@e2edev/e2e 0.5.1 → 0.6.1

package/dist/report/list-model.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"list-model.js","sourceRoot":"","sources":["../../src/report/list-model.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AASH,qDAAqD;AACrD,MAAM,UAAU,YAAY,CAAC,KAAgB;IAC3C,OAAO,KAAK,CAAC,IAAI,KAAK,OAAO,IAAI,KAAK,CAAC,IAAI,KAAK,QAAQ,CAAC;AAC3D,CAAC"}

package/dist/report/list-steps.d.ts

@@ -0,0 +1,27 @@
+/**
+ * Renders one agent step and its calls for the list reporter: the same
+ * finished step prints in a CI log, in a file block, and in the live window,
+ * each at its own indent, so nothing here knows where the line lands.
+ */
+import { type Colors } from './format.ts';
+import type { FinishedStep, ShownEvent } from './list-model.ts';
+/** One-line, quoted step label; `maxChars` clips it further to fit a row. */
+export declare function stepLabel(label: string, maxChars?: number): string;
+export interface StepLineOptions {
+    /** Text before the api, naming the test when the line prints away from it. */
+    readonly context?: string;
+    /** Printed width the line may take; the label is clipped so the tail stays on the row. */
+    readonly maxWidth?: number;
+}
+/**
+ * One finished agent step: outcome glyph, api, quoted label, then duration,
+ * model calls, and a non-passed status.
+ */
+export declare function stepLine(pc: Colors, step: FinishedStep, options?: StepLineOptions): string;
+/**
+ * One model turn or tool call under the current step. A tool call reads as
+ * the act it performed when the engine described it, else as the tool name
+ * (`tool:` prefixes come from executor tool accounting).
+ */
+export declare function eventLine(pc: Colors, event: ShownEvent): string;
+//# sourceMappingURL=list-steps.d.ts.map

package/dist/report/list-steps.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"list-steps.d.ts","sourceRoot":"","sources":["../../src/report/list-steps.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAIH,OAAO,EAOL,KAAK,MAAM,EACZ,MAAM,aAAa,CAAC;AACrB,OAAO,KAAK,EAAE,YAAY,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAehE,6EAA6E;AAC7E,wBAAgB,SAAS,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,SAAuB,GAAG,MAAM,CAGhF;AAED,MAAM,WAAW,eAAe;IAC9B,8EAA8E;IAC9E,QAAQ,CAAC,OAAO,CAAC,EAAE,MAAM,CAAC;IAC1B,0FAA0F;IAC1F,QAAQ,CAAC,QAAQ,CAAC,EAAE,MAAM,CAAC;CAC5B;AAED;;;GAGG;AACH,wBAAgB,QAAQ,CAAC,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE,YAAY,EAAE,OAAO,GAAE,eAAoB,GAAG,MAAM,CAY9F;AAaD;;;;GAIG;AACH,wBAAgB,SAAS,CAAC,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,UAAU,GAAG,MAAM,CAQ/D"}

package/dist/report/list-steps.js

@@ -0,0 +1,65 @@
+/**
+ * Renders one agent step and its calls for the list reporter: the same
+ * finished step prints in a CI log, in a file block, and in the live window,
+ * each at its own indent, so nothing here knows where the line lands.
+ */
+import { sanitizeText, truncateUtf8 } from '../internal/errors.js';
+import { collapseText } from '../internal/text.js';
+import { ellipsize, F_CHECK, F_CROSS, formatTime, formatTokens, visibleWidth, } from './format.js';
+/** Step labels stay one glanceable line; the report holds the full text. */
+const MAX_STEP_LABEL_CHARS = 72;
+/** A clipped step line keeps at least this much of its label, however narrow the terminal. */
+const MIN_STEP_LABEL_CHARS = 24;
+/** Columns a step line spends around its label: the quotes and the space before the tail. */
+const LABEL_CHROME = 3;
+/** A finished model turn. */
+const F_MODEL = '•';
+/** A tool call the model made. */
+const F_TOOL = '›';
+const F_INPUT = '↑';
+const F_OUTPUT = '↓';
+/** One-line, quoted step label; `maxChars` clips it further to fit a row. */
+export function stepLabel(label, maxChars = MAX_STEP_LABEL_CHARS) {
+    const max = Math.min(MAX_STEP_LABEL_CHARS, Math.max(MIN_STEP_LABEL_CHARS, maxChars));
+    return `"${ellipsize(collapseText(label), max)}"`;
+}
+/**
+ * One finished agent step: outcome glyph, api, quoted label, then duration,
+ * model calls, and a non-passed status.
+ */
+export function stepLine(pc, step, options = {}) {
+    const glyph = step.status === 'passed' ? pc.green(F_CHECK) : pc.red(F_CROSS);
+    const calls = step.modelCalls > 0 ? ` · ${step.modelCalls} model call${step.modelCalls === 1 ? '' : 's'}` : '';
+    const outcome = step.status === 'passed' ? '' : ` ${step.status}`;
+    const head = `${glyph} ${options.context ?? ''}${pc.dim(step.api)} `;
+    const tail = pc.dim(`${formatTime(step.durationMs)}${calls}${outcome}`);
+    const room = options.maxWidth === undefined
+        ? undefined
+        : options.maxWidth - visibleWidth(head) - visibleWidth(tail) - LABEL_CHROME;
+    return `${head}${stepLabel(step.label, room)} ${tail}`;
+}
+/**
+ * `(↑in ↓out)` for one model call, the total when only that is known, or
+ * nothing when the provider reported no usage.
