@alisio/alisio-code 0.1.0-alpha.26 → 0.1.0-alpha.28

package/dist/main.js

@@ -127,6 +127,14 @@ async function run(cmd, prompt, sessionId) {
                 process.stdout.write(`${JSON.stringify(event)}\n`);
             else if (event.type === "text_delta")
                 process.stdout.write(String(event.data.delta));
+            else if (event.type === "request_retry") {
+                const retry = event.data;
+                process.stderr.write(`\nThe model did not respond; retrying (${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")
@@ -314,7 +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, {
@@ -337,6 +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/components.js

@@ -6,7 +6,7 @@ import { attachmentCaption, MAX_ATTACHMENTS_PER_MESSAGE, } from "./attachments.j
 import { btwCurrent, btwHints, btwPosition, btwUsageLine, btwWindow, reduceBtw, } from "./btw.js";
 import { branchDisplay } from "./git-branch.js";
 import { initialQuestionState, reduceQuestions, } from "./questions.js";
-import { contextLevel, contextPercent, editSummary, entryKeyOf, fitIdentityParts, fitSegments, foldCandidateOf, formatContext, formatDuration, formatTokens, groupIdentity, groupToolEntries, humanizeToolName, isUiBlock, previewLinesFor, previewRows, REASONING_EXPAND_MAX_LINES, reasoningDurationMs, } from "./state.js";
+import { contextLevel, contextPercent, editSummary, entryKeyOf, fitIdentityParts, fitSegments, foldCandidateOf, formatContext, formatDuration, formatTokens, groupIdentity, groupToolEntries, humanizeToolName, isUiBlock, previewLinesFor, previewRows, quietFor, REASONING_EXPAND_MAX_LINES, reasoningDurationMs, runPhase, STALLED_MS, } from "./state.js";
 import { highlightCode, imageTheme, levelColor, markdownDefaultTextStyle, markdownTheme, markdownTransform, sanitizeCode, style, } from "./theme.js";
 export const SPINNER = ["⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"];
 /** Shared animation clock advanced by the app while work is running. */
@@ -437,10 +437,14 @@ export class Footer {
         const bar = pct === undefined
             ? style.gray("░".repeat(cells))
             : levelColor(level)("█".repeat(filled)) + style.gray("░".repeat(cells - filled));
+        const quiet = v.streaming ? quietFor(v, clock.now) : 0;
+        const quietNote = quiet
+            ? ` · no response for ${formatDuration(quiet)}${quiet >= STALLED_MS ? " (stuck?)" : ""}`
+            : "";
         const state = v.compacting
             ? `${SPINNER[clock.frame % SPINNER.length]} compacting`
             : v.streaming
-                ? `${SPINNER[clock.frame % SPINNER.length]} streaming ${formatDuration(clock.now - (v.runStartedAt ?? clock.now))}`
+                ? `${SPINNER[clock.frame % SPINNER.length]} ${runPhase(v).label} ${formatDuration(clock.now - (v.runStartedAt ?? clock.now))}${quietNote}`
                 : "● idle";
         const s = v.stats;
         const segments = [
@@ -449,7 +453,15 @@ export class Footer {
                 priority: 10,
                 paint: pct === undefined ? style.gray : levelColor(level),
             },
-            { text: state, priority: 9, paint: v.streaming || v.compacting ? style.cyan : style.green },
+            {
+                text: state,
+                priority: 9,
+                paint: quiet >= STALLED_MS
+                    ? style.yellow
+                    : v.streaming || v.compacting
+                        ? style.cyan
+                        : style.green,
+            },
             {
                 text: `↑${formatTokens(s.input)} ↓${formatTokens(s.output)}${s.cached ? ` ⚡${formatTokens(s.cached)}` : ""}`,
                 priority: 6,

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

@@ -343,6 +343,16 @@ export interface ViewState {
     items: TranscriptItem[];
     streaming: boolean;
     compacting: boolean;
+    /** The runner is sending a silent request again (`request_retry`); cleared by any output. */
+    retry?: {
+        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?: {
@@ -350,6 +360,8 @@ export interface ViewState {
         estimated: boolean;
     };
     runStartedAt?: number;
+    /** Epoch ms of the latest event of the live run (any type); undefined when no run is live. */
+    lastEventAt?: number;
     stats: Stats;
 }
 export declare function initialViewState(model: string, now?: number): ViewState;
@@ -360,6 +372,19 @@ export declare function reasoningDurationMs(item: {
     reasoningDoneAt?: number;
 }): number | undefined;
 export declare function reduceEvent(state: ViewState, event: RunEvent): ViewState;
+/** No event for this long: the footer says how long the run has been quiet. */
+export declare const QUIET_MS = 15000;
+/** No event for this long: the footer marks the run as possibly stuck. */
+export declare const STALLED_MS = 60000;
+export interface RunPhase {
+    label: string;
+    /** Silence is expected here (the user's turn, or a sub-agent working in its own session). */
+    silentByDesign: boolean;
+}
+/** What the live run is doing, from the transcript items and flags the events already update. */
+export declare function runPhase(view: ViewState): RunPhase;
+/** Milliseconds of silence worth reporting (0 when the run is busy, or silent by design). */
+export declare function quietFor(view: ViewState, now: number): number;
 export declare function lastAssistantText(items: TranscriptItem[]): string | undefined;
 /**
  * True when a printable key should copy the last assistant response instead of typing: bound to

