@bastani/atomic 0.9.3 → 0.9.4-alpha.10

package/dist/builtin/subagents/src/runs/shared/model-fallback.ts

@@ -117,6 +117,7 @@ export type ModelFallbackFailureKind =
 	| "provider_unavailable"
 	| "network_timeout"
 	| "model_unavailable"
+	| "request_incompatible"
 	| "cancelled"
 	| "task_failure"
 	| "unknown";
@@ -144,6 +145,7 @@ const FALLBACKABLE_FAILURE_KINDS: ReadonlySet<ModelFallbackFailureKind> = new Se
 	"provider_unavailable",
 	"network_timeout",
 	"model_unavailable",
+	"request_incompatible",
 ]);
 
 function asRecord(value: unknown): Record<string, unknown> | undefined {
@@ -222,6 +224,9 @@ function normalizeCode(value: string | number | undefined): string | undefined {
 
 function kindFromStatus(status: number | undefined): ModelFallbackFailureKind | undefined {
 	switch (status) {
+		case 400:
+		case 413: case 422:
+			return "request_incompatible";
 		case 401:
 		case 403:
 			return "auth_on_candidate_provider";
@@ -237,6 +242,15 @@ function kindFromStatus(status: number | undefined): ModelFallbackFailureKind |
 	}
 }
 
+const REQUEST_INCOMPATIBLE_CODES: ReadonlySet<string> = new Set([
+	"invalid_request", "invalid_request_error", "bad_request", "context_length_exceeded",
+	"request_too_large", "too_large", "request_entity_too_large", "max_tokens",
+	"max_context_length", "context_window_exceeded",
+]);
+function requestIncompatibleKindFromCode(code: string | number | undefined): ModelFallbackFailureKind | undefined {
+	const normalizedCode = normalizeCode(code);
+	return normalizedCode !== undefined && REQUEST_INCOMPATIBLE_CODES.has(normalizedCode) ? "request_incompatible" : undefined;
+}
 function refusalKindFromCode(code: string | number | undefined): ModelFallbackFailureKind | undefined {
 	const normalizedCode = normalizeCode(code);
 	if (normalizedCode === undefined) return undefined;
@@ -258,6 +272,14 @@ function refusalKindFromCode(code: string | number | undefined): ModelFallbackFa
 	}
 }
 
