@iowarp/clio-coder 0.3.3 → 0.3.6

package/src/interactive/turn-runtime.ts

@@ -7,6 +7,7 @@
 
 import { performance } from "node:perf_hooks";
 import type { ClioSettings } from "../core/config.js";
+import { attributedModelId } from "../core/response-model-id.js";
 import type { MiddlewareToolChoiceControl } from "../domains/middleware/index.js";
 import type { ObservabilityContract } from "../domains/observability/contract.js";
 import {
@@ -15,7 +16,7 @@ import {
 	type ProvidersContract,
 	refineRuntimeTargetWithModelHints,
 	resolveRuntimeTarget,
-	runtimeResolutionWarnings,
+	runtimeResolutionWarningsBesideThinkingNotice,
 	type TargetStatus,
 	targetRequiresAuth,
 } from "../domains/providers/index.js";
@@ -199,7 +200,12 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 				failure?.code ?? "admission-failed",
 			);
 		}
-		for (const message of runtimeResolutionWarnings(resolved.diagnostics)) {
+		// emitThinkingClampNotice prints the one combined thinking line when the
+		// dial takes effect; the thinking diagnostics are its two halves.
+		for (const message of runtimeResolutionWarningsBesideThinkingNotice(
+			resolved.diagnostics,
+			resolved.target.modelRuntime.thinking.notice,
+		)) {
 			const key = `${targetId}|${wireModelId}|${message}`;
 			if (announcedResolutionWarnings.has(key)) continue;
 			announcedResolutionWarnings.add(key);
@@ -229,6 +235,7 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 	 */
 	const TARGET_PROBE_TTL_MS = 5 * 60 * 1000;
 	let lastTargetProbe: { key: string; at: number } | null = null;
+	const targetProbesInFlight = new Map<string, Promise<void>>();
 	const ensureLiveCapabilitiesForSelectedModel = async (): Promise<void> => {
 		const settings = deps.getSettings();
 		const targetId = settings.orchestrator.target?.trim();
@@ -239,16 +246,29 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 		const runtimeDesc = deps.providers.getRuntime(target.runtime);
 		if (!runtimeDesc || runtimeDesc.tier === "cloud") return;
 		const key = `${targetId}|${wireModelId}`;
+		const inFlight = targetProbesInFlight.get(key);
+		if (inFlight) {
+			await inFlight;
+			return;
+		}
 		const now = Date.now();
 		if (lastTargetProbe?.key === key && now - lastTargetProbe.at < TARGET_PROBE_TTL_MS) return;
 		lastTargetProbe = { key, at: now };
-		let status: Awaited<ReturnType<ProvidersContract["probeTarget"]>> = null;
+		const probe = (async (): Promise<void> => {
+			let status: Awaited<ReturnType<ProvidersContract["probeTarget"]>> = null;
+			try {
+				status = await deps.providers.probeTarget(targetId);
+			} catch {
+				// Fall back to the last known target state.
+			}
+			announceColdModel(status, targetId, wireModelId);
+		})();
+		targetProbesInFlight.set(key, probe);
 		try {
-			status = await deps.providers.probeTarget(targetId);
-		} catch {
-			// Fall back to the last known target state.
