@code-yeongyu/senpi 2026.10.16-2 → 2026.10.16-3

package/dist/core/extensions/builtin/index.js

@@ -47,6 +47,7 @@ import toolPairGuardExtension from "./tool-pair-guard/index.js";
 import toolSearchExtension from "./tool-search/index.js";
 import tpsExtension from "./tps.js";
 import ttsrExtension from "./ttsr/index.js";
+import vertexModelGardenExtension from "./vertex-model-garden/index.js";
 import videoInExtension from "./video-in/index.js";
 import webfetchExtension from "./webfetch/index.js";
 import websearchExtension from "./websearch/index.js";
@@ -117,6 +118,8 @@ export const builtinExtensions = [
     { id: "claude-sdk-oauth", factory: anthropicSubscriptionExtension },
     // Registers unconditionally and reports executable/auth state through its oauth check, so it stays beside the other provider lane.
     { id: "cursor-cli-oauth", factory: cursorCliOauthExtension },
+    // Provider-only lane beside the other provider registrations; no hooks, so order is not load-bearing.
+    { id: "vertex-model-garden", factory: vertexModelGardenExtension },
     // Config reload follows settings-dependent builtins so reloads rebuild their resolved settings before catalog feeders observe them.
     { id: "config-reload", factory: configReloadExtension },
     // Shared catalog wiring loads before MCP, which feeds its tools into the shared catalog as the final builtin.

package/dist/core/extensions/builtin/prompt-preset/contract.js

@@ -1,4 +1,4 @@
-import { extractClaudeOpusVersion, extractGpt5Version, isClaudeFable5Model, isClaudeFable51Model, isClaudeHaiku55Model, isClaudeOpus5Model, isClaudeOpus55Model, isClaudeSonnet55Model, isDeepseekV4Flash0731Model, isDeepseekV4FlashModel, isDeepseekV4ProModel, isDeepseekV41FlashModel, isGlm52Model, isGlm53Model, isGpt6FamilyModel, isGrok45Model, isGrok46Model, isGrok47Model, isKimiK3Model, isKimiK26Model, isKimiK27Model, isKimiK28Model, isSWE2Model, } from "./matchers.js";
+import { extractClaudeOpusVersion, extractGpt5Version, isClaudeFable5Model, isClaudeFable51Model, isClaudeHaiku55Model, isClaudeOpus5Model, isClaudeOpus55Model, isClaudeSonnet55Model, isDeepseekV4Flash0731Model, isDeepseekV4FlashModel, isDeepseekV4ProModel, isDeepseekV41FlashModel, isGemini4ArgonModel, isGlm52Model, isGlm53Model, isGpt6FamilyModel, isGrok45Model, isGrok46Model, isGrok47Model, isKimiK3Model, isKimiK26Model, isKimiK27Model, isKimiK28Model, isSWE2Model, } from "./matchers.js";
 export function resolvePresetName(input) {
     const model = { id: input.modelID, provider: input.providerID, name: input.name };
     if (isGpt6FamilyModel(model)) {
@@ -72,6 +72,9 @@ export function resolvePresetName(input) {
     if (isGrok45Model(model)) {
         return "grok-4.5";
     }
+    if (isGemini4ArgonModel(model)) {
+        return "gemini-4-argon";
+    }
     return undefined;
 }
 export const PROMPT_PRESET_MODEL_CASES = [
@@ -104,6 +107,7 @@ export const PROMPT_PRESET_MODEL_CASES = [
     { providerID: "xai", modelID: "grok-4.7", preset: "grok-4.7" },
     { providerID: "xai", modelID: "grok-4.6", preset: "grok-4.6" },
     { providerID: "xai", modelID: "grok-4.5", preset: "grok-4.5" },
+    { providerID: "google", modelID: "gemini-4-argon", preset: "gemini-4-argon" },
 ];
 // Presets a model id can auto-resolve to. gpt-5 is a manual-only preset (settings union +
 // buildPreset, no model matcher), so it is intentionally absent here: a model-driven parity check
@@ -138,5 +142,6 @@ export const AUTO_RESOLVED_PRESET_NAMES = [
     "claude-opus-4-7",
     "claude-opus-4-6",
     "claude-opus-4-5",
+    "gemini-4-argon",
 ];
 //# sourceMappingURL=contract.js.map

