@turri-dev/cli 0.1.4-next.318a98a → 0.1.4-next.5a8c6e5

package/dist/init-config.js

@@ -110,6 +110,7 @@ export function withModels(text, lines) {
 export const DEFAULTS = {
     conformanceTests: false,
     review: false,
+    mutationTesting: true,
     allowToStopJobs: true,
     releaseAfterPulse: false,
     testingScriptsRunOn: "in-place",
@@ -143,6 +144,10 @@ export function quickStartConfig(vendor = "other", bedrock) {
         "# Whether the code and its tests are reviewed once they pass - two reviews at most, the second weighed by the arbiter.",
         `review: ${DEFAULTS.review}`,
         "",
+        "# Whether the coding agent and the reviewer can put a bug in the code for one run of the unit tests (can_it_fail),",
+        "# to see that a test catches it - on a copy of the code, or in place and put back after.",
+        `mutation-testing: ${DEFAULTS.mutationTesting}`,
+        "",
         "# Whether turri stops a running pulse the spec has moved on from. false: it lands, the change regenerated on top.",
         `allow-to-stop-jobs: ${DEFAULTS.allowToStopJobs}`,
         "",

package/dist/jellyfish.js

@@ -0,0 +1,130 @@
+import { sgr } from "./regenerate/plain-palette.js";
+/** The pixels' colours on a dark background and a light one - the tentacles darker on light, where neon does not show. */
+export const JELLYFISH_COLOURS = {
+    dark: { p: "#ff5c8d", h: "#ff7fa6", w: "#f4f7f9", o: "#ff7a3d", e: "#732553", N: "#8eebee", n: "#5ce1e6", l: "#85a3b2", g: "#3e5a6b" },
+    light: { p: "#ff5c8d", h: "#ff7fa6", w: "#f4f7f9", o: "#ff7a3d", e: "#732553", N: "#3e5a6b", n: "#1e3442", l: "#85a3b2", g: "#85a3b2" },
+};
+/**
+ * How shut the bell is, frame by frame - a quarter second each: shut quickly, held a moment, opened slowly, at rest
+ * until the next stroke.
+ */
+const SHUT = [0.1, 0.55, 0.9, 1, 0.85, 0.7, 0.56, 0.44, 0.33, 0.23, 0.14, 0.06];
+const LARGE = {
+    width: 24,
+    height: 26,
+    open: [6, 12, 16, 18, 20, 22, 22, 24, 24],
+    shut: [6, 10, 14, 16, 18, 18, 18, 20, 20],
+    glint: [[-3, 2], [-4, 2], [-5, 3], [-6, 4], [-6, 5]],
+    shine: [-2, 2],
+    gut: [".oo..oo.", "oooooooo", "oooooooo", ".oo..oo."],
+    gutRow: 4,
+    strands: [
+        { at: -8, length: 10, sway: 1.1, ink: "N" },
+        { at: -4.5, length: 13, sway: 1.3, ink: "N" },
+        { at: -1, length: 14, sway: 0.8, ink: "n" },
+        { at: 2, length: 13, sway: 0.8, ink: "n" },
+        { at: 5.5, length: 13, sway: 1.3, ink: "N" },
+        { at: 9, length: 10, sway: 1.1, ink: "N" },
+    ],
+};
+const SMALL = {
+    width: 18,
+    height: 20,
+    open: [4, 10, 12, 14, 16, 18, 18],
+    shut: [4, 8, 10, 12, 14, 14, 14],
+    glint: [[-2, 1], [-3, 2], [-4, 3]],
+    shine: [-1, 1],
+    gut: ["oo..oo", "oooooo", "oo..oo"],
+    gutRow: 3,
+    strands: [
+        { at: -6, length: 8, sway: 0.9, ink: "N" },
+        { at: -3, length: 10, sway: 1, ink: "N" },
+        { at: -0.5, length: 11, sway: 0.7, ink: "n" },
+        { at: 1.5, length: 10, sway: 0.7, ink: "n" },
+        { at: 4, length: 10, sway: 1, ink: "N" },
+        { at: 7, length: 8, sway: 0.9, ink: "N" },
+    ],
+};
+/** One frame of a design: its height in rows of its width in pixels, "." where there is none. */
+function frame(design, t) {
+    const grid = Array.from({ length: design.height }, () => Array(design.width).fill("."));
+    const centre = design.width / 2;
+    const set = (y, x, pixel, over = true) => {
+        if (y < 0 || y >= design.height || x < 0 || x >= design.width)
+            return;
+        if (over || grid[y][x] === ".")
+            grid[y][x] = pixel;
+    };
+    const shut = SHUT[t % SHUT.length];
+    // Shut, it has risen a pixel.
+    const top = shut > 0.6 ? 0 : 1;
