 128;
+function predicateInstructions(question) {
+    const meanings = [
+        question.criteria.true ? `True means: ${question.criteria.true}` : "",
+        question.criteria.false ? `False means: ${question.criteria.false}` : "",
+    ].filter(Boolean);
+    return meanings.length > 0 ? `${question.instructions}\n\n${meanings.join("\n")}` : question.instructions;
+}
+function wireQuestion(name, question) {
+    if (question.type === "choice") {
+        return {
+            type: "choice",
+            name,
+            instructions: question.instructions,
+            choices: Object.entries(question.criteria).map(([value, description]) => description ? { value, description } : { value }),
+        };
+    }
+    if (question.type === "score") {
+        return {
+            type: "score",
+            name,
+            instructions: question.instructions,
+            levels: question.criteria.map((label) => ({ label })),
+        };
+    }
+    return { type: "predicate", name, instructions: predicateInstructions(question) };
+}
+function wireInput(context) {
+    const state = JSON.stringify(context.state);
+    const images = context.images ?? [];
+    if (images.length === 0)
+        return state;
+    if (images.length > MAX_IMAGES) {
+        throw new Error(`${LABEL} accepts at most ${MAX_IMAGES} images, got ${images.length}`);
+    }
+    return [
+        {
+            role: "user",
+            content: [
+                { type: "input_text", text: state },
+                ...images.map((image) => ({
+                    type: "input_image",
+                    image_url: `data:${image.mimeType};base64,${image.data}`,
+                })),
+            ],
+        },
+    ];
+}
+function choiceProbabilities(value, id) {
+    if (!Array.isArray(value))
+        throw new Error(`${LABEL} returned invalid probabilities for ${id}`);
+    return Object.fromEntries(value.map((entry) => {
+        if (!isRecord(entry) || typeof entry.value !== "string") {
+            throw new Error(`${LABEL} returned invalid probabilities for ${id}`);
+        }
+        return [entry.value, requiredNumber(LABEL, entry.probability, `probability for ${id}.${entry.value}`)];
+    }));
+}
+function parseAnswer(id, question, answer) {
+    if (answer.type === "refusal")
+        throw new Error(`${LABEL} refused to answer ${id}`);
+    if (question.type === "choice") {
+        if (answer.type !== "choice" || typeof answer.choice !== "string") {
+            throw new Error(`${LABEL} did not return a choice answer for ${id}`);
+        }
+        return {
+            type: "choice",
+            choice: answer.choice,
+            probabilities: choiceProbabilities(answer.probabilities, id),
+            confidence: requiredNumber(LABEL, answer.confidence, `confidence for ${id}`),
+        };
+    }
+    if (question.type === "score") {
+        if (answer.type !== "score")
+            throw new Error(`${LABEL} did not return a score answer for ${id}`);
+        return {
+            type: "score",
+            score: requiredNumber(LABEL, answer.score, `score for ${id}`),
+            confidence: requiredNumber(LABEL, answer.confidence, `confidence for ${id}`),
+        };
+    }
+    if (answer.type !== "predicate")
+        throw new Error(`${LABEL} did not return a predicate answer for ${id}`);
+    return { type: "bool", probability: requiredNumber(LABEL, answer.probability, `probability for ${id}`) };
+}
+function parseAnswers(value, context) {
+    if (!Array.isArray(value))
+        throw new Error(`${LABEL} returned an unexpected response`);
+    const byName = new Map();
+    for (const answer of value) {
+        if (isRecord(answer) && typeof answer.name === "string")
+            byName.set(answer.name, answer);
+    }
+    return Object.fromEntries(Object.entries(context.questions).map(([id, question]) => {
+        const answer = byName.get(id);
+        if (!answer)
+            throw new Error(`${LABEL} did not return an answer for ${id}`);
+        return [id, parseAnswer(id, question, answer)];
+    }));
+}
+/**
+ * Cloudflare in front of api.openai.com answers 504 with an HTML page when a request runs longer
+ * than about five seconds. Large inputs, currently above roughly 600K tokens, hit this limit, and
+ * retrying the same input runs into it again, so 504 is not retried.
+ */
+const NO_RETRY_STATUSES = [504];
+function errorMessage(error) {
+    if (error.status === 504) {
+        return `${LABEL} error (504): the request timed out at the gateway. Very large inputs (above roughly 600K tokens) currently exceed its time limit.`;
+    }
+    return formatProviderError(normalizeProviderError(error), `${LABEL} error`);
+}
+/** Classification through OpenAI's Decisions API. */
+export const classify = async (model, context, options) => {
+    const output = {
+        api: model.api,
+        provider: model.provider,
+        model: model.id,
+        answers: {},
+        stopReason: "stop",
+        timestamp: Date.now(),
+    };
+    try {
+        if (model.api !== "openai-decisions")
+            throw new Error(`Unsupported classifier API: ${model.api}`);
+        const body = await postClassifierRequest(LABEL, new URL("decisions", `${model.baseUrl.replace(/\/+$/u, "")}/`), model, {
+            model: model.id,
+            input: wireInput(context),
+            questions: Object.entries(context.questions).map(([id, question]) => wireQuestion(id, question)),
+        }, options, NO_RETRY_STATUSES);
+        if (!isRecord(body))
+            throw new Error(`${LABEL} returned an unexpected response`);
+        // Set before parsing answers: a request with malformed or refused answers was still billed.
