@mjasnikovs/pi-task 0.13.18 → 0.13.20

package/dist/task/auto-orchestrator.js

@@ -11,15 +11,20 @@ import { runSingleTask } from './orchestrator.js';
 import { parseClarifyList, deriveTitle } from './parsers.js';
 import { renderInlineMarkdown, stripInlineMarkdown } from './inline-markdown.js';
 import { AUTO_CLARIFY_PROMPT, AUTO_DECOMPOSE_PROMPT } from './auto-prompts.js';
+import { isDuplicateQuestion, MAX_DUP_STRIKES, DUP_REPROMPT_HINT } from './question-dedup.js';
 import { allocateAutoId, buildAutoBody, parseDecomposeList, parseTaskList, checkOffTask, stampTaskInProgress, findResumableAuto } from './auto-io.js';
 import { writeTaskFile, readTaskFile, updateTaskFrontMatter, taskFilePath } from './task-io.js';
 import { gitCommitAll } from './auto-commit.js';
-import { runPhaseChild, USER_CANCELLED } from './child-runner.js';
+import { runPhaseChild, prependHint, USER_CANCELLED } from './child-runner.js';
 import { SessionUI, registerBridgeCommand } from '../remote/bridge.js';
 import { pushNotify } from '../remote/push.js';
 import { getConfig } from '../config/config.js';
 import { startAutoLoader } from './widget.js';
 import { getParentContextWindow, resolveContextUsage } from './context-usage.js';
+// Hard ceiling on clarify questions per feature. The loop is open-ended (it stops
+// when the model emits NONE), but a model that never says NONE would otherwise
+// barrage the user — the real mx5 run asked 10, several of them redundant.
+const MAX_CLARIFY_QUESTIONS = 8;
 // Matches pi's @-file completion token (a path after @, until whitespace).
 const MENTION_RE = /(?:^|\s)@([^\s]+)/g;
 /**
@@ -128,17 +133,39 @@ export async function planAuto(ctx, cwd, feature, deps) {
     // the real content, not a one-line "Implement @file" that reads as trivial.
     const featureForModel = await expandFeatureMentions(cwd, feature);
     const answers = [];
-    // Open-ended: keep asking until the model emits NONE or the user dismisses.
-    for (;;) {
-        const qRaw = await deps.runChild('auto-clarify', 'read', AUTO_CLARIFY_PROMPT(featureForModel, answers.join('\n')));
+    // Plain text of every question already shown, for the duplicate backstop.
+    const askedQuestions = [];
+    // Deterministic guard against a model that ignores "never re-ask": consecutive
+    // near-duplicate questions are reprompted with an explicit hint, and once it
+    // strikes out (can't produce anything novel) we stop instead of barraging the
+    // user with the same decision worded N ways. Also caps the absolute count.
+    let dupStrikes = 0;
+    let dupHint = null;
+    // Open-ended: keep asking until the model emits NONE or the user dismisses —
+    // but never past MAX_CLARIFY_QUESTIONS distinct questions.
+    while (askedQuestions.length < MAX_CLARIFY_QUESTIONS) {
+        const qRaw = await deps.runChild('auto-clarify', 'read', prependHint(dupHint, AUTO_CLARIFY_PROMPT(featureForModel, answers.join('\n'))));
         const parsed = parseClarifyList(qRaw);
         if (parsed.length === 0)
             break; // NONE / nothing left to ask
+        // Backstop: if the model re-asked a topic already settled, don't surface it.
+        // Reprompt it to move on or finish; give up after MAX_DUP_STRIKES so a model
+        // stuck on one fork can't loop forever.
+        if (isDuplicateQuestion(askedQuestions, stripInlineMarkdown(parsed[0].question))) {
+            dupStrikes++;
+            if (dupStrikes >= MAX_DUP_STRIKES)
+                break;
+            dupHint = DUP_REPROMPT_HINT;
+            continue;
+        }
+        dupStrikes = 0;
+        dupHint = null;
         const { question, suggested, alt } = parsed[0];
         // Render markdown (bold/code) for the displayed prompt; keep plain text
         // for the editable default and the persisted file.
         const shownQ = renderInlineMarkdown(question, theme);
         const plainQ = stripInlineMarkdown(question);
+        askedQuestions.push(plainQ);
         const plainSuggested = suggested === undefined ? undefined : stripInlineMarkdown(suggested);
         const plainAlt = alt === undefined ? undefined : stripInlineMarkdown(alt);
         // Identical to /task's grill dialog: a binary fork becomes a select()