+    const widths = design.open.map((open, row) => 2 * Math.round((open - (open - design.shut[row]) * shut) / 2));
+    widths.forEach((width, row) => {
+        for (let x = centre - width / 2; x < centre + width / 2; x++)
+            set(top + row, x, "p");
+    });
+    const rimWidth = widths.at(-1);
+    const rim = top + widths.length;
+    // The underside along the bottom row, the rim scalloped under it.
+    for (let x = centre - rimWidth / 2 + 2; x < centre + rimWidth / 2 - 2; x++)
+        set(rim - 1, x, "e");
+    for (let x = centre - rimWidth / 2; x < centre + rimWidth / 2; x += 2)
+        set(rim, x + 1, "p");
+    // A glint up on the left, curving with the dome, and a pixel of shine.
+    for (const [dx, dy] of design.glint)
+        set(top + dy, centre + dx, "h");
+    set(top + design.shine[1], centre + design.shine[0], "w");
+    const gutLeft = centre - Math.floor(design.gut[0].length / 2);
+    design.gut.forEach((line, dy) => [...line].forEach((pixel, dx) => pixel === "o" && set(top + design.gutRow + dy, gutLeft + dx, "o")));
+    // The tentacles: rooted under the bell, drawn in as it shuts, swaying - the further down, the later.
+    const phase = (2 * Math.PI * t) / SHUT.length;
+    design.strands.forEach((strand, i) => {
+        const root = centre - 0.5 + strand.at * (rimWidth / design.open.at(-1));
+        for (let y = 0; y < strand.length; y++) {
+            const x = Math.round(root + strand.sway * Math.sin(phase - y * 0.55 + i * 0.8) * Math.min(1, y / 3));
+            const left = strand.length - y;
+            // The tips come apart, every other pixel, fading - into code.
+            if (left <= 3 && y % 2 === 1)
+                continue;
+            set(rim + 1 + y, x, left > 5 ? strand.ink : left > 2 ? "l" : "g", false);
+            // The oral arms are frilled.
+            if (strand.ink === "n" && left > 5 && (y + t) % 3 === 0)
+                set(rim + 1 + y, x + (strand.at < 0 ? -1 : 1), "n", false);
+        }
+    });
+    return grid.map((row) => row.join(""));
+}
+export const JELLYFISH_WIDTH = LARGE.width;
+export const JELLYFISH_HEIGHT = LARGE.height;
+/** The large jellyfish's frames, a stroke: what the pictures are rendered from. */
+export const JELLYFISH_FRAMES = SHUT.map((_, t) => frame(LARGE, t));
+/** The small jellyfish's frames, a stroke: what a terminal without images draws, in half blocks. */
+export const JELLYFISH_TEXT_FRAMES = SHUT.map((_, t) => frame(SMALL, t));
+/** How long each frame shows. */
+export const JELLYFISH_FRAME_MS = 250;
+/**
+ * A frame drawn in half blocks for a terminal that shows no images: two pixels a cell, the upper one the text's colour
+ * and the lower one the background's, in as many colours as the terminal has: a column a pixel, a row every two.
+ */
+export function halfBlocks(pixels, appearance, truecolor) {
+    const colours = JELLYFISH_COLOURS[appearance];
+    const colour = (pixel) => (pixel === undefined || pixel === "." ? undefined : colours[pixel]);
+    const rows = [];
+    for (let y = 0; y < pixels.length; y += 2) {
+        let row = "";
+        for (let x = 0; x < (pixels[0]?.length ?? 0); x++) {
+            const upper = colour(pixels[y]?.[x]);
+            const lower = colour(pixels[y + 1]?.[x]);
+            if (upper === undefined && lower === undefined)
+                row += " ";
+            else if (upper === undefined)
+                row += `\x1b[${sgr(lower, false, truecolor)}m▄\x1b[39m`;
+            else if (lower === undefined)
+                row += `\x1b[${sgr(upper, false, truecolor)}m▀\x1b[39m`;
+            else
+                row += `\x1b[${sgr(upper, false, truecolor)};${sgr(lower, true, truecolor)}m▀\x1b[39;49m`;
+        }
+        rows.push(row);
+    }
+    return rows;
+}

