@iowarp/clio-coder 0.3.0 → 0.3.2

package/src/interactive/turn-context.ts

@@ -0,0 +1,855 @@
+/**
+ * Turn context ownership: the session-prompt compile cache, context-snapshot
+ * accounting, prompt-cache honesty, and compaction. `runAutoCompact` is the
+ * one compaction entry point; the pre-submit trigger, the preflight overflow
+ * guard, overflow recovery, `/context compact`, and the post-tool continuation guard
+ * all flow through it.
+ */
+
+import {
+	BusChannels,
+	type ContextActivityStatus,
+	type ContextPrunedPayload,
+	type ContextWarningPayload,
+} from "../core/bus-events.js";
+import type { ClioSettings } from "../core/config.js";
+import type { SafeEventBus } from "../core/event-bus.js";
+import type { ToolName } from "../core/tool-names.js";
+import type { ObservabilityContract } from "../domains/observability/contract.js";
+import type { CompiledSessionPrompt, SessionPromptInputs } from "../domains/prompts/compiler.js";
+import type { PromptsContract } from "../domains/prompts/contract.js";
+import type { ProvidersContract } from "../domains/providers/index.js";
+import type { LocalModelQuirks } from "../domains/providers/types/local-model-quirks.js";
+import {
+	AutoCompactionTrigger,
+	DEFAULT_COMPACTION_THRESHOLD,
+	shouldCompact,
+} from "../domains/session/compaction/auto.js";
+import type { CompactResult } from "../domains/session/compaction/compact.js";
+import { maskStaleObservations } from "../domains/session/compaction/mask-observations.js";
+import {
+	appendContextSnapshot,
+	type CaptureContextSnapshotInput,
+	type ContextSnapshot,
+	type ContextUsageBreakdown,
+	type ContextUsageSnapshot,
+	captureContextSnapshot,
+	ceilChars,
+	contentChars,
+	contextUsageSnapshot,
+	estimateAgentContextBreakdown,
+	estimateAgentContextTokens,
+	getLatestContextSnapshot,
+	reconcileSnapshot,
+	snapshotInputTokens,
+} from "../domains/session/context-accounting.js";
+import { buildContextLedger, type ContextLedger, type PromptCacheStats } from "../domains/session/context-ledger.js";
+import type { SessionContract } from "../domains/session/contract.js";
+import type { CompactionTrigger, SessionEntry } from "../domains/session/entries.js";
+import { appendPromptCompileRecord, type SessionPromptCompileRecord } from "../domains/session/prompt-manifest.js";
+import type { AgentMessage, Usage } from "../engine/types.js";
+import type { ToolRegistry } from "../tools/registry.js";
+import {
+	type BackendCacheVerdict,
+	backendCacheVerdict,
+	extractUserText,
+	runtimeSupportsTools,
+	sumRunUsage,
+	toolNamesFromAgentState,
+} from "./chat-loop-messages.js";
+import { buildReplayAgentMessagesFromTurns } from "./chat-renderer.js";
+import { renderCompactionSummaryLine } from "./renderers/compaction-summary.js";
+import type { TurnMiddleware } from "./turn-middleware.js";
+import type { AgentRuntime, ChatTurnState } from "./turn-state.js";
+
+export interface TurnContextDeps {
+	state: ChatTurnState;
+	getSettings: () => Readonly<ClioSettings>;
+	providers: ProvidersContract;
+	session?: SessionContract | undefined;
+	prompts?: PromptsContract | undefined;
+	toolRegistry?: ToolRegistry | undefined;
+	observability?: ObservabilityContract | undefined;
+	bus?: SafeEventBus | undefined;
+	readSessionEntries?: (() => ReadonlyArray<SessionEntry>) | undefined;
+	autoCompact?: ((instructions?: string, trigger?: CompactionTrigger) => Promise<CompactResult | null>) | undefined;
+	getMemorySection?: (() => string) | undefined;
+	middleware: TurnMiddleware;
+	emitNotice: (text: string) => void;
+}
+
+export interface LiveContextEstimate {
+	tokens: number;
+	contextWindow: number;
+	breakdown: ReturnType<typeof estimateAgentContextBreakdown>;
+}
+
+export interface TurnContext {
+	ensureSessionPrompt(agentRuntime: AgentRuntime): Promise<CompiledSessionPrompt | null>;
+	logPromptCompileIfPending(): void;
+	invalidateSessionPromptCache(): void;
+	addWorkingContextPaths(paths: ReadonlyArray<string>): void;
+	captureRuntimeContextSnapshot(
+		agentRuntime: AgentRuntime,
+		turnId: string,
+		compactionThreshold: number | null,
+		extra?: Partial<CaptureContextSnapshotInput>,
+	): ContextSnapshot;
+	setCurrentSnapshot(snapshot: ContextSnapshot): void;
+	persistContextSnapshot(snapshot: ContextSnapshot): void;
+	flushReconciledSnapshot(): void;
+	/** Reconcile the live snapshot against one API call's provider usage. */
+	reconcileUsage(usage: Usage): void;
+	/**
+	 * Prompt-side tokens the current snapshot accounts for. The prompt of an
+	 * in-flight call is spent whatever the operator does next, so an interrupted
+	 * turn records this rather than zero. 0 when no snapshot exists yet.