+        const usage = parseClassifierUsage(body.usage, model);
+        if (usage)
+            output.usage = usage;
+        output.answers = parseAnswers(body.answers, context);
+        return output;
+    }
+    catch (error) {
+        output.answers = {};
+        output.stopReason = options?.signal?.aborted ? "aborted" : "error";
+        output.errorMessage = errorMessage(error);
+        return output;
+    }
+};
+//# sourceMappingURL=openai-decisions.js.map

package/dist/api/openai-decisions.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"openai-decisions.js","sourceRoot":"","sources":["../../src/api/openai-decisions.ts"],"names":[],"mappings":"AAQA,OAAO,EAAE,mBAAmB,EAAE,sBAAsB,EAAE,MAAM,wBAAwB,CAAC;AACrF,OAAO,EAEN,QAAQ,EACR,oBAAoB,EACpB,qBAAqB,EACrB,cAAc,GACd,MAAM,wBAAwB,CAAC;AAEhC;;;;;;;;;;GAUG;AAEH,MAAM,KAAK,GAAG,kBAAkB,CAAC;AAEjC,sEAAsE;AACtE,MAAM,UAAU,GAAG,GAAG,CAAC;AAUvB,SAAS,qBAAqB,CAAC,QAAuD;IACrF,MAAM,QAAQ,GAAG;QAChB,QAAQ,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,eAAe,QAAQ,CAAC,QAAQ,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE;QACrE,QAAQ,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,gBAAgB,QAAQ,CAAC,QAAQ,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE;KACxE,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;IAClB,OAAO,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,QAAQ,CAAC,YAAY,OAAO,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,YAAY,CAAC;AAC3G,CAAC;AAED,SAAS,YAAY,CAAC,IAAY,EAAE,QAA4B;IAC/D,IAAI,QAAQ,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;QAChC,OAAO;YACN,IAAI,EAAE,QAAQ;YACd,IAAI;YACJ,YAAY,EAAE,QAAQ,CAAC,YAAY;YACnC,OAAO,EAAE,MAAM,CAAC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,KAAK,EAAE,WAAW,CAAC,EAAE,EAAE,CACvE,WAAW,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,WAAW,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAChD;SACD,CAAC;IACH,CAAC;IACD,IAAI,QAAQ,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;QAC/B,OAAO;YACN,IAAI,EAAE,OAAO;YACb,IAAI;YACJ,YAAY,EAAE,QAAQ,CAAC,YAAY;YACnC,MAAM,EAAE,QAAQ,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC;SACrD,CAAC;IACH,CAAC;IACD,OAAO,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,YAAY,EAAE,qBAAqB,CAAC,QAAQ,CAAC,EAAE,CAAC;AACnF,CAAC;AAED,SAAS,SAAS,CAAC,OAA0B;IAC5C,MAAM,KAAK,GAAG,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;IAC5C,MAAM,MAAM,GAAG,OAAO,CAAC,MAAM,IAAI,EAAE,CAAC;IACpC,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IACtC,IAAI,MAAM,CAAC,MAAM,GAAG,UAAU,EAAE,CAAC;QAChC,MAAM,IAAI,KAAK,CAAC,GAAG,KAAK,oBAAoB,UAAU,gBAAgB,MAAM,CAAC,MAAM,EAAE,CAAC,CAAC;IACxF,CAAC;IACD,OAAO;QACN;YACC,IAAI,EAAE,MAAM;YACZ,OAAO,EAAE;gBACR,EAAE,IAAI,EAAE,YAAY,EAAE,IAAI,EAAE,KAAK,EAAE;gBACnC,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;oBACzB,IAAI,EAAE,aAAa;oBACnB,SAAS,EAAE,QAAQ,KAAK,CAAC,QAAQ,WAAW,KAAK,CAAC,IAAI,EAAE;iBACxD,CAAC,CAAC;aACH;SACD;KACD,CAAC;AACH,CAAC;AAED,SAAS,mBAAmB,CAAC,KAAc,EAAE,EAAU;IACtD,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,GAAG,KAAK,uCAAuC,EAAE,EAAE,CAAC,CAAC;IAChG,OAAO,MAAM,CAAC,WAAW,CACxB,KAAK,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE;QACnB,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,OAAO,KAAK,CAAC,KAAK,KAAK,QAAQ,EAAE,CAAC;YACzD,MAAM,IAAI,KAAK,CAAC,GAAG,KAAK,uCAAuC,EAAE,EAAE,CAAC,CAAC;QACtE,CAAC;QACD,OAAO,CAAC,KAAK,CAAC,KAAK,EAAE,cAAc,CAAC,KAAK,EAAE,KAAK,CAAC,WAAW,EAAE,mBAAmB,EAAE,IAAI,KAAK,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC;IACxG,CAAC,CAAC,CACF,CAAC;AACH,CAAC;AAED,SAAS,WAAW,CAAC,EAAU,EAAE,QAA4B,EAAE,MAA+B;IAC7F,IAAI,MAAM,CAAC,IAAI,KAAK,SAAS;QAAE,MAAM,IAAI,KAAK,CAAC,GAAG,KAAK,sBAAsB,EAAE,EAAE,CAAC,CAAC;IACnF,IAAI,QAAQ,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;QAChC,IAAI,MAAM,CAAC,IAAI,KAAK,QAAQ,IAAI,OAAO,MAAM,CAAC,MAAM,KAAK,QAAQ,EAAE,CAAC;YACnE,MAAM,IAAI,KAAK,CAAC,GAAG,KAAK,uCAAuC,EAAE,EAAE,CAAC,CAAC;QACtE,CAAC;QACD,OAAO;YACN,IAAI,EAAE,QAAQ;YACd,MAAM,EAAE,MAAM,CAAC,MAAM;YACrB,aAAa,EAAE,mBAAmB,CAAC,MAAM,CAAC,aAAa,EAAE,EAAE,CAAC;YAC5D,UAAU,EAAE,cAAc,CAAC,KAAK,EAAE,MAAM,CAAC,UAAU,EAAE,kBAAkB,EAAE,EAAE,CAAC;SAC5E,CAAC;IACH,CAAC;IACD,IAAI,QAAQ,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;QAC/B,IAAI,MAAM,CAAC,IAAI,KAAK,OAAO;YAAE,MAAM,IAAI,KAAK,CAAC,GAAG,KAAK,sCAAsC,EAAE,EAAE,CAAC,CAAC;QACjG,OAAO;YACN,IAAI,EAAE,OAAO;YACb,KAAK,EAAE,cAAc,CAAC,KAAK,EAAE,MAAM,CAAC,KAAK,EAAE,aAAa,EAAE,EAAE,CAAC;YAC7D,UAAU,EAAE,cAAc,CAAC,KAAK,EAAE,MAAM,CAAC,UAAU,EAAE,kBAAkB,EAAE,EAAE,CAAC;SAC5E,CAAC;IACH,CAAC;IACD,IAAI,MAAM,CAAC,IAAI,KAAK,WAAW;QAAE,MAAM,IAAI,KAAK,CAAC,GAAG,KAAK,0CAA0C,EAAE,EAAE,CAAC,CAAC;IACzG,OAAO,EAAE,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,cAAc,CAAC,KAAK,EAAE,MAAM,CAAC,WAAW,EAAE,mBAAmB,EAAE,EAAE,CAAC,EAAE,CAAC;AAC1G,CAAC;AAED,SAAS,YAAY,CAAC,KAAc,EAAE,OAA0B;IAC/D,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,GAAG,KAAK,kCAAkC,CAAC,CAAC;IACvF,MAAM,MAAM,GAAG,IAAI,GAAG,EAAmC,CAAC;IAC1D,KAAK,MAAM,MAAM,IAAI,KAAK,EAAE,CAAC;QAC5B,IAAI,QAAQ,CAAC,MAAM,CAAC,IAAI,OAAO,MAAM,CAAC,IAAI,KAAK,QAAQ;YAAE,MAAM,CAAC,GAAG,CAAC,MAAM,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;IAC1F,CAAC;IACD,OAAO,MAAM,CAAC,WAAW,CACxB,MAAM,CAAC,OAAO,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,QAAQ,CAAC,EAAE,EAAE;QACxD,MAAM,MAAM,GAAG,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;QAC9B,IAAI,CAAC,MAAM;YAAE,MAAM,IAAI,KAAK,CAAC,GAAG,KAAK,iCAAiC,EAAE,EAAE,CAAC,CAAC;QAC5E,OAAO,CAAC,EAAE,EAAE,WAAW,CAAC,EAAE,EAAE,QAAQ,EAAE,MAAM,CAAC,CAAC,CAAC;IAChD,CAAC,CAAC,CACF,CAAC;AACH,CAAC;AAED;;;;GAIG;AACH,MAAM,iBAAiB,GAAG,CAAC,GAAG,CAAC,CAAC;AAEhC,SAAS,YAAY,CAAC,KAAc;IACnC,IAAK,KAAsC,CAAC,MAAM,KAAK,GAAG,EAAE,CAAC;QAC5D,OAAO,GAAG,KAAK,oIAAoI,CAAC;IACrJ,CAAC;IACD,OAAO,mBAAmB,CAAC,sBAAsB,CAAC,KAAK,CAAC,EAAE,GAAG,KAAK,QAAQ,CAAC,CAAC;AAC7E,CAAC;AAED,qDAAqD;AACrD,MAAM,CAAC,MAAM,QAAQ,GAA0C,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,EAAE;IAChG,MAAM,MAAM,GAAqB;QAChC,GAAG,EAAE,KAAK,CAAC,GAAG;QACd,QAAQ,EAAE,KAAK,CAAC,QAAQ;QACxB,KAAK,EAAE,KAAK,CAAC,EAAE;QACf,OAAO,EAAE,EAAE;QACX,UAAU,EAAE,MAAM;QAClB,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;KACrB,CAAC;IAEF,IAAI,CAAC;QACJ,IAAI,KAAK,CAAC,GAAG,KAAK,kBAAkB;YAAE,MAAM,IAAI,KAAK,CAAC,+BAA+B,KAAK,CAAC,GAAG,EAAE,CAAC,CAAC;QAClG,MAAM,IAAI,GAAG,MAAM,qBAAqB,CACvC,KAAK,EACL,IAAI,GAAG,CAAC,WAAW,EAAE,GAAG,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC,GAAG,CAAC,EAC9D,KAAK,EACL;YACC,KAAK,EAAE,KAAK,CAAC,EAAE;YACf,KAAK,EAAE,SAAS,CAAC,OAAO,CAAC;YACzB,SAAS,EAAE,MAAM,CAAC,OAAO,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,QAAQ,CAAC,EAAE,EAAE,CAAC,YAAY,CAAC,EAAE,EAAE,QAAQ,CAAC,CAAC;SAChG,EACD,OAAO,EACP,iBAAiB,CACjB,CAAC;QACF,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,GAAG,KAAK,kCAAkC,CAAC,CAAC;QACjF,4FAA4F;QAC5F,MAAM,KAAK,GAAG,oBAAoB,CAAC,IAAI,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;QACtD,IAAI,KAAK;YAAE,MAAM,CAAC,KAAK,GAAG,KAAK,CAAC;QAChC,MAAM,CAAC,OAAO,GAAG,YAAY,CAAC,IAAI,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;QACrD,OAAO,MAAM,CAAC;IACf,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,MAAM,CAAC,OAAO,GAAG,EAAE,CAAC;QACpB,MAAM,CAAC,UAAU,GAAG,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,OAAO,CAAC;QACnE,MAAM,CAAC,YAAY,GAAG,YAAY,CAAC,KAAK,CAAC,CAAC;QAC1C,OAAO,MAAM,CAAC;IACf,CAAC;AACF,CAAC,CAAC","sourcesContent":["import type {\n\tClassifierAnswer,\n\tClassifierContext,\n\tClassifierFunction,\n\tClassifierOptions,\n\tClassifierQuestion,\n\tClassifierResult,\n} from \"../types.ts\";\nimport { formatProviderError, normalizeProviderError } from \"../utils/error-body.ts\";\nimport {\n\ttype ClassifierHttpError,\n\tisRecord,\n\tparseClassifierUsage,\n\tpostClassifierRequest,\n\trequiredNumber,\n} from \"./classifier-shared.ts\";\n\n/**\n * OpenAI's Decisions API: `POST /v1/decisions` with `{ model, input, questions }`.\n * https://developers.openai.com/api/docs/guides/decisions\n *\n * The state is sent as JSON text. With images, the input becomes one user message with the state\n * as `input_text` followed by `input_image` data URLs. Questions map to Decisions types:\n * `choice` to `choice`, `score` to `score`, and `bool` to `predicate`. Predicates have no\n * criteria field, so the meanings of true and false are appended to the instructions.\n *\n * Only OpenAI API keys work: Sign in with ChatGPT tokens are rejected on this route.\n */\n\nconst LABEL = \"OpenAI Decisions\";\n\n/** The endpoint accepts at most this many image parts per request. */\nconst MAX_IMAGES = 128;\n\ninterface DecisionQuestion {\n\ttype: \"predicate\" | \"choice\" | \"score\";\n\tname: string;\n\tinstructions: string;\n\tchoices?: Array<{ value: string; description?: string }>;\n\tlevels?: Array<{ label: string }>;\n}\n\nfunction predicateInstructions(question: Extract<ClassifierQuestion, { type: \"bool\" }>): string {\n\tconst meanings = [\n\t\tquestion.criteria.true ? `True means: ${question.criteria.true}` : \"\",\n\t\tquestion.criteria.false ? `False means: ${question.criteria.false}` : \"\",\n\t].filter(Boolean);\n\treturn meanings.length > 0 ? `${question.instructions}\\n\\n${meanings.join(\"\\n\")}` : question.instructions;\n}\n\nfunction wireQuestion(name: string, question: ClassifierQuestion): DecisionQuestion {\n\tif (question.type === \"choice\") {\n\t\treturn {\n\t\t\ttype: \"choice\",\n\t\t\tname,\n\t\t\tinstructions: question.instructions,\n\t\t\tchoices: Object.entries(question.criteria).map(([value, description]) =>\n\t\t\t\tdescription ? { value, description } : { value },\n\t\t\t),\n\t\t};\n\t}\n\tif (question.type === \"score\") {\n\t\treturn {\n\t\t\ttype: \"score\",\n\t\t\tname,\n\t\t\tinstructions: question.instructions,\n\t\t\tlevels: question.criteria.map((label) => ({ label })),\n\t\t};\n\t}\n\treturn { type: \"predicate\", name, instructions: predicateInstructions(question) };\n}\n\nfunction wireInput(context: ClassifierContext): unknown {\n\tconst state = JSON.stringify(context.state);\n\tconst images = context.images ?? [];\n\tif (images.length === 0) return state;\n\tif (images.length > MAX_IMAGES) {\n\t\tthrow new Error(`${LABEL} accepts at most ${MAX_IMAGES} images, got ${images.length}`);\n\t}\n\treturn [\n\t\t{\n\t\t\trole: \"user\",\n\t\t\tcontent: [\n\t\t\t\t{ type: \"input_text\", text: state },\n\t\t\t\t...images.map((image) => ({\n\t\t\t\t\ttype: \"input_image\",\n\t\t\t\t\timage_url: `data:${image.mimeType};base64,${image.data}`,\n\t\t\t\t})),\n\t\t\t],\n\t\t},\n\t];\n}\n\nfunction choiceProbabilities(value: unknown, id: string): Record<string, number> {\n\tif (!Array.isArray(value)) throw new Error(`${LABEL} returned invalid probabilities for ${id}`);\n\treturn Object.fromEntries(\n\t\tvalue.map((entry) => {\n\t\t\tif (!isRecord(entry) || typeof entry.value !== \"string\") {\n\t\t\t\tthrow new Error(`${LABEL} returned invalid probabilities for ${id}`);\n\t\t\t}\n\t\t\treturn [entry.value, requiredNumber(LABEL, entry.probability, `probability for ${id}.${entry.value}`)];\n\t\t}),\n\t);\n}\n\nfunction parseAnswer(id: string, question: ClassifierQuestion, answer: Record<string, unknown>): ClassifierAnswer {\n\tif (answer.type === \"refusal\") throw new Error(`${LABEL} refused to answer ${id}`);\n\tif (question.type === \"choice\") {\n\t\tif (answer.type !== \"choice\" || typeof answer.choice !== \"string\") {\n\t\t\tthrow new Error(`${LABEL} did not return a choice answer for ${id}`);\n\t\t}\n\t\treturn {\n\t\t\ttype: \"choice\",\n\t\t\tchoice: answer.choice,\n\t\t\tprobabilities: choiceProbabilities(answer.probabilities, id),\n\t\t\tconfidence: requiredNumber(LABEL, answer.confidence, `confidence for ${id}`),\n\t\t};\n\t}\n\tif (question.type === \"score\") {\n\t\tif (answer.type !== \"score\") throw new Error(`${LABEL} did not return a score answer for ${id}`);\n\t\treturn {\n\t\t\ttype: \"score\",\n\t\t\tscore: requiredNumber(LABEL, answer.score, `score for ${id}`),\n\t\t\tconfidence: requiredNumber(LABEL, answer.confidence, `confidence for ${id}`),\n\t\t};\n\t}\n\tif (answer.type !