package/dist/loop-cli.js

@@ -4,6 +4,7 @@ import { resolve, sep } from "node:path";
 import { configForSpec, inheritModel, modelOnly, modelSpec, projectModels, readProjectConfig, roleModel, ROLE_THINKING, splitModelSpec, } from "./config.js";
 import { PACKAGE_ROOT } from "./environment.js";
 import { c } from "./module-report.js";
+import { describeTurnIncident, onModelEvent } from "./regenerate/activity.js";
 import { tryItCommand } from "./regenerate/delivery.js";
 import { describeDiagnosis, diagnoseFailure } from "./regenerate/diagnose.js";
 import { createJobRegistry, pulseName } from "./regenerate/jobs.js";
@@ -51,6 +52,7 @@ export async function runLoop(options) {
             ...(buildDir !== undefined ? { buildDir } : {}),
             conformanceTests: conformanceLabel(project.conformanceTests),
             review: project.review === true,
+            mutationTesting: project.mutationTesting !== false,
             models: reportModels(project, model),
         })
         : undefined;
@@ -78,16 +80,26 @@ async function runLoopReporting(options) {
                 return;
             seen.set(job.id, line);
             if (job.status === "running")
-                console.log(`${c.cyan(`▶ ${job.id}`)} ${pulseName(job)}${job.stage !== undefined ? c.dim(` - ${job.stage}`) : ""}`);
+                console.log(`${c.accent(`▶ ${job.id}`)} ${pulseName(job)}${job.stage !== undefined ? c.dim(` - ${job.stage}`) : ""}`);
         },
         onFinish: (job) => {
-            const mark = job.status === "succeeded" ? c.green("✓") : job.status === "stopped" ? c.dim("■") : c.red("✗");
+            const mark = job.status === "succeeded" ? c.success("✓") : job.status === "stopped" ? c.dim("■") : c.error("✗");
             console.log(`${mark} ${job.id} ${pulseName(job)} - ${job.status}`);
             if (job.status !== "succeeded")
                 console.log(c.dim((job.announce ?? job.report ?? job.error ?? "").trim()));
+            // Its report is not printed when it succeeds, and nobody is asked here: the spec gaps its agents met are said.
+            else
+                for (const gap of job.details?.specGaps ?? [])
+                    console.log(c.warning(`  spec gap: ${gap.question}`), c.dim(`Meanwhile: ${gap.assumption}`));
             report?.write();
         },
     });
+    // A provider failing and waited out, a turn turri ended itself: said as they happen, not left in the job logs alone.
+    onModelEvent((event) => {
+        const line = describeTurnIncident(event);
+        if (line !== undefined)
+            console.log(c.warning(`  ${event.job !== undefined ? `${event.job}: ` : ""}${line}`));
+    });
     const planner = modelSpec(projectModels(project).harness);
     const loop = createWorkLoop({
         registry,
@@ -119,7 +131,7 @@ async function runLoopReporting(options) {
         // No terminal and no --start-over: no one to ask, so the run stops here rather than build nothing and say it is done.
         if (options.startOver !== true && process.stdin.isTTY !== true) {
             const message = `${why}\nThere is no terminal to ask whether to start over: run it again with --start-over to start over without asking.`;
-            console.error(c.red(message));
+            console.error(c.error(message));
             report?.end("incomplete", message);
             process.exit(1);
         }
@@ -150,10 +162,10 @@ async function runLoopReporting(options) {
     // waiting for a start over, nothing the planner left out.
     const unfinished = moduleStates(loop.drift(plainFile)).filter((module) => module.state !== "complete");
     const verdict = failed.length > 0
-        ? c.red(c.bold(`${failed.length} job${failed.length === 1 ? "" : "s"} failed`))
+        ? c.error(c.bold(`${failed.length} job${failed.length === 1 ? "" : "s"} failed`))
         : unfinished.length > 0
-            ? c.red(c.bold(`not complete: ${unfinished.map((module) => `${module.module}${module.why !== undefined ? ` - ${module.why}` : ""}`).join("; ")}`))
-            : c.green(c.bold("all jobs done"));
+            ? c.error(c.bold(`not complete: ${unfinished.map((module) => `${module.module}${module.why !== undefined ? ` - ${module.why}` : ""}`).join("; ")}`))
+            : c.success(c.bold("all jobs done"));
     console.log(`\n${verdict}`);
     for (const job of [...ended].reverse())
         console.log(`  ${job.id} ${pulseName(job)}: ${job.status}`);
@@ -236,7 +248,7 @@ function reportModels(project, model) {
 }
 /**
  * The options every command that regenerates takes for a headless run: `--report` and `--build-dir`, and the config's
- * own keys (configKeyOptions) - `--harness-model`, `--coding-model` (`--model`), `--judge-model`, `--review` and, where
+ * own keys (configKeyOptions) - `--harness-model`, `--coding-model` (`--model`), `--judge-model`, `--review`, `--mutation-testing` and, where
  * `conformance` is true, `--conformance-tests` - read as if config.turri.yaml had them. Exits on a value it cannot read.
  */
 export function headlessOptions(options, conformance) {

