@iowarp/clio-coder 0.4.9 → 0.5.0

package/src/interactive/turn-runtime.ts

@@ -1,3 +1,4 @@
+import { resolvedRequestContext } from "../engine/context.js";
 /**
  * Turn runtime ownership: orchestrator target resolution, model synthesis,
  * live-agent construction and model hot-swap, thinking-level reconciliation,
@@ -25,8 +26,9 @@ import {
 import type { RetrySettings } from "../domains/session/retry.js";
 import type { createEngineAgent } from "../engine/agent.js";
 import { cleanupEngineSessionResources } from "../engine/ai.js";
+import { engineStreamSimple } from "../engine/api-registry.js";
 import { setGlobalDefaultMaxOutputTokens } from "../engine/apis/index.js";
-import { sanitizeLockedSynthesisMessage } from "../engine/loop-guard.js";
+import { lockedSynthesisSystemPrompt, sanitizeLockedSynthesisMessage } from "../engine/loop-guard.js";
 import {
 	patchProviderThinkingPayload,
 	patchToolChoiceNamedPayload,
@@ -34,6 +36,7 @@ import {
 } from "../engine/provider-payload.js";
 import type { AgentEvent, AgentMessage, EngineModel, Usage } from "../engine/types.js";
 import type { resolveAgentTools, ToolFinishEvent, ToolTelemetry } from "../tools/agent-tools.js";
+import { createAssistantGenerationTiming, hasAssistantGenerationDelta } from "./assistant-generation-timing.js";
 import {
 	type AssistantCallTiming,
 	type BackendCacheVerdict,
@@ -141,8 +144,8 @@ export interface TurnRuntimeDeps {
 }
 
 export interface TurnRuntime {
-	ensureRuntime(): AgentRuntime | null;
-	ensureLiveCapabilitiesForSelectedModel(): Promise<void>;
+	ensureRuntime(options?: { silent?: boolean }): AgentRuntime | null;
+	ensureLiveCapabilitiesForSelectedModel(options?: { silent?: boolean }): Promise<void>;
 	cleanupSessionResources(sessionId: string | undefined): void;
 	/**
 	 * Install on the session tool surface so admission verdicts reach the panel.
@@ -165,7 +168,18 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 	 */
 	const toolOutcomes = new Map<
 		string,
-		Pick<ToolFinishEvent, "outcome" | "reason" | "durationMs" | "ruleId" | "reasonCode" | "policySource">
+		Pick<
+			ToolFinishEvent,
+			| "outcome"
+			| "reason"
+			| "durationMs"
+			| "actionClass"
+			| "decision"
+			| "deniedPark"
+			| "ruleId"
+			| "reasonCode"
+			| "policySource"
+		>
 	>();
 	const toolTelemetry: ToolTelemetry = {
 		onFinish(event) {
@@ -174,6 +188,9 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 				outcome: event.outcome,
 				...(event.reason === undefined ? {} : { reason: event.reason }),
 				durationMs: event.durationMs,
+				...(event.actionClass === undefined ? {} : { actionClass: event.actionClass }),
+				...(event.decision === undefined ? {} : { decision: event.decision }),
+				...(event.deniedPark === true ? { deniedPark: true } : {}),
 				...(event.ruleId === undefined ? {} : { ruleId: event.ruleId }),
 				...(event.reasonCode === undefined ? {} : { reasonCode: event.reasonCode }),
 				...(event.policySource === undefined ? {} : { policySource: event.policySource }),
@@ -251,7 +268,7 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 	const TARGET_PROBE_TTL_MS = 5 * 60 * 1000;
 	let lastTargetProbe: { key: string; at: number } | null = null;
 	const targetProbesInFlight = new Map<string, Promise<void>>();
-	const ensureLiveCapabilitiesForSelectedModel = async (): Promise<void> => {
+	const ensureLiveCapabilitiesForSelectedModel = async (options?: { silent?: boolean }): Promise<void> => {
 		const settings = deps.getSettings();
 		const targetId = settings.chat.target?.trim();
 		const wireModelId = settings.chat.model?.trim();
@@ -276,7 +293,7 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 			} catch {
 				// Fall back to the last known target state.
 			}
-			announceColdModel(status, targetId, wireModelId);
+			if (!options?.silent) announceColdModel(status, targetId, wireModelId);
 		})();
 		targetProbesInFlight.set(key, probe);
 		try {
@@ -384,7 +401,7 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 	// A target id/model names the selection, not its current URL, auth, pricing
 	// or cache policy. Keep the constructed descriptor identity across calls.
 	let runtimeTargetSettings: string | null = null;
-	const ensureRuntime = (): AgentRuntime | null => {
+	const ensureRuntime = (options?: { silent?: boolean }): AgentRuntime | null => {
 		const target = readTarget();
 		if (!target) return null;
 		const targetSettings = JSON.stringify(target.target);
@@ -392,7 +409,8 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 		// Preflight and the wire budget share the session's effective settings,
 		// including overrides that never reach the saved providers snapshot.
 		setGlobalDefaultMaxOutputTokens(deps.getSettings().chat.maxOutputTokens);
-		context.emitContextWindowWarningTransition(target.runtimeResolution?.contextWindowDetails?.warning ?? null);
+		if (!options?.silent)
+			context.emitContextWindowWarningTransition(target.runtimeResolution?.contextWindowDetails?.warning ?? null);
 		if (!deps.knownTargets().has(target.target.id)) {
 			throw new TurnAdmissionError(
 				`[Clio Coder] orchestrator target=${target.target.id} unknown. Run \`clio-coder targets\` to see configured targets.`,
@@ -416,12 +434,15 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 				state.runtime.agent.state.model,
 				deps.providers.knowledgeBase,
 			);
+			// The wire adapter reads the existing model, including route controls
+			// that fresh metadata may have changed or removed for this same alias.
+			applyModelCapabilityPatch(state.runtime.agent.state.model, runtimeResolution.capabilities);
 			const desiredLevel = runtimeResolution.effectiveThinkingLevel;
 			if (state.runtime.agent.state.thinkingLevel !== desiredLevel) {
 				state.runtime.agent.state.thinkingLevel = desiredLevel;
 			}
 			state.runtime.runtimeResolution = runtimeResolution;
-			emitThinkingClampNotice(runtimeResolution);
+			if (!options?.silent) emitThinkingClampNotice(runtimeResolution);
 			return state.runtime;
 		}
 