+			await probe;
+		} finally {
+			if (targetProbesInFlight.get(key) === probe) targetProbesInFlight.delete(key);
 		}
-		announceColdModel(status, targetId, wireModelId);
 	};
 
 	/** One notice per target+model+state, so a repeated probe stays quiet. */
@@ -589,6 +609,7 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 				middlewareToolChoice.toolStarted(event.toolName);
 				deps.toolStartTimes.set(event.toolCallId, eventClock);
 				state.turnToolCalls += 1;
+				state.turnToolNames.push(event.toolName);
 			} else if (event.type === "tool_execution_end") {
 				toolsInFlight = Math.max(0, toolsInFlight - 1);
 				const startedAt = deps.toolStartTimes.get(event.toolCallId);
@@ -713,7 +734,11 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 				if (summary.hadUsage && (summary.tokens > 0 || summary.costUsd > 0)) {
 					deps.observability.recordTokens(
 						localRuntime.targetId,
-						localRuntime.wireModelId,
+						attributedModelId(
+							summary.lastResponseModelIdObservation,
+							localRuntime.wireModelId,
+							summary.lastDifferingResponseModelId,
+						),
 						summary.tokens,
 						summary.costUsd,
 						{
@@ -726,6 +751,10 @@ export function createTurnRuntime(deps: TurnRuntimeDeps): TurnRuntime {
 							apiCalls: summary.apiCalls,
 						},
 						localRuntime.runtimeResolution.costProvenance,
+						{
+							requestedModelIds: [localRuntime.wireModelId],
+							responseModelIdObservationCounts: summary.responseModelIdObservationCounts,
+						},
 					);
 				}
 				if (summary.output > 0 && firstAssistantDeltaAt !== null) {

package/src/interactive/turn-state.ts

@@ -97,6 +97,8 @@ export interface ChatTurnState {
 	toolProseAssessedChars: number;
 	/** Tool calls observed in the current turn; feeds turn_end metadata. */
 	turnToolCalls: number;
+	/** Names of those calls, in order, so turn_end can tell a listing-only turn from a working one. */
+	turnToolNames: string[];
 	/** A `[worker result]` note the operator shared entered this turn, as the prompt or as a steer; feeds turn_end metadata. */
 	turnSharedWorkerNote: boolean;
 	/** One stalled-turn nudge per user prompt; reset when a real prompt arrives. */
@@ -123,6 +125,7 @@ export function createTurnState(initialThinkingLevel: ThinkingLevel): ChatTurnSt
 		streamStallReason: null,
 		toolProseAssessedChars: 0,
 		turnToolCalls: 0,
+		turnToolNames: [],
 		turnSharedWorkerNote: false,
 		stalledTurnNudgeSpent: false,
 		pendingRequestContinuation: false,

package/src/interactive/worker-progress.ts

@@ -0,0 +1,440 @@
+/**
+ * The one projection of a worker's live progress.
+ *
+ * A dispatched run publishes its events on a single stream, and two surfaces
+ * read them: the transcript block a `/run` or `/delegate` opens, and the Fleet
+ * Runs board. Both used to interpret that stream themselves, so the same event
+ * could mean two things and the board could only manage a richer spinner while
+ * the transcript already held the answer. This module is the fold both now go
+ * through, so what an operator reads about a worker is one story told twice
+ * rather than two stories.
+ *
+ * Three rules it enforces rather than documents:
+ *
+ *   - Arguments never arrive here. A tool's name comes from the `clio_*`
+ *     telemetry events, and what the call is doing comes from the bounded
+ *     {@link CallActionDescriptor} composed at the tool-admission seam. The
+ *     `tool_execution_*` events, whose `args` carry the literal call arguments,
+ *     are not read at all.
+ *   - Reasoning content is never retained. A `thinking_delta` moves the phase
+ *     and nothing else: the operator learns that the worker is thinking, never
+ *     what it is thinking.
+ *   - Everything is bounded before it is kept. Lines, bytes, the action trail,
+ *     the distinct tool names, and the rate at which delta bytes are accepted
+ *     all have explicit caps, so neither a long run nor a hostile one can grow
+ *     what a render has to walk.
+ *
+ * Pure: no I/O, no bus, no clock of its own. `observe` takes the instant it
+ * should measure the rate window against, so every sequence below is testable.
+ */
+
+import { durableAssistantTextFromEvent, WORKER_OUTPUT_MAX_BYTES } from "../domains/dispatch/event-pump.js";
+import type { CallActionDescriptor } from "../domains/safety/call-target.js";
+import { truncateUtf8 } from "../tools/truncate-utf8.js";
+
+/**
+ * Lines of worker prose kept while a run is live. The tail is what a streaming
+ * answer is about, so the head is what gets dropped; once the run settles the
+ * body is replaced by the receipt-sealed answer and anchors at the head, which
+ * is the same rule the thinking block already follows.