package/dist/core/extensions/builtin/prompt-preset/gemini-4-argon.d.ts

@@ -0,0 +1,7 @@
+import { type BuildDynamicSystemPromptOptions } from "../../../dynamic-prompt/build.ts";
+export declare const GEMINI_4_ARGON_ROUTING_LINE_RULE = "The line is plain prose, and the tool calls that start the work follow it in the same response";
+export declare const GEMINI_4_ARGON_WORKING_TASK_OPENING = "Writing code is your superpower. A step is solved by writing one `eval` cell that does the work, not by a parade of one-off tool calls: every session tool is reachable from the cell as `tool.<name>(...)`, so the cell reads, searches, inspects, runs commands, and edits files with real control flow - a `for` over every target, an `if`/`else` per case, a `try`/`catch` per item - and fires every independent lookup AT ONCE through `parallel(thunks)`. Bias toward breadth when context is thin: pull in anything even loosely relevant in that wave instead of serially later, since wasted reads cost almost nothing and stale assumptions cost the turn. Reduce in the kernel to the facts the decision needs, keeping every failed or missing item verbatim; a result that hides a failure or a truncated tail is not evidence.\n\nSort the step before you write the cell: independent reads, searches, probes, and edits share one cell; a call whose input is a result you have not seen yet, a deploy, an approval, or a destructive command runs ONE ACTION AT A TIME, each observed before the next. Name the state a cell should produce before it runs and compare the return with that state. When the result must be SEEN rather than read - a page, an image, a layout - make one change, render it, look, then make the next. A direct tool call is the exception, for a single call that settles the step or an action that needs approval; files are read, searched, and edited through the cell's tool calls, never through shell text processing or inline scripts inside a command. Never fill missing parameters with placeholders.";
+export declare const GEMINI_4_ARGON_ERROR_RULE = "Never silently swallow errors. Retry a transient failure (timeout, rate limit, \"try again\") as is; on any other failure change the strategy or the arguments - never repeat the same failed call, and never shotgun-debug with unrelated edits.";
+export declare const GEMINI_4_ARGON_SEARCH_GROUNDING = "## Search Grounding\n\nToday's date is in the environment context and your training data ends before it: when a query is time-sensitive, form search queries from that date, and search before answering about anything \"latest\"; facts that cannot change need no search.";
+export declare function buildGemini4ArgonPrompt(options: BuildDynamicSystemPromptOptions): string;
+//# sourceMappingURL=gemini-4-argon.d.ts.map

package/dist/core/extensions/builtin/prompt-preset/gemini-4-argon.js

@@ -0,0 +1,83 @@
+// Argon system prompt (senpi#3385).
+//
+// 2026-10-15. The provider's published guidance for the previous generation, which
+// Argon continues (model docs: latest-model, prompting-strategies and "Agentic
+// workflows", function-calling "Workarounds for pre-tool text", FAQ 8),
+// says the model answers direct, concise instructions best and over-analyzes the
+// shouting style older prompts used; structured text demanded right before a tool
+// call can fail the call; the Pro line reaches for shell commands over custom
+// tools; the newer releases run diagnostic code before changing anything. The
+// default dynamic core already carries one home per rule, so this preset keeps
+// every default section and changes four spots at their source (prompt-engineering
+// A/B/C, no appended rule), plus the plain `gemini` workstation dialect:
+// - identity names the model, as the Grok preset does (C: only the harness knows
+//   which model is loaded).
+// - Intent Gate (terminal): the routing line is plain prose and the tool calls that
+//   start the work follow in the same response (B: the pre-tool-text warning and
+//   the previous generation's "narrates instead of calling" failure the 2026-02-22 omo
+//   overlays and oh-my-pi's reasoning-stall interrupt both guard against).
+// - Working the Task opening, only when `eval` is a session tool: writing code is
+//   the model's superpower, one eval cell per step with real control flow and
+//   `parallel(thunks)`, one action at a time only for unseen-input, deploy,
+//   approval and destructive calls (owner steer 2026-10-15; voice of the ultrawork
+//   skill's "Parallel execution" section; Google: 10-20 active tools, parallel
+//   calls when independent, code-first inspection).
+// - Policies: transient failures retry as is, other failures change strategy and
+//   never repeat the same call (B: the Agentic workflows template, rule 8.2).
+// - `## Search Grounding`, only when `web_search` is a session tool (C: the
+//   prompting-strategies "current day" clause; the date itself stays in the
+//   environment-context message, senpi#2093).
+// Byte-identical to the default everywhere else; prompt-presets-gemini-4-argon.test.ts
+// checks that section by section.
+import { APP_NAME } from "../../../../config.js";
+import { buildDynamicSystemPrompt, } from "../../../dynamic-prompt/build.js";
+import { buildHandoffSection } from "../../../dynamic-prompt/handoff.js";
+import { buildIntentGate } from "../../../dynamic-prompt/intent-gate.js";
+import { buildPoliciesSection } from "../../../dynamic-prompt/policies.js";
+import { buildStyleSection } from "../../../dynamic-prompt/style.js";
+import { buildVerificationSection } from "../../../dynamic-prompt/verification.js";
+import { buildWorkingTaskSection } from "../../../dynamic-prompt/working-task.js";
+export const GEMINI_4_ARGON_ROUTING_LINE_RULE = "The line is plain prose, and the tool calls that start the work follow it in the same response";
+export const GEMINI_4_ARGON_WORKING_TASK_OPENING = `Writing code is your superpower. A step is solved by writing one \`eval\` cell that does the work, not by a parade of one-off tool calls: every session tool is reachable from the cell as \`tool.<name>(...)\`, so the cell reads, searches, inspects, runs commands, and edits files with real control flow - a \`for\` over every target, an \`if\`/\`else\` per case, a \`try\`/\`catch\` per item - and fires every independent lookup AT ONCE through \`parallel(thunks)\`. Bias toward breadth when context is thin: pull in anything even loosely relevant in that wave instead of serially later, since wasted reads cost almost nothing and stale assumptions cost the turn. Reduce in the kernel to the facts the decision needs, keeping every failed or missing item verbatim; a result that hides a failure or a truncated tail is not evidence.
+
+Sort the step before you write the cell: independent reads, searches, probes, and edits share one cell; a call whose input is a result you have not seen yet, a deploy, an approval, or a destructive command runs ONE ACTION AT A TIME, each observed before the next. Name the state a cell should produce before it runs and compare the return with that state. When the result must be SEEN rather than read - a page, an image, a layout - make one change, render it, look, then make the next. A direct tool call is the exception, for a single call that settles the step or an action that needs approval; files are read, searched, and edited through the cell's tool calls, never through shell text processing or inline scripts inside a command. Never fill missing parameters with placeholders.`;
+export const GEMINI_4_ARGON_ERROR_RULE = 'Never silently swallow errors. Retry a transient failure (timeout, rate limit, "try again") as is; on any other failure change the strategy or the arguments - never repeat the same failed call, and never shotgun-debug with unrelated edits.';
+export const GEMINI_4_ARGON_SEARCH_GROUNDING = `## Search Grounding
+
+Today's date is in the environment context and your training data ends before it: when a query is time-sensitive, form search queries from that date, and search before answering about anything "latest"; facts that cannot change need no search.`;
+function hasTool(context, name) {
+    return context.tools.some((tool) => tool.name === name);
+}
+function buildGemini4ArgonCore(context) {
+    const appName = context.appName ?? APP_NAME;
+    return [
+        `You are ${appName}, a coding agent running on Gemini 4 Argon. Your work should be indistinguishable from a careful senior engineer's.`,
+        "",
+        buildIntentGate({
+            tools: context.tools,
+            surface: context.surface,
+            routingLineRule: GEMINI_4_ARGON_ROUTING_LINE_RULE,
+        }),
+        "",
+        buildWorkingTaskSection(hasTool(context, "eval") ? { opening: GEMINI_4_ARGON_WORKING_TASK_OPENING } : {}),
+        "",
+        buildVerificationSection({ surface: context.surface }),
+        "",
+        context.toolSection,
+        "",
+        buildPoliciesSection({ errorRule: GEMINI_4_ARGON_ERROR_RULE }),
+        "",
+        buildHandoffSection({ surface: context.surface }),
+        "",
+        buildStyleSection({ surface: context.surface }),
+    ].join("\n");
+}
+export function buildGemini4ArgonPrompt(options) {
+    return buildDynamicSystemPrompt({
+        ...options,
+        corePrompt: buildGemini4ArgonCore,
+        ...(options.selectedTools.includes("web_search") ? { tuningSection: GEMINI_4_ARGON_SEARCH_GROUNDING } : {}),
+        workstationDialect: "gemini",
+    });
+}
+//# sourceMappingURL=gemini-4-argon.js.map