+const CODE_KINDS_BY_KIND: ReadonlyArray<readonly [ModelFallbackFailureKind, ReadonlySet<string>]> = [
+	["auth_on_candidate_provider", new Set(["auth", "auth_required", "authentication_required", "unauthorized", "forbidden", "invalid_api_key", "missing_api_key", "invalid_key"])],
+	["network_timeout", new Set(["etimedout", "econnreset", "econnrefused", "enotfound", "eai_again", "fetch_failed", "network_error", "timeout", "timeout_error", "und_err_connect_timeout"])],
+	["rate_limit", new Set(["rate_limit", "rate_limit_exceeded", "too_many_requests", "quota_exceeded"])],
+	["cancelled", new Set(["aborterror", "aborted", "cancelled", "canceled"])],
+	["model_unavailable", new Set(["model_not_found", "model_unavailable", "model_disabled", "unknown_model"])],
+	["provider_unavailable", new Set(["provider_error", "api_error", "service_unavailable", "temporarily_unavailable", "overloaded"])],
+];
 function kindFromCode(code: string | number | undefined): ModelFallbackFailureKind | undefined {
 	const normalizedCode = normalizeCode(code);
 	if (normalizedCode === undefined) return undefined;
@@ -265,54 +287,19 @@ function kindFromCode(code: string | number | undefined): ModelFallbackFailureKi
 	if (refusalKind !== undefined) return refusalKind;
 	const httpStatusKind = kindFromStatus(integerFrom(code));
 	if (httpStatusKind !== undefined) return httpStatusKind;
-
-	switch (normalizedCode) {
-		case "auth":
-		case "auth_required":
-		case "authentication_required":
-		case "unauthorized":
-		case "forbidden":
-		case "invalid_api_key":
-		case "missing_api_key":
-		case "invalid_key":
-			return "auth_on_candidate_provider";
-		case "etimedout":
-		case "econnreset":
-		case "econnrefused":
-		case "enotfound":
-		case "eai_again":
-		case "fetch_failed":
-		case "network_error":
-		case "timeout":
-		case "timeout_error":
-		case "und_err_connect_timeout":
-			return "network_timeout";
-		case "rate_limit":
-		case "rate_limit_exceeded":
-		case "too_many_requests":
-		case "quota_exceeded":
-			return "rate_limit";
-		case "aborterror":
-		case "aborted":
-		case "cancelled":
-		case "canceled":
-			return "cancelled";
-		case "model_not_found":
-		case "model_unavailable":
-		case "model_disabled":
-		case "unknown_model":
-			return "model_unavailable";
-		case "provider_error":
-		case "api_error":
-		case "service_unavailable":
-		case "temporarily_unavailable":
-		case "overloaded":
-			return "provider_unavailable";
-		default:
-			return undefined;
-	}
-}
-
+	const requestIncompatibleKind = requestIncompatibleKindFromCode(code);
+	if (requestIncompatibleKind !== undefined) return requestIncompatibleKind;
+	return CODE_KINDS_BY_KIND.find(([_, codes]) => codes.has(normalizedCode))?.[0];
+}
+
+const REQUEST_INCOMPATIBLE_FAILURE_PATTERNS: readonly RegExp[] = [
+	/\bcontext[_\s-]?(?:length|window)(?:[_\s-]?exceeded)?\b/i,
+	/\bmax[_\s-]?(?:context|tokens?)\b/i,
+	/\b(?:request(?:[_\s-]?entity)?[_\s-]?too|too)[_\s-]?large\b/i,
+	/\b(?:unsupported|unknown|invalid)\s+(?:tool|parameter|function)\b/i,
+	/\b(?:tool|parameter|function)\s+(?:not\s+(?:supported|found|allowed)|unknown|invalid)\b/i,
+	/\b(?:invalid[_\s-]?request(?:[_\s-]?error)?|bad[_\s-]?request)\b/i,
+];
 const PROVIDER_REFUSAL_FAILURE_PATTERNS: readonly RegExp[] = [
 	/\bfinish[_\s-]?reason\b[^\n]*\bcontent[_\s-]?filter\b/i,
 	/\bcontent[_\s-]?filter(?:ed|ing)?\b/i,
@@ -335,6 +322,7 @@ function refusalKindFromMessage(message: string): ModelFallbackFailureKind | und
 function fallbackKindFromMessage(message: string, name: string | undefined): ModelFallbackFailureKind | undefined {
 	const refusalKind = refusalKindFromMessage(message);
 	if (refusalKind !== undefined) return refusalKind;
+	if (REQUEST_INCOMPATIBLE_FAILURE_PATTERNS.some((pattern) => pattern.test(message))) return "request_incompatible";
 	const nameKind = kindFromCode(name);
 	if (nameKind !== undefined) return nameKind;
 	if (!RETRYABLE_MODEL_FAILURE_PATTERNS.some((pattern) => pattern.test(message))) return undefined;
@@ -372,15 +360,21 @@ function makeSignal(
 	};
 }
 
-function fallbackSignalFromMessage(
+function fallbackSignalForMessage(
+	message: string | undefined,
 	value: unknown,
 	source: ModelFallbackFailureSource | undefined,
 ): ModelFallbackFailureSignal | undefined {
-	const message = modelFailureMessage(value);
-	if (!message.trim()) return undefined;
+	if (message === undefined || message.trim().length === 0) return undefined;
 	const kind = fallbackKindFromMessage(message, errorName(value));
 	return kind === undefined ? undefined : makeSignal(kind, value, source);
 }
+function fallbackSignalFromMessage(value: unknown, source: ModelFallbackFailureSource | undefined): ModelFallbackFailureSignal | undefined {
+	return fallbackSignalForMessage(modelFailureMessage(value), value, source);
+}
+function fallbackSignalFromDirectMessage(value: unknown, source: ModelFallbackFailureSource | undefined): ModelFallbackFailureSignal | undefined {
+	return fallbackSignalForMessage(directMessageFrom(value), value, source);
+}
 