package/dist/loop-tools.js

@@ -110,6 +110,10 @@ function plainFileOf(cwd, given, what) {
 export function asksFromScratch(words) {
     return FROM_SCRATCH.test(words);
 }
+/** Whether a /regenerate asking for its module from scratch asks for its whole requires chain with it. */
+export function asksWholeChain(words) {
+    return /\bchain\b|\bevery module\b|\ball (the )?modules\b/i.test(words);
+}
 const FROM_SCRATCH = /\b(from (scratch|nothing|zero|the (start|beginning|ground up))|start(ing)? (it )?(over|afresh)|afresh|anew|clean slate|(delete|wipe|remove|drop|throw away)\b.{0,30}\b(module|code|codebase|everything|all of it))/i;
 /** regenerate: regenerate a module's code up to its spec as it is now. */
 export function regenerateTool(options) {
@@ -131,6 +135,7 @@ export function regenerateTool(options) {
             "A change to a file the spec links - an API surface, a mapping document - is a change to the spec: regenerate sees it by itself, and regenerates every functional spec that uses what links it.",
             "Never stop pulses and start them again yourself to follow a spec change: regenerate does it.",
             "Set from_scratch only when the user asks for the module to be regenerated from scratch - started over, its code deleted - in this conversation, or agrees to it when a module's status says it has to start over: never to fix a pulse that did not land.",
+            'from_scratch "module" takes out the code where the module is built; the modules below keep theirs, and the module starts from a copy of the code below it. "chain" takes out every codebase of its requires chain, the base\'s too, and builds all of it again from the bottom: use it after modules were split, merged or moved (refactor-module), or when the user asks for the whole chain.',
         ],
         parameters: Type.Object({
             plain_file: Type.Optional(Type.String({ description: "Path to the .plain file. Omitted: the module the project's config.turri.yaml names" })),
@@ -139,8 +144,9 @@ export function regenerateTool(options) {
             functionalities: Type.Optional(Type.Array(Type.String(), {
                 description: "Functional specs already built, by id, to regenerate though their spec did not change. Omitted: what the spec now differs from the code in",
             })),
-            from_scratch: Type.Optional(Type.Boolean({
-                description: "Only when the user asks for it: the module's code and conformance tests deleted - committed first, so its lineage keeps them - and the whole module regenerated from its spec",
+            from_scratch: Type.Optional(Type.Union([Type.Literal("module"), Type.Literal("chain")], {
+                description: "Only when the user asks for it: the module's code and conformance tests deleted - committed first, so its lineage keeps them - and the whole module regenerated from its spec. " +
+                    '"chain": every module of its requires chain too, built again from the bottom',
             })),
             model: MODEL_PARAM,
             wait: WAIT_PARAM,
@@ -151,15 +157,15 @@ export function regenerateTool(options) {
             if (!check.ok)
                 throw new Error(`${relative(options.cwd, plainFile) || plainFile} does not parse, so nothing can be regenerated from it yet: ${check.error}`);
             const again = given.functionalities ?? [];
-            if (again.length > 0 && given.from_scratch === true)
+            if (again.length > 0 && given.from_scratch !== undefined)
                 throw new Error("Name functionalities or set from_scratch, not both: from scratch regenerates every one of them.");
             // Named functionalities are work on functionalities, as engine chooses a model for it; the rest, spec changes.
             const work = again.length > 0 ? "functionalities" : "spec_changes";
             const model = options.model !== undefined ? options.model(ctx, plainFile, given.model, work) : given.model;
             const loop = options.loop(ctx);
             const request = { ...(model !== undefined ? { model } : {}), ...(given.guidance !== undefined ? { guidance: given.guidance } : {}) };
-            const statuses = given.from_scratch === true
-                ? await loop.fromScratch(plainFile, request)
+            const statuses = given.from_scratch !== undefined
+                ? await loop.fromScratch(plainFile, { ...request, chain: given.from_scratch === "chain" })
                 : again.length > 0
                     ? await loop.regenerateAgain(plainFile, again, request)
                     : await loop.bringUp(plainFile, request);

