provider-events: raw provider stream recorder (LH_PROVIDER_EVENTS=1)

pi-ai 1.0's onProviderStreamEvent, at the layer our incident class lives
(GLM arg truncation, DeepSeek id resets) — env-gated (owner choice:
no app surface, off by default; raw events carry full prompts), one
JSONL per chat under data/debug/provider-events/ with a 20MB cap and a
run-start marker for record correlation.

Wired at every agent construction: teacher/prober turns (AgentBuildOptions.
chatId), jobs via LoopContext (researcher, demo-builder through
runSubagent; visual-maker + pdf-reader through their streamOnce), and
the extraction pipeline (extractor + auditor keyed by source id).
practice deliberately unwired (sunk-cost module).

config.ts

@@ -30,6 +30,9 @@ export const PATHS = {
 	 * Bounded self-contained snapshots (propositions + chains frozen at
 	 * creation), not append-only ledgers. */
 	practice: path.join(DATA_ROOT, "practice"),
+	/** Debug artifacts (env-gated): data/debug/provider-events/<chatId>.jsonl
+	 * raw provider stream recordings (src/agent/provider-events.ts). */
+	debug: path.join(DATA_ROOT, "debug"),
 } as const;
 
 // ─── Models ────────────────────────────────────────────────────────────────

src/agent/provider-events.ts

@@ -0,0 +1,119 @@
+/**
+ * provider-events — the raw-wire recorder (owner, 2026-10-04).
+ *
+ * pi-ai 1.0's onProviderStreamEvent observer fires for each PARSED
+ * provider stream event BEFORE pi normalizes it — the exact layer where
+ * our provider-quirk incident class lives and where nothing else in the
+ * app can see: GLM-flash truncating tool-call arguments mid-JSON
+ * (2026-09-25), DeepSeek resetting tool-call ids every response
+ * (2026-10-04, the Lacan session). This module turns that observer into
+ * an inspectable artifact.
+ *
+ * Gating (owner choice, 2026-10-04): env var only —
+ *   LH_PROVIDER_EVENTS=1 npm run dev
+ * Default off: raw events carry full prompts (privacy) and unbounded
+ * volume (disk). No app surface; a debugging feature, not a feature.
+ *
+ * Artifact: one JSONL file per chat at
+ *   data/debug/provider-events/<chatId>.jsonl
+ * (data/ is disposable scratch — no retention system). Each line:
+ *   {"t":<epoch ms>,"ctx":"teacher|prober|job:<role>","provider":…,"model":…,"event":<raw>}
+ * plus {"note":…} run markers. A per-file byte cap (CAPPED at ~20 MB)
+ * writes one truncation marker and stops — a runaway turn must not fill
+ * the disk. Event payloads are adapter-owned; a non-serializable event
+ * degrades to a type note, never throws (an observer exception must not
+ * kill a stream).
+ */
+
+import * as fs from "node:fs";
+import path from "node:path";
+import { PATHS } from "../../config.ts";
+
+/** Max bytes per chat's JSONL before recording stops (with a marker). */
+const CAP_BYTES = 20 * 1024 * 1024;
+
+/** Test seam (the toolTimeoutMs idiom): shrink the cap to exercise it. */
+export const providerEventsCap = { current: CAP_BYTES };
+
+/** Test seam (the toolTimeoutMs idiom): redirect the output directory. */