 function classifyAssistantRefusalSignal(
 	value: unknown,
@@ -433,6 +427,12 @@ function structuredSignal(
 		if (isRefusalSignal(causeSignal)) return causeSignal;
 		firstNestedFallbackSignal ??= causeSignal;
 	}
+	// Direct-message classification runs after nested traversal so a generic wrapper
+	// ("invalid request"/"400 bad request") cannot mask a non-retryable nested signal.
+	// Kept in the same position as the workflows classifier so the two copies stay
+	// behaviorally parallel (see test/unit/model-fallback-classifier-conformance.test.ts).
+	const directMessageSignal = fallbackSignalFromDirectMessage(value, source);
+	if (directMessageSignal !== undefined) return directMessageSignal;
 	const statusKind = kindFromStatus(statusFrom(value));
 	if (statusKind !== undefined) return makeSignal(statusKind, value, source);
 	if (codeKind !== undefined) return makeSignal(codeKind, value, source);
@@ -463,8 +463,7 @@ function messageFromUnknown(value: unknown, seen: Set<unknown>): string | undefi
 	const status = statusFrom(value);
 	if (status !== undefined) return `Model request failed with status ${status}`;
 	const code = codeFrom(value);
-	if (code !== undefined) return `Model request failed with code ${String(code)}`;
-	return undefined;
+	return code !== undefined ? `Model request failed with code ${String(code)}` : undefined;
 }
 export function modelFailureMessage(error: unknown): string {
 	const structuredMessage = messageFromUnknown(error, new Set());

package/dist/builtin/subagents/src/runs/shared/parallel-utils.ts

@@ -1,4 +1,5 @@
 import type { CodexFastModeResolvedSettings, CodexFastModeScope } from "@bastani/atomic";
+import type { ModelAttempt } from "../../shared/types.ts";
 
 export interface RunnerSubagentStep {
 	agent: string;
@@ -13,6 +14,7 @@ export interface RunnerSubagentStep {
 	fastMode?: boolean;
 	modelFastModes?: Record<string, boolean>;
 	modelCandidates?: string[];
+	modelAttempts?: ModelAttempt[];
 	codexFastModeSettings?: CodexFastModeResolvedSettings;
 	codexFastModeScope?: CodexFastModeScope;
 	tools?: string[];

package/dist/builtin/subagents/src/shared/model-info.ts

@@ -27,6 +27,22 @@ export function toModelInfo(model: RegistryModelLike): ModelInfo {
 	};
 }
 
+interface KnownProviderRegistryLike {
+	getAvailable(): ReadonlyArray<{ provider: string }>;
+	/** Older hosts may not expose `getAll()`; callers probe at runtime and fall back
+	 * to `getAvailable()`. */
+	getAll?: () => ReadonlyArray<{ provider: string }>;
+}
+
+/** Collect the distinct providers of every model the registry knows about — including
+ * providers without configured auth — falling back to `getAvailable()` when the host
+ * registry does not expose `getAll()`. Used to pre-filter spawn candidates whose
+ * provider is known but keyless. */
+export function collectKnownModelProviders(registry: KnownProviderRegistryLike): string[] {
+	const models = typeof registry.getAll === "function" ? registry.getAll() : registry.getAvailable();
+	return [...new Set(models.map((model) => model.provider))];
+}
+
 /** Resolve the effective thinking level from a model string (which may contain a known suffix like `:high`)
  * and an explicit thinking config value. Returns `undefined` when no thinking is applicable
  * (e.g. no model was specified, or the model has no suffix and no config was provided). */

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

@@ -80,6 +80,8 @@ export interface RunSyncOptions {
 	modelOverride?: string;
 	/** Registry models available for heuristic bare-model resolution */
 	availableModels?: Array<{ provider: string; id: string; fullId: string }>;
+	/** Providers known to the registry before auth filtering */
+	knownModelProviders?: string[];
 	/** Current parent-session provider to prefer for ambiguous bare model ids */
 	preferredModelProvider?: string;
 	/** Current parent-session model to try after configured fallback models */

package/dist/builtin/web-access/CHANGELOG.md

@@ -4,6 +4,36 @@ All notable changes to this project will be documented in this file.
 
