@iowarp/clio-coder 0.3.3 → 0.3.6

package/src/interactive/worker-stream.ts

@@ -8,6 +8,13 @@
  * to a bus, or renders; the chat panel owns presentation and the wiring owns
  * I/O, so every sequence below is testable as a pure fold.
  *
+ * What a worker is saying and doing is not decided here. That is the canonical
+ * {@link WorkerProgressSnapshot} projection in `worker-progress.ts`, which the
+ * Fleet Runs board reads from the same events, so the two surfaces cannot drift
+ * on the answer tail, the tool names, or what a call is touching. This module
+ * owns the half the board has no use for: which assignment an event belongs to,
+ * which attempt is current, and what the receipt sealed.
+ *
  * Three rules the fold enforces rather than documents:
  *
  *   - Tool *names* only. Names come exclusively from the `clio_*` telemetry
@@ -34,10 +41,16 @@ import type {
 	DispatchStartedPayload,
 	RunAbortedPayload,
 } from "../core/bus-events.js";
-import { durableAssistantTextFromEvent, WORKER_OUTPUT_MAX_BYTES } from "../domains/dispatch/event-pump.js";
 import type { RunKind } from "../domains/dispatch/types.js";
 import type { WorkerRunOrigin, WorkerRunRuntime, WorkerRunRuntimeKind } from "../domains/session/index.js";
-import { truncateUtf8 } from "../tools/truncate-utf8.js";
+import { createWorkerProgressFold, type WorkerProgressFold, type WorkerProgressSnapshot } from "./worker-progress.js";
+
+export {
+	boundSettledText,
+	WORKER_LIVE_TAIL_LINES,
+	WORKER_TOOL_NAME_LIMIT,
+} from "./worker-progress.js";
+export type { WorkerProgressSnapshot };
 
 export interface WorkerAttempt {
 	runId: string;
@@ -95,6 +108,8 @@ export interface WorkerEntryState {
 	droppedLines: number;
 	/** Distinct tool names in first-use order, bounded. */
 	tools: string[];
+	/** Canonical live projection. Absent only on receipt-based session replay. */
+	progress?: WorkerProgressSnapshot;
 	/** Every attempt of this assignment, oldest first. */
 	attempts: WorkerAttempt[];
 	pending: boolean;
@@ -143,23 +158,6 @@ export interface WorkerStream {
 	reset(): void;
 }
 
-/**
- * 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;
-
-/** Distinct tool names carried on an entry. The receipt's call count carries the total. */
-export const WORKER_TOOL_NAME_LIMIT = 8;
-
-const WORKER_TEXT_TRUNCATION_MARKER = "";
-
-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;
 }
@@ -202,58 +200,6 @@ export function workerTargetLabel(runtime: WorkerRunRuntime): string {
 	return runtime.targetId ?? runtime.wireModelId ?? "unknown";
 }
 
-/**
- * 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 textDelta(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 : "";
-}
-
-/** Tool name from Clio's own telemetry. Never from `tool_execution_*`, whose args are the call's arguments. */
-function toolName(event: unknown): string | undefined {
-	if (!isRecord(event)) return undefined;
-	if (event.type !== "clio_tool_start" && event.type !== "clio_tool_finish") return undefined;
-	const payload = isRecord(event.payload) ? event.payload : null;
-	return nonEmptyString(payload?.tool);
-}
-
-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 `WORKER_LIVE_TAIL_LINES` lines; report what that cost. */
-function boundLiveText(text: string): { text: string; dropped: number } {
-	const lines = text.split("\n");
-	if (lines.length <= WORKER_LIVE_TAIL_LINES) return { text, dropped: 0 };
-	const dropped = lines.length - WORKER_LIVE_TAIL_LINES;
-	return { text: lines.slice(dropped).join("\n"), dropped };
-}
-
-/** 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, WORKER_TEXT_TRUNCATION_MARKER);
-	if (bounded === text) return { text, dropped: 0 };
-	return { text: bounded, dropped: Math.max(0, lineCount(text) - lineCount(bounded)) };
-}
-
 /**
  * Terminal facts a lifecycle payload can supply on its own. Used only when the
  * sealed receipt is unreadable, so the block still reports an honest outcome
@@ -294,8 +240,21 @@ export function createWorkerStream(options: WorkerStreamOptions = {}): WorkerStr
 	const attemptByAssignment = new Map<string, number>();
 	/** Every attempt run id to its assignment, so progress and terminal events find their entry. */
 	const assignmentByRun = new Map<string, string>();
-	/** Durable answer text observed live, per attempt; the receipt wins when it is readable. */
-	const durableTextByRun = new Map<string, string>();
+	/**
+	 * The canonical progress projection per assignment. One fold spans every
+	 * attempt of an assignment, because a failover keeps streaming into the block
+	 * the operator is already reading.
+	 */
+	const progressByAssignment = new Map<string, WorkerProgressFold>();
+
+	/** Copy the canonical projection onto the entry fields the block renders. */
+	const applyProgress = (entry: WorkerEntryState, fold: WorkerProgressFold): void => {
+		const snapshot = fold.snapshot();
+		entry.text = snapshot.tailText;
+		entry.droppedLines = snapshot.droppedLines;
+		entry.tools = [...snapshot.toolNames];
+		entry.progress = snapshot;
+	};
 