package/dist/module-report.js

@@ -1,9 +1,10 @@
+import { brand } from "./brand.js";
 /** Colours for what the command line prints. */
 export const c = {
     dim: (s) => `\x1b[2m${s}\x1b[0m`,
     bold: (s) => `\x1b[1m${s}\x1b[0m`,
-    green: (s) => `\x1b[32m${s}\x1b[0m`,
-    red: (s) => `\x1b[31m${s}\x1b[0m`,
-    yellow: (s) => `\x1b[33m${s}\x1b[0m`,
-    cyan: (s) => `\x1b[36m${s}\x1b[0m`,
+    success: brand.success,
+    error: brand.error,
+    warning: brand.warning,
+    accent: brand.accent,
 };

package/dist/pi.js

@@ -4,6 +4,7 @@ import { createAgentSession, DefaultResourceLoader, getAgentDir, ModelRegistry,
 import { loadRules, renderRulesBlock } from "./rules.js";
 import { inheritModel, stackOf, readProjectConfig, ROLE_THINKING, roleModel, roleTurnLimitMs, splitModelSpec } from "./config.js";
 import { maskToolResults } from "./regenerate/masking.js";
+import { withListArguments } from "./regenerate/list-arguments.js";
 import { checkAgainstScript, ENVIRONMENT_ERROR_TOOL, EnvironmentError, environmentErrorTool, sandboxGapOf } from "./regenerate/environment-error.js";
 import { MODULES_DIR, projectRootOf, sessionsDir } from "./regenerate/paths.js";
 import { currentStopSignal, describeToolCall, describeToolResult, JobStoppedError, shownName, reportActivity, reportNow, reportAgent, reportTranscript, reportModelEvent, reportUsage, } from "./regenerate/activity.js";
@@ -146,6 +147,19 @@ export class ProviderError extends Error {
         this.name = "ProviderError";
     }
 }
+/** Why turri itself ended a session's turn under way: the abort that comes back is its own, not the provider failing. */
+const harnessStops = new WeakMap();
+/**
+ * End the agent's turn, turri's own doing - it is going round in a circle, its time is up. The turn ends as any turn
+ * does: what the agent said stands, and the harness goes on from there. Never a provider error.
+ */
+export function stopTurn(session, reason) {
+    if (session === undefined || harnessStops.has(session))
+        return;
+    harnessStops.set(session, reason);
+    reportModelEvent({ kind: "stopped", ...modelEventOf(session), reason });
+    void session.abort();
+}
 /**
  * Run one turn and return the assistant's text, streaming deltas to `onDelta`.
  *
@@ -212,17 +226,32 @@ async function askUnrestricted(session, text, onDelta) {
         catch (error) {
             if (!(error instanceof ProviderError) || !TRANSIENT_PROVIDER_ERROR.test(error.message) || attempt >= PROVIDER_RETRIES)
                 throw error;
-            const wait = PROVIDER_RETRY_DELAY_MS * 2 ** attempt;
-            reportActivity(`  the model provider is failing for now - trying again in ${Math.round(wait / 1000)}s (${attempt + 1} of ${PROVIDER_RETRIES})`);
-            debugLogOf(session)?.note(`provider failing - trying again in ${Math.round(wait / 1000)}s: ${error.message.slice(0, 200)}`);
+            const wait = providerRetryDelay(attempt);
+            const shown = `${Math.round(wait / 1000)}s (${attempt + 1} of ${PROVIDER_RETRIES})`;
+            reportActivity(`  the model provider is failing for now - trying again in ${shown}`);
+            reportNow(`the model provider is failing - trying again in ${shown}`);
+            reportModelEvent({ kind: "provider_retry", ...modelEventOf(session), error: error.message, attempt: attempt + 1, of: PROVIDER_RETRIES, waitMs: wait });
+            debugLogOf(session)?.note(`provider failing - trying again in ${shown}: ${error.message.slice(0, 200)}`);
             await waitOrStop(wait);
+            reportNow(undefined);
             prompt = "Carry on where you were: the model provider failed part-way through your last turn.";
         }
     }
 }
-/** How many times, after pi's own retries, a turn whose provider failed for a moment is tried again - and the first wait, doubled each time. */
-export const PROVIDER_RETRIES = 3;
+/**
+ * How many times, after pi's own retries, a turn whose provider failed for a moment is tried again; the first wait,
+ * doubled each time up to the longest - 15s, 30s, 1m, 2m, 4m, 5m: an outage of about 13 minutes is waited out.
+ */
+export const PROVIDER_RETRIES = 6;
 export const PROVIDER_RETRY_DELAY_MS = Number(process.env.TURRI_PROVIDER_RETRY_DELAY ?? 15) * 1000;
+export const PROVIDER_RETRY_MAX_DELAY_MS = 20 * PROVIDER_RETRY_DELAY_MS;
+/** The wait before try `attempt + 1`: doubled each time, up to the longest, with a tenth either way so parallel jobs do not all come back at once. */
+export function providerRetryDelay(attempt, random = Math.random) {
+    const base = Math.min(PROVIDER_RETRY_DELAY_MS * 2 ** attempt, PROVIDER_RETRY_MAX_DELAY_MS);
+    return Math.round(base * (0.9 + 0.2 * random()));
+}
+/** How long a request to the model may go without anything coming back before the provider is said to be stalling; 0 never. */
+export const PROVIDER_STALL_MS = Number(process.env.TURRI_PROVIDER_STALL ?? 120) * 1000;