 ## [Unreleased]
 
+## [0.9.4-alpha.6] - 2026-07-01
+
+### Changed
+
+- Aligned the web-access extension peer dependency with upstream pi TUI `^0.80.3`; no web-access extension source changes were needed for this metadata sync.
+
+## [0.9.4-alpha.5] - 2026-07-01
+
+### Changed
+
+- Published a synchronized Atomic 0.9.4-alpha.5 prerelease for the web-access extension; no web-access extension changes were made after 0.9.4-alpha.1.
+
+## [0.9.4-alpha.4] - 2026-06-30
+
+### Changed
+
+- Published a synchronized Atomic 0.9.4-alpha.4 prerelease for the web-access extension; no web-access extension changes were made after 0.9.4-alpha.1.
+
+## [0.9.4-alpha.3] - 2026-06-30
+
+### Changed
+
+- Published a synchronized Atomic 0.9.4-alpha.3 prerelease for the web-access extension; no web-access extension changes were made after 0.9.4-alpha.1.
+
+## [0.9.4-alpha.1] - 2026-06-29
+
+### Changed
+
+- Published a synchronized Atomic 0.9.4-alpha.1 prerelease for the web-access extension; no web-access extension changes were made after 0.9.3.
+
 ## [0.9.3] - 2026-06-29
 
 ### Changed

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

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/web-access",
-  "version": "0.9.3",
+  "version": "0.9.4-alpha.10",
   "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": [
@@ -31,7 +31,7 @@
   },
   "peerDependencies": {
     "@bastani/atomic": "*",
-    "@earendil-works/pi-tui": "^0.80.2"
+    "@earendil-works/pi-tui": "^0.80.3"
   },
   "peerDependenciesMeta": {
     "@bastani/atomic": {

package/dist/builtin/workflows/CHANGELOG.md

@@ -6,6 +6,60 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 
 ## [Unreleased]
 