+ */
+export const WORKER_LIVE_TAIL_LINES = 40;
+
+/**
+ * Bytes of live tail retained. The line bound alone leaves a worker that emits
+ * one enormous line unbounded, so the byte bound is the backstop: whole lines
+ * leave the head until the tail fits, and a single line larger than the bound
+ * is cut at its head.
+ */
+export const WORKER_LIVE_TAIL_MAX_BYTES = 4096;
+
+/** Distinct tool names carried on an entry. The receipt's call count carries the total. */
+export const WORKER_TOOL_NAME_LIMIT = 8;
+
+/** Finished actions kept as the recent trail, newest first. */
+export const WORKER_ACTION_TRAIL_LIMIT = 4;
+
+/**
+ * The rate window for accepted delta bytes. A worker streaming faster than an
+ * operator can read is not producing more information, only more work for the
+ * fold, so bytes past the budget in one window are counted and dropped rather
+ * than concatenated into a tail that would immediately discard them anyway.
+ */
+export const WORKER_PROGRESS_WINDOW_MS = 250;
+
+/** Delta bytes accepted per {@link WORKER_PROGRESS_WINDOW_MS}. */
+export const WORKER_PROGRESS_WINDOW_BYTES = 16_384;
+
+const TAIL_TRUNCATION_MARKER = "";
+
+/**
+ * Where a worker is in its turn. Derived from the event stream and never from
+ * reasoning content: `thinking` says only that a reasoning block is open.
+ */
+export type WorkerProgressPhase = "starting" | "thinking" | "writing" | "tool" | "waiting" | "settled";
+
+/** One tool call as an operator surface may show it: the name plus its redacted descriptor. */
+export interface WorkerAction {
+	tool: string;
+	/** Correlation identity carried by new streams. Absent on legacy events. */
+	toolCallId?: string;
+	/** Absent when the runtime emitted no descriptor, so the name stands alone. */
+	descriptor?: CallActionDescriptor;
+}
+
+/**
+ * Everything both surfaces read about a live worker, bounded. Snapshots are
+ * immutable and `revision` changes only when a visible field does, so a
+ * renderer can skip a repaint by identity rather than by diffing text.
+ */
+export interface WorkerProgressSnapshot {
+	revision: number;
+	phase: WorkerProgressPhase;
+	/** Bounded worker prose: the live tail while running, the sealed answer once settled. */
+	tailText: string;
+	/** Lines the bounds dropped from `tailText`. */
+	droppedLines: number;
+	/** Bytes refused by the rate window or cut from an oversized line. */
+	droppedBytes: number;
+	/** The most recently started call still executing; null between calls and once settled. */
+	currentAction: WorkerAction | null;
+	/** Finished calls, newest first, bounded by {@link WORKER_ACTION_TRAIL_LIMIT}. */
+	recentActions: ReadonlyArray<WorkerAction>;
+	/** Distinct tool names in first-use order, bounded by {@link WORKER_TOOL_NAME_LIMIT}. */
+	toolNames: ReadonlyArray<string>;
+	settled: boolean;
+}
+
+export interface WorkerProgressFold {
+	/**
+	 * Fold one dispatch progress event. Returns true when the snapshot changed,
+	 * so a caller can decide whether the surface needs anything at all.
+	 */
+	observe(event: unknown, nowMs?: number): boolean;
+	/**
+	 * Seal the projection. `text` is the receipt's answer where one could be
+	 * read; without it the live tail stands as the run's last honest word.
+	 */
+	settle(text?: string): boolean;
+	/**
+	 * A later attempt of the same work takes over. The tail and the trail are
+	 * history the operator is already reading and stay; the current call, the
+	 * settled mark, and the previous attempt's durable answer go, because an
+	 * answer belongs to the attempt that produced it.
+	 */
+	restart(): void;
+	snapshot(): WorkerProgressSnapshot;
+	/** The last durable assistant answer this run streamed, unbounded by the tail. */
+	durableText(): string;
+}
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+	return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+
+function nonEmptyString(value: unknown): string | undefined {
+	return typeof value === "string" && value.trim().length > 0 ? value : undefined;
+}
+
+/**
+ * Incremental assistant prose carried by one worker event.
