co-maintainer 0.5.1 → 0.5.2

package/dist/package.json

@@ -1,6 +1,6 @@
 {
     "name": "co-maintainer",
-    "version": "0.5.1",
+    "version": "0.5.2",
     "description": "An imitator of yours, A PR review tool like co-maintainer.",
     "keywords": [
         "ai",

package/dist/src/ai/anthropic.js

@@ -1,4 +1,5 @@
 import { getEnv } from "../util/runtime.js";
+import { outputLimitError } from "./provider.js";
 import { CliError, EXIT_RUNTIME, EXIT_USAGE, networkFailure, } from "../cli/error.js";
 /** Required on every Messages request and the only version the errors below
  * are written against. */
@@ -197,8 +198,12 @@ export function parseMessagesBody(json, model) {
         }
     }
     const usage = json.usage;
+    const text = texts.join("");
+    if (json.stop_reason === "max_tokens" && !text.trim() && !toolCalls.length) {
+        throw outputLimitError("anthropic", model);
+    }
     return {
-        text: texts.join(""),
+        text,
         ...(toolCalls.length ? { toolCalls } : {}),
         // Tokens read from or written to the prompt cache are input too, and with
         // the guides cached they are most of it.

package/dist/src/ai/openai.js

@@ -1,4 +1,5 @@
 import { getEnv } from "../util/runtime.js";
+import { outputLimitError, withoutRefusedPart } from "./provider.js";
 import { CliError, EXIT_RUNTIME, EXIT_USAGE, networkFailure, } from "../cli/error.js";
 /** `CM_OPENAI_URL` points at a local fake in tests. Unset means the real
  * endpoint, so production behavior is unchanged. */
@@ -145,8 +146,12 @@ function parseResponsesBody(json, model) {
         }
     }
     const usage = json.usage;
+    const text = texts.join("");
+    if (json.status === "incomplete" && !text.trim() && !toolCalls.length) {
+        throw outputLimitError("openai", model);
+    }
     return {
-        text: texts.join(""),
+        text,
         ...(toolCalls.length ? { toolCalls } : {}),
         tokensIn: Number(usage?.input_tokens ?? 0),
         tokensOut: Number(usage?.output_tokens ?? 0),
@@ -184,25 +189,19 @@ export class OpenAiProvider {
         }
     }
     async complete(request) {
-        let response = await this.post(responsesBody(this.model, request));
-        // A model may reject `text.format` next to tools. The schema is only a
-        // hint (the prompt also asks for a fenced JSON block), so one retry
-        // without it beats failing the review outright.
-        if (response.status === 400 && request.responseFormat) {
-            const detail = await response.text();
-            if (/text\.format|json_schema|structured/i.test(detail)) {
-                const { responseFormat: _dropped, ...withoutSchema } = request;
-                response = await this.post(responsesBody(this.model, withoutSchema));
-                if (response.ok) {
-                    return parseResponsesBody((await response.json()), this.model);
-                }
-                throw openAiError(this.model, response.status, await response.text());
+        let current = request;
+        for (;;) {
+            const response = await this.post(responsesBody(this.model, current));
+            if (response.ok) {
+                return parseResponsesBody((await response.json()), this.model);
             }
-            throw openAiError(this.model, response.status, detail);
-        }
-        if (!response.ok) {
-            throw openAiError(this.model, response.status, await response.text());
+            const detail = await response.text();
+            const lighter = response.status === 400
+                ? withoutRefusedPart(current, detail, /text\.format|json_schema|structured/i)
+                : undefined;
+            if (!lighter)
+                throw openAiError(this.model, response.status, detail);
+            current = lighter;
         }
-        return parseResponsesBody((await response.json()), this.model);
     }
 }

package/dist/src/ai/openrouter.js

@@ -1,4 +1,4 @@
-import { chatBody, parseChatResponse } from "./provider.js";
+import { chatBody, parseChatResponse, withoutRefusedPart } from "./provider.js";
 import { getEnv } from "../util/runtime.js";
 import { CliError, EXIT_RUNTIME, EXIT_USAGE, networkFailure, } from "../cli/error.js";
 /** `CM_OPENROUTER_URL` points the provider at a local fake in tests. It is
@@ -75,25 +75,19 @@ export class OpenRouterProvider {
         return openRouterError(this.model, status, body, this.host);
     }
     async complete(request) {
-        let response = await this.post(chatBody(this.model, request));
-        // A provider may refuse `response_format` next to `tools`. The schema is
-        // only a hint — the prompt also asks for a fenced JSON block — so one
-        // retry without it is better than failing the review.
-        if (response.status === 400 && request.responseFormat) {
-            const detail = await response.text();
-            if (/response_format|json_schema|structured/i.test(detail)) {
-                const { responseFormat: _dropped, ...withoutSchema } = request;
-                response = await this.post(chatBody(this.model, withoutSchema));
-                if (response.ok) {
-                    return parseChatResponse((await response.json()), this.host.provider, this.model);
-                }
-                throw this.fail(response.status, await response.text());
+        let current = request;
+        for (;;) {
+            const response = await this.post(chatBody(this.model, current));
+            if (response.ok) {
+                return parseChatResponse((await response.json()), this.host.provider, this.model);
             }
-            throw this.fail(response.status, detail);
-        }
-        if (!response.ok) {
-            throw this.fail(response.status, await response.text());
+            const detail = await response.text();
+            const lighter = response.status === 400
+                ? withoutRefusedPart(current, detail, /response_format|json_schema|structured/i)
+                : undefined;
+            if (!lighter)
+                throw this.fail(response.status, detail);
+            current = lighter;
         }
-        return parseChatResponse((await response.json()), this.host.provider, this.model);
     }
 }