package/dist/task/orchestrator.js

@@ -26,6 +26,7 @@ import { startWidget } from './widget.js';
 import { publishViewer, publishNotify, registerBridgeCommand, getBridge } from '../remote/bridge.js';
 import { pushNotify } from '../remote/push.js';
 import { parseVerifyBlock } from './spec-validation.js';
+import { titleForDisplay } from './parsers.js';
 import { formatTimings } from './timings.js';
 import { getParentContextWindow, resolveContextUsage } from './context-usage.js';
 // ─── Module-level state ──────────────────────────────────────────────────────
@@ -132,11 +133,13 @@ export class TaskRunner {
         // Initialise or resume the TASK file.
         let id;
         let title;
+        let label;
         let resumePhase = 'refine';
         if (this._resumeId) {
             id = this._resumeId;
             const { frontMatter } = await readTaskFile(cwd, id);
             title = frontMatter.title;
+            label = frontMatter.label;
             resumePhase = frontMatter.phase;
             await updateTaskFrontMatter(cwd, id, { state: 'in_progress' });
         }
@@ -171,6 +174,7 @@ export class TaskRunner {
         // Register as active.
         this._widgetState.taskId = id;
         this._widgetState.title = title;
+        this._widgetState.label = label;
         this._widgetState.phase = resumePhase;
         this._widgetState.startedAt = this._startedAt;
         this._deps.taskId = id;
@@ -214,7 +218,7 @@ export class TaskRunner {
                 }
                 await setTaskSection(cwd, id, phase.section, out);
                 this._pc[phase.field] = out;
-                await postCommitPhase(phase, this._pc, out);
+                await postCommitPhase(phase, this._deps, this._pc, out);
             }
             // All phases done — hand off the spec.
             await advance('done');
@@ -445,7 +449,7 @@ async function handleTaskList(_args, ctx) {
         const idx = PHASE_INDEX[fm.phase];
         const phasePart = `phase ${Math.min(idx + 1, PHASE_ORDER.length)}/${PHASE_ORDER.length} ${fm.phase}`;
         const date = fm.updated_at.replace('T', ' ').slice(0, 16);
-        lines.push(`${fm.id}  ${fm.state.padEnd(12)}  ${phasePart.padEnd(24)}  ${date}  "${fm.title}"`);
+        lines.push(`${fm.id}  ${fm.state.padEnd(12)}  ${phasePart.padEnd(24)}  ${date}  "${titleForDisplay(fm)}"`);
     }
     if (lines.length === 0)
         lines.push('(no tasks in .pi-tasks/)');

package/dist/task/parsers.d.ts

@@ -35,3 +35,20 @@ export declare function parseVerifyToolingOutput(output: string): {
     }>;
 };
 export declare function deriveTitle(refined: string): string;
+/** Max characters shown for a task's display label. */
+export declare const LABEL_MAX = 72;
+/**
+ * Collapse whitespace and clamp `s` to `max` chars, cutting on a word boundary
+ * (when one falls reasonably late) and appending an ellipsis. Returns the
+ * collapsed string unchanged when it already fits.
+ */
+export declare function truncateLabel(s: string, max?: number): string;
+/**
+ * The string to show for a task: the stored short `label` when present,
+ * otherwise a truncation of the full `title`. Both paths are clamped to
+ * LABEL_MAX so a hand-edited or legacy over-long label can't blow up a line.
+ */
+export declare function titleForDisplay(task: {
+    title: string;
+    label?: string;
+}): string;