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$/, "");
 export function formatTokens(n) {
     if (!Number.isFinite(n) || n < 0)
@@ -577,6 +577,96 @@ function updateTool(state, id, update) {
     return { ...state, items };
 }
 export function reduceEvent(state, event) {
+    const next = reduceEventItems(state, event);
+    const terminal = event.type === "run_completed" ||
+        event.type === "run_failed" ||
+        event.type === "run_cancelled" ||
+        event.type === "run_turns_exceeded";
+    if (terminal) {
+        const { lastEventAt: _ended, retry: _retry, recovery: _recovery, ...rest } = next;
+        return rest;
+    }
+    return { ...next, lastEventAt: Date.parse(event.timestamp) || Date.now() };
+}
+/** No event for this long: the footer says how long the run has been quiet. */
+export const QUIET_MS = 15_000;
+/** No event for this long: the footer marks the run as possibly stuck. */
+export const STALLED_MS = 60_000;
+/** A short, safe label from tool arguments: a title or a file name, never commands or URLs. */
+function shortArg(args) {
+    const input = parseArgs(args);
+    if (!input)
+        return "";
+    const first = Array.isArray(input.inputs) ? input.inputs[0] : {};
+    const value = [input.title, input.fileName, input.path, input.file_path, first?.path].find((v) => typeof v === "string" && v.trim() !== "");
+    if (!value)
+        return "";
+    return truncatePlain(oneLine(value.replace(/\\/g, "/").split("/").filter(Boolean).pop() ?? ""), 40);
+}
+/** What the live run is doing, from the transcript items and flags the events already update. */
+export function runPhase(view) {
+    if (view.compacting)
+        return { label: "compacting context", silentByDesign: false };
+    if (view.retry)
+        return {
+            label: `the model did not respond; retrying (${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}/${view.recovery.of})`,
+            silentByDesign: false,
+        };
+    const tool = view.items.findLast((i) => i.kind === "tool" && i.status !== "ok" && i.status !== "error");
+    if (tool?.kind === "tool") {
+        const name = tool.name.replace(/^p_[0-9a-f]{10}_/, "");
+        const arg = shortArg(tool.args);
+        const detail = arg ? ` (${arg})` : "";
+        if (tool.status === "approval")
+            return {
+                label: `waiting for your approval: ${humanizeToolName(name)}`,
+                silentByDesign: true,
+            };
+        if (name === "ask_user_question")
+            return { label: "waiting for your answer", silentByDesign: true };
+        if (name === "task" || name === "task_wait")
+            return { label: "a sub-agent is working", silentByDesign: true };
+        if (name === "python_run")
+            return { label: `running Python${detail}`, silentByDesign: false };
+        if (name === "data_inspect" || name === "data_query")
+            return { label: `reading your data${detail}`, silentByDesign: false };
+        if (name === "artifact_create")
+            return { label: `publishing the artifact${detail}`, silentByDesign: false };
+        return { label: `running ${humanizeToolName(name)}${detail}`, silentByDesign: false };
+    }
+    const last = view.items.at(-1);
+    if (last?.kind === "assistant" && !last.done)
+        return {
+            label: last.text ? "writing the answer" : "thinking",
+            silentByDesign: false,
+        };
+    return { label: "waiting for the model", silentByDesign: false };
+}
+/** Milliseconds of silence worth reporting (0 when the run is busy, or silent by design). */
+export function quietFor(view, now) {
+    if (view.lastEventAt === undefined || runPhase(view).silentByDesign)
+        return 0;
+    const idle = Math.max(0, now - view.lastEventAt);
+    return idle >= QUIET_MS ? idle : 0;
+}
+/** What the retried request produced (or the end of the turn) ends the retry label. */
+const RETRY_ENDERS = new Set([
+    "text_delta",
+    "reasoning_delta",
+    "turn_completed",
+    "tool_started",
+    "compaction_started",
+]);
+function reduceEventItems(state, event) {
+    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 ?? {});
     const at = Date.parse(event.timestamp);
     switch (event.type) {
@@ -679,6 +769,10 @@ export function reduceEvent(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":
             return { ...state, compacting: true };
         case "compaction_completed": {
@@ -720,7 +814,9 @@ export function reduceEvent(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.26",
+  "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.22",
+    "@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.18",
-    "@alisio/server": "0.1.0-alpha.6",
+    "@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"
   },