package/dist/src/ai/provider.d.ts

@@ -7,5 +7,14 @@ export { providerKeyEnv } from "../config.ts";
 /** Hetzner was dropped in 0.5.1. Naming it beats a generic "must be one of". */
 export declare function rejectRetiredProvider(value: string | undefined): void;
 export declare function createAiProvider(options: Options, model: string): AiProvider | undefined;
+/** A 400 that names one optional part of the request earns one retry without
+ * it: the JSON schema, which the prompt asks for anyway, or "thinking off",
+ * which a model that always thinks refuses. Undefined when nothing is left to
+ * drop, so the caller reports the error. */
+export declare function withoutRefusedPart(request: AiRequest, detail: string, schemaPattern: RegExp): AiRequest | undefined;
+/** The model stopped at its output limit before writing a word. Thinking counts
+ * toward that limit, so a thinking model can spend all of it there. Without
+ * this the empty answer read as a normal one and failed later as "not JSON". */
+export declare function outputLimitError(provider: string, model: string): Error;
 export declare function parseChatResponse(json: Json, provider: AiResponse["provider"], model: string): AiResponse;
 export declare function chatBody(model: string, request: AiRequest): Json;

package/dist/src/ai/provider.js

@@ -38,6 +38,28 @@ export function createAiProvider(options, model) {
     }
     die(`${options.ai} support is not finished yet. Use openrouter or none.`);
 }
+/** A 400 that names one optional part of the request earns one retry without
+ * it: the JSON schema, which the prompt asks for anyway, or "thinking off",
+ * which a model that always thinks refuses. Undefined when nothing is left to
+ * drop, so the caller reports the error. */
+export function withoutRefusedPart(request, detail, schemaPattern) {
+    if (request.responseFormat && schemaPattern.test(detail)) {
+        const { responseFormat: _dropped, ...rest } = request;
+        return rest;
+    }
+    if (request.reasoningEffort === "none" &&
+        /reasoning|effort|thinking/i.test(detail)) {
+        const { reasoningEffort: _dropped, ...rest } = request;
+        return rest;
+    }
+    return undefined;
+}
+/** The model stopped at its output limit before writing a word. Thinking counts
+ * toward that limit, so a thinking model can spend all of it there. Without
+ * this the empty answer read as a normal one and failed later as "not JSON". */
+export function outputLimitError(provider, model) {
+    return new Error(`${provider} model ${model} reached its output limit before writing an answer. Thinking counts toward that limit.`);
+}
 export function parseChatResponse(json, provider, model) {
     const choice = Array.isArray(json.choices)
         ? json.choices[0]
@@ -61,8 +83,14 @@ export function parseChatResponse(json, provider, model) {
         })
             .filter((call) => call.id && call.function.name)
         : undefined;
+    const text = String(message?.content ?? "");
+    if (choice?.finish_reason === "length" &&
+        !text.trim() &&
+        !toolCalls?.length) {
+        throw outputLimitError(provider, model);
+    }
     return {
-        text: String(message?.content ?? ""),
+        text,
         ...(toolCalls?.length ? { toolCalls } : {}),
         tokensIn: Number(usage?.prompt_tokens ?? 0),
         tokensOut: Number(usage?.completion_tokens ?? 0),

package/dist/src/knowledge/synthesis.js

@@ -157,6 +157,9 @@ function extractRequest(prompt, maxTokens) {
         job: "extract_unit",
         prompt,
         maxTokens,
+        // Pulling facts out of a text needs no thinking, and on a thinking model
+        // the thinking used the whole limit and left the answer empty.
+        reasoningEffort: "none",
         system: "Extract only repository-specific facts. Never invent files, commands, or process.",
     };
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "co-maintainer",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "description": "An imitator of yours, A PR review tool like co-maintainer.",
   "keywords": [
     "ai",