 	/** The entry a run id addresses, and only while that run is the entry's current attempt. */
 	const currentEntryForRun = (runId: unknown): WorkerEntryState | undefined => {
@@ -315,11 +274,10 @@ export function createWorkerStream(options: WorkerStreamOptions = {}): WorkerStr
 		// Terminal truth is the sealed answer. A run whose receipt could not be
 		// read keeps whatever durable text its own stream produced, and a run that
 		// produced none keeps the live tail rather than blanking the block.
-		const settled = nonEmptyString(facts.text) ?? nonEmptyString(durableTextByRun.get(entry.runId));
-		if (settled !== undefined) {
-			const bounded = boundSettledText(settled);
-			entry.text = bounded.text;
-			entry.droppedLines = bounded.dropped;
+		const fold = progressByAssignment.get(entry.assignmentId);
+		if (fold !== undefined) {
+			fold.settle(nonEmptyString(facts.text));
+			applyProgress(entry, fold);
 		}
 		return { kind: "updated", entry };
 	};
@@ -343,10 +301,16 @@ export function createWorkerStream(options: WorkerStreamOptions = {}): WorkerStr
 				delete existing.receipt;
 				existing.attempts.push({ runId, targetLabel: workerTargetLabel(runtime) });
 				assignmentByRun.set(runId, assignmentId);
+				const progress = progressByAssignment.get(assignmentId);
+				if (progress !== undefined) {
+					progress.restart();
+					applyProgress(existing, progress);
+				}
 				return { kind: "updated", entry: existing };
 			}
 			if (payload.requestOrigin !== "user" && payload.requestOrigin !== "agent") return null;
 			attemptByAssignment.set(assignmentId, payload.attempt);
+			const progress = createWorkerProgressFold();
 			const entry: WorkerEntryState = {
 				assignmentId,
 				runId,
@@ -356,49 +320,26 @@ export function createWorkerStream(options: WorkerStreamOptions = {}): WorkerStr
 				text: "",
 				droppedLines: 0,
 				tools: [],
+				progress: progress.snapshot(),
 				attempts: [{ runId, targetLabel: workerTargetLabel(runtime) }],
 				pending: true,
 				...(payload.parentToolCallId !== undefined ? { parentToolCallId: payload.parentToolCallId } : {}),
 			};
 			byAssignment.set(assignmentId, entry);
 			assignmentByRun.set(runId, assignmentId);
+			progressByAssignment.set(assignmentId, progress);
 			return { kind: "created", entry };
 		},
 