+	 */
+	promptSideTokens(): number;
+	liveContextEstimate(agentRuntime: AgentRuntime, pendingUserText?: string): LiveContextEstimate;
+	refreshAgentMessagesFromSession(agentRuntime: AgentRuntime): ReadonlyArray<SessionEntry>;
+	runAutoCompact(
+		agentRuntime: AgentRuntime,
+		force: boolean,
+		instructions?: string,
+		triggerOverride?: CompactionTrigger,
+		pendingUserText?: string,
+	): Promise<boolean>;
+	postToolContinuationGuard(
+		agentRuntime: AgentRuntime,
+		signal?: AbortSignal,
+	): Promise<
+		| {
+				context: { systemPrompt: string; messages: AgentMessage[]; tools: AgentRuntime["agent"]["state"]["tools"] };
+				model: AgentRuntime["agent"]["state"]["model"];
+				thinkingLevel: AgentRuntime["agent"]["state"]["thinkingLevel"];
+		  }
+		| undefined
+	>;
+	contextUsage(): ContextUsageSnapshot;
+	contextLedger(): ContextLedger;
+	emitContextWindowWarningTransition(warning: string | null): void;
+	/** Consume disturbances since the last settled run (T3.3 honesty). */
+	consumeExpectedColdReasons(runtimeId: string): void;
+	/** Prompt-cache record for one persisted assistant call, with cold-reason stamp. */
+	promptCachePayloadForAssistant(usage: Usage): Record<string, unknown>;
+	/** Record the settled run's cache summary for /context. */
+	noteRunCacheSummary(messages: ReadonlyArray<AgentMessage>, runFirstCallVerdict: BackendCacheVerdict | null): void;
+	resetForSession(): void;
+	dispose(): void;
+}
+
+export function createTurnContext(deps: TurnContextDeps): TurnContext {
+	const { state, middleware } = deps;
+	const compactionTrigger = new AutoCompactionTrigger<CompactResult | null>();
+
+	let currentContextSnapshot: ContextSnapshot | null = null;
+	let lastCompactionEvent: { stage: string; tokensBefore: number; tokensAfter: number; trigger: string } | null = null;
+	// Last settled run's provider cache usage plus whether the compiled system
+	// prompt was reused. Shown together in /context so "prompt reused" can
+	// never imply provider cache reuse the backend did not report.
+	let lastPromptCache: PromptCacheStats | null = null;
+	let lastSystemPromptReused = false;
+	// The session system prompt, compiled once per session. Recompiles happen
+	// only on explicit events: the compile key (target, model, safety level,
+	// session id) changes, or a config hot-reload invalidates the cache. A
+	// recompile that changes the prompt text appends a "promptRecompiled"
+	// ledger entry so a cold provider cache is always explainable.
+	let sessionPrompt: CompiledSessionPrompt | null = null;
+	let sessionPromptKey: string | null = null;
+	const sessionWorkingContextPaths = new Set<string>();
+	let pendingPromptLogEntry: SessionPromptCompileRecord | null = null;
+
+	// Cache-disturbance honesty (T3.3). Dispatch traffic and history
+	// compaction invalidate a single-slot local backend's prefix cache.
+	// Accumulate every disturbance since the last settled run; the next
+	// submit consumes the set, stamps `promptCache.expectedColdReasons` on
+	// its first assistant entry, and shows one dim notice.
+	const pendingColdReasons = new Set<string>();
+	let runExpectedColdReasons: string[] = [];
+	let stampColdReasonsPending = false;
+	const unsubscribeColdReasonSources = [
+		...[BusChannels.DispatchStarted, BusChannels.DispatchCompleted, BusChannels.DispatchFailed].map(
+			(channel) =>
+				deps.bus?.on(channel, () => {
+					pendingColdReasons.add("dispatch");
+				}) ?? null,
+		),
+		...[BusChannels.CompactionBegin, BusChannels.CompactionEnd].map(
+			(channel) =>
+				deps.bus?.on(channel, () => {
+					pendingColdReasons.add("compaction");
+				}) ?? null,
+		),
+	];
+
+	/**
+	 * True when the in-memory snapshot has been reconciled against provider
+	 * usage since it was last persisted. A tool-calling turn reconciles once
+	 * per API call; only the final reconciled state is written to the JSONL
+	 * ledger, when the run settles. Any persist (turn submit, compaction,
+	 * flush) clears the flag because it writes the current snapshot state.
+	 */
+	let snapshotPersistPending = false;
+
+	/**
+	 * Publish the window-resolution warning only on transitions (appeared,
+	 * changed, cleared). ensureRuntime runs on every submit; re-emitting the
+	 * same state each turn would spam every ContextWarning subscriber.