+ */
+function tokenSplit(event) {
+    if (event.inputTokens !== undefined && event.outputTokens !== undefined) {
+        return ` (${F_INPUT}${formatTokens(event.inputTokens)} ${F_OUTPUT}${formatTokens(event.outputTokens)})`;
+    }
+    return event.count !== undefined && event.count > 0 ? ` (${formatTokens(event.count)} tokens)` : '';
+}
+/**
+ * One model turn or tool call under the current step. A tool call reads as
+ * the act it performed when the engine described it, else as the tool name
+ * (`tool:` prefixes come from executor tool accounting).
+ */
+export function eventLine(pc, event) {
+    if (event.kind === 'model') {
+        return pc.dim(`${F_MODEL} Thinking (${formatTime(event.durationMs)})${tokenSplit(event)}`);
+    }
+    const name = sanitizeText(event.name ?? 'engine').replace(/^tool:/, '');
+    const act = event.detail === undefined ? name : sanitizeText(collapseText(event.detail));
+    const failed = event.status === 'passed' ? '' : ` ${pc.red(F_CROSS)}`;
+    return `${pc.dim(`${F_TOOL} ${truncateUtf8(act, 60)} (${formatTime(event.durationMs)})`)}${failed}`;
+}
+//# sourceMappingURL=list-steps.js.map

package/dist/report/list-steps.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"list-steps.js","sourceRoot":"","sources":["../../src/report/list-steps.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,EAAE,YAAY,EAAE,YAAY,EAAE,MAAM,uBAAuB,CAAC;AACnE,OAAO,EAAE,YAAY,EAAE,MAAM,qBAAqB,CAAC;AACnD,OAAO,EACL,SAAS,EACT,OAAO,EACP,OAAO,EACP,UAAU,EACV,YAAY,EACZ,YAAY,GAEb,MAAM,aAAa,CAAC;AAGrB,4EAA4E;AAC5E,MAAM,oBAAoB,GAAG,EAAE,CAAC;AAChC,8FAA8F;AAC9F,MAAM,oBAAoB,GAAG,EAAE,CAAC;AAChC,6FAA6F;AAC7F,MAAM,YAAY,GAAG,CAAC,CAAC;AACvB,6BAA6B;AAC7B,MAAM,OAAO,GAAG,GAAG,CAAC;AACpB,kCAAkC;AAClC,MAAM,MAAM,GAAG,GAAG,CAAC;AACnB,MAAM,OAAO,GAAG,GAAG,CAAC;AACpB,MAAM,QAAQ,GAAG,GAAG,CAAC;AAErB,6EAA6E;AAC7E,MAAM,UAAU,SAAS,CAAC,KAAa,EAAE,QAAQ,GAAG,oBAAoB;IACtE,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,oBAAoB,EAAE,IAAI,CAAC,GAAG,CAAC,oBAAoB,EAAE,QAAQ,CAAC,CAAC,CAAC;IACrF,OAAO,IAAI,SAAS,CAAC,YAAY,CAAC,KAAK,CAAC,EAAE,GAAG,CAAC,GAAG,CAAC;AACpD,CAAC;AASD;;;GAGG;AACH,MAAM,UAAU,QAAQ,CAAC,EAAU,EAAE,IAAkB,EAAE,OAAO,GAAoB,EAAE;IACpF,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,EAAE,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC;IAC7E,MAAM,KAAK,GACT,IAAI,CAAC,UAAU,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,UAAU,cAAc,IAAI,CAAC,UAAU,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;IACnG,MAAM,OAAO,GAAG,IAAI,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;IAClE,MAAM,IAAI,GAAG,GAAG,KAAK,IAAI,OAAO,CAAC,OAAO,IAAI,EAAE,GAAG,EAAE,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC;IACrE,MAAM,IAAI,GAAG,EAAE,CAAC,GAAG,CAAC,GAAG,UAAU,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,KAAK,GAAG,OAAO,EAAE,CAAC,CAAC;IACxE,MAAM,IAAI,GACR,OAAO,CAAC,QAAQ,KAAK,SAAS;QAC5B,CAAC,CAAC,SAAS;QACX,CAAC,CAAC,OAAO,CAAC,QAAQ,GAAG,YAAY,CAAC,IAAI,CAAC,GAAG,YAAY,CAAC,IAAI,CAAC,GAAG,YAAY,CAAC;IAChF,OAAO,GAAG,IAAI,GAAG,SAAS,CAAC,IAAI,CAAC,KAAK,EAAE,IAAI,CAAC,IAAI,IAAI,EAAE,CAAC;AACzD,CAAC;AAED;;;GAGG;AACH,SAAS,UAAU,CAAC,KAAiB;IACnC,IAAI,KAAK,CAAC,WAAW,KAAK,SAAS,IAAI,KAAK,CAAC,YAAY,KAAK,SAAS,EAAE,CAAC;QACxE,OAAO,KAAK,OAAO,GAAG,YAAY,CAAC,KAAK,CAAC,WAAW,CAAC,IAAI,QAAQ,GAAG,YAAY,CAAC,KAAK,CAAC,YAAY,CAAC,GAAG,CAAC;IAC1G,CAAC;IACD,OAAO,KAAK,CAAC,KAAK,KAAK,SAAS,IAAI,KAAK,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,YAAY,CAAC,KAAK,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC,CAAC,EAAE,CAAC;AACtG,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,SAAS,CAAC,EAAU,EAAE,KAAiB;IACrD,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;QAC3B,OAAO,EAAE,CAAC,GAAG,CAAC,GAAG,OAAO,cAAc,UAAU,CAAC,KAAK,CAAC,UAAU,CAAC,IAAI,UAAU,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;IAC7F,CAAC;IACD,MAAM,IAAI,GAAG,YAAY,CAAC,KAAK,CAAC,IAAI,IAAI,QAAQ,CAAC,CAAC,OAAO,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC;IACxE,MAAM,GAAG,GAAG,KAAK,CAAC,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,YAAY,CAAC,YAAY,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC;IACzF,MAAM,MAAM,GAAG,KAAK,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC;IACtE,OAAO,GAAG,EAAE,CAAC,GAAG,CAAC,GAAG,MAAM,IAAI,YAAY,CAAC,GAAG,EAAE,EAAE,CAAC,KAAK,UAAU,CAAC,KAAK,CAAC,UAAU,CAAC,GAAG,CAAC,GAAG,MAAM,EAAE,CAAC;AACtG,CAAC"}