package/dist/core/extensions/builtin/prompt-preset/matchers.d.ts

@@ -51,6 +51,8 @@ export declare const CLAUDE_SONNET_55_MARKERS: readonly ["sonnet-5-5", "sonnet-5
 export declare function isClaudeSonnet55Model(modelId: string): boolean;
 export declare const CLAUDE_HAIKU_55_MARKERS: readonly ["haiku-5-5", "haiku-5.5"];
 export declare function isClaudeHaiku55Model(modelId: string): boolean;
+export declare function hasGemini4ArgonSignal(value: string): boolean;
+export declare function isGemini4ArgonModel(model: ModelWithPromptPresetMetadata): boolean;
 export type ClaudeOpusVersion = "claude-opus-4-8" | "claude-opus-4-7" | "claude-opus-4-6" | "claude-opus-4-5";
 export declare function extractClaudeOpusVersion(modelId: string): ClaudeOpusVersion | undefined;
 //# sourceMappingURL=matchers.d.ts.map

package/dist/core/extensions/builtin/prompt-preset/matchers.js

@@ -191,6 +191,18 @@ export function isClaudeHaiku55Model(modelId) {
     const normalized = normalizeModelId(modelId);
     return CLAUDE_HAIKU_55_MARKERS.some((marker) => normalized.includes(marker));
 }
+// Argon id shapes verified against the generated catalogs (2026-10-15):
+// gemini-4-argon (google, google-vertex, github-copilot, opencode, venice),
+// google/gemini-4-argon and its :batch row (openrouter, vercel-ai-gateway),
+// opencode-go's gemini-4-argon-contributor, the Antigravity -low/-medium/-high
+// SKUs, and the catalog display names of the plain and batch rows.
+// Bare gemini-4, other generations of the family and unknown siblings stay out.
+export function hasGemini4ArgonSignal(value) {
+    return /(?:^|[/@:._-])gemini[._-]?4[._-]argon(?:$|[/@:._-])/.test(normalizeModelId(value));
+}
+export function isGemini4ArgonModel(model) {
+    return hasGemini4ArgonSignal(model.id) || (model.name !== undefined && hasGemini4ArgonSignal(model.name));
+}
 export function extractClaudeOpusVersion(modelId) {
     const normalized = normalizeModelId(modelId);
     if (normalized.includes("opus-4-8")) {