+## [0.9.4-alpha.9] - 2026-07-02
+
+### Changed
+
+- Changed the Claude Fable 5 reasoning level back from `high` to `xhigh` across all builtin workflow model chains (`ralph` prompt-engineer/reviewer-a/reviewer-b/reviewer-c, `goal` reviewer, `deep-research-codebase` planner, and `open-claude-design`), covering both the native `anthropic/claude-fable-5` entries and their OpenRouter mirrors; this reverts the `xhigh` → `high` change shipped in 0.9.4-alpha.8.
+- Curated all builtin workflow model chains against Atomic's agentic-coding benchmark (pass@1 / avg cost per task, 2026-07-02) under a role-based placement principle: **reviewers get best-in-class verification quality, every other stage gets the best measured performance-per-dollar**. Reviewer stages keep `anthropic/claude-fable-5:xhigh` (70% / $13.41, `ralph` reviewer-a and the `goal` reviewer), `openai-codex/gpt-5.5:xhigh` (67% / $7.23, reviewer-b), and `zai/glm-5.2:xhigh` (44% / $3.92, reviewer-c for third-family diversity). Non-reviewer stages moved to perf-per-dollar primaries: `ralph` prompt-engineer and the `deep-research-codebase` planner from `claude-fable-5:xhigh` to `openai-codex/gpt-5.5:xhigh` (−3pts for half the cost, with `fable-5:xhigh` as the first cross-family fallback), and the `deep-research-codebase` explorer from unbenchmarked `gpt-5.4-mini:low` to the measured cheap-tier value point `openai-codex/gpt-5.5:low` ($1.20 / task, 28 steps, 9.4k output tokens), backed only by measured fallbacks (`claude-opus-4.8:low` at 41% / $2.29, then `glm-5.2:high`). Unbenchmarked models (`gpt-5.4-mini`, `claude-haiku-4.5`) and unmeasured levels (`gemini-3.5-flash:low`) are removed entirely — every chain entry now corresponds to a benchmark datapoint. Dropped strictly dominated candidates from every chain: `claude-sonnet-5` (40–54% at $4.08–$26.40 per task with up to 268 steps), `claude-sonnet-4.6` (30% / $5.52), `gemini-3.1-pro` (12% / $9.48), and `gemini-3.5-flash` (37% / $7.34 with 276k output tokens). Moved `claude-opus-4.8` fallback entries from `:xhigh` to `:high` — its cost/value point (52% / $4.28 vs 54% / $8.01) — everywhere except `open-claude-design`, which keeps an Anthropic-led chain (`fable-5:xhigh` primary, Opus at `:xhigh`) since design quality, not $/task, is its objective. Added `anthropic/claude-fable-5:low` (60% / $3.76) as the Anthropic candidate in the medium-tier `ralph` research/orchestrator and `goal` worker chains.
+- Normalized all GLM-5.2 chain entries to the model's two real reasoning tiers: GLM-5.2's thinking-level map collapses `minimal`/`low`/`medium`/`high` to its "high" effort and `xhigh` to "max", so the previous `zai/glm-5.2:medium` entries in the `ralph` research/orchestrator and `goal` worker chains silently ran (and billed) at the "high" tier under a misleading label. Native `zai`/`zai-coding-cn` entries in medium-tier chains now say `:high` explicitly, judgment-tier chains keep `:xhigh` (→ "max"), and the `openrouter/z-ai/glm-5.2` mirror — which only maps `:xhigh` — is now always `:xhigh` instead of unsupported `:medium`.
+
+## [0.9.4-alpha.8] - 2026-07-02
+
+### Changed
+
+- Changed the Claude Fable 5 reasoning level from `xhigh` to `high` across all builtin workflow model chains (`ralph` prompt-engineer/reviewer-a/reviewer-b/reviewer-c, `goal` reviewer, `deep-research-codebase` planner, and `open-claude-design`), covering both the native `anthropic/claude-fable-5` entries and their OpenRouter mirrors.
+
+### Fixed
+
+- Fixed the builtin `goal` reviewer and `deep-research-codebase` planner model fallback chains missing the OpenRouter mirror of their primary model: `openrouter/anthropic/claude-fable-5` is now the first OpenRouter fallback candidate in both chains, matching the ordering already used by the `ralph` prompt-engineer/reviewer and `open-claude-design` chains.
+
+## [0.9.4-alpha.7] - 2026-07-02
+
+### Fixed
+
+- Fixed workflow stage model fallback so request/context incompatibility failures (HTTP 400/413/422 bad/unprocessable/payload-too-large request, unsupported tool/parameter, context-length/context-window overflow, `invalid_request`/`bad_request`/`too_large` errors) advance the chain to the next candidate instead of stopping. When none of the configured candidates can serve the request, the stage now falls back to the current user-selected model. Refusals, content-filter/safety blocks, cancellations, and task failures still stop the chain ([#1580](https://github.com/bastani-inc/atomic/issues/1580)).
+
+## [0.9.4-alpha.6] - 2026-07-01
+
+### Changed
+
+- Aligned the workflows extension peer dependency with upstream pi TUI `^0.80.3`; no workflows extension source changes were needed for this metadata sync.
+
+## [0.9.4-alpha.5] - 2026-07-01
+
+### Changed
+
+- Published a synchronized Atomic 0.9.4-alpha.5 prerelease for the workflows extension; no workflow changes were made after 0.9.4-alpha.3.
+
+## [0.9.4-alpha.4] - 2026-06-30
+
+### Changed
+
+- Published a synchronized Atomic 0.9.4-alpha.4 prerelease for the workflows extension; no workflow changes were made after 0.9.4-alpha.3.
+
+## [0.9.4-alpha.3] - 2026-06-30
+
+### Fixed
+
+- Classified generic provider `Connection error.` / `fetch failed` transport outages as explicit workflow transport errors so stage fallback metadata can distinguish them from ordinary provider/model outages while preserving the normal next-model fallback order.
+
+## [0.9.4-alpha.1] - 2026-06-29
+
+### Fixed
+
+- Fixed workflow lifecycle notices for structured workflow outputs whose returned `status` is `failed` or `blocked`: the runtime now records those runs as failed/blocked instead of successful completions, preserves their returned output, carries blocked summaries into lifecycle notice reasons, marks returned failures non-resumable with explicit terminal failure metadata, and emits failure/blocked lifecycle cards rather than success/checkmark notices.
+
 ## [0.9.3] - 2026-06-29
 