+export const providerEventsDir = { current: path.join(PATHS.debug, "provider-events") };
+
+/** The model shape the observer receives (structural — pi passes Model). */
+export interface ProviderEventModel {
+	provider?: string;
+	id?: string;
+}
+
+export type ProviderEventObserver = (data: unknown, model: ProviderEventModel) => void;
+
+/** Per-file accounting so the cap marker writes exactly once. */
+const files = new Map<string, { bytes: number; capped: boolean }>();
+
+function fileFor(chatId: string): string {
+	// chat ids are app-minted (chat_<base36>) — sanitize against any future
+	// id scheme putting path-hostile characters in here.
+	const safe = chatId.replace(/[^a-zA-Z0-9_-]/g, "_");
+	return path.join(providerEventsDir.current, `${safe}.jsonl`);
+}
+
+function writeLine(file: string, line: Record<string, unknown>): void {
+	const state = files.get(file);
+	if (state?.capped) return;
+	const text = `${JSON.stringify({ t: Date.now(), ...line })}\n`;
+	try {
+		const before = state?.bytes ?? (fs.existsSync(file) ? fs.statSync(file).size : 0);
+		fs.mkdirSync(path.dirname(file), { recursive: true });
+		fs.appendFileSync(file, text, "utf-8");
+		const bytes = before + Buffer.byteLength(text);
+		if (bytes > providerEventsCap.current) {
+			files.set(file, { bytes, capped: true });
+			fs.appendFileSync(file, `${JSON.stringify({ t: Date.now(), note: `cap ${providerEventsCap.current} bytes reached — recording stopped` })}\n`, "utf-8");
+		} else {
+			files.set(file, { bytes, capped: false });
+		}
+	} catch {
+		// The debug recorder must never take a run down with it.
+	}
+}
+
+/** Encode an adapter-owned event safely: JSON when it cooperates, a type
+ *  note when it does not (cyclic/BigInt payloads). */
+function encodeEvent(data: unknown): unknown {
+	try {
+		JSON.stringify(data);
+		return data;
+	} catch {
+		return { __nonSerializable: typeof data };
+	}
+}
+
+/** True when recording is on (read per call — toggling needs no restart
+ *  beyond the next turn's agent construction). */
+export function providerEventsEnabled(): boolean {
+	return !!process.env.LH_PROVIDER_EVENTS;
+}
+
+/**
+ * The observer for one agent run, or undefined when disabled. Pass the
+ * result as the Agent's onProviderStreamEvent (both construction sites —
+ * teacher-agent's baseAgentOptions and pi-run's runSubagent — build it
+ * from a chatId in hand).
+ */
+export function providerEventObserver(chatId: string, context: string): ProviderEventObserver | undefined {
+	if (!providerEventsEnabled()) return undefined;
+	return (data, model) => {
+		writeLine(fileFor(chatId), {
+			ctx: context,
+			provider: model?.provider,
+			model: model?.id,
+			event: encodeEvent(data),
+		});
+	};
+}
+
+/** A run marker (start/end/abort notes) so events correlate with the
+ *  persisted record; no-op when disabled. */
+export function providerEventNote(chatId: string, context: string, text: string): void {
+	if (!providerEventsEnabled()) return;
+	writeLine(fileFor(chatId), { ctx: context, note: text });
+}