package/dist/core/extensions/builtin/prompt-preset/presets.js

@@ -13,6 +13,7 @@ import { buildDeepseekV41FlashPrompt } from "./deepseek-v4-1-flash.js";
 import { buildDeepseekV4FlashPrompt } from "./deepseek-v4-flash.js";
 import { buildDeepseekV4Flash0731Prompt } from "./deepseek-v4-flash-0731.js";
 import { buildDeepseekV4ProPrompt } from "./deepseek-v4-pro.js";
+import { buildGemini4ArgonPrompt } from "./gemini-4-argon.js";
 import { buildGlm52Prompt } from "./glm-5-2.js";
 import { buildGlm53Prompt } from "./glm-5-3.js";
 import { buildGpt52Prompt } from "./gpt-5.2.js";
@@ -74,6 +75,8 @@ export function buildPreset(name, options) {
             return { name, prompt: buildGrok46Prompt(options) };
         case "grok-4.5":
             return { name, prompt: buildGrok45Prompt(options) };
+        case "gemini-4-argon":
+            return { name, prompt: buildGemini4ArgonPrompt(options) };
         case "kimi-k3":
             return { name, prompt: buildKimiK3Prompt(options) };
         case "kimi-k2-8":

package/dist/core/extensions/builtin/prompt-preset/settings.d.ts

@@ -1,5 +1,5 @@
 import type { SettingsManager } from "../../../settings-manager.ts";
-export type PromptPresetName = "auto" | "claude-fable-5" | "claude-fable-5-1" | "claude-opus-5-5" | "claude-opus-5" | "claude-sonnet-5-5" | "claude-haiku-5-5" | "claude-opus-4-8" | "claude-opus-4-7" | "claude-opus-4-6" | "claude-opus-4-5" | "deepseek-v4-flash" | "deepseek-v4-flash-0731" | "deepseek-v4-1-flash" | "deepseek-v4-pro" | "glm-5.2" | "glm-5.3" | "grok-4.5" | "grok-4.6" | "grok-4.7" | "kimi-k3" | "kimi-k2-8" | "kimi-k2-7" | "kimi-k2-6" | "gpt-5" | "gpt-5.2" | "gpt-5.3-codex" | "gpt-5.4" | "gpt-5.5" | "gpt-5.6" | "gpt-6-astra";
+export type PromptPresetName = "auto" | "claude-fable-5" | "claude-fable-5-1" | "claude-opus-5-5" | "claude-opus-5" | "claude-sonnet-5-5" | "claude-haiku-5-5" | "claude-opus-4-8" | "claude-opus-4-7" | "claude-opus-4-6" | "claude-opus-4-5" | "deepseek-v4-flash" | "deepseek-v4-flash-0731" | "deepseek-v4-1-flash" | "deepseek-v4-pro" | "glm-5.2" | "glm-5.3" | "gemini-4-argon" | "grok-4.5" | "grok-4.6" | "grok-4.7" | "kimi-k3" | "kimi-k2-8" | "kimi-k2-7" | "kimi-k2-6" | "gpt-5" | "gpt-5.2" | "gpt-5.3-codex" | "gpt-5.4" | "gpt-5.5" | "gpt-5.6" | "gpt-6-astra";
 export interface PromptPresetSettings {
     promptPreset: PromptPresetName;
 }

package/dist/core/extensions/builtin/prompt-preset/settings.js

@@ -16,6 +16,7 @@ export const VALID_PRESETS = new Set([
     "deepseek-v4-pro",
     "glm-5.2",
     "glm-5.3",
+    "gemini-4-argon",
     "grok-4.5",
     "grok-4.6",
     "grok-4.7",

package/dist/core/extensions/builtin/vertex-model-garden/auth.d.ts

@@ -0,0 +1,16 @@
+import type { AuthContext, AuthResult } from "@earendil-works/pi-ai";
+/** The ADC file `gcloud auth application-default login` writes, same as the builtin google-vertex provider. */
+export declare const VERTEX_MODEL_GARDEN_ADC_PATH = "~/.config/gcloud/application_default_credentials.json";
+export declare const VERTEX_MODEL_GARDEN_ADC_SOURCE = "gcloud application default credentials";
+/**
+ * Availability through the shared Google Application Default Credentials chain.
+ * This is deliberately ambient (senpi#2327): the credentials exist for other
+ * gcloud tooling too, so the lane is selectable but never outranks a provider the
+ * user configured, and it reports nothing while the chain is absent — the failure
+ * belongs at selection time, not at the first request.
+ */
+export declare function resolveVertexModelGardenAmbient(input: {
+    ctx: AuthContext;
+    signal: AbortSignal;
+}): Promise<AuthResult | undefined>;
+//# sourceMappingURL=auth.d.ts.map