 ### Breaking Changes

package/dist/builtin/workflows/README.md

@@ -31,18 +31,18 @@ Adding workflow files under `.atomic/workflows/` (project scope) or `~/.atomic/a
 
 ### Workflow lifecycle notifications
 
-Workflow lifecycle notices are enabled by default. They send steer prompts into the main chat/model context when a run completes or fails. Awaiting-input prompts are tracked for dedupe/restore, but they do not wake the main chat agent. Configure lifecycle tracking in the same extension config file:
+Workflow lifecycle notices are enabled by default. They send steer prompts into the main chat/model context when a run completes, fails, or ends blocked. Awaiting-input prompts are tracked for dedupe/restore, but they do not wake the main chat agent. Configure lifecycle tracking in the same extension config file:
 
 ```json
 {
   "workflowNotifications": {
     "enabled": true,
-    "notifyOn": ["completed", "failed", "awaiting_input"]
+    "notifyOn": ["completed", "failed", "blocked", "awaiting_input"]
   }
 }
 ```
 
-Set `enabled` to `false` to disable all lifecycle notices, or narrow `notifyOn` to a non-empty list of selected events. Completion and failure lifecycle notices are emitted for top-level workflow runs, use steer delivery, and wake an idle model so the lifecycle update enters the model context when it happens. Nested child workflow completion/failure is reflected inside the expanded parent graph instead of producing separate top-level completion cards. Awaiting-input states are tracked for dedupe/restore, but workflows do not enqueue main-chat `/workflow connect` cards for them; prompt state remains visible through workflow status/connect surfaces, avoiding stale actionable cards if a prompt resolves while the main chat is streaming.
+Set `enabled` to `false` to disable all lifecycle notices, or narrow `notifyOn` to a non-empty list of selected events. Completion, failure, and blocked lifecycle notices are emitted for top-level workflow runs, use steer delivery, and wake an idle model so the lifecycle update enters the model context when it happens. Nested child workflow completion/failure is reflected inside the expanded parent graph instead of producing separate top-level completion cards. Awaiting-input states are tracked for dedupe/restore, but workflows do not enqueue main-chat `/workflow connect` cards for them; prompt state remains visible through workflow status/connect surfaces, avoiding stale actionable cards if a prompt resolves while the main chat is streaming.
 
 When a stage human-in-the-loop prompt is answered from the workflow TUI/stage chat, workflows also emits a separate display-only `workflows:hil-answer-notice` custom message. It records the answer for user-visible audit, but it does not wake the main agent, enter LLM context, or authorize answering later workflow prompts. Answers sent by the main-chat `workflow` tool do not emit this notice because the tool result already tells the main agent what happened.
 