package/dist/task/parsers.js

@@ -221,3 +221,39 @@ export function deriveTitle(refined) {
     }
     return '(untitled)';
 }
+// ─── Display label ───────────────────────────────────────────────────────────
+//
+// `title` is stored in full (a refine GOAL paragraph can be 1000+ chars — see
+// the "no truncation at storage" contract in deriveTitle). These helpers shrink
+// it for status surfaces (widget head, /task list) WITHOUT touching what's
+// stored: the pipeline always reads the full title. A model-compressed `label`
+// is preferred when present (see title-label.ts); otherwise we fall back to a
+// deterministic, word-boundary truncation so a label-less task still reads
+// cleanly.
+/** Max characters shown for a task's display label. */
+export const LABEL_MAX = 72;
+/**
+ * Collapse whitespace and clamp `s` to `max` chars, cutting on a word boundary
+ * (when one falls reasonably late) and appending an ellipsis. Returns the
+ * collapsed string unchanged when it already fits.
+ */
+export function truncateLabel(s, max = LABEL_MAX) {
+    const flat = s.replace(/\s+/g, ' ').trim();
+    if (flat.length <= max)
+        return flat;
+    const slice = flat.slice(0, max - 1);
+    const lastSpace = slice.lastIndexOf(' ');
+    const cut = lastSpace > max * 0.6 ? slice.slice(0, lastSpace) : slice;
+    return cut.trimEnd() + '…';
+}
+/**
+ * The string to show for a task: the stored short `label` when present,
+ * otherwise a truncation of the full `title`. Both paths are clamped to
+ * LABEL_MAX so a hand-edited or legacy over-long label can't blow up a line.
+ */
+export function titleForDisplay(task) {
+    const label = task.label?.trim();
+    if (label)
+        return truncateLabel(label);
+    return truncateLabel(task.title);
+}

package/dist/task/phases.d.ts

@@ -52,4 +52,4 @@ export declare function phaseCompose(deps: PhaseDeps, refined: string, research:
 export declare function phaseCritique(deps: PhaseDeps, spec: string, refined: string, qa: string): Promise<string>;
 export declare function critiqueWithFallback(d: PhaseDeps, p: PhaseContext): Promise<string>;
 export declare const PHASES: PhaseConfig[];
-export declare function postCommitPhase(phase: PhaseConfig, pc: PhaseContext, out: string): Promise<void>;
+export declare function postCommitPhase(phase: PhaseConfig, deps: PhaseDeps, pc: PhaseContext, out: string): Promise<void>;

