@bastani/atomic 0.9.14-alpha.4 → 0.9.14-alpha.5

package/CHANGELOG.md

@@ -2,6 +2,19 @@
 
 ## [Unreleased]
 
+## [0.9.14-alpha.5] - 2026-08-19
+
+### Added
+
+- Added HumanLayer's bundled `show-me` skill for visual explanations, diagrams, code-shape sketches, and focused HTML artifacts. Sourced from https://github.com/humanlayer/skills and distributed under the MIT License.
+
+### Fixed
+
+- Interactive startup now exits quietly when Ctrl+C stops the engine during first-paint binding, without leaking expected in-flight RPC transport failures.
+- Bare `/` and slash-command drafts typed while explicitly loaded extension and workflow packages finish loading now remain in the interactive editor until Enter is pressed.
+- Queue-control RPC frames now remain reachable during long-running prompts. When all overlapping remote `clear_queue` calls fail, Atomic restores the pre-clear host queue only if no later `queue_update` has supplied engine truth, including an empty queue ([#2516](https://github.com/bastani-inc/atomic/issues/2516)).
+- The `ask_user_question` dialog no longer draws one preview row at the wrong horizontal offset. In side-by-side layout the preview box border landed 11-13 columns early on the row carrying the `❯` cursor, cutting into the option label, and the damage followed the cursor between options. The active-row mark that caused it now terminates its APC string with ST as ECMA-48 requires instead of BEL, and both interactive renderers strip the mark centrally before painting, so it no longer reaches the terminal from any host — including workflow stage chats and other custom-UI mounts that do not go through a reserving bottom overlay.
+
 ## [0.9.14-alpha.4] - 2026-08-18
 
 ### Changed

package/dist/builtin/intercom/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/intercom",
-  "version": "0.9.14-alpha.4",
+  "version": "0.9.14-alpha.5",
   "private": true,
   "description": "Atomic extension providing a private coordination channel between parent and child agent sessions. Fork of: https://github.com/nicobailon/pi-intercom",
   "contributors": [

package/dist/builtin/mcp/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/mcp",
-  "version": "0.9.14-alpha.4",
+  "version": "0.9.14-alpha.5",
   "private": true,
   "description": "Atomic extension that adapts MCP (Model Context Protocol) servers into the coding agent. Fork of: https://github.com/nicobailon/pi-mcp-adapter",
   "contributors": [

package/dist/builtin/subagents/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [0.9.14-alpha.5] - 2026-08-19
+
+### Added
+
+- Added pure progress-trend evidence and injected control score-series support that can raise wall-clock attention priority without creating failure states, suppressing signals, or scheduling model calls ([#2489](https://github.com/bastani-inc/atomic/issues/2489)).
+
 ## [0.9.14-alpha.3] - 2026-08-17
 
 ### Changed

package/dist/builtin/subagents/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/subagents",
-  "version": "0.9.14-alpha.4",
+  "version": "0.9.14-alpha.5",
   "private": true,
   "description": "Atomic extension for delegating tasks to subagents with parallel execution. Fork of: https://github.com/nicobailon/pi-subagents",
   "contributors": [

package/dist/builtin/subagents/src/extension/schemas.ts

@@ -93,6 +93,11 @@ const ControlOverrides = Type.Object({
 			description: "Consecutive mutating-tool failures before escalating to needs_attention (default: 3)",
 		}),
 	),
+	progressScores: Type.Optional(
+		Type.Array(Type.Number(), {
+			description: "Optional injected progress scores used only to raise attention priority; never a run outcome.",
+		}),
+	),
 	notifyOn: Type.Optional(
 		Type.Array(Type.String({ enum: ["active_long_running", "needs_attention"] }), {
 			description:

package/dist/builtin/subagents/src/runs/shared/long-running-guard.ts

@@ -1,4 +1,5 @@
 import type { ResolvedControlConfig } from "../../shared/types.js";
+import { progressAwareThreshold } from "./progress-trend.js";
 
 interface LongRunningNoticeMetrics {
 	startedAt: number;
@@ -121,7 +122,8 @@ export function nextLongRunningTrigger(
 	config: ResolvedControlConfig,
 	metrics: LongRunningNoticeMetrics,
 ): LongRunningTriggerReason | undefined {
-	if (metrics.now - metrics.startedAt >= config.activeNoticeAfterMs) return "time_threshold";
+	const activeNoticeAfterMs = progressAwareThreshold(config.activeNoticeAfterMs, config.progressScores);
+	if (metrics.now - metrics.startedAt >= activeNoticeAfterMs) return "time_threshold";
 	if (config.activeNoticeAfterTurns !== undefined && metrics.turns >= config.activeNoticeAfterTurns)
 		return "turn_threshold";
 	if (config.activeNoticeAfterTokens !== undefined && metrics.tokens >= config.activeNoticeAfterTokens)