package/dist/core/extensions/builtin/vertex-model-garden/auth.js

@@ -0,0 +1,19 @@
+/** The ADC file `gcloud auth application-default login` writes, same as the builtin google-vertex provider. */
+export const VERTEX_MODEL_GARDEN_ADC_PATH = "~/.config/gcloud/application_default_credentials.json";
+export const VERTEX_MODEL_GARDEN_ADC_SOURCE = "gcloud application default credentials";
+/**
+ * Availability through the shared Google Application Default Credentials chain.
+ * This is deliberately ambient (senpi#2327): the credentials exist for other
+ * gcloud tooling too, so the lane is selectable but never outranks a provider the
+ * user configured, and it reports nothing while the chain is absent — the failure
+ * belongs at selection time, not at the first request.
+ */
+export async function resolveVertexModelGardenAmbient(input) {
+    input.signal.throwIfAborted();
+    const path = (await input.ctx.env("GOOGLE_APPLICATION_CREDENTIALS")) ?? VERTEX_MODEL_GARDEN_ADC_PATH;
+    input.signal.throwIfAborted();
+    if (!(await input.ctx.fileExists(path)))
+        return undefined;
+    return { auth: {}, source: VERTEX_MODEL_GARDEN_ADC_SOURCE, ambient: true };
+}
+//# sourceMappingURL=auth.js.map

package/dist/core/extensions/builtin/vertex-model-garden/index.d.ts

@@ -0,0 +1,6 @@
+import type { ExtensionAPI } from "../../types.ts";
+import { type ResolveVertexAccess } from "./transport.ts";
+export declare const VERTEX_MODEL_GARDEN_PROVIDER_ID = "vertex-garden";
+export declare function registerVertexModelGarden(pi: ExtensionAPI, resolveAccess?: ResolveVertexAccess): void;
+export default function vertexModelGardenExtension(pi: ExtensionAPI): void;
+//# sourceMappingURL=index.d.ts.map

package/dist/core/extensions/builtin/vertex-model-garden/index.js

@@ -0,0 +1,27 @@
+import { resolveVertexModelGardenAmbient } from "./auth.js";
+import { VERTEX_MODEL_GARDEN_MODELS } from "./models.js";
+import { defaultResolveVertexAccess, streamVertexModelGarden } from "./transport.js";
+export const VERTEX_MODEL_GARDEN_PROVIDER_ID = "vertex-garden";
+// The request URL is built per call from the ADC-resolved project; this only satisfies registration.
+const REGISTRATION_BASE_URL = "https://aiplatform.googleapis.com/v1";
+export function registerVertexModelGarden(pi, resolveAccess = defaultResolveVertexAccess) {
+    pi.registerProvider(VERTEX_MODEL_GARDEN_PROVIDER_ID, {
+        name: "Vertex AI Model Garden",
+        api: "openai-completions",
+        baseUrl: REGISTRATION_BASE_URL,
+        // Ambient like the builtin google-vertex ADC lane (senpi#2327): available
+        // while the ADC chain is present, never an explicitly configured provider,
+        // and never the automatic default over one the user configured.
+        ambientAuth: {
+            name: "Google Cloud credentials (ADC)",
+            resolve: resolveVertexModelGardenAmbient,
+        },
+        models: VERTEX_MODEL_GARDEN_MODELS,
+        streamSimple: (model, context, options) => streamVertexModelGarden(model, context, options, resolveAccess),
+        fallbackEligible: () => false,
+    });
+}
+export default function vertexModelGardenExtension(pi) {
+    registerVertexModelGarden(pi);
+}
+//# sourceMappingURL=index.js.map

package/dist/core/extensions/builtin/vertex-model-garden/models.d.ts

@@ -0,0 +1,3 @@
+import type { ProviderModelConfig } from "../../types.ts";
+export declare const VERTEX_MODEL_GARDEN_MODELS: ProviderModelConfig[];
+//# sourceMappingURL=models.d.ts.map

package/dist/core/extensions/builtin/vertex-model-garden/models.js

@@ -0,0 +1,34 @@
+// This starter list covers open models through the global OpenAI-compatible endpoint.
+// Marketplace partner models are excluded: they are card-billed and promotional credits do not apply.
+// Prices are the model card's published MaaS list prices, USD per million tokens
+// (https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing,
+// "Deepseek's models" and "GLM's models"); DeepSeek V4.1 Flash is listed at no charge through 2026-10-16.
+// These endpoints accept no thinking parameters, so no level may be offered as steerable.
+const NO_THINKING_CONTROLS = {
+    off: null,
+    minimal: null,
+    low: null,
+    medium: null,
+    high: null,
+    xhigh: null,
+    max: null,
+};
+const OPENAPI_COMPAT = { supportsDeveloperRole: false, supportsReasoningEffort: false };
+function garden(id, name, cost, contextWindow) {
+    return {
+        id,
+        name,
+        reasoning: true,
+        thinkingLevelMap: NO_THINKING_CONTROLS,
+        input: ["text"],
+        cost,
+        contextWindow,
+        maxTokens: 32_000,
+        compat: OPENAPI_COMPAT,
+    };
+}
+export const VERTEX_MODEL_GARDEN_MODELS = [
+    garden("deepseek-ai/deepseek-v4.1-flash-maas", "DeepSeek V4.1 Flash (Vertex)", { input: 0.3, output: 1.2, cacheRead: 0.03, cacheWrite: 0 }, 1_000_000),
+    garden("zai-org/glm-5.2-maas", "GLM 5.2 (Vertex)", { input: 1.4, output: 4.4, cacheRead: 0.14, cacheWrite: 0 }, 200_000),
+];
+//# sourceMappingURL=models.js.map