package/dist/report/list.d.ts

@@ -1,8 +1,9 @@
 /**
- * Human-readable list reporter (spec 06-cli.md), styled after vitest's
- * default reporter: one block per test file and target, a `Failed Tests`
- * section with code frames, and a padded summary. On a TTY a live window
- * below the log shows the running files, their tests, and the counters.
+ * Human-readable list reporter, styled after vitest's default reporter: one
+ * block per test file and target with each test's agent steps nested, a
+ * `Failed Tests` section with code frames, and a padded summary. On a TTY a
+ * live window below the log shows the running tree (`RunningTree`) and the
+ * counters.
  */
 import type { RunEventFact } from '../run/events.ts';
 export interface ListReporterOutput {
@@ -25,6 +26,7 @@ export declare class ListReporter {
     private readonly output;
     private readonly pc;
     private readonly window;
+    private readonly tree;
     private readonly separator;
     private projectRoot;
     /** Selected targets in declaration order; decides each badge's color. */
@@ -33,8 +35,10 @@ export declare class ListReporter {
     private startedAt;
     /** File groups keyed `${target}\u0000${file}`, in first-seen order. */
     private readonly groups;
-    /** Pairs executing right now, keyed by `pairKey`, in start order. */
-    private readonly running;
+    /** Pairs that started and whose result is still to come, keyed by `pairKey`, in start order. */
+    private readonly pairs;
+    /** Whether a live window paints; without one, finished steps stream permanently. */
+    private readonly live;
     private readonly failures;
     /**
      * Serial groups whose members' results are still to come, by group id. The
@@ -66,18 +70,19 @@ export declare class ListReporter {
     /**
      * A worker began one test-target pair: show it in the live window. A serial
      * member replaces the previous member of its group, which has finished
-     * executing even though its result only arrives with the whole group.
+     * executing even though its result only arrives with the whole group. A
+     * serial group retries as a whole, so a member starts once per group
+     * attempt; its steps from earlier attempts stay with it, as an ordinary
+     * test's do.
      */
     private testStarted;
     /**
-     * Streams step progress of a running pair. With a single test running, its
-     * agent steps print permanently and chronologically under a header naming
-     * the file and test - the scrollback of a long agentic run reads without
-     * `--debug`. Deterministic steps are fast and many, so they only ever show
-     * in the live window. With parallel tests the stream would interleave, so
-     * each pair instead shows its current step and latest calls transiently in
-     * the live window, and finished agent steps print with the test they
-     * belong to.
+     * Tracks step progress of a running pair. With a live window the pair's
+     * finished agent steps accumulate under its row there and print once, nested
+     * under its line, when the file block prints - so every name appears exactly
+     * once in the scrollback. Without a window (CI logs) nothing is transient, so
+     * each finished agent step prints at once, prefixed with its test. Deterministic
+     * steps are fast and many, so they only ever show in the live window.
      */
     private step;
     /** Holds a serial group until each of its members' results has read its details. */
@@ -95,8 +100,11 @@ export declare class ListReporter {
     private fileCounters;
     /**
      * Prints one file's block: the file line with its counts, then - when the
-     * file failed, had a flaky pass, streamed its steps, or is the run's only
-     * file - one line per test, as vitest lists tests for a failed module.
+     * file failed, had a flaky pass, or is the run's only file - one line per
+     * test with its finished agent steps nested, as vitest lists tests for a
+     * failed module. With a live window the block is also the only permanent
+     * record of a file's agent steps, so a file that has any lists its tests
+     * too; without one the steps already printed as they finished.