package/dist/builtin/subagents/src/runs/shared/progress-trend.ts

@@ -0,0 +1,69 @@
+/** Pure progress-trend evidence shared by subagent control decisions. */
+
+export const DEFAULT_TREND_WINDOW = 3;
+export const DEFAULT_RISE_DELTA = 1.5;
+export const DEFAULT_FALL_DELTA = -1.5;
+export const FLAT_LOW_SCORE_CEILING = 8;
+export const PROGRESS_ATTENTION_THRESHOLD_FRACTION = 0.5;
+
+export type TrendConfig = {
+	window?: number;
+	riseDelta?: number;
+	fallDelta?: number;
+};
+
+export type Trend = "rising" | "flat" | "regressing";
+
+export type TrendResult = {
+	trend: Trend;
+	evidence: {
+		series: readonly number[];
+		window: number;
+		delta: number;
+	};
+};
+
+function mean(values: readonly number[]): number {
+	if (values.length === 0) return 0;
+	return values.reduce((total, value) => total + value, 0) / values.length;
+}
+
+/**
+ * Classify a score series using the same deterministic hysteresis as workflows.
+ * The result is evidence only: it carries no run action or terminal status.
+ */
+export function classify_trend(series: readonly number[], config: TrendConfig = {}): TrendResult {
+	const window = config.window ?? DEFAULT_TREND_WINDOW;
+	const riseDelta = config.riseDelta ?? DEFAULT_RISE_DELTA;
+	const fallDelta = config.fallDelta ?? DEFAULT_FALL_DELTA;
+	if (series.length < window + 1) {
+		return { trend: "flat", evidence: { series, window, delta: 0 } };
+	}
+	const sample = series.slice(-2 * window);
+	const usableLength = sample.length % 2 === 1 ? sample.length - 1 : sample.length;
+	const halfLength = usableLength / 2;
+	const leading = sample.slice(0, halfLength);
+	const trailing = sample.slice(sample.length - halfLength);
+	const delta = halfLength === 0 ? 0 : mean(trailing) - mean(leading);
+	const trend: Trend = delta >= riseDelta ? "rising" : delta <= fallDelta ? "regressing" : "flat";
+	return { trend, evidence: { series, window, delta } };
+}
+
+function hasFiniteProgressScores(series: readonly number[] | undefined): series is readonly number[] {
+	return series !== undefined && series.length > 0 && series.every((score) => Number.isFinite(score));
+}
+
+export function hasProgressAttentionSignal(series: readonly number[] | undefined): boolean {
+	if (!hasFiniteProgressScores(series)) return false;
+	const result = classify_trend(series);
+	const latest = series.at(-1);
+	return (
+		result.trend === "regressing" ||
+		(result.trend === "flat" && latest !== undefined && latest <= FLAT_LOW_SCORE_CEILING)
+	);
+}
+
+/** Lower a wall-clock threshold only when progress evidence raises priority. */
+export function progressAwareThreshold(baseMs: number, series: readonly number[] | undefined): number {
+	return hasProgressAttentionSignal(series) ? Math.max(1, baseMs * PROGRESS_ATTENTION_THRESHOLD_FRACTION) : baseMs;
+}

package/dist/builtin/subagents/src/runs/shared/subagent-control.ts

@@ -6,6 +6,7 @@ import type {
 	ControlNotificationChannel,
 	ResolvedControlConfig,
 } from "../../shared/types.js";
+import { progressAwareThreshold } from "./progress-trend.js";
 
 const CONTROL_EVENT_TYPES: ControlEventType[] = ["active_long_running", "needs_attention"];
 const CONTROL_NOTIFICATION_CHANNELS: ControlNotificationChannel[] = ["event", "intercom"];
@@ -34,6 +35,12 @@ function parseControlList<T extends string>(value: unknown, allowed: readonly T[
 	return parsed.length > 0 ? Array.from(new Set(parsed)) : undefined;
 }
 
+function parseProgressScores(value: unknown): number[] | undefined {
+	if (!Array.isArray(value)) return undefined;
+	const parsed = value.filter((entry): entry is number => typeof entry === "number" && Number.isFinite(entry));
+	return parsed.length > 0 || value.length === 0 ? parsed : undefined;
+}
+
 export function resolveControlConfig(globalConfig?: ControlConfig, override?: ControlConfig): ResolvedControlConfig {
 	const enabled = override?.enabled ?? globalConfig?.enabled ?? DEFAULT_CONTROL_CONFIG.enabled;
 	const needsAttentionAfterMs =
@@ -52,6 +59,8 @@ export function resolveControlConfig(globalConfig?: ControlConfig, override?: Co
 		parsePositiveInt(override?.failedToolAttemptsBeforeAttention) ??
 		parsePositiveInt(globalConfig?.failedToolAttemptsBeforeAttention) ??
 		DEFAULT_CONTROL_CONFIG.failedToolAttemptsBeforeAttention;
+	const progressScores =
+		parseProgressScores(override?.progressScores) ?? parseProgressScores(globalConfig?.progressScores);
 	const notifyOn =
 		parseControlList(override?.notifyOn, CONTROL_EVENT_TYPES) ??
 		parseControlList(globalConfig?.notifyOn, CONTROL_EVENT_TYPES) ??
@@ -67,6 +76,7 @@ export function resolveControlConfig(globalConfig?: ControlConfig, override?: Co
 		activeNoticeAfterTurns,
 		activeNoticeAfterTokens,
 		failedToolAttemptsBeforeAttention,
+		...(progressScores === undefined ? {} : { progressScores: [...progressScores] }),
 		notifyOn: [...notifyOn],
 		notifyChannels: [...notifyChannels],
 	};
@@ -82,7 +92,8 @@ export function deriveActivityState(input: {
 	const now = input.now ?? Date.now();
 	const lastActivity = input.lastActivityAt ?? input.startedAt;
 	const ageMs = Math.max(0, now - lastActivity);
-	return ageMs > input.config.needsAttentionAfterMs ? "needs_attention" : undefined;
+	const attentionThreshold = progressAwareThreshold(input.config.needsAttentionAfterMs, input.config.progressScores);
+	return ageMs > attentionThreshold ? "needs_attention" : undefined;
 }
 