== \"predicate\") throw new Error(`${LABEL} did not return a predicate answer for ${id}`);\n\treturn { type: \"bool\", probability: requiredNumber(LABEL, answer.probability, `probability for ${id}`) };\n}\n\nfunction parseAnswers(value: unknown, context: ClassifierContext): Record<string, ClassifierAnswer> {\n\tif (!Array.isArray(value)) throw new Error(`${LABEL} returned an unexpected response`);\n\tconst byName = new Map<string, Record<string, unknown>>();\n\tfor (const answer of value) {\n\t\tif (isRecord(answer) && typeof answer.name === \"string\") byName.set(answer.name, answer);\n\t}\n\treturn Object.fromEntries(\n\t\tObject.entries(context.questions).map(([id, question]) => {\n\t\t\tconst answer = byName.get(id);\n\t\t\tif (!answer) throw new Error(`${LABEL} did not return an answer for ${id}`);\n\t\t\treturn [id, parseAnswer(id, question, answer)];\n\t\t}),\n\t);\n}\n\n/**\n * Cloudflare in front of api.openai.com answers 504 with an HTML page when a request runs longer\n * than about five seconds. Large inputs, currently above roughly 600K tokens, hit this limit, and\n * retrying the same input runs into it again, so 504 is not retried.\n */\nconst NO_RETRY_STATUSES = [504];\n\nfunction errorMessage(error: unknown): string {\n\tif ((error as Partial<ClassifierHttpError>).status === 504) {\n\t\treturn `${LABEL} error (504): the request timed out at the gateway. Very large inputs (above roughly 600K tokens) currently exceed its time limit.`;\n\t}\n\treturn formatProviderError(normalizeProviderError(error), `${LABEL} error`);\n}\n\n/** Classification through OpenAI's Decisions API. */\nexport const classify: ClassifierFunction<ClassifierOptions> = async (model, context, options) => {\n\tconst output: ClassifierResult = {\n\t\tapi: model.api,\n\t\tprovider: model.provider,\n\t\tmodel: model.id,\n\t\tanswers: {},\n\t\tstopReason: \"stop\",\n\t\ttimestamp: Date.now(),\n\t};\n\n\ttry {\n\t\tif (model.api !== \"openai-decisions\") throw new Error(`Unsupported classifier API: ${model.api}`);\n\t\tconst body = await postClassifierRequest(\n\t\t\tLABEL,\n\t\t\tnew URL(\"decisions\", `${model.baseUrl.replace(/\\/+$/u, \"\")}/`),\n\t\t\tmodel,\n\t\t\t{\n\t\t\t\tmodel: model.id,\n\t\t\t\tinput: wireInput(context),\n\t\t\t\tquestions: Object.entries(context.questions).map(([id, question]) => wireQuestion(id, question)),\n\t\t\t},\n\t\t\toptions,\n\t\t\tNO_RETRY_STATUSES,\n\t\t);\n\t\tif (!isRecord(body)) throw new Error(`${LABEL} returned an unexpected response`);\n\t\t// Set before parsing answers: a request with malformed or refused answers was still billed.\n\t\tconst usage = parseClassifierUsage(body.usage, model);\n\t\tif (usage) output.usage = usage;\n\t\toutput.answers = parseAnswers(body.answers, context);\n\t\treturn output;\n\t} catch (error) {\n\t\toutput.answers = {};\n\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\toutput.errorMessage = errorMessage(error);\n\t\treturn output;\n\t}\n};\n"]}

package/dist/api/openai-decisions.lazy.d.ts

@@ -0,0 +1,3 @@
+import type { ProviderClassifier } from "../types.ts";
+export declare const openAIDecisionsApi: () => ProviderClassifier;
+//# sourceMappingURL=openai-decisions.lazy.d.ts.map

package/dist/api/openai-decisions.lazy.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"openai-decisions.lazy.d.ts","sourceRoot":"","sources":["../../src/api/openai-decisions.lazy.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,aAAa,CAAC;AAEtD,eAAO,MAAM,kBAAkB,QAAO,kBAGpC,CAAC"}

package/dist/api/openai-decisions.lazy.js

@@ -0,0 +1,4 @@
+export const openAIDecisionsApi = () => ({
+    classify: async (model, context, options) => (await import("./openai-decisions.js")).classify(model, context, options),
+});
+//# sourceMappingURL=openai-decisions.lazy.js.map

package/dist/api/openai-decisions.lazy.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"openai-decisions.lazy.js","sourceRoot":"","sources":["../../src/api/openai-decisions.lazy.ts"],"names":[],"mappings":"AAEA,MAAM,CAAC,MAAM,kBAAkB,GAAG,GAAuB,EAAE,CAAC,CAAC;IAC5D,QAAQ,EAAE,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,EAAE,CAC3C,CAAC,MAAM,MAAM,CAAC,uBAAuB,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,CAAC;CAC1E,CAAC,CAAC","sourcesContent":["import type { ProviderClassifier } from \"../types.ts\";\n\nexport const openAIDecisionsApi = (): ProviderClassifier => ({\n\tclassify: async (model, context, options) =>\n\t\t(await import(\"./openai-decisions.ts\")).classify(model, context, options),\n});\n"]}

package/dist/api/openai-responses-shared.d.ts

@@ -1,6 +1,7 @@
 import type { Tool as OpenAITool, ResponseCreateParamsStreaming, ResponseInput, ResponseStreamEvent } from "openai/resources/responses/responses.js";
 import type { Api, AssistantMessage, Model, StreamOptions, Tool, TranscriptContext, Usage } from "../types.ts";
 import type { AssistantMessageEventStream } from "../utils/event-stream.ts";
+import { type TranscriptToolSupport } from "../utils/transcript.ts";
 export interface OpenAIResponsesStreamOptions {
     onProviderStreamEvent?: StreamOptions["onProviderStreamEvent"];
     serviceTier?: ResponseCreateParamsStreaming["service_tier"];
@@ -16,6 +17,16 @@ export interface ConvertResponsesMessagesOptions {
     supportsAdditionalTools?: boolean;
     supportsToolSearch?: boolean;
     toolOptions?: ConvertResponsesToolsOptions;
+    /**
+     * Effort for the reply being generated. When set, the thinking level is sent as
+     * configuration_update items instead of a changing reasoning.effort, which would invalidate the
+     * prompt cache. Each item sets the effort from that point on: one goes before every earlier
+     * reply with the level recorded on it (`providerThinkingLevel`), and one with this effort goes
+     * at the end. Earlier input therefore never changes. OpenAI documents the item before user
+     * messages only. After tool results it was verified on gpt-6-luna (October 2026) to be accepted
+     * and to apply to the next reply.