@@ -446,7 +467,7 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 			const effectiveThinkingLevel = target.runtimeResolution.effectiveThinkingLevel;
 			state.runtime.agent.state.thinkingLevel = effectiveThinkingLevel;
 			state.runtime.runtimeResolution = target.runtimeResolution;
-			emitThinkingClampNotice(target.runtimeResolution);
+			if (!options?.silent) emitThinkingClampNotice(target.runtimeResolution);
 			persistence.appendModelChangeEntry(target);
 			ensureReasoningProbe(target);
 			return state.runtime;
@@ -473,6 +494,15 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 		let apiCallStartedAt: number | null = null;
 		let apiCallFirstDeltaAt: number | null = null;
 		const handle = deps.createAgent({
+			transcriptStreamFn: (currentModel, currentContext, options) => {
+				if (!state.synthesisToolLock) return engineStreamSimple(currentModel, currentContext, options);
+				const request = resolvedRequestContext(currentContext);
+				return engineStreamSimple(
+					currentModel,
+					{ ...request, systemPrompt: lockedSynthesisSystemPrompt(request.systemPrompt ?? "") },
+					options,
+				);
+			},
 			initialState: {
 				systemPrompt: fallbackIdentityPrompt(),
 				model,
@@ -530,7 +560,7 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 			wireModelId: target.wireModelId,
 			runtimeResolution: target.runtimeResolution,
 		};
-		emitThinkingClampNotice(target.runtimeResolution);
+		if (!options?.silent) emitThinkingClampNotice(target.runtimeResolution);
 