      */
     private printGroup;
     /** One test row under its file line, plus the first error line for a failure. */
@@ -106,13 +114,7 @@ export declare class ListReporter {
      * files, tests, model usage, run errors, start time, and elapsed time.
      */
     private summaryRows;
-    /**
-     * The live window: running files and tests, then the summary. The block
-     * must fit the screen, because repainting more rows than the terminal has
-     * breaks the cursor-up erase: rows left after the summary go to the running
-     * tests, each test's calls share what remains once every test has its own
-     * line, and tests that still do not fit fold into one `more running` marker.
-     */
+    /** The live window: the running tree, then the summary. */
     private renderWindow;
     /** vitest's `Failed Tests` section: a banner, then each failure with its code frame. */
     private printFailures;

package/dist/report/list.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"list.d.ts","sourceRoot":"","sources":["../../src/report/list.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAOH,OAAO,KAAK,EAAE,YAAY,EAA8B,MAAM,kBAAkB,CAAC;AA0DjF,MAAM,WAAW,kBAAkB;IACjC,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IAC1B,uEAAuE;IACvE,GAAG,CAAC,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;CAC1B;AAED,MAAM,WAAW,mBAAmB;IAClC,kEAAkE;IAClE,IAAI,CAAC,EAAE,OAAO,CAAC;IACf,kFAAkF;IAClF,MAAM,CAAC,EAAE,OAAO,CAAC;CAClB;AA+HD;;;;GAIG;AACH,qBAAa,YAAY;IA0BrB,OAAO,CAAC,QAAQ,CAAC,MAAM;IAzBzB,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAS;IAC5B,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAa;IACpC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAS;IACnC,OAAO,CAAC,WAAW,CAAqB;IACxC,yEAAyE;IACzE,OAAO,CAAC,OAAO,CAAyB;IACxC,iFAAiF;IACjF,OAAO,CAAC,SAAS,CAAc;IAC/B,uEAAuE;IACvE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAgC;IACvD,qEAAqE;IACrE,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAkC;IAC1D,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAiB;IAC1C;;;;OAIG;IACH,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAsE;IACpG,yFAAyF;IACzF,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAyB;IAChD,gFAAgF;IAChF,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAgB;IAEzC,YACmB,MAAM,GAAE,kBAAmC,EAC5D,OAAO,GAAE,mBAAwB,EAMlC;IAED;;;OAGG;IACH,QAAQ,CAAC,MAAM,UAAW,YAAY,KAAG,IAAI,CAiC3C;IAEF,oEAAoE;IACpE,OAAO,CAAC,KAAK;IAMb,mFAAmF;IACnF,OAAO,CAAC,KAAK;IAQb,yEAAyE;IACzE,OAAO,CAAC,WAAW;IAMnB,OAAO,CAAC,KAAK;IAUb,OAAO,CAAC,UAAU;IAiBlB,OAAO,CAAC,IAAI;IAOZ,kFAAkF;IAClF,OAAO,CAAC,cAAc;IAUtB;;;;OAIG;IACH,OAAO,CAAC,WAAW;IAoBnB;;;;;;;;;OASG;IACH,OAAO,CAAC,IAAI;IAkDZ,oFAAoF;IACpF,OAAO,CAAC,WAAW;IAInB;;;;OAIG;IACH,OAAO,CAAC,SAAS;IASjB,OAAO,CAAC,YAAY;IAyBpB,wFAAwF;IACxF,OAAO,CAAC,YAAY;IAOpB,uFAAuF;IACvF,OAAO,CAAC,YAAY;IASpB;;;;OAIG;IACH,OAAO,CAAC,UAAU;IA0ClB,iFAAiF;IACjF,OAAO,CAAC,QAAQ;IA+BhB;;;OAGG;IACH,OAAO,CAAC,WAAW;IAmBnB;;;;;;OAMG;IACH,OAAO,CAAC,YAAY;IAsDpB,wFAAwF;IACxF,OAAO,CAAC,aAAa;IA0BrB,gEAAgE;IAChE,OAAO,CAAC,WAAW;IAOnB,wEAAwE;IACxE,OAAO,CAAC,oBAAoB;IAU5B;;;;OAIG;IACH,OAAO,CAAC,WAAW;IAgBnB,OAAO,CAAC,WAAW;CAyBpB"}
+{"version":3,"file":"list.d.ts","sourceRoot":"","sources":["../../src/report/list.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAMH,OAAO,KAAK,EAAE,YAAY,EAA8B,MAAM,kBAAkB,CAAC;AAoDjF,MAAM,WAAW,kBAAkB;IACjC,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IAC1B,uEAAuE;IACvE,GAAG,CAAC,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;CAC1B;AAED,MAAM,WAAW,mBAAmB;IAClC,kEAAkE;IAClE,IAAI,CAAC,EAAE,OAAO,CAAC;IACf,kFAAkF;IAClF,MAAM,CAAC,EAAE,OAAO,CAAC;CAClB;AA0DD;;;;GAIG;AACH,qBAAa,YAAY;IA6BrB,OAAO,CAAC,QAAQ,CAAC,MAAM;IA5BzB,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAS;IAC5B,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAa;IACpC,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAc;IACnC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAS;IACnC,OAAO,CAAC,WAAW,CAAqB;IACxC,yEAAyE;IACzE,OAAO,CAAC,OAAO,CAAyB;IACxC,iFAAiF;IACjF,OAAO,CAAC,SAAS,CAAc;IAC/B,uEAAuE;IACvE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAgC;IACvD,gGAAgG;IAChG,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAkC;IACxD,oFAAoF;IACpF,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAU;IAC/B,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAiB;IAC1C;;;;OAIG;IACH,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAsE;IACpG,yFAAyF;IACzF,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAyB;IAChD,gFAAgF;IAChF,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAgB;IAEzC,YACmB,MAAM,GAAE,kBAAmC,EAC5D,OAAO,GAAE,mBAAwB,EAQlC;IAED;;;OAGG;IACH,QAAQ,CAAC,MAAM,UAAW,YAAY,KAAG,IAAI,CAiC3C;IAEF,oEAAoE;IACpE,OAAO,CAAC,KAAK;IAMb,mFAAmF;IACnF,OAAO,CAAC,KAAK;IAQb,yEAAyE;IACzE,OAAO,CAAC,WAAW;IAMnB,OAAO,CAAC,KAAK;IAUb,OAAO,CAAC,UAAU;IAoBlB,OAAO,CAAC,IAAI;IAOZ,kFAAkF;IAClF,OAAO,CAAC,cAAc;IAUtB;;;;;;;OAOG;IACH,OAAO,CAAC,WAAW;IAmBnB;;;;;;;OAOG;IACH,OAAO,CAAC,IAAI;IA2CZ,oFAAoF;IACpF,OAAO,CAAC,WAAW;IAInB;;;;OAIG;IACH,OAAO,CAAC,SAAS;IASjB,OAAO,CAAC,YAAY;IA6BpB,wFAAwF;IACxF,OAAO,CAAC,YAAY;IAOpB,uFAAuF;IACvF,OAAO,CAAC,YAAY;IASpB;;;;;;;OAOG;IACH,OAAO,CAAC,UAAU;IA2ClB,iFAAiF;IACjF,OAAO,CAAC,QAAQ;IA+BhB;;;OAGG;IACH,OAAO,CAAC,WAAW;IAmBnB,2DAA2D;IAC3D,OAAO,CAAC,YAAY;IAKpB,wFAAwF;IACxF,OAAO,CAAC,aAAa;IA0BrB,gEAAgE;IAChE,OAAO,CAAC,WAAW;IAOnB,wEAAwE;IACxE,OAAO,CAAC,oBAAoB;IAU5B;;;;OAIG;IACH,OAAO,CAAC,WAAW;IAgBnB,OAAO,CAAC,WAAW;CAyBpB"}

package/dist/report/list.js

@@ -1,63 +1,35 @@
 /**
- * Human-readable list reporter (spec 06-cli.md), styled after vitest's
- * default reporter: one block per test file and target, a `Failed Tests`
- * section with code frames, and a padded summary. On a TTY a live window
- * below the log shows the running files, their tests, and the counters.
+ * Human-readable list reporter, styled after vitest's default reporter: one
+ * block per test file and target with each test's agent steps nested, a
+ * `Failed Tests` section with code frames, and a padded summary. On a TTY a
+ * live window below the log shows the running tree (`RunningTree`) and the
+ * counters.
  */
 import path from 'node:path';
 import picocolors from 'picocolors';
-import { sanitizeText, truncateUtf8 } from '../internal/errors.js';
 import { packageVersion } from '../internal/package-version.js';
-import { collapseText } from '../internal/text.js';
 import { codeFrame, userFrame } from './code-frame.js';
-import { addUsage, aiSegment, bounded, ellipsize, emptyCounters, emptyUsage, fileOutcome, formatClock, formatTime, formatTokens, padTitle, rule, stateString, statusBucket, stepsUsage, sumUsage, tally, terminalColumns, terminalRows, } from './format.js';
+import { addUsage, aiSegment, bounded, ellipsize, emptyCounters, emptyUsage, F_CHECK, F_CROSS, F_POINTER, fileOutcome, formatClock, formatTime, padTitle, rule, stateString, statusBucket, stepsUsage, sumUsage, tally, terminalColumns, usageText, } from './format.js';
+import { isShownEvent } from './list-model.js';
+import { stepLine } from './list-steps.js';
 import { LiveWindow } from './live-window.js';
-/** Step labels stay one glanceable line; the report holds the full text. */
-const MAX_STEP_LABEL_BYTES = 72;
-const F_POINTER = '❯';
-const F_CHECK = '✓';
-const F_CROSS = '×';
+import { RunningTree } from './running-tree.js';
 const F_DOWN = '↓';
 const F_RIGHT = '→';
-const F_DOWN_RIGHT = '↳';
-const F_TREE_MIDDLE = '├──';
-const F_TREE_END = '└──';
 /** Indentation under a badge line, matching vitest's banner padding. */
 const BADGE_PADDING = '      ';
-/** Badge backgrounds, assigned to targets in declaration order. */
-const BADGE_COLORS = ['bgYellow', 'bgCyan', 'bgGreen', 'bgMagenta'];
+/** Indentation of a test line under its file line. */
+const TEST_INDENT = '   ';
+/** Indentation of a finished agent step under its test line in a file block. */
+const STEP_INDENT = '     ';
 /**
- * Rows the live window spends outside the running tests and the summary: its
- * blank lines, the `… more running` marker, and a margin above the prompt.