 export function buildControlEvent(input: {

package/dist/builtin/subagents/src/shared/types-results.ts

@@ -55,6 +55,7 @@ export interface ControlConfig {
 	activeNoticeAfterTurns?: number;
 	activeNoticeAfterTokens?: number;
 	failedToolAttemptsBeforeAttention?: number;
+	progressScores?: number[];
 	notifyOn?: ControlEventType[];
 	notifyChannels?: ControlNotificationChannel[];
 }
@@ -66,6 +67,7 @@ export interface ResolvedControlConfig {
 	activeNoticeAfterTurns?: number;
 	activeNoticeAfterTokens?: number;
 	failedToolAttemptsBeforeAttention: number;
+	progressScores?: number[];
 	notifyOn: ControlEventType[];
 	notifyChannels: ControlNotificationChannel[];
 }

package/dist/builtin/web-access/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/web-access",
-  "version": "0.9.14-alpha.4",
+  "version": "0.9.14-alpha.5",
   "private": true,
   "description": "Atomic extension for web search, URL fetching, GitHub repo cloning, PDF/video extraction. Fork of: https://github.com/nicobailon/pi-web-access",
   "contributors": [

package/dist/builtin/workflows/CHANGELOG.md

@@ -6,11 +6,51 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 
 ## [Unreleased]
 