@@ -736,7 +736,7 @@ Config-based discovery (`~/.atomic/agent/extensions/workflow/config.json` or `.a
   },
   "workflowNotifications": {
     "enabled": true,
-    "notifyOn": ["completed", "failed", "awaiting_input"]
+    "notifyOn": ["completed", "failed", "blocked", "awaiting_input"]
   }
 }
 ```

package/dist/builtin/workflows/builtin/deep-research-codebase-utils.ts

@@ -31,41 +31,43 @@ export interface DeepResearchCodebaseResult {
 
 export const FILE_ONLY_OUTPUT = "file-only" satisfies WorkflowOutputMode;
 
+// Chains curated from Atomic's agentic-coding benchmark (see ralph-models.ts
+// for the frontier data and placement principle: only reviewers get
+// best-in-class models; the planner is a perf-per-dollar stage).
 export const PLANNER_MODEL_CONFIG = {
-  model: "anthropic/claude-fable-5:xhigh",
+  model: "openai-codex/gpt-5.5:xhigh",
   fallbackModels: [
-    "openai-codex/gpt-5.5:xhigh",
     "github-copilot/gpt-5.5:xhigh",
     "openai/gpt-5.5:xhigh",
-    "github-copilot/claude-opus-4.8 (1m):xhigh",
-    "anthropic/claude-opus-4-8:xhigh",
+    "anthropic/claude-fable-5:xhigh",
+    "github-copilot/claude-opus-4.8 (1m):high",
+    "anthropic/claude-opus-4-8:high",
     "zai/glm-5.2:xhigh",
     "zai-coding-cn/glm-5.2:xhigh",
-    "github-copilot/gemini-3.5-flash (1m):high",
-    "google/gemini-3.5-flash:high",
-    "google-vertex/gemini-3.5-flash:high",
-    "github-copilot/gemini-3.1-pro-preview (1m):high",
-    "google/gemini-3.1-pro-preview:high",
-    "google-vertex/gemini-3.1-pro-preview:high",
     "openrouter/openai/gpt-5.5:xhigh",
-    "openrouter/anthropic/claude-opus-4-8:xhigh",
-    "openrouter/z-ai/glm-5.2:xhigh",
-    "openrouter/google/gemini-3.5-flash:high",
-    "openrouter/google/gemini-3.1-pro-preview:high"
+    "openrouter/anthropic/claude-fable-5:xhigh",
+    "openrouter/anthropic/claude-opus-4-8:high",
+    "openrouter/z-ai/glm-5.2:xhigh"
   ],
   excludedTools: ["ask_user_question"],
 } as const;
 
+// gpt-5.5:low is the measured cheap-tier diamond ($1.20/task, 28 steps, 9.4k
+// output tokens). Benchmark-measured models only: opus-4.8:low (41%/$2.29)
+// and glm-5.2:high (36%/$2.84) back it up; unbenchmarked minis/haiku/flash
+// are excluded.
 export const EXPLORER_MODEL_CONFIG = {
-  model: "openai-codex/gpt-5.4-mini:low",
+  model: "openai-codex/gpt-5.5:low",
   fallbackModels: [
-    "github-copilot/gpt-5.4-mini:low",
-    "openai/gpt-5.4-mini:low",
-    "github-copilot/claude-haiku-4.5:low",
-    "anthropic/claude-haiku-4-5:low",
-    "github-copilot/gemini-3.5-flash (1m):low",
-    "google/gemini-3.5-flash:low",
-    "google-vertex/gemini-3.5-flash:low",
+    "github-copilot/gpt-5.5:low",
+    "openai/gpt-5.5:low",
+    "github-copilot/claude-opus-4.8 (1m):low",
+    "anthropic/claude-opus-4-8:low",
+    "zai/glm-5.2:high",
+    "zai-coding-cn/glm-5.2:high",
+    "openrouter/openai/gpt-5.5:low",
+    "openrouter/anthropic/claude-opus-4-8:low",
+    "openrouter/z-ai/glm-5.2:xhigh",
   ],
   excludedTools: ["ask_user_question"],
 } as const;

package/dist/builtin/workflows/builtin/goal-runner.ts

@@ -92,26 +92,22 @@ export async function runGoalWorkflow(ctx: GoalRunnerContext, options: GoalWorkf
     const comparisonBaseBranch = normalizeBranchInput(inputs.base_branch, "origin/main");
     const { ledger, ledgerPath, artifactDir } = await createGoalLedger(objective);
 
+    // Chains curated from Atomic's agentic-coding benchmark (see
+    // ralph-models.ts for the frontier data and drop rationale).
     const workerModelConfig = {
       model: "openai-codex/gpt-5.5:medium",