 		progress(payload): WorkerStreamChange | null {
 			const entry = currentEntryForRun(payload.runId);
 			if (entry === undefined || !entry.pending) return null;
-			const event = payload.event;
-			let changed = false;
-
-			const delta = textDelta(event);
-			if (delta.length > 0) {
-				const bounded = boundLiveText(entry.text + delta);
-				entry.text = bounded.text;
-				entry.droppedLines += bounded.dropped;
-				changed = true;
-			}
-
-			// A durable message_end is the answer as the receipt will seal it. An
-			// ACP peer that streamed no deltas reaches the transcript only here.
-			const durable = durableAssistantTextFromEvent(event);
-			if (durable.trim().length > 0) {
-				durableTextByRun.set(entry.runId, durable);
-				if (delta.length === 0 && entry.text.trim().length === 0) {
-					const bounded = boundLiveText(durable);
-					entry.text = bounded.text;
-					entry.droppedLines += bounded.dropped;
-				}
-				changed = true;
-			}
-
-			const tool = toolName(event);
-			if (tool !== undefined && !entry.tools.includes(tool) && entry.tools.length < WORKER_TOOL_NAME_LIMIT) {
-				entry.tools.push(tool);
-				changed = true;
-			}
-
-			return changed ? { kind: "updated", entry } : null;
+			const fold = progressByAssignment.get(entry.assignmentId);
+			// Everything the block shows about the stream comes from the canonical
+			// projection, so an event that moved nothing visible repaints nothing.
+			if (fold === undefined || !fold.observe(payload.event)) return null;
+			applyProgress(entry, fold);
+			return { kind: "updated", entry };
 		},
 
 		completed(payload): WorkerStreamChange | null {
@@ -440,7 +381,7 @@ export function createWorkerStream(options: WorkerStreamOptions = {}): WorkerStr
 			byAssignment.clear();
 			attemptByAssignment.clear();
 			assignmentByRun.clear();
-			durableTextByRun.clear();
+			progressByAssignment.clear();
 		},
 	};
 }

package/src/tools/agent-tools.ts

@@ -22,12 +22,14 @@ import { type SkillActivation, skillActivationFromToolDetails } from "../core/sk
 import type { ToolName } from "../core/tool-names.js";
 import { ToolNames } from "../core/tool-names.js";
 import type { ResolvedRuntimeTarget } from "../domains/providers/index.js";
+import { type CallActionDescriptor, describeCallAction } from "../domains/safety/call-target.js";
 import type { SafetyDecision } from "../domains/safety/contract.js";
 import { formatModelRejection } from "../domains/safety/rejection-feedback.js";
 import { validateEngineToolArguments } from "../engine/ai.js";
 import type { AgentTool, AgentToolResult, AgentToolUpdateCallback } from "../engine/types.js";
 import { applyToolProfile, type ToolProfileName } from "./profiles.js";
 import type { ToolInvokeOptions, ToolRegistry, ToolResult, ToolSpec } from "./registry.js";
+import { isDispositionedToolResultError, toolResultContextText } from "./result-disposition.js";
 
 /**
  * Lightweight per-call observability hook. Default no-op so unused
@@ -44,6 +46,12 @@ export type ToolOutcome = "ok" | "error" | "blocked";
 
 export interface ToolStartEvent {
 	tool: string;
+	/**
+	 * The engine's id for this call, when the producer has one. New producers
+	 * put the same id on start and finish so concurrent calls of one tool remain
+	 * independently attributable. Optional for direct callers and old streams.
+	 */
+	toolCallId?: string;
 	posture: "operating";
 	/**
 	 * Wall-clock anchor for the instant the call began, for a human correlating
@@ -52,6 +60,14 @@ export interface ToolStartEvent {
 	 * event is measured on that process's monotonic clock instead.
 	 */
 	startedAt: number;
+	/**
+	 * What this call is doing, as a bounded redacted descriptor composed here
+	 * where the arguments are trusted. This is the only account of a worker's
+	 * arguments that may cross the NDJSON stdout seam; the arguments themselves
+	 * never do, so an operator surface reads a verb and an object it can show,
+	 * never an argument object it would have to sanitize itself.
+	 */
+	action?: CallActionDescriptor;
 }
 