-## [0.9.14-alpha.4] - 2026-08-18
+## [0.9.14-alpha.5] - 2026-08-19
+
+### Added
+
+- Added optional run-budget declarations on workflow config, authored definitions, and `workflow` tool runs or resumes. `maxDurationMs`, `maxTokens`, `maxCost`, and `warnAtPercent` resolve per field with run overrides taking precedence over definition and config values; `0` disables a dimension. Invalid declarations now fail validation, and `budget_exceeded` is reserved as a resumable returned blocked status; these declarations provide the resolution core for the duration enforcement below ([#2212](https://github.com/bastani-inc/atomic/issues/2212)).
+- Added duration enforcement for workflow budgets at stage and durable-tool boundaries. Runs emit one user-visible `budget_warning` lifecycle notice at the default 80% threshold or configured threshold, receive one wrap-up turn only when a frontier stage turn is already live, and stop on a resumable `budget_exceeded` blocked rail with a duration report; no new stage is created for wrap-up, no-live-turn boundaries leave the allowance unused, paused time is excluded, elapsed time carries across resume, a raised resume budget continues prior spend, the wrap-up turn's own usage is recorded as `wrapUpUsage` when available without tree-wide token/cost aggregation, root duration scope is enforced inside child workflows and wins simultaneous child exhaustion, and child-scoped exhaustion returns to the parent without stopping it ([#2212](https://github.com/bastani-inc/atomic/issues/2212)).
+- Added tree-wide token and cost budget metering across nested workflow runs and stage retries. Uncached input and output tokens feed `maxTokens`, cache reads and writes remain reported counters, `usage.cost` feeds `maxCost`, and persisted accounting baselines prevent duplicate charges across boundaries and resume ([#2212](https://github.com/bastani-inc/atomic/issues/2212)).
+- Added HumanLayer's bundled `show-me` skill for visual explanations, diagrams, code-shape sketches, and focused HTML artifacts. Sourced from https://github.com/humanlayer/skills and distributed under the MIT License.
+- Updated the bundled `create-spec` skill to choose context-driven Path A or clarify-first Path B, inspect local precedent before design decisions, compare materially different alternatives, and use show-me shape-first visuals (pseudocode, call trees, component/file trees, Mermaid, diffs, and optional focused HTML). Working-method guidance is drawn from Dillon Mulroy's [tech-spec skill](https://github.com/dmmulroy/skills/blob/main/tech-spec/SKILL.md).
+### Breaking Changes
+
+- `adversarial-verification` now accepts optional per-criterion `criteria`, `accept_mean`, and bounded `reask_limit` inputs, and returns deterministic graded mean/veto results through `mean_score` and `score_table_path`; callers must migrate from the removed `result`, `verifier_artifact_paths`, and `artifact_dir` outputs. Invalid verifier reports are recorded as invalid markers and re-asked instead of becoming fail votes ([#2487](https://github.com/bastani-inc/atomic/issues/2487)).
+- Replaced `bracket_path`/`bracket.json` and the knockout tournament schedule with `comparisons_path`/`comparisons.json` on the soft-scored pivot-pairing schedule. Judge stages now use deterministic `judge-<a>-<b>-<criterion>-r<rep>` identities, and the reducer stage is named `comparisons-reducer`.
+- Replaced `bracket_path`/`bracket.json` and the knockout tournament schedule with `comparisons_path`/`comparisons.json` on the soft-scored pivot-pairing schedule. Judge stages now use deterministic `judge-<a>-<b>-<criterion>-r<rep>` identities, and the reducer stage is named `comparisons-reducer`.
+- `adversarial-verification` now accepts optional per-criterion `criteria`, `accept_mean`, and bounded `reask_limit` inputs, and returns deterministic graded mean/veto results through `mean_score` and `score_table_path`; callers must migrate from the removed `result`, `verifier_artifact_paths`, and `artifact_dir` outputs. Invalid verifier reports are recorded as invalid markers and re-asked instead of becoming fail votes ([#2487](https://github.com/bastani-inc/atomic/issues/2487)).
+
+### Added
+
+- Added a shared `verification-criteria` module (`parse_rubric`, `normalize_criteria`, `select_criteria`, `VERIFICATION_SCALE`, `decide_verification`) so verification builtins can score one criterion at a time on an anchored 1–20 scale and accept only when a quorum mean clears the threshold with no veto finding. Unparseable reports cannot become scores and cannot shift the mean ([#2487](https://github.com/bastani-inc/atomic/issues/2487)).
+- Added prefix-cache-aware verification prompts with a byte-identical shared head, UTF-8 bounded candidate inlining with whole-family read fallback, and warm-first verifier fan-out scheduling that preserves input order while releasing later phases after warm failures ([#2493](https://github.com/bastani-inc/atomic/issues/2493)).
+- Added optional run-budget declarations on workflow config, authored definitions, and `workflow` tool runs or resumes. `maxDurationMs`, `maxTokens`, `maxCost`, and `warnAtPercent` resolve per field with run overrides taking precedence over definition and config values; `0` disables a dimension. Invalid declarations now fail validation, and `budget_exceeded` is reserved as a resumable returned blocked status; these declarations provide the resolution core for the duration enforcement below ([#2212](https://github.com/bastani-inc/atomic/issues/2212)).