+ * Badge backgrounds, assigned to targets in declaration order. Bright variants
+ * because GitHub Actions renders plain yellow as dark brown, which swallows the
+ * black label.
  */
-const WINDOW_CHROME_ROWS = 6;
-/** Rows one running test takes before its calls: the test line and its current step. */
-const WINDOW_ROWS_PER_TEST = 2;
+const BADGE_COLORS = ['bgYellowBright', 'bgCyanBright', 'bgGreenBright', 'bgMagentaBright'];
 /** Indentation of the `→ first error line` under a failed test in a file block. */
 const ERROR_GLANCE_INDENT = 7;
-/** One-line, quoted step label bounded for the live view. */
-function stepLabel(label) {
-    const flat = collapseText(label);
-    const shown = truncateUtf8(flat, MAX_STEP_LABEL_BYTES);
-    return `"${shown}${shown === flat ? '' : '…'}"`;
-}
-/**
- * Live tail for one step event. Model and engine calls are the ones worth a
- * glance (`tool:` prefixes come from executor tool accounting); polls and
- * policy decisions stay quiet.
- */
-function eventTail(event) {
-    if (event.kind === 'model') {
-        const tokens = event.count !== undefined && event.count > 0 ? ` · ${formatTokens(event.count)} tokens` : '';
-        return `model turn ${formatTime(event.durationMs)}${tokens}`;
-    }
-    if (event.kind === 'engine') {
-        const name = sanitizeText(event.name ?? 'engine').replace(/^tool:/, '');
-        const failed = event.status === 'passed' ? '' : ` ${F_CROSS}`;
-        return `${truncateUtf8(name, 40)} ${formatTime(event.durationMs)}${failed}`;
-    }
-    return undefined;
-}
 /** The identity of one test-target pair across `test-started`, `step`, and `test-finished`. */
 function pairKey(testId, target) {
     return `${testId}@${target}`;
@@ -104,6 +76,7 @@ export class ListReporter {
     output;
     pc;
     window;
+    tree;
     separator;
     projectRoot;
     /** Selected targets in declaration order; decides each badge's color. */
@@ -112,8 +85,10 @@ export class ListReporter {
     startedAt = new Date();
     /** File groups keyed `${target}\u0000${file}`, in first-seen order. */
     groups = new Map();
-    /** Pairs executing right now, keyed by `pairKey`, in start order. */
-    running = new Map();
+    /** Pairs that started and whose result is still to come, keyed by `pairKey`, in start order. */
+    pairs = new Map();
+    /** Whether a live window paints; without one, finished steps stream permanently. */
+    live;
     failures = [];
     /**
      * Serial groups whose members' results are still to come, by group id. The
@@ -128,8 +103,10 @@ export class ListReporter {
     constructor(output = DEFAULT_OUTPUT, options = {}) {
         this.output = output;
         const live = (options.live ?? process.stdout.isTTY === true) && output.raw !== undefined;
+        this.live = live;
         this.pc = picocolors.createColors(options.colors ?? (live || picocolors.isColorSupported));
         this.separator = this.pc.dim(' > ');
+        this.tree = new RunningTree(this.pc, (target) => this.badge(target));
         this.window = new LiveWindow(live ? output.raw?.bind(output) : undefined, () => this.renderWindow());
     }
     /**
@@ -194,7 +171,7 @@ export class ListReporter {
         const key = `${target}\u0000${file}`;
         let group = this.groups.get(key);
         if (group === undefined) {
-            group = { file, target, planned: undefined, lines: [], streamed: false, printed: false };
+            group = { file, target, planned: undefined, lines: [], printed: false };
             this.groups.set(key, group);
         }
         return group;
@@ -211,6 +188,9 @@ export class ListReporter {
         if (event.ci)
             details.push('CI');
         this.output.write(BADGE_PADDING + pc.dim(details.join(' · ')));
+        if (event.model !== undefined) {
+            this.output.write(BADGE_PADDING + pc.dim(`model ${bounded(event.model)}`));
+        }
         this.output.write('');
         this.window.start();
     }
@@ -229,79 +209,80 @@ export class ListReporter {
     /**
      * A worker began one test-target pair: show it in the live window. A serial
      * member replaces the previous member of its group, which has finished
-     * executing even though its result only arrives with the whole group.
+     * executing even though its result only arrives with the whole group. A
+     * serial group retries as a whole, so a member starts once per group
+     * attempt; its steps from earlier attempts stay with it, as an ordinary
+     * test's do.