package/dist/core/extensions/builtin/vertex-model-garden/transport.d.ts

@@ -0,0 +1,14 @@
+import { type Api, type AssistantMessageEventStream, type Model, type SimpleStreamOptions, type TranscriptContext } from "@earendil-works/pi-ai";
+import { type VertexAdcAccess } from "@earendil-works/pi-ai/api/google-vertex-adc.lazy";
+export declare const VERTEX_MODEL_GARDEN_LOCATION = "global";
+export type ResolveVertexAccess = (options?: SimpleStreamOptions) => Promise<VertexAdcAccess>;
+/**
+ * The project comes from its own environment (`options.env`, then the ADC quota
+ * project, then `getProjectId()`), never from the API-key slot: a stored secret
+ * must not end up in the request URL, and the project may legitimately come from
+ * ADC alone.
+ */
+export declare const defaultResolveVertexAccess: ResolveVertexAccess;
+export declare function vertexHost(location: string): string;
+export declare function streamVertexModelGarden(model: Model<Api>, context: TranscriptContext, options: SimpleStreamOptions | undefined, resolveAccess?: ResolveVertexAccess): AssistantMessageEventStream;
+//# sourceMappingURL=transport.d.ts.map

package/dist/core/extensions/builtin/vertex-model-garden/transport.js

@@ -0,0 +1,35 @@
+import { getApiProvider, } from "@earendil-works/pi-ai";
+import { getVertexAdcAccess } from "@earendil-works/pi-ai/api/google-vertex-adc.lazy";
+import { lazyStream } from "@earendil-works/pi-ai/api/lazy";
+export const VERTEX_MODEL_GARDEN_LOCATION = "global";
+/**
+ * The project comes from its own environment (`options.env`, then the ADC quota
+ * project, then `getProjectId()`), never from the API-key slot: a stored secret
+ * must not end up in the request URL, and the project may legitimately come from
+ * ADC alone.
+ */
+export const defaultResolveVertexAccess = (options) => getVertexAdcAccess({ env: options?.env });
+export function vertexHost(location) {
+    return location === "global" ? "aiplatform.googleapis.com" : `${location}-aiplatform.googleapis.com`;
+}
+function buildVertexOpenApiBaseUrl(project, location = VERTEX_MODEL_GARDEN_LOCATION) {
+    return `https://${vertexHost(location)}/v1/projects/${encodeURIComponent(project)}/locations/${location}/endpoints/openapi`;
+}
+function requireApi(api) {
+    const provider = getApiProvider(api);
+    if (!provider)
+        throw new Error(`API provider ${api} is not registered`);
+    return provider;
+}
+export function streamVertexModelGarden(model, context, options, resolveAccess = defaultResolveVertexAccess) {
+    return lazyStream(model, async () => {
+        const access = await resolveAccess(options);
+        const target = { ...model, baseUrl: buildVertexOpenApiBaseUrl(access.project) };
+        return requireApi("openai-completions").streamSimple(target, context, {
+            ...options,
+            apiKey: access.accessToken,
+            headers: { ...options?.headers, "x-goog-user-project": access.quotaProject },
+        });
+    });
+}
+//# sourceMappingURL=transport.js.map