+- Added duration enforcement for workflow budgets at stage and durable-tool boundaries. Runs emit one user-visible `budget_warning` lifecycle notice at the default 80% threshold or configured threshold, receive one wrap-up turn only when a frontier stage turn is already live, and stop on a resumable `budget_exceeded` blocked rail with a duration report; no new stage is created for wrap-up, no-live-turn boundaries leave the allowance unused, paused time is excluded, elapsed time carries across resume, a raised resume budget continues prior spend, the wrap-up turn's own usage is recorded as `wrapUpUsage` when available without tree-wide token/cost aggregation, root duration scope is enforced inside child workflows and wins simultaneous child exhaustion, and child-scoped exhaustion returns to the parent without stopping it ([#2212](https://github.com/bastani-inc/atomic/issues/2212)).
+- Added optional run-budget declarations on workflow config, authored definitions, and `workflow` tool runs. `maxDurationMs`, `maxTokens`, `maxCost`, and `warnAtPercent` resolve per field with run overrides taking precedence over definition and config values; `0` disables a dimension. Invalid declarations now fail validation, and `budget_exceeded` is reserved as a resumable returned blocked status. This core slice only resolves and validates budgets; it does not meter or stop runs yet ([#2212](https://github.com/bastani-inc/atomic/issues/2212)).
+- Added a shared `verification-criteria` module (`parse_rubric`, `normalize_criteria`, `select_criteria`, `VERIFICATION_SCALE`, `decide_verification`) so verification builtins can score one criterion at a time on an anchored 1–20 scale and accept only when a quorum mean clears the threshold with no veto finding. Unparseable reports cannot become scores and cannot shift the mean ([#2487](https://github.com/bastani-inc/atomic/issues/2487)).
+- Added prefix-cache-aware verification prompts with a byte-identical shared head, UTF-8 bounded candidate inlining with whole-family read fallback, and warm-first verifier fan-out scheduling that preserves input order while releasing later phases after warm failures ([#2493](https://github.com/bastani-inc/atomic/issues/2493)).
+- Added Goal and Ralph review-round re-verification for low-confidence, single-reviewer blocking findings, using fresh 1–20 scoring stages, two-tier demotion bars, and durable per-repeat audit evidence while leaving `stop_review_loop` authoritative.
+- Added pure, seeded selection math for probabilistic pivot tournaments: deterministic Hamiltonian ring and deduplicated pivot planning, criterion/repeat slot-swap jobs, normalized Bradley–Terry soft wins, count-normalized pivot selection, and complete candidate rankings ([#2488](https://github.com/bastani-inc/atomic/issues/2488)).
+- Added tournament `n_evaluations`, `pivots`, `seed`, `criteria`, and heterogeneous `models` inputs, plus `ranking` and `seed` outputs and the auditable `comparisons.json` score ledger with invalid-report and budget records ([#2488](https://github.com/bastani-inc/atomic/issues/2488)).
+- `adversarial-verification` now fans out one schema-validated 1–20 score for every criterion/verifier pair, applies mean-plus-veto acceptance, bounds invalid-report re-asks, and consolidates confirmed findings into repair guidance ([#2487](https://github.com/bastani-inc/atomic/issues/2487)).
+- Added pure `fold_usage` accounting over every model attempt (including retries), returning `UsageTotals { calls, input, output, cacheRead, cacheWrite, cost, turns, cacheHitRate }` with a derived cache-hit rate; tournament comparisons and adversarial verification round summaries now record per-phase and total usage blocks ([#2494](https://github.com/bastani-inc/atomic/issues/2494)).
+- Added observation-grounded progress scoring with batched checkpoint repeats, null-safe invalid-repeat handling, and a pure hysteretic trend classifier that reports evidence without an action ([#2489](https://github.com/bastani-inc/atomic/issues/2489)).
+- Added Goal and Ralph per-round convergence evidence series with V7 trend classification, folded usage totals, and escalation-only `needs_human` evidence that never approves or terminates a review loop ([#2490](https://github.com/bastani-inc/atomic/issues/2490)).
+- Added reviewer calibration rules that prioritize observed output over agent narration, plus audit-only 1–20 `criterion_scores` in Goal and Ralph review decisions without changing the authoritative closure gate ([#2494](https://github.com/bastani-inc/atomic/issues/2494)).
+- Added loop-until-done progress consumers that record per-iteration advisory score curves, final trends, and calibration disclaimers in completion and exhaustion reports without changing stop decisions ([#2489](https://github.com/bastani-inc/atomic/issues/2489)).
+- Added observation-grounded progress scoring with batched checkpoint repeats, null-safe invalid-repeat handling, and a pure hysteretic trend classifier that reports evidence without an action ([#2489](https://github.com/bastani-inc/atomic/issues/2489)).
+- Added loop-until-done progress consumers that record per-iteration advisory score curves, final trends, and calibration disclaimers in completion and exhaustion reports without changing stop decisions ([#2489](https://github.com/bastani-inc/atomic/issues/2489)).
+- Added pure, seeded selection math for probabilistic pivot tournaments: deterministic Hamiltonian ring and deduplicated pivot planning, criterion/repeat slot-swap jobs, normalized Bradley–Terry soft wins, count-normalized pivot selection, and complete candidate rankings ([#2488](https://github.com/bastani-inc/atomic/issues/2488)).
+- Added tournament `n_evaluations`, `pivots`, `seed`, `criteria`, and heterogeneous `models` inputs, plus `ranking` and `seed` outputs and the auditable `comparisons.json` score ledger with invalid-report and budget records ([#2488](https://github.com/bastani-inc/atomic/issues/2488)).
 