 		// Stall watchdog. `agent_start` arms a timer for `retry.streamStallMs`;
 		// every later engine event counts as progress and pushes the deadline
@@ -607,15 +637,31 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 		handle.agent.prepareNextTurn = async (signal?: AbortSignal) => {
 			stallSuspendDepth += 1;
 			try {
-				return await context.postToolContinuationGuard(localRuntime, signal);
+				if (signal?.aborted || state.activeInterruptReason !== null) return undefined;
+				const contextChanged = await middleware.prepareToolContinuation(localRuntime, signal);
+				const update = await context.postToolContinuationGuard(localRuntime, signal, contextChanged);
+				if (signal?.aborted || state.activeInterruptReason !== null) return undefined;
+				// Compaction owns any replacement snapshot. Without one, publish the
+				// already-accounted reminder through context, not message events that
+				// would manufacture a new operator turn and reset task authority.
+				return (
+					update ??
+					(contextChanged
+						? {
+								context: {
+									messages: [...localRuntime.agent.state.messages],
+									tools: [...localRuntime.agent.state.tools],
+								},
+							}
+						: undefined)
+				);
 			} finally {
 				stallSuspendDepth -= 1;
 				lastActivityAt = performance.now();
 			}
 		};
 
-		let streamStartedAt: number | null = null;
-		let firstAssistantDeltaAt: number | null = null;
+		const generationTiming = createAssistantGenerationTiming();
 		// First call of the run is the one whose verdict says whether the
 		// backend reused the session prefix; later calls in a tool loop are
 		// trivially warm.
@@ -679,6 +725,13 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 						? {}
 						: {
 								outcome: admission.outcome,
+								// The registry's per-call classification and admission
+								// decision. Headless receipts count safety decisions and
+								// tell a successful write from any other success with
+								// these, the same facts a worker's receipt is built from.
+								...(admission.actionClass === undefined ? {} : { actionClass: admission.actionClass }),
+								...(admission.decision === undefined ? {} : { decision: admission.decision }),
+								...(admission.deniedPark === true ? { deniedPark: true } : {}),
 								...(admission.ruleId === undefined ? {} : { ruleId: admission.ruleId }),
 								...(admission.reasonCode === undefined ? {} : { reasonCode: admission.reasonCode }),
 								...(admission.policySource === undefined ? {} : { policySource: admission.policySource }),
@@ -746,10 +799,9 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 					publicEvent = { ...enrichedEvent, messages: reported } as typeof enrichedEvent;
 				}
 			}
+			if (publicEvent) generationTiming.record(publicEvent, eventClock);
 			if (publicEvent?.type === "agent_start") {
 				runStartMessageCount = localRuntime.agent.state.messages.length;
-				streamStartedAt = eventClock;
-				firstAssistantDeltaAt = null;
 				apiCallStartedAt = null;
 				apiCallFirstDeltaAt = null;
 				runFirstCallVerdict = null;
@@ -768,15 +820,7 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 			// at the turn that issued it instead of for the rest of the run.
 			if (publicEvent?.type === "turn_end") toolsInFlight = 0;
 			if (publicEvent?.type === "message_update") {
-				const assistantEvent = publicEvent.assistantMessageEvent as { type?: string; delta?: unknown };
-				const hasDelta =
-					assistantEvent.type === "toolcall_start" ||
-					((assistantEvent.type === "text_delta" ||
-						assistantEvent.type === "thinking_delta" ||
-						assistantEvent.type === "toolcall_delta") &&
-						typeof assistantEvent.delta === "string" &&
-						assistantEvent.delta.length > 0);
-				if (hasDelta && firstAssistantDeltaAt === null) firstAssistantDeltaAt = eventClock;
+				const hasDelta = hasAssistantGenerationDelta(publicEvent.assistantMessageEvent);
 				if (hasDelta && apiCallFirstDeltaAt === null) {
 					apiCallFirstDeltaAt = eventClock;
 					// The timer may be armed for the longer first-token window; the
@@ -818,13 +862,14 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 						},
 					);
 				}
-				if (summary.output > 0 && firstAssistantDeltaAt !== null) {
-					const durationMs = Math.round(Math.max(1, eventClock - firstAssistantDeltaAt));
+				const timing = generationTiming.snapshot(eventClock);
+				if (summary.output > 0 && timing) {
+					const durationMs = Math.round(timing.durationMs);
 					deps.observability.recordTokenThroughput({
 						tokensPerSecond: summary.output / (durationMs / 1000),
 						outputTokens: summary.output,
 						durationMs,
-						...(streamStartedAt !== null ? { ttftMs: Math.round(firstAssistantDeltaAt - streamStartedAt) } : {}),
+						ttftMs: Math.round(timing.ttftMs),
 						providerId: localRuntime.targetId,
 						modelId: localRuntime.wireModelId,
 						recordedAt: eventAt,
@@ -912,7 +957,7 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 					const input = typeof usage.input === "number" ? usage.input : 0;
 					const cacheRead = typeof usage.cacheRead === "number" ? usage.cacheRead : 0;
 					const backendTimings = (enrichedEvent.message as { backendTimings?: BackendCompletionTimings }).backendTimings;
-					runFirstCallVerdict = backendCacheVerdict(input, cacheRead, backendTimings);
+					runFirstCallVerdict = backendCacheVerdict(input, cacheRead, backendTimings, usage.cacheReadReported);
 				}
 				if (usage) {
 					context.reconcileUsage(usage);