 /** A provider error that passes: overloaded, rate limited, unreachable, timed out. */
 const TRANSIENT_PROVIDER_ERROR = /\b(429|500|502|503|504|529)\b|overloaded|high demand|rate.?limit|too many requests|unavailable|fetch failed|network|connection|timed? ?out|ECONNRESET|socket|terminated/i;
 /** Wait, unless the job is stopped first. */
@@ -251,7 +280,7 @@ async function askOnce(session, text, onDelta) {
     // Calls to one tool failing one after another - arguments it refuses, or errors - with nothing going through.
     let failing = 0;
     let failingTool;
-    let stoppedFailing = false;
+    harnessStops.delete(session);
     reportNow("waiting for the model");
     // Every entry says which agent it is from, and what it runs on - for reading one agent at a time.
     const agent = agentOf(session);
@@ -263,7 +292,30 @@ async function askOnce(session, text, onDelta) {
     if (agent !== undefined)
         reportAgent({ label: agent.label, ...(agent.model !== undefined ? { model: agent.model } : {}), turns: 1 });
     reportModelEvent({ kind: "turn", ...modelEventOf(session) });
+    // The provider stalling: a request to the model that nothing has come back on - not a token, not an error - for
+    // long. Said once for each silence; the turn goes on waiting, its own time limit ending it if it must.
+    let toolsRunning = 0;
+    let stall;
+    const watchForStall = (afterMs = PROVIDER_STALL_MS) => {
+        if (stall !== undefined)
+            clearTimeout(stall);
+        stall = toolsRunning > 0 || PROVIDER_STALL_MS <= 0 ? undefined : setTimeout(() => {
+            reportModelEvent({ kind: "provider_stall", ...modelEventOf(session), waitMs: PROVIDER_STALL_MS });
+            debugLogOf(session)?.note(`the provider has sent nothing back for ${seconds(PROVIDER_STALL_MS)}`);
+        }, afterMs);
+    };
     const unsubscribe = session.subscribe((event) => {
+        if (event.type === "tool_execution_start")
+            toolsRunning += 1;
+        if (event.type === "tool_execution_end")
+            toolsRunning = Math.max(0, toolsRunning - 1);
+        // pi tries a failed request again itself first: the provider failing is said then, not only once pi gives up.
+        if (event.type === "auto_retry_start") {
+            reportModelEvent({ kind: "provider_retry", ...modelEventOf(session), error: event.errorMessage, attempt: event.attempt, of: event.maxAttempts, waitMs: event.delayMs });
+            watchForStall(event.delayMs + PROVIDER_STALL_MS);
+        }
+        else
+            watchForStall();
         if (event.type === "message_update") {
             const update = event.assistantMessageEvent;
             if (update.type === "text_delta") {
@@ -308,14 +360,13 @@ async function askOnce(session, text, onDelta) {
         if (event.type === "tool_execution_end") {
             failing = event.isError ? (event.toolName === failingTool ? failing + 1 : 1) : 0;
             failingTool = event.isError ? event.toolName : undefined;
-            if (failing >= FAILURES_BEFORE_STOP && !stoppedFailing) {
+            if (failing >= FAILURES_BEFORE_STOP && !harnessStops.has(session)) {
                 // A call its tool refuses never reaches the tool, so no guard on the tools sees it: an agent can
                 // make the same bad call, worded a little differently, thousands of times. Its turn ends here, and
                 // the harness goes on from there, as after any turn.
-                stoppedFailing = true;
                 reportActivity(`  stopped the agent's turn: ${failing} ${shownName(event.toolName)} calls failed in a row`);
                 debugLogOf(session)?.note(`turn stopped: ${failing} ${event.toolName} calls failed in a row`);
-                void session.abort();
+                stopTurn(session, `${failing} ${shownName(event.toolName)} calls failed in a row`);
             }
             const output = resultText(event.result);
             transcript({ kind: "result", text: clipText(output, MAX_TOOL_RESULT), error: event.isError, full: output });
@@ -376,7 +427,7 @@ async function askOnce(session, text, onDelta) {
         stopped = true;
         reportActivity(`  stopped the agent's turn: its ${formatLimit(limitMs)} are up`);
         debug?.note(`turn stopped: its ${formatLimit(limitMs)} are up${how}`);
-        void session.abort();
+        stopTurn(session, `its ${formatLimit(limitMs)} were up`);
     };
     const atNextRequest = timed
         ? session.subscribe((event) => {
@@ -392,6 +443,7 @@ async function askOnce(session, text, onDelta) {
     }
     // What the model is sent, and only that: its task is its system prompt, which the log already holds whole.
     debug?.prompt(text);
+    watchForStall();
     try {
         // Its images, when its conversation does not hold them: the first turn, and the first after a compaction.