src/agent/subagents/demo-builder.ts

@@ -40,6 +40,8 @@ export interface DemoBuilderOptions {
 	abortSignal?: AbortSignal;
 	/** Money channel: one call per agent step, folded into JobRecord.usage. */
 	onUsage?: (usage: Usage) => void;
+	/** Debug: raw provider-event recording key (LH_PROVIDER_EVENTS=1). */
+	chatId?: string;
 }
 
 const INSTRUCTIONS = `You are a demo builder — a specialist sub-agent inside a learning application. You
@@ -209,6 +211,7 @@ export async function runDemoBuilder(input: DemoBuilderInput, opts: DemoBuilderO
 			abortSignal: opts.abortSignal,
 			onUsage: (usage) => opts.onUsage?.(usage),
 			onToolStart: (toolName) => opts.onActivity?.(toolName),
+			chatId: opts.chatId,
 		});
 		finalText = result.text;
 	} catch (err) {

src/agent/subagents/extraction.ts

@@ -372,6 +372,7 @@ export function sourceDigest(graph: DerivedGraph, sourceId: string): string {
 /** The extractor over ONE session transcript (the chat front door). */
 function runExtractor(
 	engine: PiEngine | undefined,
+	sourceId: string,
 	prompt: string,
 	existingIds: ReadonlySet<string>,
 	signal: AbortSignal | undefined,
@@ -382,6 +383,9 @@ function runExtractor(
 	// over the MIXED namespace: local ids + existing graph ids.
 	return runSubmitAgent<FindingsSubmit>({
 		role: "extractor",
+		// Debug recorder key: the source id (the run's ledger key — the
+		// chat's chatId for chat runs, the src id for notes).
+		chatId: sourceId,
 		toolName: "submit_findings",
 		toolDescription: "Submit the session findings.",
 		parameters: findingsSchema,
@@ -401,6 +405,7 @@ function runExtractor(
  *  front door): same single-submit shape, notes-mode prompt + schema. */
 function runNotesExtractor(
 	engine: PiEngine | undefined,
+	sourceId: string,
 	prompt: string,
 	existingIds: ReadonlySet<string>,
 	signal: AbortSignal | undefined,
@@ -408,6 +413,8 @@ function runNotesExtractor(
 ): Promise<NotesFindings> {
 	return runSubmitAgent<FindingsSubmit>({
 		role: "extractor",
+		// Debug recorder key: the notes' source id (the ledger key).
+		chatId: sourceId,
 		toolName: "submit_findings",
 		toolDescription: "Submit the notes findings.",
 		parameters: findingsSchema,
@@ -467,6 +474,7 @@ Include the "links" field even when empty — "links": [] when nothing links. NE
  *  tools plus the one submit call. */
 function runAuditor(
 	engine: PiEngine | undefined,
+	sourceId: string,
 	prompt: string,
 	signal: AbortSignal | undefined,
 	onStep: (usage: Usage, model: string, step: number, finishReason: string, tools: string) => void,
@@ -474,6 +482,8 @@ function runAuditor(
 ): Promise<AuditVerdicts> {
 	return runSubmitAgent({
 		role: "auditor",
+		// Debug recorder key: same source id as the extractor run.
+		chatId: sourceId,
 		toolName: "submit_review",
 		toolDescription: "Submit the identity verdicts ONLY (links) — never the propositions themselves.",
 		parameters: auditSchema,
@@ -804,6 +814,7 @@ export function createExtractionPipeline(deps: ExtractionPipelineDeps): Extracti
 
 		const findings = await runExtractor(
 			stageEngine("extractor", deps.extractor),
+			sourceId,
 			prompt,
 			new Set(graph.nodes.map((n) => n.id)),
 			signal,
@@ -816,7 +827,7 @@ export function createExtractionPipeline(deps: ExtractionPipelineDeps): Extracti
 		return findings;
 	}
 
-	async function audit(jobKey: string, findings: SessionFindings): Promise<AuditVerdicts> {
+	async function audit(jobKey: string, sourceId: string, findings: SessionFindings): Promise<AuditVerdicts> {
 		const signal = runningExtractions.get(jobKey)?.controller.signal;
 		signal?.throwIfAborted(); // an abort during the extract→audit gap ends here
 		const graph = getGraph(graphStore);
@@ -831,6 +842,7 @@ export function createExtractionPipeline(deps: ExtractionPipelineDeps): Extracti
 		// exploration tools instead of an injected dump.
 		const auditResult = await runAuditor(
 			stageEngine("auditor", deps.auditor),
+			sourceId,
 			`Proposed findings from the extractor (the source itself is not available to you — the extractor is fully trusted; the propositions only — the reasoning chains are the merger's business, owner 2026-09-30):\n\n` +
 				`${JSON.stringify({ propositions: findings.propositions }, null, 2)}\n\n` +
 				`The current knowledge graph has ${graph.nodes.length} node(s). Explore it with your tools and link what links.`,
@@ -1017,7 +1029,7 @@ export function createExtractionPipeline(deps: ExtractionPipelineDeps): Extracti
 							// stands AFTER every earlier job's merge.
 							return withAuditLock(jobKey, "chat", undefined, async () => {
 								setPhase(jobKey, "auditing", "chat", undefined, part);
-								const auditResult = await audit(jobKey, findings);
+								const auditResult = await audit(jobKey, chatId, findings);
 								setPhase(jobKey, "merging", "chat", undefined, part);
 								// SESSION-DELETION GUARD (wave 6): a chat deleted mid-run must never
 								// merge. The deletion route aborts this run first and runs its purge
@@ -1131,6 +1143,7 @@ export function createExtractionPipeline(deps: ExtractionPipelineDeps): Extracti
 		const t0 = performance.now();
 		const findings = await runNotesExtractor(
 			stageEngine("extractor", deps.extractor),
+			input.id,
 			prompt,
 			new Set(graph.nodes.map((n) => n.id)),
 			signal,
@@ -1191,7 +1204,7 @@ export function createExtractionPipeline(deps: ExtractionPipelineDeps): Extracti
 							// per chunk so cross-source runs interleave fairly.
 							return withAuditLock(jobKey, "notes", input.filePath, async () => {
 								setPhase(jobKey, "auditing", "notes", input.filePath, part);
-								const auditResult = await audit(jobKey, findings);
+								const auditResult = await audit(jobKey, input.id, findings);
 								setPhase(jobKey, "merging", "notes", input.filePath, part);
 								const merged = mergeFindings(
 									id,