package/src/interactive/turn-state.ts

@@ -11,6 +11,7 @@
 
 import type { ClioSettings } from "../core/config.js";
 import type { PendingSkillToolPolicy } from "../core/skill-activation.js";
+import type { TurnConstraints } from "../core/turn-constraints.js";
 import type {
 	ResolvedRuntimeTarget,
 	RuntimeDescriptor,
@@ -127,6 +128,7 @@ export interface ChatTurnState {
 	/** A middleware request_continuation is waiting to resubmit after settle. */
 	pendingRequestContinuation: boolean;
 	currentPendingSkillPolicy: PendingSkillToolPolicy | undefined;
+	currentTurnConstraints: TurnConstraints | undefined;
 	/**
 	 * Tool surface an activated skill narrowed, held across the operator's
 	 * later turns. Replaced when the operator activates a different skill,
@@ -160,6 +162,7 @@ export function createTurnState(initialThinkingLevel: ThinkingLevel): ChatTurnSt
 		stalledTurnNudgeSpent: false,
 		pendingRequestContinuation: false,
 		currentPendingSkillPolicy: undefined,
+		currentTurnConstraints: undefined,
 		activeSkillSurface: undefined,
 		currentAskUserPolicy: undefined,
 		lastRunSnapshot: null,

package/src/interactive/welcome-art.ts

@@ -0,0 +1,36 @@
+/** Five-row block lettering, following the visual style of the Warpio CLI wordmark. */
+const LETTERS: Record<string, readonly string[]> = {
+	C: [" ████", "██   ", "██   ", "██   ", " ████"],
+	L: ["██   ", "██   ", "██   ", "██   ", "█████"],
+	I: ["███", " █ ", " █ ", " █ ", "███"],
+	O: [" ███ ", "██ ██", "██ ██", "██ ██", " ███ "],
+	D: ["████ ", "██ ██", "██ ██", "██ ██", "████ "],
+	E: ["█████", "██   ", "████ ", "██   ", "█████"],
+	R: ["████ ", "██ ██", "████ ", "██ ██", "██ ██"],
+};
+
+const WIDE_LETTERS: Record<string, readonly string[]> = {
+	C: [" ██████", "██     ", "██     ", "██     ", " ██████"],
+	L: ["██     ", "██     ", "██     ", "██     ", "███████"],
+	I: ["██", "██", "██", "██", "██"],
+	O: [" ██████ ", "██    ██", "██    ██", "██    ██", " ██████ "],
+	D: ["██████ ", "██   ██", "██   ██", "██   ██", "██████ "],
+	E: ["███████", "██     ", "█████  ", "██     ", "███████"],
+	R: ["██████ ", "██   ██", "██████ ", "██   ██", "██   ██"],
+	" ": [" ", " ", " ", " ", " "],
+};
+function stacked(font: Record<string, readonly string[]>): string[] {
+	const word = (text: string) =>
+		Array.from({ length: 5 }, (_, row) => [...text].map((letter) => font[letter]?.[row] ?? "").join(" "));
+	const clio = word("CLIO");
+	const coder = word("CODER");
+	const width = coder[0]?.length ?? 0;
+	return [
+		...clio.map((line) => line.padStart(line.length + Math.floor((width - line.length) / 2)).padEnd(width)),
+		" ".repeat(width),
+		...coder,
+	];
+}
+
+export const WELCOME_WORDMARK = stacked(LETTERS);
+export const WELCOME_WORDMARK_WIDE = stacked(WIDE_LETTERS);