      */
     testStarted(event) {
         if (event.serialId !== undefined) {
-            for (const [key, test] of this.running) {
-                if (test.serialId === event.serialId && test.group.target === event.target) {
-                    this.running.delete(key);
-                }
+            for (const test of this.pairs.values()) {
+                if (test.serialId === event.serialId && test.group.target === event.target)
+                    test.executing = false;
             }
         }
-        this.running.set(pairKey(event.testId, event.target), {
+        const key = pairKey(event.testId, event.target);
+        this.pairs.set(key, {
             group: this.group(event.file, event.target),
             serialId: event.serialId,
             title: bounded(event.title),
             startedMs: Date.now(),
-            step: undefined,
-            events: [],
-            streaming: false,
+            executing: true,
+            current: undefined,
+            steps: this.pairs.get(key)?.steps ?? [],
         });
         this.window.redraw();
     }
     /**
-     * Streams step progress of a running pair. With a single test running, its
-     * agent steps print permanently and chronologically under a header naming
-     * the file and test - the scrollback of a long agentic run reads without
-     * `--debug`. Deterministic steps are fast and many, so they only ever show
-     * in the live window. With parallel tests the stream would interleave, so
-     * each pair instead shows its current step and latest calls transiently in
-     * the live window, and finished agent steps print with the test they
-     * belong to.
+     * Tracks step progress of a running pair. With a live window the pair's
+     * finished agent steps accumulate under its row there and print once, nested
+     * under its line, when the file block prints - so every name appears exactly
+     * once in the scrollback. Without a window (CI logs) nothing is transient, so
+     * each finished agent step prints at once, prefixed with its test. Deterministic
+     * steps are fast and many, so they only ever show in the live window.
      */
     step(event) {
-        const { pc } = this;
-        const running = this.running.get(pairKey(event.testId, event.target));
+        const running = this.pairs.get(pairKey(event.testId, event.target));
         if (running === undefined)
             return;
         const { progress } = event;
         switch (progress.phase) {
             case 'start': {
-                running.step = `${progress.api} ${stepLabel(progress.label)}`;
-                running.events = [];
-                if (progress.kind === 'agent' && this.running.size === 1 && !running.streaming) {
-                    running.streaming = true;
-                    const { group } = running;
-                    group.streamed = true;
-                    this.print(` ${pc.yellow(F_POINTER)} ${this.badge(group.target)} ${pc.dim(bounded(group.file))}${this.separator}${running.title}`);
-                }
+                const { api, label, kind } = progress;
+                running.current = { api, label, kind, events: [], turnStart: 0 };
                 this.window.redraw();
                 break;
             }
             case 'event': {
-                const tail = eventTail(progress.event);
-                if (running.step === undefined || tail === undefined)
+                const { current } = running;
+                if (current === undefined || !isShownEvent(progress.event))
                     break;
-                running.events.push(tail);
+                if (progress.event.kind === 'model') {
+                    current.events.splice(current.turnStart, 0, progress.event);
+                    current.turnStart = current.events.length;
+                }
+                else {
+                    current.events.push(progress.event);
+                }
                 this.window.redraw();
                 break;
             }
             case 'end': {
-                running.step = undefined;
-                running.events = [];
+                running.current = undefined;
                 if (progress.kind !== 'agent') {
                     this.window.redraw();
                     break;
                 }
-                const glyph = progress.status === 'passed' ? pc.green(F_CHECK) : pc.red(F_CROSS);
-                const calls = progress.modelCalls > 0
-                    ? ` · ${progress.modelCalls} model call${progress.modelCalls === 1 ? '' : 's'}`
-                    : '';
-                const outcome = progress.status === 'passed' ? '' : ` ${progress.status}`;
-                const context = running.streaming ? '' : `${pc.dim(running.title)}${this.separator}`;
-                this.print(`   ${glyph} ${context}${pc.dim(progress.api)} ${stepLabel(progress.label)} ` +
-                    pc.dim(`${formatTime(progress.durationMs)}${calls}${outcome}`));
+                const { api, label, status, durationMs, modelCalls } = progress;
+                const step = { api, label, status, durationMs, modelCalls };
+                if (this.live) {
+                    running.steps.push(step);
+                    this.window.redraw();
+                }
+                else {
+                    const context = `${this.pc.dim(running.title)}${this.separator}`;
+                    this.print(`${TEST_INDENT}${stepLine(this.pc, step, { context })}`);
+                }
                 break;
             }
         }
@@ -331,12 +312,14 @@ export class ListReporter {
         // counted them.
         if (!result.selected)
             return;
-        this.running.delete(pairKey(result.test.id, result.target.name));
+        const key = pairKey(result.test.id, result.target.name);
+        const steps = this.pairs.get(key)?.steps ?? [];
+        this.pairs.delete(key);
         const group = this.group(result.test.file, result.target.name);
         const { durationMs, usage, error } = this.detailsOf(result);
         addUsage(this.runUsage, usage);
         const title = bounded(result.test.titlePath.join(' > '));
-        group.lines.push({
+        const line = {
             title,
             declarationIndex: result.test.declarationIndex,
             status: result.status,
@@ -344,7 +327,9 @@ export class ListReporter {
             usage,
             firstErrorLine: error === undefined ? undefined : bounded(error.message).split('\n')[0],
             skipReason: result.skip === undefined ? undefined : bounded(result.skip.reason),
-        });
+            steps,
+        };
+        group.lines.push(line);
         if (statusBucket(result.status) === 'failed') {
             this.failures.push({ group, title, status: result.status, error });
         }
@@ -372,8 +357,11 @@ export class ListReporter {
     }
     /**
      * Prints one file's block: the file line with its counts, then - when the
-     * file failed, had a flaky pass, streamed its steps, or is the run's only
-     * file - one line per test, as vitest lists tests for a failed module.