+ *
+ * Native workers slim `message_update` down to `assistantMessageEvent.delta`
+ * before it crosses the NDJSON seam (src/worker/event-projection.ts); the ACP
+ * mapper emits a top-level `text_delta` carrying `text`. Both spellings are
+ * read here so ACP peers need no separate UI path.
+ */
+function workerTextDelta(event: unknown): string {
+	if (!isRecord(event)) return "";
+	if (event.type === "message_update") {
+		const assistantEvent = isRecord(event.assistantMessageEvent) ? event.assistantMessageEvent : null;
+		if (assistantEvent?.type !== "text_delta") return "";
+		return typeof assistantEvent.delta === "string" ? assistantEvent.delta : "";
+	}
+	if (event.type !== "text_delta") return "";
+	if (typeof event.delta === "string") return event.delta;
+	return typeof event.text === "string" ? event.text : "";
+}
+
+/** Whether this event opens or continues a reasoning block. Its content is never read. */
+function isThinkingEvent(event: Record<string, unknown>): boolean {
+	if (event.type === "thinking_delta") return true;
+	if (event.type !== "message_update") return false;
+	const assistantEvent = isRecord(event.assistantMessageEvent) ? event.assistantMessageEvent : null;
+	return assistantEvent?.type === "thinking_delta";
+}
+
+/** The action a `clio_tool_start`/`clio_tool_finish` event names, or null when it is neither. */
+function toolEventAction(event: Record<string, unknown>): WorkerAction | null {
+	if (event.type !== "clio_tool_start" && event.type !== "clio_tool_finish") return null;
+	const payload = isRecord(event.payload) ? event.payload : null;
+	const tool = nonEmptyString(payload?.tool);
+	if (tool === undefined) return null;
+	const toolCallId = nonEmptyString(payload?.toolCallId);
+	const raw = isRecord(payload?.action) ? (payload.action as Partial<CallActionDescriptor>) : undefined;
+	const verb = nonEmptyString(raw?.verb);
+	if (verb === undefined) return { tool, ...(toolCallId !== undefined ? { toolCallId } : {}) };
+	const object = nonEmptyString(raw?.object);
+	return {
+		tool,
+		...(toolCallId !== undefined ? { toolCallId } : {}),
+		descriptor: {
+			verb,
+			...(object !== undefined ? { object } : {}),
+			...(raw?.truncated === true ? { truncated: true } : {}),
+		},
+	};
+}
+
+function lineCount(text: string): number {
+	if (text.length === 0) return 0;
+	let lines = 1;
+	for (let index = 0; index < text.length; index += 1) {
+		if (text[index] === "\n") lines += 1;
+	}
+	return lines;
+}
+
+/** Keep the newest bytes of one oversized line, cutting on a UTF-8 boundary. */
+function keepTailBytes(text: string, maxBytes: number): string {
+	const buffer = Buffer.from(text, "utf8");
+	if (buffer.byteLength <= maxBytes) return text;
+	let cut = buffer.byteLength - maxBytes;
+	while (cut < buffer.byteLength) {
+		const byte = buffer[cut];
+		if (byte === undefined || (byte & 0xc0) !== 0x80) break;
+		cut += 1;
+	}
+	return buffer.subarray(cut).toString("utf8");
+}
+
+/**
+ * Bound the live tail: newest lines first, then whole lines from the head until
+ * the byte cap is met, and finally a head cut on a single line larger than the
+ * cap on its own.