       fallbackModels: [
           "github-copilot/gpt-5.5:medium",
           "openai/gpt-5.5:medium",
+          "anthropic/claude-fable-5:low",
           "github-copilot/claude-opus-4.8 (1m):medium",
           "anthropic/claude-opus-4-8:medium",
-          "zai/glm-5.2:medium",
-          "zai-coding-cn/glm-5.2:medium",
-          "github-copilot/gemini-3.5-flash (1m):medium",
-          "google/gemini-3.5-flash:medium",
-          "google-vertex/gemini-3.5-flash:medium",
-          "github-copilot/gemini-3.1-pro-preview (1m):medium",
-          "google/gemini-3.1-pro-preview:medium",
-          "google-vertex/gemini-3.1-pro-preview:medium",
+          "zai/glm-5.2:high",
+          "zai-coding-cn/glm-5.2:high",
           "openrouter/openai/gpt-5.5:medium",
+          "openrouter/anthropic/claude-fable-5:low",
           "openrouter/anthropic/claude-opus-4-8:medium",
-          "openrouter/z-ai/glm-5.2:medium",
-          "openrouter/google/gemini-3.5-flash:medium",
-          "openrouter/google/gemini-3.1-pro-preview:medium"
+          "openrouter/z-ai/glm-5.2:xhigh"
       ],
       tools: goalRunnerTools,
     };
@@ -122,21 +118,14 @@ export async function runGoalWorkflow(ctx: GoalRunnerContext, options: GoalWorkf
           "openai-codex/gpt-5.5:xhigh",
           "github-copilot/gpt-5.5:xhigh",
           "openai/gpt-5.5:xhigh",
-          "github-copilot/claude-opus-4.8 (1m):xhigh",
-          "anthropic/claude-opus-4-8:xhigh",
+          "github-copilot/claude-opus-4.8 (1m):high",
+          "anthropic/claude-opus-4-8:high",
           "zai/glm-5.2:xhigh",
           "zai-coding-cn/glm-5.2:xhigh",
-          "github-copilot/gemini-3.5-flash (1m):high",
-          "google/gemini-3.5-flash:high",
-          "google-vertex/gemini-3.5-flash:high",
-          "github-copilot/gemini-3.1-pro-preview (1m):high",
-          "google/gemini-3.1-pro-preview:high",
-          "google-vertex/gemini-3.1-pro-preview:high",
+          "openrouter/anthropic/claude-fable-5:xhigh",
           "openrouter/openai/gpt-5.5:xhigh",
-          "openrouter/anthropic/claude-opus-4-8:xhigh",
-          "openrouter/z-ai/glm-5.2:xhigh",
-          "openrouter/google/gemini-3.5-flash:high",
-          "openrouter/google/gemini-3.1-pro-preview:high"
+          "openrouter/anthropic/claude-opus-4-8:high",
+          "openrouter/z-ai/glm-5.2:xhigh"
       ],
       tools: goalRunnerTools,
       schema: reviewDecisionSchema,

package/dist/builtin/workflows/builtin/open-claude-design-runner.ts

@@ -74,6 +74,9 @@ export async function runOpenClaudeDesignWorkflow(ctx: OpenClaudeDesignContext):
     });
   }
 
+  // Anthropic-heavy chain for design taste; sonnet-5/sonnet-4.6 dropped as
+  // strictly dominated in Atomic's benchmark (see ralph-models.ts). Opus stays
+  // at :xhigh here — visual quality, not $/task, is the objective for design.
   const designModelConfig = {
     model: "anthropic/claude-fable-5:xhigh",
     fallbackModels: [
@@ -81,12 +84,9 @@ export async function runOpenClaudeDesignWorkflow(ctx: OpenClaudeDesignContext):
       "anthropic/claude-opus-4-8:xhigh",
       "zai/glm-5.2:xhigh",
       "zai-coding-cn/glm-5.2:xhigh",
-      "github-copilot/claude-sonnet-4.6 (1m):high",
-      "anthropic/claude-sonnet-4-6:high",
       "openrouter/anthropic/claude-fable-5:xhigh",
       "openrouter/anthropic/claude-opus-4-8:xhigh",
-      "openrouter/z-ai/glm-5.2:xhigh",
-      "openrouter/anthropic/claude-sonnet-4-6:high"
+      "openrouter/z-ai/glm-5.2:xhigh"
     ],
   };
   // Phase 1: combined discovery + init — one stage interviews the user via