package/dist/core/extensions/types.d.ts

@@ -10,7 +10,7 @@
 import type { ContextEvent } from "./context-event.ts";
 export type { ContextEvent } from "./context-event.ts";
 import type { AgentMessage, AgentTool, AgentToolCallOutcome, AgentToolResult, AgentToolUpdateCallback, ThinkingLevel, ToolExecutionMode } from "@earendil-works/pi-agent-core";
-import type { AnyModel, Api, AssistantMessageEvent, AssistantMessageEventStream, ClassifierApi, ConstrainedSamplingConfig, Context, FreeformToolFormat, ImageApi, ImageContent, JsonValue, Message, Model, OAuthCredentials, OAuthLoginCallbacks, Provider, ProviderClassifier, ProviderHeaders, ProviderId, ProviderImages, RefreshModelsContext, SimpleStreamOptions, TextContent, ToolResultMessage, TranscriptContext, Usage } from "@earendil-works/pi-ai";
+import type { AnyModel, Api, AssistantMessageEvent, AssistantMessageEventStream, AuthContext, AuthResult, ClassifierApi, ConstrainedSamplingConfig, Context, FreeformToolFormat, ImageApi, ImageContent, JsonValue, Message, Model, OAuthCredentials, OAuthLoginCallbacks, Provider, ProviderClassifier, ProviderHeaders, ProviderId, ProviderImages, RefreshModelsContext, SimpleStreamOptions, TextContent, ToolResultMessage, TranscriptContext, Usage } from "@earendil-works/pi-ai";
 import type { AutocompleteItem, AutocompleteProvider, Component, EditorComponent, EditorTheme, ImageProtocol, KeyId, OverlayHandle, OverlayOptions, TUI } from "@earendil-works/pi-tui";
 import type { Static, TSchema } from "typebox";
 import type { Theme } from "../../modes/interactive/theme/theme.ts";
@@ -1219,6 +1219,16 @@ export interface AgentEndEvent {
     willRetry?: boolean;
     /** Present when the host can attribute the abort to a user action or internal operation. */
     abortSource?: "user" | "system" | "provider";
+    /**
+     * Present on a system abort the turn deadline lease caused (`turn_inactivity_ms` or
+     * `turn_max_runtime_ms`). A deadline stop is never auto-continued: goal continuations,
+     * monitor backstops and hidden nudges hold off until admitted user input clears the hold.
+     */
+    deadline?: {
+        kind: "inactivity" | "max_runtime";
+        limitMs: number;
+        elapsedMs: number;
+    };
 }
 export type AgentActivityOutcome = "completed" | "aborted" | "error";
 export interface CustomEntryDraft {
@@ -2095,6 +2105,27 @@ export interface ProviderConfig {
      * Use context.publish({ persist: entry }) when the catalog should persist across sessions.
      */
     refreshModels?(context: RefreshModelsContext): Promise<ProviderModelConfig[]>;
+    /**
+     * Ambient api-key auth resolved from a shared cloud credential chain (Google
+     * ADC, AWS env) rather than a credential configured for this provider. The
+     * provider counts as configured only while `resolve` yields a result, and the
+     * ambient marker keeps it from outranking a provider the user configured when
+     * the automatic default model is chosen (senpi#2327). Ignored when `apiKey` or
+     * credential headers already supply auth.
+     */
+    ambientAuth?: {
+        /** Display name, e.g. "Google Cloud credentials". */
+        name?: string;
+        /**
+         * Resolve from the ambient chain. `undefined` means the chain currently
+         * supplies no credential, so the provider stays unavailable. A returned
+         * result should carry `ambient: true`.
+         */
+        resolve(input: {
+            ctx: AuthContext;
+            signal: AbortSignal;
+        }): Promise<AuthResult | undefined>;
+    };
     /** OAuth provider for /login support. The `id` is set automatically from the provider name. */
     oauth?: {
         /** Display name for the provider in login UI. */

package/dist/core/provider-api-key-auth.d.ts

@@ -1,10 +1,18 @@
-import { type ApiKeyAuth, type ModelAuth, type Provider, type ProviderHeaders } from "@earendil-works/pi-ai";
+import { type ApiKeyAuth, type AuthContext, type AuthResult, type ModelAuth, type Provider, type ProviderHeaders } from "@earendil-works/pi-ai";
 import type { ModelsJsonProvider } from "./model-config.ts";
+interface AmbientAuthExtensionInput {
+    name?: string;
+    resolve(input: {
+        ctx: AuthContext;
+        signal: AbortSignal;
+    }): Promise<AuthResult | undefined>;
+}
 interface ProviderAuthExtensionInput {
     apiKey?: string;
     headers?: Record<string, string>;
     authHeader?: boolean;
     oauth?: unknown;
+    ambientAuth?: AmbientAuthExtensionInput;
 }
 export declare function composeApiKeyAuth(providerId: string, base: Provider | undefined, config: ModelsJsonProvider | undefined, extension: ProviderAuthExtensionInput | undefined): ApiKeyAuth | undefined;
 export declare function configuredApiKey(config: ModelsJsonProvider | undefined, extension: ProviderAuthExtensionInput | undefined): string | undefined;