+     */
+    configurationUpdateEffort?: string;
 }
 export interface ConvertResponsesToolsOptions {
     strict?: boolean | null;
@@ -23,6 +34,21 @@ export interface ConvertResponsesToolsOptions {
     supportsOpenAIGrammarTools?: boolean;
     toolSearchResult?: boolean;
 }
+/**
+ * How a Responses transport expresses tool changes.
+ *
+ * `additional_tools` items accumulate: a later item adds tools and replaces same-name
+ * definitions from earlier items, but it cannot replace a top-level tool and cannot remove any
+ * tool (verified with strict schemas on the OpenAI, Codex, and Copilot backends, October 2026).
+ * Like Codex, every tool is therefore declared in `additional_tools` items, starting with one
+ * at the head of the input, and removals are announced as text. Synthetic tool search can
+ * only add tools. Without mid-conversation system messages there is nothing to anchor.
+ */
+export declare function getResponsesToolSupport(options: {
+    supportsMidConvoSystemMessages?: boolean;
+    supportsAdditionalTools?: boolean;
+    supportsToolSearch?: boolean;
+}): TranscriptToolSupport;
 export declare function convertResponsesMessages<TApi extends Api>(model: Model<TApi>, context: TranscriptContext, allowedToolCallProviders: ReadonlySet<string>, options?: ConvertResponsesMessagesOptions): ResponseInput;
 export declare function convertResponsesTools(tools: readonly Tool[], options?: ConvertResponsesToolsOptions): OpenAITool[];
 export declare function processResponsesStream<TApi extends Api>(openaiStream: AsyncIterable<ResponseStreamEvent>, output: AssistantMessage, stream: AssistantMessageEventStream, model: Model<TApi>, options?: OpenAIResponsesStreamOptions): Promise<void>;

package/dist/api/openai-responses-shared.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai-responses-shared.d.ts","sourceRoot":"","sources":["../../src/api/openai-responses-shared.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EACX,IAAI,IAAI,UAAU,EAClB,6BAA6B,EAC7B,aAAa,EAQb,mBAAmB,EAEnB,MAAM,yCAAyC,CAAC;AAEjD,OAAO,KAAK,EACX,GAAG,EACH,gBAAgB,EAEhB,KAAK,EAEL,aAAa,EAKb,IAAI,EAEJ,iBAAiB,EACjB,KAAK,EACL,MAAM,aAAa,CAAC;AACrB,OAAO,KAAK,EAAE,2BAA2B,EAAE,MAAM,0BAA0B,CAAC;AAoF5E,MAAM,WAAW,4BAA4B;IAC5C,qBAAqB,CAAC,EAAE,aAAa,CAAC,uBAAuB,CAAC,CAAC;IAC/D,WAAW,CAAC,EAAE,6BAA6B,CAAC,cAAc,CAAC,CAAC;IAC5D,0BAA0B,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzD,kBAAkB,CAAC,EAAE,CACpB,mBAAmB,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,EAC9E,kBAAkB,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,KACzE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,CAAC;IAC/D,uBAAuB,CAAC,EAAE,CACzB,KAAK,EAAE,KAAK,EACZ,WAAW,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,KAClE,IAAI,CAAC;CACV;AAED,MAAM,WAAW,+BAA+B;IAC/C,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,0BAA0B,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzD,0GAA0G;IAC1G,8BAA8B,CAAC,EAAE,OAAO,CAAC;IACzC,uBAAuB,CAAC,EAAE,OAAO,CAAC;IAClC,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,WAAW,CAAC,EAAE,4BAA4B,CAAC;CAC3C;AAED,MAAM,WAAW,4BAA4B;IAC5C,MAAM,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC;IACxB,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,0BAA0B,CAAC,EAAE,OAAO,CAAC;IACrC,gBAAgB,CAAC,EAAE,OAAO,CAAC;CAC3B;AAMD,wBAAgB,wBAAwB,CAAC,IAAI,SAAS,GAAG,EACxD,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,iBAAiB,EAC1B,wBAAwB,EAAE,WAAW,CAAC,MAAM,CAAC,EAC7C,OAAO,CAAC,EAAE,+BAA+B,GACvC,aAAa,CA4Mf;AAMD,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,SAAS,IAAI,EAAE,EAAE,OAAO,CAAC,EAAE,4BAA4B,GAAG,UAAU,EAAE,CAqClH;AAoCD,wBAAsB,sBAAsB,CAAC,IAAI,SAAS,GAAG,EAC5D,YAAY,EAAE,aAAa,CAAC,mBAAmB,CAAC,EAChD,MAAM,EAAE,gBAAgB,EACxB,MAAM,EAAE,2BAA2B,EACnC,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,CAAC,EAAE,4BAA4B,GACpC,OAAO,CAAC,IAAI,CAAC,CAmWf"}
+{"version":3,"file":"openai-responses-shared.d.ts","sourceRoot":"","sources":["../../src/api/openai-responses-shared.