@alisio/alisio-code 0.1.0-alpha.27 → 0.1.0-alpha.28

package/dist/main.js

@@ -131,6 +131,10 @@ async function run(cmd, prompt, sessionId) {
                 const retry = event.data;
                 process.stderr.write(`\nThe model did not respond; retrying (${retry.attempt}/${retry.of})…\n`);
             }
+            else if (event.type === "truncation_recovery") {
+                const recovery = event.data;
+                process.stderr.write(`\nThe response was cut off by the output limit; retrying in smaller steps (${recovery.attempt}/${recovery.of})…\n`);
+            }
             else if (event.type === "tool_started")
                 process.stderr.write(`\n→ ${event.data.name}\n`);
             else if (event.type === "artifact_published")
@@ -318,7 +322,7 @@ program
     console.log('To generate AGENTS.md for this project, run /init inside alisio (or: alisio run "/init" --allow-write).');
 });
 program.command("doctor").action(async (_opts, cmd) => {
-    const { findWorkspace, loadConfigWithProvenance, overridesSavedProviderProfile, ProviderSettingsStore, RIPGREP_INSTALL_HINT, which, } = await import("@alisio/core");
+    const { DEFAULT_MAX_OUTPUT_TOKENS, findWorkspace, loadConfigWithProvenance, MAX_AUTO_OUTPUT_TOKENS, overridesSavedProviderProfile, ProviderSettingsStore, RIPGREP_INSTALL_HINT, which, } = await import("@alisio/core");
     const o = options(cmd);
     const workspace = await findWorkspace(o.cwd ?? process.cwd());
     const { config, provenance } = await loadConfigWithProvenance(workspace, {
@@ -341,6 +345,14 @@ program.command("doctor").action(async (_opts, cmd) => {
         workspace,
         git: (await which("git")) ?? null,
         ripgrep: (await which("rg")) ?? null,
+        // The per-request output budget is resolved when a run starts, from the model's catalog;
+        // doctor does not contact the provider, so it reports what the configuration decides.
+        outputTokens: config.limits.maxOutputTokens !== undefined
+            ? { limit: config.limits.maxOutputTokens, source: "set by you (limits.maxOutputTokens)" }
+            : {
+                limit: null,
+                source: `the model catalog's declared limit (capped at ${MAX_AUTO_OUTPUT_TOKENS}), else ${DEFAULT_MAX_OUTPUT_TOKENS}`,
+            },
         provider: {
             id: useSaved ? saved.profile.provider : "openai-compatible",
             baseURL: useSaved ? saved.profile.values.baseURL : config.provider.baseURL,

package/dist/tui/settings.d.ts

@@ -50,7 +50,8 @@ export interface SettingsConfigView {
     };
     limits: {
         maxTurns: number;
-        maxOutputTokens: number;
+        /** Only present when the user set it; unset means "the model's declared limit, else 16384". */
+        maxOutputTokens?: number | undefined;
         maxContextChars: number;
         timeoutMs: number;
     };

package/dist/tui/settings.js

@@ -12,7 +12,7 @@ export const WEBSEARCH_PROVIDERS = [
 export const defaultConfig = {
     compaction: { auto: true, threshold: 0.85, keepTurns: 2, maxOutputTokens: 16_000 },
     context: { claudeMdFallback: false, maxBytes: 32 * 1024 },
-    limits: { maxTurns: 100, maxOutputTokens: 16_384, maxContextChars: 800_000, timeoutMs: 300_000 },
+    limits: { maxTurns: 100, maxContextChars: 800_000, timeoutMs: 300_000 },
     tui: { paddingX: 1, contentPaddingX: 2, skillSlashCommands: true },
     mcp: { allow: false },
     websearch: { provider: undefined },
@@ -125,9 +125,10 @@ export const SETTINGS_DEFINITIONS = [
         label: "Agent max output tokens",
         category: "Limits",
         valueType: "number",
-        values: [1_024, 2_048, 4_096, 8_192, 16_384],
-        read: (config) => config.limits.maxOutputTokens,
-        description: "Per-call output token budget for agent turns. A cut response shows a notice suggesting a higher value. Applied from the next run.",
+        values: [1_024, 2_048, 4_096, 8_192, 16_384, 32_768, 65_536],
+        // Unset: each request uses the model's declared limit (capped at 65536), else 16384.
+        read: (config) => config.limits.maxOutputTokens ?? 16_384,
+        description: "Per-call output token budget for agent turns. While you leave it unchanged, each request uses the output limit the model's catalog declares (capped at 65536), or 16384 when it declares none; a value set here always wins. A cut response shows a notice suggesting a higher value. Applied from the next run.",
     },
     {
         id: "limits.maxContextChars",