package/dist/task/phases.js

@@ -14,7 +14,9 @@ import { gatherExternalContext } from './external-context.js';
 import { REFINE_PROMPT, RESEARCH_FILES_PROMPT, RESEARCH_APIS_PROMPT, RESEARCH_CONTEXT_PROMPT, RESEARCH_TOOLING_PROMPT, GRILL_GEN_PROMPT, GRILL_AUTO_ANSWER_PROMPT, GRILL_AUTO_FORMAT_HINT, COMPOSE_PROMPT, CRITIQUE_PROMPT, CRITIQUE_TRIAGE_PROMPT, VERIFY_TOOLING_PROMPT, MAX_GRILL_QUESTIONS, appendNoThink } from './prompts.js';
 import { setTaskSection, updateTaskFrontMatter } from './task-io.js';
 import { renderInlineMarkdown, stripInlineMarkdown } from './inline-markdown.js';
+import { isDuplicateQuestion, MAX_DUP_STRIKES, DUP_REPROMPT_HINT } from './question-dedup.js';
 import { parseGrillQuestions, parseAutoAnswer, autoAnswerHasTag, parseVerifyToolingOutput, deriveTitle } from './parsers.js';
+import { compressTitle } from './title-label.js';
 import { parseVerifyBlock, validateSpecShape, stripSpecPreamble, isCritiqueClean } from './spec-validation.js';
 import { runPhaseChild, runPhaseWithLoopGuard, runWithEmphasisRetry, prependHint, USER_CANCELLED } from './child-runner.js';
 import { SessionUI } from '../remote/bridge.js';
@@ -314,15 +316,37 @@ export async function phaseGrill(deps, ctx, widgetState, refined, research) {
     const ui = new SessionUI(ctx);
     const out = []; // human-facing Q&A transcript (with auto-worker debug lines)
     const qa = []; // compact Q&A fed back into the next question
+    const askedQuestions = []; // plain text of each question, for the dup backstop
+    // Deterministic backstop against a model that ignores "never re-ask": a
+    // near-duplicate question is reprompted (not auto-answered or shown), and after
+    // MAX_DUP_STRIKES consecutive dups the loop stops. Each question otherwise runs
+    // the research-backed auto-answer, which fans out doc/fetch/search workers — so
+    // a re-ask isn't just an extra prompt, it's an expensive recompute. Capped at
+    // MAX_GRILL_QUESTIONS so a model that never emits NONE can't run unbounded.
+    let dupStrikes = 0;
+    let dupHint = null;
     // Open-ended: keep asking until the model emits NONE or the user dismisses.
-    for (let n = 0;; n++) {
+    for (let n = 0; n < MAX_GRILL_QUESTIONS; n++) {
         const tGenStart = Date.now();
-        const raw = await runPhaseWithLoopGuard(deps, 'grill-gen', 'read', hint => prependHint(hint, GRILL_GEN_PROMPT(refined, research, qa.join('\n'))));
+        const genHint = dupHint;
+        const raw = await runPhaseWithLoopGuard(deps, 'grill-gen', 'read', hint => prependHint(hint, prependHint(genHint, GRILL_GEN_PROMPT(refined, research, qa.join('\n')))));
         deps.recordSubStep?.('gen', Date.now() - tGenStart);
         const questions = parseGrillQuestions(raw);
         if (questions.length === 0)
             break; // NONE / nothing left to ask
         const q = questions[0];
+        // Suppress a re-asked decision before paying for its auto-answer.
+        if (isDuplicateQuestion(askedQuestions, stripInlineMarkdown(q))) {
+            dupStrikes++;
+            if (dupStrikes >= MAX_DUP_STRIKES)
+                break;
+            dupHint = DUP_REPROMPT_HINT;
+            n--;
+            continue;
+        }
+        dupStrikes = 0;
+        dupHint = null;
+        askedQuestions.push(stripInlineMarkdown(q));
         widgetState.lastLine = `auto-answering Q${n + 1}…`;
         const tAutoStart = Date.now();
         const auto = await phaseAutoAnswer(deps, refined, research, q);
@@ -504,10 +528,16 @@ export const PHASES = [
     },
     { name: 'critique', section: 'spec', field: 'spec', run: critiqueWithFallback }
 ];
-export async function postCommitPhase(phase, pc, out) {
+export async function postCommitPhase(phase, deps, pc, out) {
     if (phase.name !== 'refine')
         return;
     const title = deriveTitle(out);
     pc.widgetState.title = title;
+    // Compress the (often paragraph-long) title into a short display label. This
+    // is best-effort and self-falls-back to a truncation, so it never blocks the
+    // pipeline — but persist title first so a label failure can't lose the title.
     await updateTaskFrontMatter(pc.cwd, pc.id, { title });
+    const label = await compressTitle(deps, title);
+    pc.widgetState.label = label;
+    await updateTaskFrontMatter(pc.cwd, pc.id, { label });
 }

package/dist/task/prompts.d.ts

@@ -4,7 +4,7 @@
  * Each template is a pure function: inputs → prompt string.  No I/O, no side
  * effects, trivially testable.
  */
-export declare const MAX_GRILL_QUESTIONS = 10;
+export declare const MAX_GRILL_QUESTIONS = 20;
 /**
  * Qwen3 "soft switch": placing `/no_think` in the prompt disables the model's
  * <think> reasoning trace for that turn and persists across the tool-call loop
@@ -28,6 +28,13 @@ export declare const MAX_GRILL_QUESTIONS = 10;
 export declare const NO_THINK = "/no_think";
 /** Append the Qwen3 `/no_think` soft switch to a prompt. See {@link NO_THINK}. */
 export declare function appendNoThink(prompt: string): string;
+/**
+ * Compress a verbose task title into a short status-bar label. Pure judgment, no
+ * tools — the child reads only the title we hand it. The output is sanitised and
+ * length-clamped by the caller (title-label.ts), so the prompt optimises for the
+ * right *content* (identifying nouns) rather than trusting the model on length.
+ */
+export declare const COMPRESS_LABEL_PROMPT: (title: string, maxChars: number) => string;
 declare const REFINE_PROMPT: (raw: string) => string;
 declare const RESEARCH_READ_ONLY_CONSTRAINT = "IMPORTANT: You are ONLY allowed to READ. Do NOT create, modify, or delete any files. Use the read, grep, find, and ls tools to inspect the repo.";
 declare const RESEARCH_FILES_PROMPT: (refined: string) => string;