 ### Fixed
 
 - Opened `ctx.tool` detail now matches the main-chat tool block: `Box(1,1)` inner padding, a blank row under the `$` header, full-width alignment with the orchestrator bars, unpainted canvas above and below the card, host `toolTitle`/`toolOutput` colors, a dim `ctrl+o Expand` hint, and second-resolution `Took`/`Elapsed` timing.
+- Tournament criteria now accepts every documented V1 shape, including markdown rubrics, records, string lists, and CriterionInput objects with optional ids and names.
+- Criterion ids are preserved in the ledger while judge artifact filenames are sanitized, collision-free, and contained under the artifact directory.
+- Corrected the planned comparison budget for oversized pivot counts and recorded each comparison row's own soft preference separately from pair aggregates.
 
 ## [0.9.14-alpha.3] - 2026-08-17
 

package/dist/builtin/workflows/README.md

@@ -677,10 +677,11 @@ Raw stage-chat prompt answer replay is live-memory only. `StageSnapshot.promptAn
 ```json
 {
   "name": "workflow",
-  "description": "Run named builtin, project, user, or package workflows; custom definitions may import reusable project/package workflows or builtin definitions from @bastani/workflows/builtin and nest them with ctx.workflow(...), including deeper composition within the configured maxDepth; when workflow execution fits but another shape would better achieve the task, author a custom TypeScript workflow({...}) inline with normal coding tools, reload it, and run it; after successfully creating and reloading a newly authored custom workflow, report the folder containing its generated code as 'Custom workflow created. You can inspect its code at: <workflow-folder-path>'; do this only for newly created custom workflows, never builtin or pre-existing workflows; discover with list/get/inputs/models, list session runs with status (no runId; statusFilter narrows the list), inspect status/stages/stage details, send prompt answers or steering only while the root workflow is nonterminal, pause/resume/interrupt/quit runs, and reload workflow resources. For primitive prompt answers, use booleans or the documented confirm labels, exact case-insensitive select labels or 1-based indexes, and text strings for input/editor; an invalid answer remains pending and returns guidance instead of choosing a default. For large stage handoffs, write context to files/artifacts, pass paths via reads, and prompt downstream agents to 'Read the file at <path>...' instead of injecting large previous text. Wrap critical parts of run inputs and steering messages in <keepContext>...</keepContext> so compaction preserves them verbatim in the stages that inherit them; tag role constraints, prohibitions, must-hold criteria, and identifiers, not background or bulk reference material. For transcripts, prefer status/stages/stage to get sessionFile/transcriptPath, quote the exact path without rewriting separators (Windows backslashes are valid), then search it with rg/grep and read small ranges; transcript is path-only by default when sessionFile/transcriptPath exists, explicit tail/limit returns bounded previews, and missing transcript paths fall back to a small preview. Use action 'models' to inspect models in the configured catalog; the result is a configured-auth snapshot showing what's present in the registry with configured authentication, not proof of credentials, entitlements, OAuth freshness, or live provider access. When authoring a workflow that should dynamically select a model, first call workflow({ action: 'models' }) to inspect the configured catalog, then select from the returned provider/id entries considering the isCurrent marker and available thinking levels.",
+  "description": "Run named builtin, project, user, or package workflows; custom definitions may import reusable project/package workflows or builtin definitions from @bastani/workflows/builtin and nest them with ctx.workflow(...), including deeper composition within the configured maxDepth; when workflow execution fits but another shape would better achieve the task, author a custom TypeScript workflow({...}) inline with normal coding tools, reload it, and run it; after successfully creating and reloading a newly authored custom workflow, report the folder containing its generated code as 'Custom workflow created. You can inspect its code at: <workflow-folder-path>'; do this only for newly created custom workflows, never builtin or pre-existing workflows; discover with list/get/inputs/models, list session runs with status (no runId; statusFilter narrows the list), inspect status/stages/stage details, send prompt answers or steering only while the root workflow is nonterminal, pause/resume/interrupt/quit runs, and reload workflow resources. For action 'run' and 'resume', budget accepts per-field duration, token, cost, and warning overrides; fields resolve over the workflow declaration and config, and 0 disables a field. For primitive prompt answers, use booleans or the documented confirm labels, exact case-insensitive select labels or 1-based indexes, and text strings for input/editor; an invalid answer remains pending and returns guidance instead of choosing a default. For large stage handoffs, write context to files/artifacts, pass paths via reads, and prompt downstream agents to 'Read the file at <path>...' instead of injecting large previous text. Wrap critical parts of run inputs and steering messages in <keepContext>...</keepContext> so compaction preserves them verbatim in the stages that inherit them; tag role constraints, prohibitions, must-hold criteria, and identifiers, not background or bulk reference material. For transcripts, prefer status/stages/stage to get sessionFile/transcriptPath, quote the exact path without rewriting separators (Windows backslashes are valid), then search it with rg/grep and read small ranges; transcript is path-only by default when sessionFile/transcriptPath exists, explicit tail/limit returns bounded previews, and missing transcript paths fall back to a small preview. Use action 'models' to inspect models in the configured catalog; the result is a configured-auth snapshot showing what's present in the registry with configured authentication, not proof of credentials, entitlements, OAuth freshness, or live provider access. When authoring a workflow that should dynamically select a model, first call workflow({ action: 'models' }) to inspect the configured catalog, then select from the returned provider/id entries considering the isCurrent marker and available thinking levels.",
   "parameters": {