package/dist/tui/state.d.ts

@@ -348,6 +348,11 @@ export interface ViewState {
         attempt: number;
         of: number;
     };
+    /** The response was cut by the output limit and is requested again (`truncation_recovery`). */
+    recovery?: {
+        attempt: number;
+        of: number;
+    };
     model: string;
     /** Context size in tokens: provider-reported, or estimated (`~`). */
     context?: {

package/dist/tui/state.js

@@ -2,7 +2,7 @@
  * Pure presentation logic for the TUI: formatting, command parsing and the reduction of
  * versioned runner events into a view model. No terminal or pi-tui imports here.
  */
-import { CommandCatalog } from "@alisio/core";
+import { CommandCatalog, describeOutputLimitSource } from "@alisio/core";
 const trimZero = (value) => value.replace(/\.0$/, "");
 export function formatTokens(n) {
     if (!Number.isFinite(n) || n < 0)
@@ -583,7 +583,7 @@ export function reduceEvent(state, event) {
         event.type === "run_cancelled" ||
         event.type === "run_turns_exceeded";
     if (terminal) {
-        const { lastEventAt: _ended, retry: _retry, ...rest } = next;
+        const { lastEventAt: _ended, retry: _retry, recovery: _recovery, ...rest } = next;
         return rest;
     }
     return { ...next, lastEventAt: Date.parse(event.timestamp) || Date.now() };
@@ -612,6 +612,11 @@ export function runPhase(view) {
             label: `the model did not respond; retrying (${view.retry.attempt}/${view.retry.of})`,
             silentByDesign: false,
         };
+    if (view.recovery)
+        return {
+            label: `the response was cut off; retrying in smaller steps (${view.recovery.attempt}/${view.recovery.of})`,
+            silentByDesign: false,
+        };
     const tool = view.items.findLast((i) => i.kind === "tool" && i.status !== "ok" && i.status !== "error");
     if (tool?.kind === "tool") {
         const name = tool.name.replace(/^p_[0-9a-f]{10}_/, "");
@@ -658,8 +663,8 @@ const RETRY_ENDERS = new Set([
     "compaction_started",
 ]);
 function reduceEventItems(state, event) {
-    if (state.retry && RETRY_ENDERS.has(event.type)) {
-        const { retry: _retry, ...rest } = state;
+    if ((state.retry || state.recovery) && RETRY_ENDERS.has(event.type)) {
+        const { retry: _retry, recovery: _recovery, ...rest } = state;
         return reduceEventItems(rest, event);
     }
     const d = (event.data ?? {});
@@ -764,6 +769,8 @@ function reduceEventItems(state, event) {
             }
             return ended;
         }
+        case "truncation_recovery":
+            return { ...state, recovery: { attempt: Number(d.attempt) || 1, of: Number(d.of) || 1 } };
         case "request_retry":
             return { ...state, retry: { attempt: Number(d.attempt) || 1, of: Number(d.of) || 1 } };
         case "compaction_started":
@@ -807,7 +814,9 @@ function reduceEventItems(state, event) {
         case "response_truncated":
             return addItem(state, {
                 kind: "notice",
-                text: "Response cut by max output tokens — the answer may be incomplete. Raise limits.maxOutputTokens (/settings → Agent max output tokens) to allow longer answers.",
+                text: `Response cut by max output tokens${typeof d.maxOutputTokens === "number"
+                    ? ` (${d.maxOutputTokens} tokens, ${describeOutputLimitSource(d.source)})`
+                    : ""} — the answer may be incomplete. Raise limits.maxOutputTokens (/settings → Agent max output tokens) to allow longer answers.`,
             });
         case "model_changed":
             return {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@alisio/alisio-code",
-  "version": "0.1.0-alpha.27",
+  "version": "0.1.0-alpha.28",
   "description": "Alisio: an extensible, provider-agnostic coding-agent harness for your terminal. TUI, OpenAI-compatible providers, permissioned local tools, context compaction, persistent memory and a typed plugin SDK.",
   "author": "Gustavo Gutiérrez",
   "license": "MIT",
@@ -45,12 +45,12 @@
     "alisio": "./dist/main.js"
   },
   "dependencies": {
-    "@alisio/core": "0.1.0-alpha.23",
+    "@alisio/core": "0.1.0-alpha.24",
     "@alisio/plugin-memory": "0.1.0-alpha.9",
-    "@alisio/plugin-openai-compatible": "0.1.0-alpha.9",
+    "@alisio/plugin-openai-compatible": "0.1.0-alpha.10",
     "@alisio/plugin-subagents": "0.1.0-alpha.11",
-    "@alisio/sdk": "0.1.0-alpha.19",
-    "@alisio/server": "0.1.0-alpha.7",
+    "@alisio/sdk": "0.1.0-alpha.20",
+    "@alisio/server": "0.1.0-alpha.8",
     "@earendil-works/pi-tui": "0.87.1",
     "commander": "15.0.0"
   },