+	 */
+	let lastContextWindowWarning: string | null = null;
+
+	const persistContextSnapshot = (snapshot: ContextSnapshot): void => {
+		const currentSession = deps.session?.current();
+		if (currentSession) appendContextSnapshot(currentSession, snapshot);
+		snapshotPersistPending = false;
+	};
+
+	/**
+	 * Capture a context snapshot from the runtime's live agent state. All
+	 * capture sites (turn submit, both compaction paths) flow through this
+	 * helper so window resolution and category decomposition stay identical.
+	 */
+	const captureRuntimeContextSnapshot = (
+		agentRuntime: AgentRuntime,
+		turnId: string,
+		compactionThreshold: number | null,
+		extra: Partial<CaptureContextSnapshotInput> = {},
+	): ContextSnapshot => {
+		const details = agentRuntime.runtimeResolution.contextWindowDetails;
+		// The snapshot row belongs to the named Clio session whose ledger it is
+		// appended to; the engine agent's own sessionId is unset in practice.
+		return captureContextSnapshot({
+			sessionId: deps.session?.current()?.id ?? agentRuntime.agent.sessionId ?? "unknown",
+			turnId,
+			providerId: agentRuntime.targetId,
+			runtimeId: agentRuntime.runtimeId,
+			modelId: agentRuntime.wireModelId,
+			systemPrompt: agentRuntime.agent.state.systemPrompt,
+			conversationMessages: agentRuntime.agent.state.messages,
+			activeToolSchemas: agentRuntime.agent.state.tools,
+			desiredContextWindow: details.desiredContextWindow,
+			effectiveContextWindow: details.effectiveContextWindow,
+			contextWindowSource: details.contextWindowSource,
+			compactionThreshold,
+			...extra,
+		});
+	};
+
+	/**
+	 * Output tokens of the in-flight response. While streaming, estimate from
+	 * the partial assistant tail; once the turn settles, the reconciled
+	 * snapshot carries the provider-reported value.
+	 */
+	const liveStreamingOutputTokens = (): number => {
+		if (!state.runtime) return 0;
+		if (state.streaming) {
+			const messages = state.runtime.agent.state.messages;
+			const lastMsg = messages[messages.length - 1] as { role?: string; payload?: unknown; content?: unknown } | undefined;
+			if (lastMsg && lastMsg.role === "assistant") {
+				return ceilChars(contentChars(lastMsg.payload ?? lastMsg.content));
+			}
+			return 0;
+		}
+		return currentContextSnapshot?.categories.streaming || 0;
+	};
+
+	/**
+	 * Tokens for the submitted text that the snapshot has not yet counted.
+	 * The turn snapshot is captured before the user message joins the
+	 * conversation, so until the provider reconciles (or a fresh capture sees
+	 * the text in the conversation) the pending input occupies window space
+	 * that no category covers. Zero once reconciled or once the text landed.
+	 */
+	const pendingUserInputTokens = (): number => {
+		const snapshot = currentContextSnapshot;
+		if (!snapshot?.pendingUserInput) return 0;
+		if (snapshot.sources.total === "reconciled") return 0;
+		if (snapshot.turnId !== "pending") {
+			const pending = snapshot.pendingUserInput;
+			const landed = (snapshot.conversationMessages ?? []).some(
+				(message) => extractUserText(message as AgentMessage) === pending,
+			);
+			if (landed) return 0;
+		}
+		return ceilChars(snapshot.pendingUserInput.length);
+	};
+
+	const liveContextEstimate = (agentRuntime: AgentRuntime, pendingUserText?: string): LiveContextEstimate => {
+		const contextWindow = agentRuntime.runtimeResolution.contextWindowDetails.effectiveContextWindow;
+		const estimateInput = {
+			systemPrompt: agentRuntime.agent.state.systemPrompt,
+			messages: agentRuntime.agent.state.messages,
+			tools: agentRuntime.agent.state.tools,
+			...(pendingUserText !== undefined ? { pendingUserText } : {}),
+		};
+		return {
+			tokens: estimateAgentContextTokens(estimateInput),
+			contextWindow,
+			breakdown: estimateAgentContextBreakdown(estimateInput),
+		};
+	};
+
+	const refreshAgentMessagesFromSession = (agentRuntime: AgentRuntime): ReadonlyArray<SessionEntry> => {
+		const refreshedEntries = deps.readSessionEntries?.() ?? [];
+		agentRuntime.agent.state.messages = buildReplayAgentMessagesFromTurns(refreshedEntries, {
+			...(state.lastTurnId ? { activeLeafTurnId: state.lastTurnId } : {}),
+		});
+		state.replayedContextMessages = [];
+		return refreshedEntries;
+	};
+
+	// Compaction rides the context island as a single-phase "compaction"
+	// activity (rendered as "Context Compact"). Each stage brackets its work
+	// with a started/completed pair; a throwing summary emits failed. The
+	// island already knows this kind and phase, and `deps.bus` is optional.