+ */
+function boundLiveTail(text: string): { text: string; droppedLines: number; droppedBytes: number } {
+	const lines = text.split("\n");
+	let droppedLines = 0;
+	if (lines.length > WORKER_LIVE_TAIL_LINES) {
+		droppedLines = lines.length - WORKER_LIVE_TAIL_LINES;
+		lines.splice(0, droppedLines);
+	}
+	while (lines.length > 1 && Buffer.byteLength(lines.join("\n"), "utf8") > WORKER_LIVE_TAIL_MAX_BYTES) {
+		lines.shift();
+		droppedLines += 1;
+	}
+	const joined = lines.join("\n");
+	const bytes = Buffer.byteLength(joined, "utf8");
+	if (bytes <= WORKER_LIVE_TAIL_MAX_BYTES) return { text: joined, droppedLines, droppedBytes: 0 };
+	const kept = keepTailBytes(joined, WORKER_LIVE_TAIL_MAX_BYTES);
+	return { text: kept, droppedLines, droppedBytes: bytes - Buffer.byteLength(kept, "utf8") };
+}
+
+/** Head-anchored byte bound for a settled answer, matching the receipt's own bound. */
+export function boundSettledText(text: string): { text: string; dropped: number } {
+	const bounded = truncateUtf8(text, WORKER_OUTPUT_MAX_BYTES, TAIL_TRUNCATION_MARKER);
+	if (bounded === text) return { text, dropped: 0 };
+	return { text: bounded, dropped: Math.max(0, lineCount(text) - lineCount(bounded)) };
+}
+
+const EMPTY_SNAPSHOT: WorkerProgressSnapshot = {
+	revision: 0,
+	phase: "starting",
+	tailText: "",
+	droppedLines: 0,
+	droppedBytes: 0,
+	currentAction: null,
+	recentActions: [],
+	toolNames: [],
+	settled: false,
+};
+
+export function createWorkerProgressFold(): WorkerProgressFold {
+	let revision = 0;
+	let phase: WorkerProgressPhase = "starting";
+	let tailText = "";
+	let droppedLines = 0;
+	let droppedBytes = 0;
+	let currentAction: WorkerAction | null = null;
+	const recentActions: WorkerAction[] = [];
+	const toolNames: string[] = [];
+	let settled = false;
+	let durable = "";
+	/** Starts still awaiting a finish, oldest first. */
+	const pendingActions: WorkerAction[] = [];
+	/** New streams resolve exact starts through this index. */
+	const pendingActionsById = new Map<string, WorkerAction>();
+	let windowStartMs = 0;
+	let windowBytes = 0;
+	let cached: WorkerProgressSnapshot | null = EMPTY_SNAPSHOT;
+
+	const touch = (): true => {
+		revision += 1;
+		cached = null;
+		return true;
+	};
+
+	const setPhase = (next: WorkerProgressPhase): boolean => {
+		if (phase === next) return false;
+		phase = next;
+		return touch();
+	};
+
+	/**
+	 * Admit delta bytes against the rate window. Returns the text to append,
+	 * which is empty once the window's budget is spent; the refused bytes are
+	 * counted so the surface can say the stream outran it rather than quietly
+	 * showing less.
+	 */
+	const admitDelta = (delta: string, nowMs: number): string => {
+		if (nowMs - windowStartMs >= WORKER_PROGRESS_WINDOW_MS) {
+			windowStartMs = nowMs;
+			windowBytes = 0;
+		}
+		const bytes = Buffer.byteLength(delta, "utf8");
+		if (windowBytes + bytes <= WORKER_PROGRESS_WINDOW_BYTES) {
+			windowBytes += bytes;
+			return delta;
+		}
+		windowBytes = WORKER_PROGRESS_WINDOW_BYTES;
+		droppedBytes += bytes;
+		return "";
+	};
+
+	const appendTail = (text: string): boolean => {
+		if (text.length === 0) return false;
+		const bounded = boundLiveTail(tailText + text);
+		tailText = bounded.text;
+		droppedLines += bounded.droppedLines;
+		droppedBytes += bounded.droppedBytes;
+		return touch();
+	};
+
+	const noteToolName = (tool: string): boolean => {
+		if (toolNames.includes(tool) || toolNames.length >= WORKER_TOOL_NAME_LIMIT) return false;
+		toolNames.push(tool);
+		return touch();
+	};
+
+	const removePendingAction = (action: WorkerAction): void => {
+		const index = pendingActions.lastIndexOf(action);
+		if (index >= 0) pendingActions.splice(index, 1);
+		if (action.toolCallId !== undefined && pendingActionsById.get(action.toolCallId) === action) {
+			pendingActionsById.delete(action.toolCallId);
+		}
+	};
+
+	const lastPendingActionForTool = (tool: string): WorkerAction | undefined => {
+		for (let index = pendingActions.length - 1; index >= 0; index -= 1) {
+			const candidate = pendingActions[index];
+			if (candidate?.tool === tool) return candidate;
+		}
+		return undefined;
+	};
+
+	return {
+		observe(event: unknown, nowMs = Date.now()): boolean {
+			if (!isRecord(event) || settled) return false;
+			let changed = false;
+
+			const delta = workerTextDelta(event);
+			if (delta.length > 0) {
+				changed = appendTail(admitDelta(delta, nowMs)) || changed;
+				changed = setPhase("writing") || changed;
+			} else if (isThinkingEvent(event)) {
+				// The phase moves and nothing else: reasoning content is never kept.