         const images = sessionImages.get(session);
@@ -412,19 +464,24 @@ async function askOnce(session, text, onDelta) {
             clearTimeout(timer);
         if (backstop !== undefined)
             clearTimeout(backstop);
+        if (stall !== undefined)
+            clearTimeout(stall);
         atNextRequest?.();
         unsubscribe();
         stop?.removeEventListener("abort", onStop);
         reportNow(undefined);
     }
-    if (stopped)
-        out = `${out.trim()}\n\n[the turn was stopped: its ${formatLimit(limitMs)} were up]`.trim();
+    // turri ended the turn itself - its time was up, the agent went round in a circle: the abort that ended it is its own.
+    const stoppedFor = harnessStops.get(session);
+    harnessStops.delete(session);
+    if (stoppedFor !== undefined)
+        out = `${out.trim()}\n\n[the turn was stopped: ${stoppedFor}]`.trim();
     if (stop?.aborted) {
         debug?.note(`turn stopped on request after ${seconds(Date.now() - turnStarted)}`);
         throw new JobStoppedError();
     }
-    // A turn ended for its time is ended, whatever the provider was doing then.
-    if (providerError !== undefined && !stopped) {
+    // A turn turri ended is ended, whatever the provider was doing then: what comes back of the abort is not the provider failing.
+    if (providerError !== undefined && stoppedFor === undefined) {
         debug?.note(`turn ended with a provider error after ${seconds(Date.now() - turnStarted)}: ${providerError}`);
         throw new ProviderError(`${session.model?.id ?? "model"}: ${providerError}`);
     }
@@ -614,7 +671,7 @@ export function guardRepeats(tools, session, budget) {
                 refused += 1;
                 if (refused > REFUSED_BEFORE_STOP) {
                     reportActivity(`  stopped the agent's turn: ${refused} calls to tools this turn does not have`);
-                    void current?.abort();
+                    stopTurn(current, `${refused} calls to tools this turn does not have`);
                     throw new Error(`Stopped: ${refused} calls to tools this turn does not have.`);
                 }
                 throw new Error(`Not run: your work is done and checked - this turn is for ${[...only].join(" and ")} alone, and nothing else is ` +
@@ -624,7 +681,7 @@ export function guardRepeats(tools, session, budget) {
             if (budget !== undefined && !budget.finish.includes(tool.name) && calls > budget.calls) {
                 if (calls > budget.calls * 2) {
                     reportActivity(`  stopped the agent's turn: ${calls} tool calls, past its budget of ${budget.calls}`);
-                    void session()?.abort();
+                    stopTurn(session(), `${calls} tool calls, past its budget of ${budget.calls}`);
                     throw new Error(`Stopped: ${calls} tool calls, past the budget of ${budget.calls}.`);
                 }
                 throw new Error(`Not run: you have made ${budget.calls} tool calls, which is all you have. Decide now with what you know: call ` +
@@ -652,7 +709,7 @@ export function guardRepeats(tools, session, budget) {
             }
             if (repeats > REPEATS_BEFORE_STOP) {
                 reportActivity(`  stopped the agent's turn: it made the same ${tool.name} call ${repeats} times with nothing changed`);
-                void session()?.abort();
+                stopTurn(session(), `the same ${tool.name} call ${repeats} times with nothing changed`);
                 throw new Error(`Stopped: the same ${tool.name} call ${repeats} times with nothing changed.`);
             }
             if (repeats > REPEATS_ALLOWED) {
@@ -695,7 +752,7 @@ export async function createExecutor(opts) {
         modelRuntime: await subAgentModelRuntime(),
         tools: hasShell ? [...opts.tools, ENVIRONMENT_ERROR_TOOL] : opts.tools,
         // What an agent is shown never holds a credential the project passed: each value is its name (masking.ts).
-        customTools: guardRepeats(maskToolResults(withEnvironmentErrors), () => session, opts.budget),
+        customTools: guardRepeats(maskToolResults(withListArguments(withEnvironmentErrors)), () => session, opts.budget),
         sessionManager,
         resourceLoader: await subAgentResources(opts.cwd, opts.rules === true, [opts.system, opts.task].filter((part) => part !== undefined && part !== "").join("\n\n") || undefined),
     });
@@ -814,7 +871,7 @@ export async function createJudge(opts) {
         model,
         modelRuntime,
         tools: [...(opts.extraTools ?? []), "submit_verdict"],
-        customTools: [opts.verdictTool],
+        customTools: withListArguments([opts.verdictTool]),
         sessionManager: SessionManager.inMemory(opts.cwd),
         // Judges assess specs against the rules, so they have them too.
         resourceLoader: await subAgentResources(opts.cwd, true, opts.task),