+	const emitCompactionActivity = (status: ContextActivityStatus, message: string): void => {
+		deps.bus?.emit(BusChannels.ContextActivity, {
+			kind: "compaction",
+			phase: "done",
+			status,
+			message,
+			at: Date.now(),
+		});
+	};
+	const compactionFailureMessage = (error: unknown): string =>
+		`compaction failed: ${error instanceof Error ? error.message : String(error)}`;
+
+	const recordCompactionUsage = (agentRuntime: AgentRuntime, result: CompactResult): void => {
+		const usage = result.usage;
+		if (!usage || !deps.observability) return;
+		if (usage.totalTokens <= 0 && usage.cost.total <= 0) return;
+		deps.observability.recordTokens(
+			agentRuntime.targetId,
+			agentRuntime.wireModelId,
+			usage.totalTokens,
+			usage.cost.total,
+			{
+				input: usage.input,
+				output: usage.output,
+				cacheRead: usage.cacheRead,
+				cacheWrite: usage.cacheWrite,
+				reasoningTokens: usage.reasoning,
+				totalTokens: usage.totalTokens,
+				apiCalls: Math.max(1, Math.round(usage.apiCalls)),
+			},
+			agentRuntime.runtimeResolution.costProvenance,
+		);
+	};
+
+	/**
+	 * Two-mechanism context protection. When pressure crosses the single
+	 * threshold, first mask the bodies of tool observations older than
+	 * `excludeLastTurns` (cheap, no LLM call). If pressure stays above the
+	 * threshold, delegate to the pi-style LLM compaction path: append a
+	 * compaction summary entry, then replay from the session view.
+	 *
+	 * `force = true` skips the pressure check and the mask pre-stage and runs
+	 * the LLM summary directly. Used for `/context compact`, CLIO_CODER_FORCE_COMPACT=1,
+	 * and overflow recovery.
+	 */
+	const runAutoCompact = async (
+		agentRuntime: AgentRuntime,
+		force: boolean,
+		instructions?: string,
+		triggerOverride?: CompactionTrigger,
+		pendingUserText?: string,
+	): Promise<boolean> => {
+		if (!deps.autoCompact || !deps.readSessionEntries) return false;
+		const settings = deps.getSettings();
+		const cfg = settings.compaction;
+		const autoEnabled = cfg?.auto !== false;
+		if (!force && !autoEnabled) return false;
+
+		const compactionThreshold = cfg?.threshold ?? DEFAULT_COMPACTION_THRESHOLD;
+		const trigger: CompactionTrigger = triggerOverride ?? (force ? "force" : "auto");
+
+		if (!force) {
+			const estimate = liveContextEstimate(agentRuntime, pendingUserText);
+			const verdict = shouldCompact(estimate.tokens, compactionThreshold, estimate.contextWindow);
+			if (!verdict.shouldCompact) return false;
+
+			// Mechanism B pre-stage: mask stale observations before paying for
+			// an LLM summary. History rewrites here invalidate the backend
+			// prefix cache; the "compaction" expectedColdReasons stamp from the
+			// CompactionBegin/End subscription explains the next cold turn.
+			if (deps.session?.current()) {
+				const beforeSnapshotId = currentContextSnapshot?.snapshotId ?? null;
+				let masked: ReturnType<typeof maskStaleObservations>;
+				try {
+					masked = maskStaleObservations(deps.readSessionEntries() ?? [], cfg?.excludeLastTurns ?? 6);
+				} catch (error) {
+					middleware.fireCompactionHook("mask_observations", trigger, estimate.tokens);
+					deps.bus?.emit(BusChannels.CompactionBegin, { trigger, at: Date.now() });
+					emitCompactionActivity("started", "compacting context (mask stage)");
+					emitCompactionActivity("failed", compactionFailureMessage(error));
+					deps.bus?.emit(BusChannels.CompactionEnd, { trigger, at: Date.now() });
+					throw error;
+				}
+				if (masked.changed) {
+					middleware.fireCompactionHook("mask_observations", trigger, estimate.tokens);
+					deps.bus?.emit(BusChannels.CompactionBegin, { trigger, at: Date.now() });
+					emitCompactionActivity("started", "compacting context (mask stage)");
+					deps.session.replaceEntries(masked.entries);
+					refreshAgentMessagesFromSession(agentRuntime);
+					deps.bus?.emit(BusChannels.CompactionEnd, { trigger, at: Date.now() });
+
+					const postMaskSnapshot = captureRuntimeContextSnapshot(
+						agentRuntime,
+						state.activeUserTurnId || "compaction",
+						compactionThreshold,
+					);
+					currentContextSnapshot = postMaskSnapshot;
+					persistContextSnapshot(postMaskSnapshot);
+
+					const tokensAfterMask = snapshotInputTokens(postMaskSnapshot);
+					lastCompactionEvent = {
+						stage: "mask_observations",
+						tokensBefore: estimate.tokens,
+						tokensAfter: tokensAfterMask,
+						trigger,
+					};
+					deps.bus?.emit(BusChannels.ContextPruned, {
+						stage: "mask_observations",
+						pressure: verdict.pressure,
+						tokensBefore: estimate.tokens,
+						tokensAfter: tokensAfterMask,