package/dist/task/prompts.js

@@ -4,7 +4,10 @@
  * Each template is a pure function: inputs → prompt string.  No I/O, no side
  * effects, trivially testable.
  */
-export const MAX_GRILL_QUESTIONS = 10;
+// Caps both the per-response question count (parseGrillQuestions/parseClarifyList)
+// and the grill loop's total iterations, so a model that never emits NONE can't
+// run unbounded.
+export const MAX_GRILL_QUESTIONS = 20;
 /**
  * Qwen3 "soft switch": placing `/no_think` in the prompt disables the model's
  * <think> reasoning trace for that turn and persists across the tool-call loop
@@ -30,6 +33,22 @@ export const NO_THINK = '/no_think';
 export function appendNoThink(prompt) {
     return `${prompt}\n\n${NO_THINK}`;
 }
+/**
+ * Compress a verbose task title into a short status-bar label. Pure judgment, no
+ * tools — the child reads only the title we hand it. The output is sanitised and
+ * length-clamped by the caller (title-label.ts), so the prompt optimises for the
+ * right *content* (identifying nouns) rather than trusting the model on length.
+ */
+export const COMPRESS_LABEL_PROMPT = (title, maxChars) => `Shorten the following task title into a brief label for a one-line status display.
+
+Rules:
+- Output ONLY the label text. No preamble, no quotes, no markdown, no trailing period.
+- One line, at most ${maxChars} characters.
+- Keep the most identifying nouns: the component, file, feature, or domain names.
+- Drop filler, constraints, verbatim file lists, and any "spec: @..." reference.
+
+TITLE:
+${title}`;
 const REFINE_PROMPT = (raw) => `You receive a user's task description for an AI coding agent. Rewrite it to be unambiguous and actionable.
 
 Output structure (four sections, exact headings, in this order):

package/dist/task/question-dedup.d.ts

@@ -0,0 +1,28 @@
+/**
+ * Detects when a freshly-generated clarifying question is a near-duplicate of one
+ * already asked. The clarify/grill loops re-call the generator after every answer
+ * and rely on a prompt instruction ("Never re-ask something already answered") to
+ * stop repeats. A weaker local model ignores that and re-asks the same fork worded
+ * differently (observed: a "Bun bundler vs Vite" question asked four ways across a
+ * single run). This is a deterministic backstop so a model that won't self-police
+ * can't barrage the user with the same decision.
+ *
+ * Pure logic, no I/O — trivially unit-testable, and validated against the actual
+ * repeated questions from a real /task-auto run (see question-dedup.test.ts).
+ */
+/**
+ * Lowercase, drop markdown emphasis/backticks, split on non-alphanumeric, then
+ * drop stopwords and 1-char tokens. `pg_trgm` → `pg`,`trgm`; `bun:sql` → `bun`,`sql`.
+ */
+export declare function tokenizeQuestion(q: string): Set<string>;
+/** Jaccard similarity of two token sets: |A∩B| / |A∪B|. 0 when either is empty. */
+export declare function jaccard(a: Set<string>, b: Set<string>): number;
+export declare const DEFAULT_DUP_THRESHOLD = 0.12;
+/**
+ * True when `next` is a near-duplicate of any question in `prior` (Jaccard of
+ * salient tokens ≥ threshold). Compares against the actual question text already
+ * asked, so a re-worded re-ask of an already-answered fork is caught.
+ */
+export declare function isDuplicateQuestion(prior: readonly string[], next: string, threshold?: number): boolean;
+export declare const MAX_DUP_STRIKES = 2;
+export declare const DUP_REPROMPT_HINT: string;