 export interface ToolFinishEvent {
@@ -99,10 +115,10 @@ type WorkerToolOkDetails = { kind: "ok" } & Record<string, unknown>;
 function projectToolResult(result: ToolResult): WorkerAgentToolResult {
 	const details = { ...(result.details ?? {}), kind: result.kind } as { kind: "ok" } | { kind: "error" };
 	if (result.kind === "error") {
-		return { content: [{ type: "text", text: result.message }], details };
+		return { content: [{ type: "text", text: toolResultContextText(result) }], details };
 	}
 	return {
-		content: [{ type: "text", text: result.output }],
+		content: [{ type: "text", text: toolResultContextText(result) }],
 		details,
 		...(result.terminate === true ? { terminate: true } : {}),
 	};
@@ -119,6 +135,8 @@ interface RunValidatedToolCallInput {
 	signal?: AbortSignal;
 	telemetry?: ToolTelemetry;
 	invokeOptions?: Partial<ToolInvokeOptions>;
+	/** Agent loops can preserve canonical failure envelopes and classify them after execution. */
+	returnDispositionedErrors?: boolean;
 }
 
 async function runValidatedToolCall(input: RunValidatedToolCallInput): Promise<WorkerAgentToolResult> {
@@ -127,11 +145,21 @@ async function runValidatedToolCall(input: RunValidatedToolCallInput): Promise<W
 	// spans the call, so `durationMs` survives a clock correction mid-tool.
 	const startedAt = Date.now();
 	const startedAtClock = performance.now();
-	telemetry?.onStart?.({ tool: spec.name, posture: "operating", startedAt });
-	// Stamped onto every finish event so a consumer can join this authoritative
-	// outcome to the engine's `tool_execution_end` for the same call.
+	// Stamped onto both lifecycle events so a concurrent finish resolves only
+	// the start for its own engine call. Direct callers may legitimately omit it.
 	const callId = input.invokeOptions?.toolCallId;
 	const withCallId = callId === undefined ? {} : { toolCallId: callId };
+	// The one seam that holds validated arguments and publishes telemetry, so it
+	// is where the redacted descriptor is composed. Everything downstream of here
+	// sees the descriptor and never the arguments.
+	const action = describeCallAction(spec.name, args);
+	telemetry?.onStart?.({
+		tool: spec.name,
+		...withCallId,
+		posture: "operating",
+		startedAt,
+		...(action !== null ? { action } : {}),
+	});
 	const invokeOpts: ToolInvokeOptions = {};
 	if (input.invokeOptions) Object.assign(invokeOpts, input.invokeOptions);
 	if (signal) invokeOpts.signal = signal;
@@ -183,13 +211,16 @@ async function runValidatedToolCall(input: RunValidatedToolCallInput): Promise<W
 			reason: verdict.result.message,
 			decision: verdict.decision,
 		});
-		throw new Error(verdict.result.message);
+		if (input.returnDispositionedErrors === true && isDispositionedToolResultError(verdict.result)) {
+			return projectToolResult(verdict.result);
+		}
+		throw new Error(toolResultContextText(verdict.result));
 	}
 	const toolDetails = isRecord(verdict.result.details) ? verdict.result.details : {};
 	const skillActivation =
 		spec.name === ToolNames.Context ? skillActivationFromToolDetails(toolDetails, input.invokeOptions?.turnId) : null;
 	const result: AgentToolResult<WorkerToolOkDetails> = {
-		content: [{ type: "text", text: verdict.result.output }],
+		content: [{ type: "text", text: toolResultContextText(verdict.result) }],
 		details: { ...toolDetails, kind: "ok" },
 	};
 	if (verdict.result.terminate === true) {
@@ -291,6 +322,7 @@ function toAgentTool(
 				spec,
 				args: params as Record<string, unknown>,
 				registry,
+				returnDispositionedErrors: true,
 			};
 			if (signal) callInput.signal = signal;
 			if (telemetry) callInput.telemetry = telemetry;

package/src/tools/ask-user.ts

@@ -5,6 +5,7 @@ import { Type } from "typebox";
 import { ToolNames } from "../core/tool-names.js";
 import { clioStateDir } from "../core/xdg.js";
 import { type AutonomyExposure, DEFAULT_AUTONOMY_EXPOSURE } from "../domains/safety/autonomy.js";
+import { classifyDecisionPresentation, decisionFactsForAnswer } from "../domains/safety/decision-presentation.js";
 import { StringEnum } from "../engine/ai.js";
 import type {
 	AskUserToolPolicy,
@@ -57,6 +58,7 @@ export interface AskUserResult {
 