+						maskedObservations: masked.maskedObservations,
+						maskedThinkingBlocks: masked.maskedThinkingBlocks,
+						maskedThinkingChars: masked.maskedThinkingChars,
+						trigger,
+						snapshotIdBefore: beforeSnapshotId,
+						snapshotIdAfter: postMaskSnapshot.snapshotId,
+						at: Date.now(),
+					} satisfies ContextPrunedPayload);
+					emitCompactionActivity("completed", `${masked.maskedObservations} observations masked`);
+					const thinkingNote =
+						masked.maskedThinkingBlocks > 0
+							? `, ${masked.maskedThinkingBlocks} thinking blocks dropped (~${masked.maskedThinkingChars} chars)`
+							: "";
+					deps.emitNotice(
+						`[context engine] mask_observations: ${masked.maskedObservations} observations masked${thinkingNote}; ~${estimate.tokens} tokens -> ~${tokensAfterMask} tokens`,
+					);
+
+					const after = liveContextEstimate(agentRuntime, pendingUserText);
+					if (!shouldCompact(after.tokens, compactionThreshold, after.contextWindow).shouldCompact) return true;
+				}
+			}
+		}
+
+		middleware.fireCompactionHook("llm_summary", trigger);
+		deps.bus?.emit(BusChannels.CompactionBegin, { trigger, at: Date.now() });
+		emitCompactionActivity("started", "compacting context (summary)");
+		let result: CompactResult | null = null;
+		const beforeSnapshotId = currentContextSnapshot?.snapshotId ?? null;
+		try {
+			result = await compactionTrigger.fire(() => (deps.autoCompact ?? (async () => null))(instructions, trigger));
+		} catch (error) {
+			emitCompactionActivity("failed", compactionFailureMessage(error));
+			throw error;
+		} finally {
+			deps.bus?.emit(BusChannels.CompactionEnd, { trigger, at: Date.now() });
+		}
+		if (!result || result.summary.length === 0) {
+			emitCompactionActivity("completed", "nothing to compact");
+			return false;
+		}
+
+		// The summarization call spends tokens on the same target the turn would
+		// have. The ledger entry carries its usage for a later reseed; this is the
+		// live sink, so `/cost` and the footer move the moment /context compact
+		// returns instead of staying byte-identical to before it ran.
+		recordCompactionUsage(agentRuntime, result);
+
+		refreshAgentMessagesFromSession(agentRuntime);
+
+		const postCompactSnapshot = captureRuntimeContextSnapshot(
+			agentRuntime,
+			state.activeUserTurnId || "compaction",
+			compactionThreshold,
+		);
+		currentContextSnapshot = postCompactSnapshot;
+		persistContextSnapshot(postCompactSnapshot);
+
+		const tokensAfter = snapshotInputTokens(postCompactSnapshot);
+		lastCompactionEvent = {
+			stage: "llm_summary",
+			tokensBefore: result.tokensBefore,
+			tokensAfter,
+			trigger,
+		};
+		deps.bus?.emit(BusChannels.ContextPruned, {
+			stage: "llm_summary",
+			tokensBefore: result.tokensBefore,
+			tokensAfter,
+			trigger,
+			snapshotIdBefore: beforeSnapshotId,
+			snapshotIdAfter: postCompactSnapshot.snapshotId,
+			at: Date.now(),
+		} satisfies ContextPrunedPayload);
+		emitCompactionActivity("completed", `compacted ~${result.tokensBefore} -> ~${tokensAfter} tokens`);
+
+		deps.emitNotice(
+			renderCompactionSummaryLine({
+				messagesSummarized: result.messagesSummarized,
+				summaryChars: result.summary.length,
+				tokensBefore: result.tokensBefore,
+				isSplitTurn: result.isSplitTurn,
+			}),
+		);
+		return true;
+	};
+
+	const toolResultTail = (agentRuntime: AgentRuntime): boolean => {
+		const messages = agentRuntime.agent.state.messages;
+		const tail = messages[messages.length - 1] as AgentMessage | undefined;
+		return !!tail && typeof tail === "object" && tail !== null && "role" in tail && tail.role === "toolResult";
+	};
+
+	const continuationContextUpdate = (agentRuntime: AgentRuntime) => ({
+		context: {
+			systemPrompt: agentRuntime.agent.state.systemPrompt,
+			messages: [...agentRuntime.agent.state.messages],
+			tools: [...agentRuntime.agent.state.tools],
+		},
+		model: agentRuntime.agent.state.model,
+		thinkingLevel: agentRuntime.agent.state.thinkingLevel,
+	});
+
+	return {
+		captureRuntimeContextSnapshot,
+		persistContextSnapshot,
+		liveContextEstimate,
+		refreshAgentMessagesFromSession,
+		runAutoCompact,
+
+		setCurrentSnapshot(snapshot: ContextSnapshot): void {
+			currentContextSnapshot = snapshot;
+		},
+
+		flushReconciledSnapshot(): void {
+			if (!snapshotPersistPending || !currentContextSnapshot) return;
+			persistContextSnapshot(currentContextSnapshot);
+		},
+
+		reconcileUsage(usage: Usage): void {
+			if (!currentContextSnapshot) return;
+			// Reconcile in memory on every API call so the live meters
+			// track usage; persistence waits for the run to settle.