+     * file failed, had a flaky pass, or is the run's only file - one line per
+     * test with its finished agent steps nested, as vitest lists tests for a
+     * failed module. With a live window the block is also the only permanent
+     * record of a file's agent steps, so a file that has any lists its tests
+     * too; without one the steps already printed as they finished.
      */
     printGroup(group) {
         if (group.printed)
@@ -409,14 +397,16 @@ export class ListReporter {
         this.print(parts.join(' '));
         const verbose = counts.failed > 0 ||
             counts.flaky > 0 ||
-            group.streamed ||
-            this.groups.size === 1;
+            this.groups.size === 1 ||
+            (this.live && group.lines.some((line) => line.steps.length > 0));
         if (!verbose)
             return;
         const ordered = group.lines.toSorted((a, b) => a.declarationIndex - b.declarationIndex);
         for (const line of ordered) {
             for (const rendered of this.testLine(line))
                 this.print(rendered);
+            for (const step of line.steps)
+                this.print(`${STEP_INDENT}${stepLine(pc, step)}`);
         }
     }
     /** One test row under its file line, plus the first error line for a failure. */
@@ -461,7 +451,7 @@ export class ListReporter {
             padTitle(pc, 'Test Files') + (files.total === 0 ? pc.dim('no test files') : stateString(pc, files)),
             padTitle(pc, 'Tests') + (tests.total === 0 ? pc.dim('no tests executed') : stateString(pc, tests)),
         ];
-        const ai = aiSegment(this.runUsage);
+        const ai = usageText(this.runUsage);
         if (ai !== undefined)
             rows.push(padTitle(pc, 'AI') + `${ai} · ${this.runUsage.calls} model calls`);
         if (this.errors.length > 0) {
@@ -472,68 +462,10 @@ export class ListReporter {
         rows.push(padTitle(pc, 'Duration') + formatTime(Date.now() - this.startedAt.getTime()));
         return rows;
     }
-    /**
-     * The live window: running files and tests, then the summary. The block
-     * must fit the screen, because repainting more rows than the terminal has
-     * breaks the cursor-up erase: rows left after the summary go to the running
-     * tests, each test's calls share what remains once every test has its own
-     * line, and tests that still do not fit fold into one `more running` marker.
-     */
+    /** The live window: the running tree, then the summary. */
     renderWindow() {
-        const { pc } = this;
-        const now = Date.now();
-        const active = new Map();
-        for (const test of this.running.values()) {
-            const tests = active.get(test.group);
-            if (tests === undefined)
-                active.set(test.group, [test]);
-            else
-                tests.push(test);
-        }
-        const summary = this.summaryRows();
-        let budget = terminalRows() - summary.length - WINDOW_CHROME_ROWS;
-        const spare = budget - active.size - this.running.size * WINDOW_ROWS_PER_TEST;
-        const maxEvents = Math.max(0, Math.floor(spare / Math.max(1, this.running.size)));
-        const lines = [''];
-        let hidden = 0;
-        for (const [group, tests] of active) {
-            if (budget < 2) {
-                hidden += tests.length;
-                continue;
-            }
-            const progress = pc.dim(` ${group.lines.length}/${group.planned ?? '?'}`);
-            lines.push(`${pc.bold(pc.yellow(` ${F_POINTER} `))}${this.badge(group.target)} ${bounded(group.file)}${progress}`);
-            budget -= 1;
-            tests.forEach((test, index) => {
-                if (budget < 1) {
-                    hidden += 1;
-                    return;
-                }
-                const glyph = index === tests.length - 1 ? F_TREE_END : F_TREE_MIDDLE;
-                const elapsed = pc.bold(pc.yellow(formatTime(Math.max(0, now - test.startedMs))));
-                lines.push(`${pc.bold(pc.yellow(`   ${glyph} `))}${test.title} ${elapsed}`);
-                budget -= 1;
-                if (test.step === undefined)
-                    return;
-                const detail = [`       ${pc.dim(`${F_DOWN_RIGHT} ${test.step}`)}`];
-                const overflow = test.events.length - maxEvents;
-                if (overflow > 0) {
-                    detail.push(`         ${pc.dim(`… ${overflow} earlier call${overflow === 1 ? '' : 's'}`)}`);
-                }
-                for (const tail of overflow > 0 ? test.events.slice(overflow) : test.events) {
-                    detail.push(`         ${pc.dim(tail)}`);
-                }
-                const shown = detail.slice(0, budget);
-                lines.push(...shown);
-                budget -= shown.length;
-            });
-        }
-        if (hidden > 0)
-            lines.push(pc.dim(`   … ${hidden} more running`));
-        if (active.size > 0)
-            lines.push('');
-        lines.push(...summary, '');
-        return lines;
+        const running = [...this.pairs.values()].filter((test) => test.executing);
+        return this.tree.render(running, this.summaryRows(), Date.now());
     }
     /** vitest's `Failed Tests` section: a banner, then each failure with its code frame. */
     printFailures() {