 export interface AskUserCall {
 	action: AskUserAction;
+	exposure?: AutonomyExposure;
 	mode?: AskUserMode;
 	questions?: AskUserQuestion[];
 	decisions?: AskUserDecision[];
@@ -284,6 +286,7 @@ export function normalizeAskUserCall(args: Record<string, unknown>): { call?: As
 	return {
 		call: {
 			action: "ask",
+			exposure: exposure.exposure ?? DEFAULT_AUTONOMY_EXPOSURE,
 			...(mode.mode ? { mode: mode.mode } : {}),
 			questions: questions.questions,
 			...(summary ? { summary } : {}),
@@ -319,6 +322,7 @@ function createStandalonePolicy(options?: ToolInvokeOptions): AskUserToolPolicy
 		updatedAt: now,
 		...(options?.sessionId ? { sessionId: options.sessionId } : {}),
 		...(options?.turnId ? { turnId: options.turnId } : {}),
+		exposure: DEFAULT_AUTONOMY_EXPOSURE,
 		rounds: [],
 		decisions: [],
 		inFlight: false,
@@ -335,6 +339,13 @@ function hydratePolicy(policy: AskUserToolPolicy, options?: ToolInvokeOptions):
 	if (options?.turnId && !policy.turnId) policy.turnId = options.turnId;
 }
 
+/** Once an interview is outward, a later local round cannot lower its replay tier. */
+function mergePolicyExposure(policy: AskUserToolPolicy, exposure: AutonomyExposure): AutonomyExposure {
+	const merged = policy.exposure === "outward" || exposure === "outward" ? "outward" : "local";
+	policy.exposure = merged;
+	return merged;
+}
+
 function transcriptQuestions(questions: ReadonlyArray<AskUserQuestion>): AskUserTranscriptQuestion[] {
 	return questions.map((question) => ({
 		question: question.question,
@@ -455,6 +466,7 @@ async function persistAskUserTranscript(
 			...(policy.endedAt ? { endedAt: policy.endedAt } : {}),
 			...(policy.sessionId ? { sessionId: policy.sessionId } : {}),
 			...(policy.turnId ? { turnId: policy.turnId } : {}),
+			exposure: policy.exposure ?? DEFAULT_AUTONOMY_EXPOSURE,
 			...(policy.summary ? { summary: policy.summary } : {}),
 			decisions: policy.decisions,
 			rounds: policy.rounds,
@@ -478,6 +490,7 @@ function compactInterview(
 		rounds: policy.rounds.length,
 		max_rounds: policy.maxCalls,
 		transcript_path: policy.transcriptPath ?? null,
+		exposure: policy.exposure ?? DEFAULT_AUTONOMY_EXPOSURE,
 		latest_answers: latestAnswers,
 		decisions: policy.decisions,
 		...(policy.summary ? { summary: policy.summary } : {}),
@@ -591,6 +604,7 @@ export function createAskUserTool(deps: AskUserToolDeps = {}): ToolSpec {
 				return completeInterview(policy, "complete", options, call.summary, call.decisions ?? []);
 			}
 			const questions = call.questions ?? [];
+			const exposure = mergePolicyExposure(policy, call.exposure ?? DEFAULT_AUTONOMY_EXPOSURE);
 			if (call.max_rounds !== undefined) {
 				const nextLimit = policy.callCount === 0 ? call.max_rounds : Math.max(policy.maxCalls, call.max_rounds);
 				policy.maxCalls = Math.max(nextLimit, policy.callCount + 1);
@@ -632,7 +646,13 @@ export function createAskUserTool(deps: AskUserToolDeps = {}): ToolSpec {
 			if (call.summary) policy.summary = call.summary;
 			const handler = deps.askUser ?? defaultAskUserHandler;
 			try {
-				const result = normalizeAskUserResult(questions, await handler(questions, options));
+				const result = normalizeAskUserResult(
+					questions,
+					await handler(questions, {
+						...options,
+						decisionPresentation: classifyDecisionPresentation(decisionFactsForAnswer(exposure)),
+					}),
+				);
 				const answeredAt = new Date().toISOString();
 				policy.callCount += 1;
 				policy.askedQuestionKeys.add(key);