+			currentContextSnapshot = reconcileSnapshot(currentContextSnapshot, usage);
+			snapshotPersistPending = true;
+		},
+
+		promptSideTokens(): number {
+			return currentContextSnapshot ? snapshotInputTokens(currentContextSnapshot) : 0;
+		},
+
+		/**
+		 * Ensure the session system prompt is compiled and applied to the live
+		 * agent. Compiles only when the compile key (target, model, safety
+		 * level, session id) changes or a config hot-reload invalidated the
+		 * cache; every other submit reuses the cached prompt byte-for-byte. A
+		 * compile whose text differs from the previous prompt queues a
+		 * "promptRecompiled" ledger entry (written once the session exists).
+		 */
+		async ensureSessionPrompt(agentRuntime: AgentRuntime): Promise<CompiledSessionPrompt | null> {
+			if (!deps.prompts) return null;
+			const settings = deps.getSettings();
+			const autonomy = settings.autonomy ?? "auto-edit";
+			const sessionId = deps.session?.current()?.id ?? "";
+			const workingContextKey = [...sessionWorkingContextPaths].sort().join("\0");
+			const key = `${agentRuntime.targetId}|${agentRuntime.wireModelId}|${autonomy}|${sessionId}|${workingContextKey}`;
+			if (sessionPrompt && sessionPromptKey === key) {
+				lastSystemPromptReused = true;
+				return sessionPrompt;
+			}
+			const modelState = agentRuntime.agent.state.model as
+				| (typeof agentRuntime.agent.state.model & { clio?: { quirks?: LocalModelQuirks } })
+				| undefined;
+			const contextWindow = typeof modelState?.contextWindow === "number" ? modelState.contextWindow : null;
+			const guidance = modelState?.clio?.quirks?.thinking?.guidance;
+			// Per-tool prompt hints come from registry metadata, derived once from
+			// the frozen surface per compile. The compiler renders them sorted by
+			// tool name, so the compiled text stays byte-stable for a given surface.
+			const toolNames = toolNamesFromAgentState(agentRuntime.agent.state.tools);
+			const toolPromptHints = toolNames.flatMap((name) => {
+				const hint = deps.toolRegistry?.get(name as ToolName)?.metadata?.promptHint;
+				return hint ? [{ tool: name, hint }] : [];
+			});
+			const sessionInputs: SessionPromptInputs = {
+				provider: agentRuntime.targetId,
+				model: agentRuntime.wireModelId,
+				contextWindow,
+				providerSupportsTools: runtimeSupportsTools(agentRuntime),
+				toolNames,
+				...(guidance ? { thinkingGuidance: guidance } : {}),
+				...(toolPromptHints.length > 0 ? { toolPromptHints } : {}),
+			};
+			if (deps.getMemorySection) {
+				try {
+					const memorySection = deps.getMemorySection();
+					if (memorySection.length > 0) sessionInputs.memorySection = memorySection;
+				} catch (err) {
+					deps.emitNotice(
+						`[Clio Coder] memory load failed; continuing without memory injection: ${err instanceof Error ? err.message : String(err)}`,
+					);
+				}
+			}
+			try {
+				const result = await deps.prompts.compileSessionPrompt({
+					sessionInputs,
+					autonomy,
+					cwd: process.cwd(),
+					workingContextPaths: [...sessionWorkingContextPaths],
+				});
+				const previousHash = sessionPrompt?.systemPromptHash ?? null;
+				const changed = agentRuntime.agent.state.systemPrompt !== result.systemPrompt;
+				if (changed) {
+					agentRuntime.agent.state.systemPrompt = result.systemPrompt;
+					pendingPromptLogEntry = {
+						at: new Date().toISOString(),
+						previousHash,
+						systemPromptHash: result.systemPromptHash,
+						tokenEstimate: result.tokenEstimate,
+						thinkingLevel: agentRuntime.agent.state.thinkingLevel ?? null,
+						projectPreload: result.projectPreload ?? null,
+						sections: result.sections.map((s) => ({ id: s.id, tokenEstimate: s.tokenEstimate })),
+						fragments: result.fragmentManifest.map((f) => ({
+							id: f.id,
+							relPath: f.relPath,
+							contentHash: f.contentHash,
+							dynamic: f.dynamic,
+						})),
+					};
+				}
+				lastSystemPromptReused = !changed;
+				sessionPrompt = result;
+				sessionPromptKey = key;
+				return result;
+			} catch (err) {
+				deps.emitNotice(
+					`[Clio Coder] prompt compile failed; using fallback identity: ${err instanceof Error ? err.message : String(err)}`,
+				);
+				return null;
+			}
+		},
+
+		/**
+		 * Write the queued prompt-compile ledger entry and its full manifest
+		 * record. Deferred until after the user turn is appended so the session
+		 * is guaranteed to exist. The current.jsonl entry stays hash-only; the
+		 * section/fragment breakdown goes to the prompt-manifest.jsonl sibling.