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EACX,IAAI,IAAI,UAAU,EAElB,6BAA6B,EAC7B,aAAa,EAQb,mBAAmB,EAEnB,MAAM,yCAAyC,CAAC;AAEjD,OAAO,KAAK,EACX,GAAG,EACH,gBAAgB,EAEhB,KAAK,EAEL,aAAa,EAKb,IAAI,EAEJ,iBAAiB,EACjB,KAAK,EACL,MAAM,aAAa,CAAC;AACrB,OAAO,KAAK,EAAE,2BAA2B,EAAE,MAAM,0BAA0B,CAAC;AAK5E,OAAO,EAA6C,KAAK,qBAAqB,EAAE,MAAM,wBAAwB,CAAC;AAgF/G,MAAM,WAAW,4BAA4B;IAC5C,qBAAqB,CAAC,EAAE,aAAa,CAAC,uBAAuB,CAAC,CAAC;IAC/D,WAAW,CAAC,EAAE,6BAA6B,CAAC,cAAc,CAAC,CAAC;IAC5D,0BAA0B,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzD,kBAAkB,CAAC,EAAE,CACpB,mBAAmB,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,EAC9E,kBAAkB,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,KACzE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,CAAC;IAC/D,uBAAuB,CAAC,EAAE,CACzB,KAAK,EAAE,KAAK,EACZ,WAAW,EAAE,6BAA6B,CAAC,cAAc,CAAC,GAAG,SAAS,KAClE,IAAI,CAAC;CACV;AAED,MAAM,WAAW,+BAA+B;IAC/C,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,0BAA0B,CAAC,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACzD,0GAA0G;IAC1G,8BAA8B,CAAC,EAAE,OAAO,CAAC;IACzC,uBAAuB,CAAC,EAAE,OAAO,CAAC;IAClC,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,WAAW,CAAC,EAAE,4BAA4B,CAAC;IAC3C;;;;;;;;OAQG;IACH,yBAAyB,CAAC,EAAE,MAAM,CAAC;CACnC;AAED,MAAM,WAAW,4BAA4B;IAC5C,MAAM,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC;IACxB,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,0BAA0B,CAAC,EAAE,OAAO,CAAC;IACrC,gBAAgB,CAAC,EAAE,OAAO,CAAC;CAC3B;AAMD;;;;;;;;;GASG;AACH,wBAAgB,uBAAuB,CAAC,OAAO,EAAE;IAChD,8BAA8B,CAAC,EAAE,OAAO,CAAC;IACzC,uBAAuB,CAAC,EAAE,OAAO,CAAC;IAClC,kBAAkB,CAAC,EAAE,OAAO,CAAC;CAC7B,GAAG,qBAAqB,CAGxB;AAkBD,wBAAgB,wBAAwB,CAAC,IAAI,SAAS,GAAG,EACxD,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,iBAAiB,EAC1B,wBAAwB,EAAE,WAAW,CAAC,MAAM,CAAC,EAC7C,OAAO,CAAC,EAAE,+BAA+B,GACvC,aAAa,CA4Of;AAMD,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,SAAS,IAAI,EAAE,EAAE,OAAO,CAAC,EAAE,4BAA4B,GAAG,UAAU,EAAE,CAqClH;AA2BD,wBAAsB,sBAAsB,CAAC,IAAI,SAAS,GAAG,EAC5D,YAAY,EAAE,aAAa,CAAC,mBAAmB,CAAC,EAChD,MAAM,EAAE,gBAAgB,EACxB,MAAM,EAAE,2BAA2B,EACnC,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,CAAC,EAAE,4BAA4B,GACpC,OAAO,CAAC,IAAI,CAAC,CAgVf"}

package/dist/api/openai-responses-shared.js

@@ -1,10 +1,10 @@
 import { calculateCost } from "../models.js";
 import { shortHash } from "../utils/hash.js";
-import { parseStreamingJson } from "../utils/json-parse.js";
 import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
+import { ToolCallArgumentsStream } from "../utils/streaming-json.js";
 import { getSystemMessageText, renderSystemMessageUpdate } from "../utils/text.js";
 import { resolveTranscript, resolveTranscriptTools } from "../utils/transcript.js";
-import { appendGrammarToolInputJsonDelta, getGrammarToolInput, getJsonSchemaToolParameters, resolveGrammarConstrainedSampling, resolveJsonSchemaStrictSampling, } from "./constrained-sampling.js";
+import { appendGrammarToolInputJsonDelta, closeGrammarToolInputJson, getGrammarToolInput, getJsonSchemaToolParameters, resolveGrammarConstrainedSampling, resolveJsonSchemaStrictSampling, } from "./constrained-sampling.js";
 import { transformMessages } from "./transform-messages.js";
 // =============================================================================
 // Utilities
@@ -46,7 +46,7 @@ function convertToolResultOutput(model, content) {
         .map((c) => c.text)
         .join("\n");
     const images = content.filter((c) => c.type === "image");
-    const hasText = textResult.length > 0;
+    const hasText = textResult.trim().length > 0;
     if (images.length === 0 || !model.input.includes("image")) {
         return sanitizeSurrogates(hasText ? textResult : images.length > 0 ? "(see attached image)" : "(no tool output)");
     }
@@ -66,6 +66,34 @@ function convertToolResultOutput(model, content) {
 // =============================================================================
 // Message conversion
 // =============================================================================
+/**
+ * How a Responses transport expresses tool changes.
+ *
+ * `additional_tools` items accumulate: a later item adds tools and replaces same-name
+ * definitions from earlier items, but it cannot replace a top-level tool and cannot remove any
+ * tool (verified with strict schemas on the OpenAI, Codex, and Copilot backends, October 2026).
+ * Like Codex, every tool is therefore declared in `additional_tools` items, starting with one
+ * at the head of the input, and removals are announced as text. Synthetic tool search can
+ * only add tools. Without mid-conversation system messages there is nothing to anchor.
+ */
+export function getResponsesToolSupport(options) {
+    if (options.supportsMidConvoSystemMessages && options.supportsAdditionalTools)
+        return "changes";
+    return options.supportsToolSearch ? "additions" : "none";
+}
+/**
+ * Precedes every additional_tools item after the first. Without it, models read the latest item as
+ * the complete tool set once a removal was announced; Codex precedes its incremental tool batches
+ * with the same guidance. Restating the tools that are still unavailable keeps smaller models
+ * (gpt-5.6-luna, gpt-6-luna) from calling them again after later loads (verified October 2026).