     "workflow": "string (optional) — workflow ID or normalized name",
     "inputs": "object (optional) — key/value map of workflow inputs",
+    "budget": "optional run-only object: maxDurationMs/maxTokens non-negative integers; maxCost/warnAtPercent non-negative numbers; 0 disables a field",
     "action": "'run' | 'list' | 'get' | 'inputs' | 'models' | 'status' | 'stages' | 'stage' | 'transcript' | 'send' | 'pause' | 'interrupt' | 'quit' | 'resume' | 'reload'",
     "runId": "optional full 36-character run id; prefixes are rejected; control actions default to the active run where safe; use '--all' or all:true for pause/interrupt/quit all",
     "stageId": "optional exact stage id or exact stage name for stage-scoped actions; prefixes and partial names are rejected; cannot be combined with all:true",
@@ -772,10 +773,10 @@ The six common patterns ship as full builtins and are discoverable/runnable by n
 |---|---|---|---|
 | `classify-and-act` | structured classifier → deterministic action; HIL fallback for low confidence | `prompt`; `categories` (3 defaults), `confidence_threshold=0.75` | result, category/confidence, classification/action paths |
 | `fan-out-and-synthesize` | partition → bounded artifact fan-out → evidence synthesis barrier | `prompt`; `max_branches=4`, `max_concurrency=4` | result, partitions, branch/synthesis/manifest paths |
-| `adversarial-verification` | worker → fresh rubric verifiers → reducer → bounded repair | `task`; `verifier_count=3`, `max_repairs=2` | result, approval, repairs, candidate/review/verifier paths |
-| `generate-and-filter` | candidate fan-out → dedupe/filter → optional judge → shortlist | `prompt`; `num_candidates=8`, `shortlist_size=3`, `use_judge=true`, `max_concurrency=4` | result, shortlist and candidate/filter/judge/final/manifest paths |
-| `tournament` | independent attempts → order-balanced pairwise judges → bracket reducer | `prompt`; `num_attempts=4`, `max_concurrency=4` | result, winner, attempt/judge/bracket paths |
-| `loop-until-done` | durable ledger → iteration/evaluator loop → complete or inspectable exhaustion | `prompt`; `max_iterations=5` | result/status, ledger, iteration/evaluation paths, remaining work |
+| `adversarial-verification` | worker → per-criterion fresh rubric verifiers → deterministic mean+veto reducer → findings consolidation / bounded repair; decompose rubrics into named criteria and aggregate by mean plus explicit veto, never a unanimity AND; `criteria.md` uses `## Criteria` with non-empty `### Name {#id}` bodies and the shared `verification-criteria` doors; consolidator cannot approve | `task`; `criteria` (record or criteria.md markdown, defaults to task_fit/evidence/completeness), `verifier_count=3` (1–5), `max_repairs=2` (0–5), `accept_mean=14`, `reask_limit=1`; invalid reports are re-asked within the bound and never become fail votes; normal calls per round: criteria.length × verifier_count | `approved`, `mean_score`, `score_table_path` (`verification-summary-<round>.json` with scores, mean, invalidCount, decision, and usage), `repairs_completed`, `candidate_path`, `review_report_path`, `remaining_work` |
+| `generate-and-filter` | candidate fan-out → dedupe/filter → optional judge → shortlist; use graded per-criterion scores, Bradley–Terry preference from score gaps, and slot-swap repeats when ranking candidates; custom judges can reuse the shared `verification-criteria` module and `criteria.md` rubric format | `prompt`; `num_candidates=8`, `shortlist_size=3`, `use_judge=true`, `max_concurrency=4` | result, shortlist and candidate/filter/judge/final/manifest paths |
+| `tournament` | independent attempts → seeded ring and pivot-round soft scoring with order-balanced pairwise judges → full ranking/comparisons reducer; derive a Bradley–Terry preference from graded score gaps and repeat with A/B slot swaps; `comparisons.json` preserves per-job score or invalid rows, pair aggregates, ranking, and planned/executed budget records | `prompt`; `num_attempts=4`, `max_concurrency=4`, `n_evaluations=2`, `pivots=1`, `seed=0`, optional `criteria` (markdown rubric, record, string list, or `CriterionInput` list)/`models` | result, winner, `attempt_artifact_paths`, `judge_artifact_paths`, `comparisons_path`, `ranking`, `seed` |
+| `loop-until-done` | durable ledger → iteration/evaluator loop → complete or inspectable exhaustion; each scored iteration records `progress` (`score`, `perRepeat`, `trend`, `window`) and the ledger emits `progress_curve`, `final_trend`, and `progress_disclaimer`; trend is advisory and never a kill switch | `prompt`; `max_iterations=5`, `progress_scoring=true`, `progress_repeats=1` | result/status, ledger, iteration/evaluation paths, remaining work, `progress_curve`, `final_trend`, `progress_disclaimer` |
 
 All six are exported from `@bastani/workflows/builtin` as definitions (`classifyAndAct`, `fanOutAndSynthesize`, `adversarialVerification`, `generateAndFilter`, `tournament`, and `loopUntilDone`). Import and nest them through `ctx.workflow(definition, { inputs, stageName })`; nested calls respect `maxDepth`. Prefer composition over copying their prompts or graphs: children contribute their stages, dedicated prompts, gates, artifacts, HIL nodes, and declared outputs to the expanded parent graph.
 

package/dist/builtin/workflows/builtin/adversarial-verification-prompts.ts

@@ -1,3 +1,5 @@
+import type { Criterion } from "./verification-criteria.js";
+
 const GROUNDED_REPORTING = "Before reporting progress, audit each claim against a tool result from this session. Report only work you can point to evidence for; say so explicitly when something is unverified.";
 const READABLE_REPORT = "Lead with the outcome. Keep facts, decisions, caveats, and next steps; drop background and repetition. Use complete, readable sentences rather than compressed fragments.";
 
@@ -5,14 +7,20 @@ export function renderWorkerPrompt(task: string): string {
   return `<role>\nYou produce a candidate solution for independent verification.\n</role>\n\n<success_criteria>\nThe task is complete and important claims have observable support. Preserve concrete evidence and state every validation performed.\n</success_criteria>\n\n<stop_rules>\nStop when the candidate is complete and validated where practical, or state the evidence still missing.\n</stop_rules>\n\n<output_format>\nA self-contained candidate with actions taken, evidence, validation, and remaining risks. ${READABLE_REPORT}\n${GROUNDED_REPORTING}\n</output_format>\n\n<objective>\n${task}\n</objective>`;
 }
 
-export function renderVerifierPrompt(task: string, candidatePath: string, rubricPath: string): string {