+		 */
+		logPromptCompileIfPending(): void {
+			const currentMeta = deps.session?.current();
+			if (!pendingPromptLogEntry || !currentMeta) return;
+			const entry = pendingPromptLogEntry;
+			pendingPromptLogEntry = null;
+			try {
+				deps.session?.appendEntry({
+					kind: "custom",
+					customType: "promptRecompiled",
+					parentTurnId: state.lastTurnId,
+					data: {
+						previousHash: entry.previousHash,
+						hash: entry.systemPromptHash,
+						tokenEstimate: entry.tokenEstimate,
+					},
+				});
+			} catch {
+				// Ledger logging is diagnostics, not control flow; never abort a turn.
+			}
+			appendPromptCompileRecord(currentMeta, entry);
+		},
+
+		invalidateSessionPromptCache(): void {
+			sessionPromptKey = null;
+		},
+
+		addWorkingContextPaths(paths: ReadonlyArray<string>): void {
+			for (const path of paths) sessionWorkingContextPaths.add(path);
+		},
+
+		async postToolContinuationGuard(agentRuntime: AgentRuntime, signal?: AbortSignal) {
+			if (signal?.aborted || !toolResultTail(agentRuntime)) return undefined;
+			const before = liveContextEstimate(agentRuntime);
+			if (before.contextWindow <= 0 || before.tokens <= 0) return undefined;
+
+			const settings = deps.getSettings();
+			const threshold = settings.compaction?.threshold ?? DEFAULT_COMPACTION_THRESHOLD;
+			const verdict = shouldCompact(before.tokens, threshold, before.contextWindow);
+			let compacted = false;
+			if (verdict.shouldCompact) {
+				try {
+					compacted = await runAutoCompact(agentRuntime, false, undefined, "auto");
+				} catch (err) {
+					throw new Error(
+						`[Clio Coder] post-tool context guard could not compact before continuation: ${err instanceof Error ? err.message : String(err)}`,
+					);
+				}
+			}
+
+			const after = liveContextEstimate(agentRuntime);
+			if (after.tokens >= after.contextWindow) {
+				throw new Error(
+					`[Clio Coder] post-tool context guard stopped continuation before provider call: estimated ${after.tokens} tokens exceeds reported context window ${after.contextWindow}. Use /context compact, narrower reads, or a follow-up turn with smaller observations.`,
+				);
+			}
+			return compacted ? continuationContextUpdate(agentRuntime) : undefined;
+		},
+
+		contextUsage(): ContextUsageSnapshot {
+			if (!state.runtime) return contextUsageSnapshot(null, 0);
+			const effectiveWindow = state.runtime.runtimeResolution.contextWindowDetails.effectiveContextWindow;
+			if (!currentContextSnapshot) {
+				return contextUsageSnapshot(null, effectiveWindow);
+			}
+
+			const pendingTokens = pendingUserInputTokens();
+			const totalUsed = snapshotInputTokens(currentContextSnapshot) + pendingTokens + liveStreamingOutputTokens();
+			const breakdown: ContextUsageBreakdown = {
+				systemPromptTokens: currentContextSnapshot.categories.system,
+				messageTokens: currentContextSnapshot.categories.messages,
+				pendingUserTokens: pendingTokens,
+				toolSchemaTokens: currentContextSnapshot.categories.tools,
+			};
+			return contextUsageSnapshot(totalUsed > 0 ? totalUsed : null, effectiveWindow, breakdown);
+		},
+
+		contextLedger(): ContextLedger {
+			const settings = deps.getSettings();
+			const compactionThreshold = settings.compaction?.threshold ?? null;
+			const compactionAuto = settings.compaction?.auto !== false;
+			if (!state.runtime) {
+				return buildContextLedger({
+					provider: settings.orchestrator?.target ?? null,
+					model: settings.orchestrator?.model ?? null,
+					contextWindow: 0,
+					compactionThreshold,
+					compactionAuto,
+				});
+			}
+			const windowDetails = state.runtime.runtimeResolution.contextWindowDetails;
+			const effectiveWindow = windowDetails.effectiveContextWindow;
+
+			const provider = state.runtime.targetId ?? settings.orchestrator?.target ?? null;
+			const model = state.runtime.wireModelId ?? settings.orchestrator?.model ?? null;
+
+			if (!currentContextSnapshot) {
+				return buildContextLedger({
+					provider,
+					model,
+					contextWindow: effectiveWindow,
+					contextWindowSource: windowDetails.contextWindowSource,
+					toolCount: state.runtime.agent.state.tools.length,
+					compactionThreshold,
+					compactionAuto,
+					promptCache: lastPromptCache,
+				});
+			}
+
+			const streamingOutput = liveStreamingOutputTokens();
+			const pendingTokens = pendingUserInputTokens();
+			const totalUsed = snapshotInputTokens(currentContextSnapshot) + pendingTokens + streamingOutput;
+			const measured = currentContextSnapshot.sources.total === "reconciled";
+
+			return buildContextLedger({
+				provider,
+				model,
+				contextWindow: effectiveWindow,
+				contextWindowSource: windowDetails.contextWindowSource,
+				compactionThreshold,
+				compactionAuto,
+				systemPromptTokens: currentContextSnapshot.categories.system,
+				toolSchemaTokens: currentContextSnapshot.categories.tools,
+				// Persisted snapshots strip the captured schemas; fall back to
+				// the live agent state after a session resume.