+ */
+function renderIncrementalToolsNotice(unavailable) {
+    const notice = "This is an incremental tools update. Previously declared tools remain available unless explicitly marked unavailable. A tool declared again here is available again, and its latest definition replaces the earlier one.";
+    return unavailable.size > 0 ? `${notice}\nStill unavailable: ${[...unavailable].join(", ")}.` : notice;
+}
+function renderToolRemovalNotice(names) {
+    return `The following tools are no longer available. Do not call them:\n${names.map((name) => `- ${name}`).join("\n")}`;
+}
 export function convertResponsesMessages(model, context, allowedToolCallProviders, options) {
     const normalizedContext = resolveTranscript(context, options?.supportsMidConvoSystemMessages);
     const messages = [];
@@ -94,20 +122,30 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
         return `${normalizedCallId}|${normalizedItemId}`;
     };
     const transformedMessages = transformMessages(normalizedContext.messages, model, normalizeToolCallId);
-    const transcriptTools = resolveTranscriptTools(normalizedContext.messages, (options?.supportsAdditionalTools ?? false) || (options?.supportsToolSearch ?? false));
+    const toolSupport = getResponsesToolSupport(options ?? {});
+    const transcriptTools = resolveTranscriptTools(normalizedContext.messages, toolSupport);
+    let declaredAdditionalTools = false;
+    /** Tools announced as removed and not declared again since. */
+    const unavailableTools = new Set();
     const appendSystemToolAdditions = (message, seed) => {
-        const tools = transcriptTools.anchorsAdditions ? (message.toolsAdded ?? []) : [];
+        const tools = message.toolsAdded ?? [];
+        for (const tool of tools)
+            unavailableTools.delete(tool.name);
         if (tools.length === 0)
             return;
-        if (options?.supportsAdditionalTools) {
+        if (toolSupport === "changes") {
+            if (declaredAdditionalTools) {
+                messages.push({ role: instructionRole, content: renderIncrementalToolsNotice(unavailableTools) });
+            }
+            declaredAdditionalTools = true;
             messages.push({
                 type: "additional_tools",
                 role: "developer",
-                tools: convertResponsesTools(tools, options.toolOptions),
+                tools: convertResponsesTools(tools, options?.toolOptions),
             });
             return;
         }
-        if (!options?.supportsToolSearch)
+        if (toolSupport !== "additions")
             return;
         const names = tools.map((tool) => tool.name);
         const callId = `pi_tool_load_${shortHash(`${seed}:${names.join(",")}`)}`;
@@ -123,21 +161,41 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
             call_id: callId,
             execution: "client",
             status: "completed",
-            tools: convertResponsesTools(tools, { ...options.toolOptions, toolSearchResult: true }),
+            tools: convertResponsesTools(tools, { ...options?.toolOptions, toolSearchResult: true }),
         });
     };
     const includeInitialSystemMessage = options?.includeSystemPrompt ?? true;
+    const pushConfigurationUpdate = (effort) => {
+        messages.push({
+            type: "configuration_update",
+            reasoning: { effort: effort },
+        });
+    };
     const compat = model.compat;
     const instructionRole = model.reasoning && compat?.supportsDeveloperRole !== false ? "developer" : "system";
     let msgIndex = 0;
     let sourceIndex = 0;
+    let systemIndex = 0;
     for (const msg of transformedMessages) {
         const isLeadingSystemMessage = sourceIndex++ === 0 && msg.role === "system";
         if (msg.role === "system") {
-            if (!isLeadingSystemMessage)
+            // transformMessages keeps system messages in order, so their position matches the transcript.
+            const anchorsToolChanges = systemIndex++ >= transcriptTools.foldedSystemMessages;
+            if (anchorsToolChanges)
                 appendSystemToolAdditions(msg, `system:${msgIndex}`);
             if (!isLeadingSystemMessage || includeInitialSystemMessage) {
-                const text = isLeadingSystemMessage ? getSystemMessageText(msg) : renderSystemMessageUpdate(msg);
+                const parts = [isLeadingSystemMessage ? getSystemMessageText(msg) : renderSystemMessageUpdate(msg)];
+                if (anchorsToolChanges) {
+                    // A redefinition is a removal plus an addition of the same name; only true removals are announced.
+                    const added = new Set((msg.toolsAdded ?? []).map((tool) => tool.name));
+                    const removed = (msg.toolsRemoved ?? []).map((tool) => tool.name).filter((name) => !added.has(name));
+                    if (removed.length > 0)
+                        parts.push(renderToolRemovalNotice(removed));
+                    // Recorded after this message's additions, whose notice precedes this removal notice.
+                    for (const name of removed)
+                        unavailableTools.add(name);
+                }
+                const text = parts.filter((part) => part.length > 0).join("\n\n");
                 if (text.length > 0) {
                     messages.push({ role: instructionRole, content: sanitizeSurrogates(text) });
                 }
@@ -247,6 +305,11 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
             }
             if (output.length === 0)
                 continue;
+            if (options?.configurationUpdateEffort !== undefined &&
+                isSameModel &&
+                msg.providerThinkingLevel !== undefined) {
+                pushConfigurationUpdate(msg.providerThinkingLevel);
+            }
             messages.push(...output);
         }
         else if (msg.role === "toolResult") {
@@ -270,6 +333,8 @@ export function convertResponsesMessages(model, context, allowedToolCallProvider
         if (!isLeadingSystemMessage)
             msgIndex++;
     }
+    if (options?.configurationUpdateEffort !== undefined)
+        pushConfigurationUpdate(options.configurationUpdateEffort);
     return messages;
 }
 // =============================================================================
@@ -309,24 +374,11 @@ export function convertResponsesTools(tools, options) {
         return functionTool;
     });
 }
-function getCustomToolCallInput(block) {
-    const property = block.customInput?.property;
-    if (property === undefined)
-        return "";
-    const value = block.arguments[property];
-    return typeof value === "string" ? value : "";
-}
-function appendCustomToolCallInput(block, nextInput, close) {
-    const customInput = block.customInput;
-    if (!customInput)
-        return undefined;
-    const delta = appendGrammarToolInputJsonDelta(customInput.jsonBuffer, customInput.property, nextInput, close);
-    block.arguments = { [customInput.property]: nextInput };
-    return delta;
-}
 export async function processResponsesStream(openaiStream, output, stream, model, options) {
     let sawTerminalResponseEvent = false;
     const outputSlots = new Map();
+    // Tool calls whose output_item.done has not arrived yet.