-  return `<artifacts>\nRead the complete candidate at ${candidatePath} and rubric at ${rubricPath}.\n</artifacts>\n\n<role>\nYou are an independent adversarial verifier. Find blockers; do not rewrite the candidate.\n</role>\n\n<success_criteria>\nA pass requires concrete evidence for every rubric item; absence of evidence is not evidence of correctness.\n</success_criteria>\n\n<evidence_rules>\nTest important claims where practical. For each conclusion, report the command run, observed output, and file:line evidence where applicable. Report precise blocking findings. ${GROUNDED_REPORTING}\n</evidence_rules>\n\n<stop_rules>\nStop after every rubric item has pass evidence or a precise blocking finding.\n</stop_rules>\n\n<output_format>\nCall structured_output with verdict (pass or fail), concise evidence, and blocking_findings. ${READABLE_REPORT}\n</output_format>\n\n<objective>\nVerify the candidate against the task and rubric. Task: ${task}\n</objective>`;
+export function renderVerifierPrompt(task: string, candidatePath: string, criteriaPath: string, criterion: Criterion): string {
+  return `<artifacts>\nRead the complete candidate at ${candidatePath} and resolved criteria at ${criteriaPath}.\n</artifacts>\n\n<role>\nYou are an independent adversarial verifier. Score exactly one criterion and do not rewrite the candidate.\n</role>\n\n<criterion>\nid: ${criterion.id}\nname: ${criterion.name}\ninstructions: ${criterion.description}\n</criterion>\n\n<scale>\nUse an integer score from 1 to 20: 1 = certainly fails … 10 = borderline … 20 = verified correct.\n</scale>\n\n<evidence_rules>\nTest important claims where practical. Evidence must cite observable support; file findings cite file:line where applicable. Report precise findings and use severity veto only for a finding that unconditionally prevents acceptance. ${GROUNDED_REPORTING}\n</evidence_rules>\n\n<output_format>\nCall structured_output with exactly criterion_id (set to ${criterion.id}), score, evidence (string array), and findings (objects with finding and severity veto, blocking, or note). ${READABLE_REPORT}\n</output_format>\n\n<objective>\nVerify criterion ${criterion.id} for the candidate produced for: ${task}\n</objective>`;
 }
 
-export function renderReducerPrompt(task: string, candidatePath: string, verifierPaths: readonly string[], repairsCompleted: number, maxRepairs: number): string {
-  return `<artifacts>\nCandidate: ${candidatePath}\nVerifier reports: ${verifierPaths.join(", ")}\n</artifacts>\n\n<role>\nYou reduce independent verification reports into one deterministic next action.\n</role>\n\n<decision_rules>\nAccept only when all material rubric requirements have pass evidence. Request repair when findings are actionable and repair budget remains (${repairsCompleted}/${maxRepairs}). Reject when evidence proves the candidate cannot satisfy the task or the repair budget is exhausted. Preserve unresolved blockers verbatim.\n</decision_rules>\n\n<success_criteria>\nThe decision follows the evidence and repair bound, with every unresolved blocker preserved.\n</success_criteria>\n\n<stop_rules>\nStop after selecting exactly one decision: accept, reject, or repair.\n</stop_rules>\n\n<output_format>\nCall structured_output with decision (accept, reject, or repair), a concise evidence-based rationale, and remaining_work. ${READABLE_REPORT}\n${GROUNDED_REPORTING}\n</output_format>\n\n<objective>\nDecide whether the candidate for ${task} is accepted, rejected, or needs repair.\n</objective>`;
+export function renderConsolidatorPrompt(
+  task: string,
+  candidatePath: string,
+  scorePaths: readonly string[],
+  repairsCompleted: number,
+  maxRepairs: number,
+): string {
+  return `<artifacts>\nCandidate: ${candidatePath}\nConfirmed criterion score reports: ${scorePaths.join(", ")}\n</artifacts>\n\n<role>\nYou consolidate confirmed verifier findings into actionable repair guidance. You do not decide whether verification is approved; the deterministic verification gate already made that decision.\n</role>\n\n<decision_rules>\nPreserve every confirmed veto or blocking finding in remaining_work verbatim. Explain concrete repairs and validation steps. Repair budget: ${repairsCompleted}/${maxRepairs}.\n</decision_rules>\n\n<output_format>\nCall structured_output with repair_guidance (a concise actionable repair plan) and remaining_work (the unresolved findings as strings). Never emit an approval or rejection decision. ${READABLE_REPORT}\n${GROUNDED_REPORTING}\n</output_format>\n\n<objective>\nConsolidate findings for the candidate produced for: ${task}\n</objective>`;
 }
 
 export function renderRepairPrompt(task: string, candidatePath: string, reviewPath: string): string {
-  return `<artifacts>\nRead the current candidate at ${candidatePath} and reducer report at ${reviewPath}.\n</artifacts>\n\n<role>\nYou repair a candidate using independent blocking findings.\n</role>\n\n<success_criteria>\nEvery actionable blocker is addressed, relevant validation is rerun, and valid prior work is retained.\n</success_criteria>\n\n<evidence_rules>\nDo not dismiss a finding without contrary evidence. Report commands run, observed output, and file:line evidence where applicable. ${GROUNDED_REPORTING}\n</evidence_rules>\n\n<stop_rules>\nStop when all actionable blockers are repaired and validated, or list any blocker that remains with the missing evidence.\n</stop_rules>\n\n<output_format>\nA complete replacement candidate with repair summary, evidence, validation, and remaining risks. ${READABLE_REPORT}\n</output_format>\n\n<objective>\nRepair the candidate for: ${task}\n</objective>`;
+  return `<artifacts>\nRead the current candidate at ${candidatePath} and consolidated findings at ${reviewPath}.\n</artifacts>\n\n<role>\nYou repair a candidate using independent blocking findings.\n</role>\n\n<success_criteria>\nEvery actionable blocker is addressed, relevant validation is rerun, and valid prior work is retained.\n</success_criteria>\n\n<evidence_rules>\nDo not dismiss a finding without contrary evidence. Report commands run, observed output, and file:line evidence where applicable. ${GROUNDED_REPORTING}\n</evidence_rules>\n\n<stop_rules>\nStop when all actionable blockers are repaired and validated, or list any blocker that remains with the missing evidence.\n</stop_rules>\n\n<output_format>\nA complete replacement candidate with repair summary, evidence, validation, and remaining risks. ${READABLE_REPORT}\n</output_format>\n\n<objective>\nRepair the candidate for: ${task}\n</objective>`;
 }