package/dist/task/question-dedup.js

@@ -0,0 +1,172 @@
+/**
+ * Detects when a freshly-generated clarifying question is a near-duplicate of one
+ * already asked. The clarify/grill loops re-call the generator after every answer
+ * and rely on a prompt instruction ("Never re-ask something already answered") to
+ * stop repeats. A weaker local model ignores that and re-asks the same fork worded
+ * differently (observed: a "Bun bundler vs Vite" question asked four ways across a
+ * single run). This is a deterministic backstop so a model that won't self-police
+ * can't barrage the user with the same decision.
+ *
+ * Pure logic, no I/O — trivially unit-testable, and validated against the actual
+ * repeated questions from a real /task-auto run (see question-dedup.test.ts).
+ */
+// Function words plus the boilerplate the prompt makes every question share
+// ("should we use", "or should", "this determines whether the task", …). Stripping
+// these keeps the comparison on the salient nouns that actually identify a topic
+// (vite, bundler, sharp, multer) instead of the scaffolding common to all of them.
+const STOPWORDS = new Set([
+    'the',
+    'a',
+    'an',
+    'and',
+    'or',
+    'of',
+    'to',
+    'in',
+    'on',
+    'for',
+    'with',
+    'as',
+    'by',
+    'at',
+    'be',
+    'is',
+    'are',
+    'should',
+    'we',
+    'use',
+    'used',
+    'using',
+    'it',
+    'its',
+    'this',
+    'that',
+    'these',
+    'those',
+    'into',
+    'from',
+    'than',
+    'then',
+    'so',
+    'if',
+    'whether',
+    'which',
+    'what',
+    'how',
+    'do',
+    'does',
+    'add',
+    'added',
+    'also',
+    'only',
+    'just',
+    'like',
+    'such',
+    'them',
+    'they',
+    'their',
+    'there',
+    'here',
+    'will',
+    'would',
+    'can',
+    'could',
+    'instead',
+    'rather',
+    'either',
+    'both',
+    'each',
+    'all',
+    'any',
+    'some',
+    'more',
+    'most',
+    'over',
+    'up',
+    'out',
+    'no',
+    'not',
+    'without',
+    'before',
+    'after',
+    'while',
+    'during',
+    'team',
+    'want',
+    'wants',
+    'approach',
+    'simpler',
+    'simple',
+    'specified',
+    'mentioned',
+    'set',
+    'sets',
+    'task',
+    'tasks',
+    'whole',
+    'entire',
+    'single',
+    'separate',
+    'new',
+    'extra',
+    'changes',
+    'change',
+    'determines',
+    'critical',
+    'because',
+    'mode',
+    'forks',
+    'fork',
+    'header',
+    'stack',
+    'lighter',
+    'built'
+]);
+/**
+ * Lowercase, drop markdown emphasis/backticks, split on non-alphanumeric, then
+ * drop stopwords and 1-char tokens. `pg_trgm` → `pg`,`trgm`; `bun:sql` → `bun`,`sql`.
+ */
+export function tokenizeQuestion(q) {
+    const cleaned = q.toLowerCase().replace(/\*\*/g, ' ').replace(/`/g, ' ');
+    const tokens = cleaned.split(/[^a-z0-9]+/).filter(t => t.length > 1 && !STOPWORDS.has(t));
+    return new Set(tokens);
+}
+/** Jaccard similarity of two token sets: |A∩B| / |A∪B|. 0 when either is empty. */
+export function jaccard(a, b) {
+    if (a.size === 0 || b.size === 0)
+        return 0;
+    let inter = 0;
+    for (const t of a)
+        if (b.has(t))
+            inter++;
+    const union = a.size + b.size - inter;
+    return union === 0 ? 0 : inter / union;
+}
+// Validated against the real mx5 run (question-dedup.test.ts): the re-asked
+// SPA-build/serve forks scored 0.152 / 0.275 / 0.421 against an earlier question,
+// while every genuinely-distinct subsystem question topped out at 0.077. 0.12 sits
+// in that gap — above the distinct ceiling, below the lowest true duplicate — so
+// it catches the repeats without suppressing a real question.
+export const DEFAULT_DUP_THRESHOLD = 0.12;
+/**
+ * True when `next` is a near-duplicate of any question in `prior` (Jaccard of
+ * salient tokens ≥ threshold). Compares against the actual question text already
+ * asked, so a re-worded re-ask of an already-answered fork is caught.
+ */
+export function isDuplicateQuestion(prior, next, threshold = DEFAULT_DUP_THRESHOLD) {
+    const nextTokens = tokenizeQuestion(next);
+    if (nextTokens.size === 0)
+        return false;
+    return prior.some(p => jaccard(tokenizeQuestion(p), nextTokens) >= threshold);
+}
+// After this many consecutive near-duplicate questions, the question loop stops:
+// a model that can't produce a novel question after being told to move on has
+// nothing genuinely new left to ask. Shared by /task-auto's clarify loop and
+// /task's grill loop.
+export const MAX_DUP_STRIKES = 2;
+// Reprompt injected when a duplicate is caught — tells the model the topic is
+// settled and to either move to a different open fork or finish.
+export const DUP_REPROMPT_HINT = 'You just re-asked a decision that is ALREADY answered above. Do not ask about '
+    + 'it again, even reworded. Ask about a DIFFERENT subsystem or fork that remains '
+    + 'genuinely open, or output the termination token if nothing decision-changing '
+    + 'is left.';