+    const unfinishedToolCalls = new Set();
     const reasoningBlocksById = new Map();
     const applyMessagePhaseStopReason = (item) => {
         if (item.type === "message" && item.phase === "final_answer") {
@@ -371,19 +423,21 @@ export async function processResponsesStream(openaiStream, output, stream, model
             return slot;
         }
         if (item.type === "function_call") {
+            const args = new ToolCallArgumentsStream();
             const block = {
                 type: "toolCall",
                 id: `${item.call_id}|${item.id}`,
                 name: item.name,
-                arguments: {},
+                arguments: item.arguments ? args.append(item.arguments) : {},
                 ...(item.namespace !== undefined ? { namespace: item.namespace } : {}),
-                partialJson: item.arguments || "",
             };
             output.content.push(block);
+            unfinishedToolCalls.add(block);
             const slot = {
                 type: "toolCall",
                 block,
                 contentIndex: output.content.length - 1,
+                input: { kind: "json", arguments: args },
             };
             outputSlots.set(outputIndex, slot);
             stream.push({ type: "toolcall_start", contentIndex: slot.contentIndex, partial: output });
@@ -398,16 +452,14 @@ export async function processResponsesStream(openaiStream, output, stream, model
                 name: item.name,
                 arguments: { [inputProperty]: input },
                 ...(item.namespace !== undefined ? { namespace: item.namespace } : {}),
-                customInput: {
-                    property: inputProperty,
-                    jsonBuffer: { input: "", started: false, closed: false },
-                },
             };
             output.content.push(block);
+            unfinishedToolCalls.add(block);
             const slot = {
                 type: "toolCall",
                 block,
                 contentIndex: output.content.length - 1,
+                input: customInput(inputProperty, input),
             };
             outputSlots.set(outputIndex, slot);
             stream.push({ type: "toolcall_start", contentIndex: slot.contentIndex, partial: output });
@@ -567,36 +619,31 @@ export async function processResponsesStream(openaiStream, output, stream, model
         }
         else if (event.type === "response.function_call_arguments.delta") {
             const slot = getSlot(event.output_index, "toolCall");
-            if (!slot || slot.block.partialJson === undefined)
+            if (slot?.input.kind !== "json")
                 continue;
-            slot.block.partialJson += event.delta;
-            slot.block.arguments = parseStreamingJson(slot.block.partialJson);
+            slot.block.arguments = slot.input.arguments.append(event.delta);
             pushToolCallDelta(slot, event.delta);
         }
         else if (event.type === "response.function_call_arguments.done") {
             const slot = getSlot(event.output_index, "toolCall");
-            if (!slot || slot.block.partialJson === undefined)
+            if (slot?.input.kind !== "json")
                 continue;
-            const previousPartialJson = slot.block.partialJson;
-            slot.block.partialJson = event.arguments;
-            slot.block.arguments = parseStreamingJson(slot.block.partialJson);
-            if (event.arguments.startsWith(previousPartialJson)) {
-                const delta = event.arguments.slice(previousPartialJson.length);
-                if (delta.length > 0)
-                    pushToolCallDelta(slot, delta);
-            }
+            const replaced = slot.input.arguments.replace(event.arguments);
+            slot.block.arguments = replaced.arguments;
+            if (replaced.delta)
+                pushToolCallDelta(slot, replaced.delta);
         }
         else if (event.type === "response.custom_tool_call_input.delta") {
             const slot = getSlot(event.output_index, "toolCall");
-            if (!slot || !slot.block.customInput)
+            if (slot?.input.kind !== "custom")
                 continue;
-            pushToolCallDelta(slot, appendCustomToolCallInput(slot.block, getCustomToolCallInput(slot.block) + event.delta, false));
+            pushToolCallDelta(slot, appendCustomInput(slot.block, slot.input, event.delta));
         }
         else if (event.type === "response.custom_tool_call_input.done") {
             const slot = getSlot(event.output_index, "toolCall");
-            if (!slot || !slot.block.customInput)
+            if (slot?.input.kind !== "custom")
                 continue;
-            pushToolCallDelta(slot, appendCustomToolCallInput(slot.block, event.input, true));
+            pushToolCallDelta(slot, closeCustomInput(slot.block, slot.input, event.input));
         }
         else if (event.type === "response.output_item.done") {
             const item = event.item;
@@ -628,15 +675,13 @@ export async function processResponsesStream(openaiStream, output, stream, model
                 });
                 outputSlots.delete(event.output_index);
             }
-            else if (item.type === "function_call" &&
-                slot?.type === "toolCall" &&
-                slot.block.partialJson !== undefined) {
-                slot.block.arguments = parseStreamingJson(item.arguments || slot.block.partialJson || "{}");
+            else if (item.type === "function_call" && slot?.type === "toolCall" && slot.input.kind === "json") {
+                if (item.arguments)
+                    slot.input.arguments.replace(item.arguments);
+                Object.assign(slot.block, slot.input.arguments.finish());
                 if (item.namespace !== undefined)
                     slot.block.namespace = item.namespace;
-                // Finalize in-place and strip the scratch buffer so replay only
-                // carries parsed arguments.
-                delete slot.block.partialJson;
+                unfinishedToolCalls.delete(slot.block);
                 stream.push({
                     type: "toolcall_end",
                     contentIndex: slot.contentIndex,
@@ -645,11 +690,11 @@ export async function processResponsesStream(openaiStream, output, stream, model
                 });
                 outputSlots.delete(event.output_index);
             }
-            else if (item.type === "custom_tool_call" && slot?.type === "toolCall" && slot.block.customInput) {
-                pushToolCallDelta(slot, appendCustomToolCallInput(slot.block, item.input ?? getCustomToolCallInput(slot.block), true));
+            else if (item.type === "custom_tool_call" && slot?.type === "toolCall" && slot.input.kind === "custom") {
+                pushToolCallDelta(slot, closeCustomInput(slot.block, slot.input, item.input));
                 if (item.namespace !== undefined)
                     slot.block.namespace = item.namespace;
-                delete slot.block.customInput;
+                unfinishedToolCalls.delete(slot.block);
                 stream.push({
                     type: "toolcall_end",
                     contentIndex: slot.contentIndex,
@@ -683,18 +728,39 @@ export async function processResponsesStream(openaiStream, output, stream, model
     }
     // The agent runs every tool call in the final message. Refuse to hand over calls whose
     // output_item.done never arrived: their arguments may be cut off or mixed up, e.g. when a
-    // non-compliant server omits output_index. Finished calls have their scratch buffers removed.
+    // non-compliant server omits output_index.
     if (output.stopReason === "toolUse") {
-        for (const block of output.content) {
-            if (block.type !== "toolCall")
-                continue;
-            const toolCall = block;
-            if (toolCall.partialJson !== undefined || toolCall.customInput !== undefined) {
-                throw new Error(`OpenAI Responses stream completed with an unfinished tool call: ${toolCall.name} (${toolCall.id})`);
-            }
+        for (const toolCall of unfinishedToolCalls) {
+            throw new Error(`OpenAI Responses stream completed with an unfinished tool call: ${toolCall.name} (${toolCall.id})`);
         }
     }
 }
+/** Stream state of a custom tool call that starts with `initialInput`. */
+function customInput(property, initialInput) {
+    const buffer = { input: "", started: false, closed: false };
+    return {
+        kind: "custom",
+        property,
+        buffer,
+        unsent: appendGrammarToolInputJsonDelta(buffer, property, initialInput) ?? "",
+    };
+}
+/** Append to a streaming custom tool input; returns the matching piece of its JSON arguments text. */
+function appendCustomInput(block, input, delta) {
+    return sendCustomInput(block, input, appendGrammarToolInputJsonDelta(input.buffer, input.property, delta));
+}
+/** Close a custom tool input with the provider's complete input, if it sent one. */
+function closeCustomInput(block, input, finalInput) {
+    return sendCustomInput(block, input, closeGrammarToolInputJson(input.buffer, input.property, finalInput));
+}
+function sendCustomInput(block, input, jsonDelta) {
+    block.arguments = { [input.property]: input.buffer.input };
+    if (input.unsent === "")
+        return jsonDelta;
+    const sent = input.unsent + (jsonDelta ?? "");
+    input.unsent = "";
+    return sent;
+}
 function mapStopReason(status, incompleteReason) {
     if (!status)
         return { stopReason: "stop" };