+				toolCount: currentContextSnapshot.activeToolSchemas?.length ?? state.runtime.agent.state.tools.length,
+				messageTokens: currentContextSnapshot.categories.messages,
+				agentsTokens: currentContextSnapshot.categories.agents,
+				skillsTokens: currentContextSnapshot.categories.skills,
+				memoryTokens: currentContextSnapshot.categories.memory,
+				projectTokens: currentContextSnapshot.categories.project,
+				projectPreload: sessionPrompt?.projectPreload?.label ?? null,
+				projectHandbookFiles: sessionPrompt?.projectHandbookFiles ?? null,
+				pendingTokens,
+				streamingTokens: streamingOutput,
+				liveTotalTokens: totalUsed > 0 ? totalUsed : null,
+				measured,
+				lastCompaction: lastCompactionEvent,
+				promptCache: lastPromptCache,
+			});
+		},
+
+		emitContextWindowWarningTransition(warning: string | null): void {
+			if (warning === lastContextWindowWarning) return;
+			lastContextWindowWarning = warning;
+			deps.bus?.emit(BusChannels.ContextWarning, { warning } satisfies ContextWarningPayload);
+		},
+
+		consumeExpectedColdReasons(runtimeId: string): void {
+			// Cache-disturbance honesty (T3.3): consume disturbances since
+			// the last settled run. Only single-slot local backends lose their
+			// prefix cache to interleaved work, so only local-native targets
+			// stamp reasons and notify; other tiers just clear the set.
+			runExpectedColdReasons = [];
+			stampColdReasonsPending = false;
+			if (pendingColdReasons.size > 0) {
+				const reasons = [...pendingColdReasons];
+				pendingColdReasons.clear();
+				if (deps.providers.getRuntime(runtimeId)?.tier === "local-native") {
+					runExpectedColdReasons = reasons;
+					stampColdReasonsPending = true;
+					deps.emitNotice(`[context engine] backend prefix cache likely cold this turn: ${reasons.join(", ")}`);
+				}
+			}
+		},
+
+		promptCachePayloadForAssistant(usage: Usage): Record<string, unknown> {
+			// Per-call prompt-cache record (T3.2): provider-reported numbers only,
+			// classified by backendCacheVerdict in chat-loop-messages.ts. The
+			// run's first persisted call also carries any expected-cold reasons.
+			const input = typeof usage.input === "number" ? usage.input : 0;
+			const cacheRead = typeof usage.cacheRead === "number" ? usage.cacheRead : 0;
+			const cacheWrite = typeof usage.cacheWrite === "number" ? usage.cacheWrite : 0;
+			const promptCache: Record<string, unknown> = {
+				input,
+				cacheRead,
+				cacheWrite,
+				backendVerdict: backendCacheVerdict(input, cacheRead),
+			};
+			if (stampColdReasonsPending && runExpectedColdReasons.length > 0) {
+				promptCache.expectedColdReasons = [...runExpectedColdReasons];
+				stampColdReasonsPending = false;
+			}
+			return promptCache;
+		},
+
+		noteRunCacheSummary(messages, runFirstCallVerdict): void {
+			const cacheSummary = sumRunUsage(messages);
+			if (cacheSummary.hadUsage) {
+				lastPromptCache = {
+					shellReused: lastSystemPromptReused,
+					cacheReadTokens: cacheSummary.cacheRead > 0 || cacheSummary.cacheWrite > 0 ? cacheSummary.cacheRead : null,
+					cacheWriteTokens: cacheSummary.cacheRead > 0 || cacheSummary.cacheWrite > 0 ? cacheSummary.cacheWrite : null,
+					uncachedInputTokens: cacheSummary.input,
+					backendVerdict: runFirstCallVerdict,
+				};
+			}
+		},
+
+		resetForSession(): void {
+			lastPromptCache = null;
+			lastSystemPromptReused = false;
+			sessionPromptKey = null;
+			sessionWorkingContextPaths.clear();
+			pendingPromptLogEntry = null;
+			const session = deps.session?.current();
+			currentContextSnapshot = session ? getLatestContextSnapshot(session) : null;
+		},
+
+		dispose(): void {
+			for (const unsubscribe of unsubscribeColdReasonSources) unsubscribe?.();
+		},
+	};
+}