package/dist/task/task-parsers.js

@@ -12,6 +12,7 @@ const FRONT_MATTER_KEYS = [
     'created_at',
     'updated_at',
     'title',
+    'label',
     'reason'
 ];
 // ─── Front matter ────────────────────────────────────────────────────────────
@@ -20,7 +21,9 @@ export function emitFrontMatter(fm) {
     for (const k of FRONT_MATTER_KEYS) {
         const v = fm[k];
         if (v === undefined || v === '') {
-            if (k === 'reason')
+            // Optional fields are omitted entirely when empty rather than emitted
+            // as a blank `key:` line.
+            if (k === 'reason' || k === 'label')
                 continue;
         }
         lines.push(`${k}: ${typeof v === 'string' ? v : String(v)}`);
@@ -50,6 +53,7 @@ export function parseFrontMatter(content) {
         created_at: obj.created_at,
         updated_at: obj.updated_at ?? obj.created_at,
         title: obj.title ?? '',
+        label: obj.label || undefined,
         reason: obj.reason || undefined
     };
 }

package/dist/task/task-types.d.ts

@@ -13,6 +13,13 @@ export interface TaskFrontMatter {
     created_at: string;
     updated_at: string;
     title: string;
+    /**
+     * Short, human-readable display label compressed from `title`. Visual only —
+     * `title` always holds the full text the pipeline reads. Absent on pre-label
+     * tasks and on small tasks whose title is already short; display falls back to
+     * a deterministic truncation of `title` when missing. See title-label.ts.
+     */
+    label?: string;
     reason?: string;
 }
 export declare const PHASE_ORDER: PhaseName[];

package/dist/task/title-label.d.ts

@@ -0,0 +1,25 @@
+/**
+ * Display-label compression.
+ *
+ * A refined GOAL paragraph becomes the stored `title`, which the pipeline reads
+ * in full — but it can run to 1000+ characters, which makes the widget head and
+ * `/task list` unreadable. `compressTitle` spawns a tool-less local-model child
+ * to distill that title into a short, identifying label that we store ALONGSIDE
+ * (never replacing) the full title, purely for display. Every failure mode —
+ * model error, empty output, leaked tool call, abort, an over-long answer —
+ * degrades to the deterministic `truncateLabel`, so a label always comes back
+ * and the pipeline never blocks on this best-effort step.
+ */
+import type { PhaseDeps } from './child-runner.js';
+/**
+ * Reduce a raw model completion to a single clean label line: first non-empty
+ * line, stripped of surrounding quotes/backticks and a leading "label:" echo,
+ * whitespace collapsed. Returns '' when nothing usable remains.
+ */
+export declare function sanitizeLabel(raw: string): string;
+/**
+ * Compress `title` into a short display label via a local-model child. Always
+ * resolves (never throws): on any failure it returns `truncateLabel(title)`.
+ * The returned label is guaranteed to be at most LABEL_MAX characters.
+ */
+export declare function compressTitle(deps: PhaseDeps, title: string): Promise<string>;