package/dist/core/provider-api-key-auth.js

@@ -8,6 +8,14 @@ export function composeApiKeyAuth(providerId, base, config, extension) {
     const rawHeaders = configuredHeaders(config, extension);
     const headerSource = headerAuthResolutionSource(config?.headers, extension?.headers);
     const authHeader = extension?.authHeader ?? config?.authHeader ?? false;
+    const ambient = extension?.ambientAuth;
+    if (!inherited && rawKey === undefined && !headerSource && ambient) {
+        // A shared cloud credential chain (Google ADC, AWS env) owns the lane: an
+        // explicitly configured key, credential headers, or a base provider still
+        // win, and a resolution marked ambient never outranks a configured
+        // provider in automatic default selection (senpi#2327).
+        return ambientApiKeyAuth(providerId, ambient, rawHeaders, headerSource, authHeader);
+    }
     if (!inherited && rawKey === undefined && !headerSource && oauth) {
         // An OAuth provider with no key, no credential headers and no base normally
         // gets no api-key auth at all. Providers with ambient credentials still
@@ -101,6 +109,32 @@ function ambientOnlyAuth(providerId, oauth, rawHeaders, headerSource, authHeader
         resolve,
     };
 }
+/**
+ * Api-key adapter for extension providers whose credentials live in a shared
+ * cloud credential chain (Google ADC, AWS env). No login is offered — the chain
+ * is the only source — and `Models.checkAuth` resolves this auth, so the
+ * `ambient` marker reaches the provider's auth status.
+ */
+function ambientApiKeyAuth(providerId, ambient, rawHeaders, headerSource, authHeader) {
+    const resolve = async (input) => {
+        const result = await ambient.resolve({ ctx: input.ctx, signal: input.signal });
+        if (!result)
+            return undefined;
+        const explicitEnv = { ...(input.credential?.env ?? {}), ...(result.env ?? {}) };
+        const headerEnv = await configContextEnv(Object.values(rawHeaders ?? {}), input.ctx, explicitEnv);
+        const headers = await resolveHeadersOrThrow(rawHeaders, `provider "${providerId}"`, headerEnv);
+        return {
+            ...result,
+            auth: withConfiguredAuth(result.auth, headers, authHeader),
+            source: result.source ?? headerSource,
+        };
+    };
+    return {
+        name: ambient.name ?? "Ambient credentials",
+        ambientOnly: true,
+        resolve,
+    };
+}
 export function configuredApiKey(config, extension) {
     return extension?.apiKey ?? config?.apiKey;
 }

package/dist/core/provider-composer.d.ts

@@ -80,6 +80,15 @@ export interface ProviderConfigInput {
     extraBody?: Record<string, unknown>;
     authHeader?: boolean;
     oauth?: ExtensionOAuthConfig;
+    /**
+     * Ambient api-key auth resolved from a shared cloud credential chain (Google
+     * ADC, AWS env) rather than a credential configured for this provider. The
+     * provider counts as configured only while `resolve` yields a result, and the
+     * ambient marker keeps it from outranking a provider the user configured when
+     * the automatic default model is chosen (senpi#2327). Ignored when `apiKey`,
+     * credential headers, or a base provider already supply auth.
+     */
+    ambientAuth?: ProviderAmbientAuth;
     retryPolicy?: RetryPolicyProfile;
     models?: ProviderModelConfig[];
     refreshModels?(context: RefreshModelsContext): Promise<ProviderModelConfig[]>;
@@ -92,6 +101,20 @@ export interface ProviderConfigInput {
      */
     fallbackEligible?(): boolean;
 }
+/** Ambient api-key auth hook for extension providers (see `ProviderConfigInput.ambientAuth`). */
+export interface ProviderAmbientAuth {
+    /** Display name, e.g. "Google Cloud credentials". */
+    name?: string;
+    /**
+     * Resolve from the ambient chain. `undefined` means the chain currently
+     * supplies no credential, so the provider stays unavailable. A returned
+     * result should carry `ambient: true`.
+     */
+    resolve(input: {
+        ctx: AuthContext;
+        signal: AbortSignal;
+    }): Promise<AuthResult | undefined>;
+}
 export type AuthStatus = {
     configured: boolean;
     source?: "stored" | "runtime" | "environment" | "fallback" | "models_json_key" | "models_json_command" | HeaderAuthStatusSource;

package/dist/core/settings-manager.d.ts

@@ -108,6 +108,8 @@ export interface Settings {
         maxPerUserInput?: number;
         maxToolFreePerMinute?: number;
     };
+    turn_inactivity_ms?: number;
+    turn_max_runtime_ms?: number;
     todo?: TodoSettings;
     recommendedModels?: string[];
     favoriteModels?: string[];
@@ -274,6 +276,8 @@ export declare class SettingsManager {
         maxPerUserInput: number;
         maxToolFreePerMinute: number;
     };
+    getTurnInactivityMs(): number | undefined;
+    getTurnMaxRuntimeMs(): number | undefined;
     getTodoFirstTurnPlan(): TodoFirstTurnPlan;
     getTodoTurnEndBackstop(): boolean;
     isProjectTrusted(): boolean;

package/dist/core/settings-manager.js

@@ -717,6 +717,12 @@ export class SettingsManager {
             maxToolFreePerMinute: limit(configured?.maxToolFreePerMinute, MAX_TOOL_FREE_ENGINE_TURNS_PER_WINDOW),
         };
     }
+    getTurnInactivityMs() {
+        return this.settings.turn_inactivity_ms;
+    }
+    getTurnMaxRuntimeMs() {
+        return this.settings.turn_max_runtime_ms;
+    }
     getTodoFirstTurnPlan() {
         const configured = this.settings.todo?.firstTurnPlan;
         return configured === "remind" || configured === "off" ? configured : "force";

package/dist/core/settings-shapes.d.ts

@@ -49,4 +49,8 @@ export interface OpenAISettings {
 export interface ProviderConcurrencySettings {
     maxConcurrency?: number;
 }
+export interface TurnDeadlineSettings {
+    turn_inactivity_ms?: number;
+    turn_max_runtime_ms?: number;
+}
 //# sourceMappingURL=settings-shapes.d.ts.map