package/dist/question-dialog.js

@@ -25,6 +25,7 @@ export class QuestionDialog {
         }
         else {
             this.input = undefined;
+            this.selected = Math.min(Math.max(0, options.selected ?? 0), options.options.length - 1);
         }
     }
     get focused() {

package/dist/regenerate/activity.js

@@ -90,6 +90,27 @@ export function onModelEvent(listener) {
     modelListeners.add(listener);
     return () => modelListeners.delete(listener);
 }
+/**
+ * The model provider failing or stalling, in a line for the person: the chat's notification, a headless run's console.
+ * A turn turri stopped itself is the work going its normal course - kept in the activity, the pulse's log and the run
+ * report, not said.
+ */
+export function describeTurnIncident(event) {
+    const who = event.role ?? event.label ?? "an agent";
+    const model = event.model !== undefined ? ` (${event.model})` : "";
+    if (event.kind === "provider_retry") {
+        const wait = Math.round((event.waitMs ?? 0) / 1000);
+        return `The model provider failed${model}: ${(event.error ?? "").slice(0, 160)} - trying ${who}'s request again in ${wait}s (${event.attempt} of ${event.of}).`;
+    }
+    if (event.kind === "provider_stall") {
+        return `The model provider has sent nothing back to ${who}${model} for ${formatSilence(event.waitMs ?? 0)} - still waiting.`;
+    }
+    return undefined;
+}
+function formatSilence(ms) {
+    const seconds = Math.round(ms / 1000);
+    return seconds >= 60 && seconds % 60 === 0 ? `${seconds / 60}m` : seconds >= 60 ? `${Math.floor(seconds / 60)}m ${seconds % 60}s` : `${seconds}s`;
+}
 export function reportModelEvent(event) {
     if (modelListeners.size === 0)
         return;