package/dist/task/title-label.js

@@ -0,0 +1,51 @@
+/**
+ * Display-label compression.
+ *
+ * A refined GOAL paragraph becomes the stored `title`, which the pipeline reads
+ * in full — but it can run to 1000+ characters, which makes the widget head and
+ * `/task list` unreadable. `compressTitle` spawns a tool-less local-model child
+ * to distill that title into a short, identifying label that we store ALONGSIDE
+ * (never replacing) the full title, purely for display. Every failure mode —
+ * model error, empty output, leaked tool call, abort, an over-long answer —
+ * degrades to the deterministic `truncateLabel`, so a label always comes back
+ * and the pipeline never blocks on this best-effort step.
+ */
+import { runPhaseChild } from './child-runner.js';
+import { LABEL_MAX, truncateLabel } from './parsers.js';
+import { COMPRESS_LABEL_PROMPT, appendNoThink } from './prompts.js';
+/**
+ * Reduce a raw model completion to a single clean label line: first non-empty
+ * line, stripped of surrounding quotes/backticks and a leading "label:" echo,
+ * whitespace collapsed. Returns '' when nothing usable remains.
+ */
+export function sanitizeLabel(raw) {
+    const firstLine = raw.split('\n').find(l => l.trim().length > 0) ?? '';
+    return firstLine
+        .replace(/^\s*label\s*:\s*/i, '')
+        .replace(/^["'`]+|["'`]+$/g, '')
+        .replace(/\s+/g, ' ')
+        .trim();
+}
+/**
+ * Compress `title` into a short display label via a local-model child. Always
+ * resolves (never throws): on any failure it returns `truncateLabel(title)`.
+ * The returned label is guaranteed to be at most LABEL_MAX characters.
+ */
+export async function compressTitle(deps, title) {
+    const fallback = truncateLabel(title);
+    // Already short enough — no point paying for a model call.
+    if (title.replace(/\s+/g, ' ').trim().length <= LABEL_MAX)
+        return fallback;
+    try {
+        const raw = await runPhaseChild(deps, 'compress-label', '', appendNoThink(COMPRESS_LABEL_PROMPT(title, LABEL_MAX)));
+        const cleaned = sanitizeLabel(raw);
+        if (cleaned.length === 0)
+            return fallback;
+        // Clamp regardless of what the model returned; truncateLabel is a no-op
+        // when it already fits.
+        return truncateLabel(cleaned);
+    }
+    catch {
+        return fallback;
+    }
+}

package/dist/task/widget.d.ts

@@ -9,6 +9,8 @@ import { type PhaseName, type TaskState } from './task-types.js';
 export interface WidgetState {
     taskId: string;
     title: string;
+    /** Short display label compressed from `title`; falls back to a truncation when absent. */
+    label?: string;
     phase: PhaseName;
     startedAt: number;
     lastLine?: string;

package/dist/task/widget.js

@@ -5,6 +5,7 @@
  * context usage, and the latest child-process line.
  */
 import { PHASE_INDEX, PHASE_ORDER } from './task-types.js';
+import { titleForDisplay } from './parsers.js';
 import { setTaskWidget } from '../remote/session-state.js';
 // ─── Constants ───────────────────────────────────────────────────────────────
 export const WIDGET_KEY = 'pi-tasks';
@@ -71,7 +72,7 @@ function lastLineTrailer(lastLine, theme) {
 }
 export function buildWidgetLines(s, theme) {
     const elapsed = formatDuration(Date.now() - s.startedAt);
-    const head = `${s.taskId} · ${s.title}`;
+    const head = `${s.taskId} · ${titleForDisplay(s)}`;
     const idx = PHASE_INDEX[s.phase];
     const total = PHASE_ORDER.length;
     const stepNum = Math.min(idx + 1, total);

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@mjasnikovs/pi-task",
-    "version": "0.13.18",
+    "version": "0.13.20",
     "description": "Deterministic spec-orchestration for local models, with a bundled real-time remote web view and web/docs/fetch/worker subagent tools.",
     "type": "module",
     "main": "./dist/index.js",