+				changed = setPhase("thinking") || changed;
+			}
+
+			// A durable message_end is the answer as the receipt will seal it. An
+			// ACP peer that streamed no deltas reaches a surface only here.
+			const durableText = durableAssistantTextFromEvent(event);
+			if (durableText.trim().length > 0) {
+				durable = durableText;
+				if (delta.length === 0 && tailText.trim().length === 0) {
+					changed = appendTail(durableText) || changed;
+				}
+			}
+
+			const action = toolEventAction(event);
+			if (action !== null && event.type === "clio_tool_start") {
+				if (action.toolCallId !== undefined) {
+					const duplicate = pendingActionsById.get(action.toolCallId);
+					if (duplicate !== undefined) removePendingAction(duplicate);
+					pendingActionsById.set(action.toolCallId, action);
+				}
+				pendingActions.push(action);
+				currentAction = action;
+				noteToolName(action.tool);
+				touch();
+				setPhase("tool");
+				changed = true;
+			} else if (action !== null && event.type === "clio_tool_finish") {
+				// An id is authoritative when both sides carry it. If either side came
+				// from an older or incomplete producer, retain the historical name match.
+				const finished =
+					(action.toolCallId !== undefined ? pendingActionsById.get(action.toolCallId) : undefined) ??
+					lastPendingActionForTool(action.tool) ??
+					action;
+				removePendingAction(finished);
+				currentAction = pendingActions[pendingActions.length - 1] ?? null;
+				recentActions.unshift(finished);
+				if (recentActions.length > WORKER_ACTION_TRAIL_LIMIT) recentActions.length = WORKER_ACTION_TRAIL_LIMIT;
+				noteToolName(finished.tool);
+				touch();
+				setPhase(currentAction === null ? "waiting" : "tool");
+				changed = true;
+			}
+
+			if (event.type === "agent_start") changed = setPhase("starting") || changed;
+			return changed;
+		},
+
+		settle(text?: string): boolean {
+			const sealed = nonEmptyString(text) ?? nonEmptyString(durable);
+			if (sealed !== undefined) {
+				const bounded = boundSettledText(sealed);
+				tailText = bounded.text;
+				droppedLines = bounded.dropped;
+			}
+			currentAction = null;
+			pendingActions.length = 0;
+			pendingActionsById.clear();
+			settled = true;
+			phase = "settled";
+			touch();
+			return true;
+		},
+
+		restart(): void {
+			currentAction = null;
+			pendingActions.length = 0;
+			pendingActionsById.clear();
+			durable = "";
+			settled = false;
+			phase = "starting";
+			touch();
+		},
+
+		snapshot(): WorkerProgressSnapshot {
+			if (cached !== null) return cached;
+			cached = {
+				revision,
+				phase,
+				tailText,
+				droppedLines,
+				droppedBytes,
+				currentAction: currentAction === null ? null : { ...currentAction },
+				recentActions: recentActions.map((entry) => ({ ...entry })),
+				toolNames: [...toolNames],
+				settled,
+			};
+			return cached;
+		},
+
+		durableText(): string {
+			return durable;
+		},
+	};
+}