src/agent/subagents/pdf-reader.ts

@@ -2,6 +2,7 @@ import type { AssistantMessage, Message, ToolResultMessage, Usage, Tool } from "
 import { normalizeContext } from "@earendil-works/pi-ai";
 import { Type } from "typebox";
 import { piEngineFor, type PiEngine, type ThinkingLevel } from "../models.ts";
+import { providerEventObserver, type ProviderEventObserver } from "../provider-events.ts";
 import type { PdfResult, PdfToolInput } from "../../core/contracts.ts";
 import { getHeadless, type HeadlessApi } from "./headless.ts";
 import { pdfExtract, pdfInfo, pdfProbe, pdfSearch, type PdfSearchHit } from "./pdf-ops.ts";
@@ -66,6 +67,8 @@ export interface PdfReaderOptions {
 	onActivity?: (line: string) => void;
 	headless?: HeadlessApi;
 	maxIterations?: number;
+	/** Debug: raw provider-event recording key (LH_PROVIDER_EVENTS=1). */
+	chatId?: string;
 }
 
 export interface PdfReaderInput extends PdfToolInput {}
@@ -199,7 +202,15 @@ export async function runPdfReaderAgent(
 
 	try {
 		for (let iteration = 0; iteration < maxIterations; iteration++) {
-			const result = await streamOnce(engine, INSTRUCTIONS, messages, toolDeclarations, reasoning, opts.abortSignal);
+			const result = await streamOnce(
+				engine,
+				INSTRUCTIONS,
+				messages,
+				toolDeclarations,
+				reasoning,
+				opts.abortSignal,
+				opts.chatId ? providerEventObserver(opts.chatId, "job:pdfReader") : undefined,
+			);
 			// Money channel: each iteration is one model call.
 			opts.onUsage?.(result.usage);
 			messages = [...messages, result];
@@ -392,6 +403,7 @@ async function streamOnce(
 	tools: Tool[],
 	reasoning: ThinkingLevel | undefined,
 	abortSignal?: AbortSignal,
+	onProviderStreamEvent?: ProviderEventObserver,
 ): Promise<AssistantMessage> {
 	// 0.87.1: provider-facing contexts are TranscriptContexts — the prompt
 	// and tool declarations fold into a leading system message.
@@ -400,6 +412,7 @@ async function streamOnce(
 		reasoning: reasoning && reasoning !== "off" ? reasoning : undefined,
 		apiKey: process.env.OPENROUTER_API_KEY || undefined,
 		signal: abortSignal,
+		onProviderStreamEvent: onProviderStreamEvent as never,
 	} as never);
 	for await (const event of stream) {
 		if (event.type === "done") return event.message;

src/agent/subagents/pi-run.ts

@@ -1,6 +1,7 @@
 import { Agent, type AgentEvent, type AgentMessage, type AgentTool } from "@earendil-works/pi-agent-core";
 import type { AssistantMessage, ToolResultMessage, Usage } from "@earendil-works/pi-ai";
 import { piEngineFor, type PiEngine } from "../models.ts";
+import { providerEventObserver } from "../provider-events.ts";
 import type { ModelRole } from "../../../config.ts";
 
 /**
@@ -37,6 +38,10 @@ export interface SubagentRunOptions {
 	onAssistantMessage?: (message: AssistantMessage, step: number) => void;
 	/** Extra stop rule (extractor: stop after a clean submit). */
 	shouldStop?: (message: AssistantMessage, toolResults: ToolResultMessage[]) => boolean;
+	/** Debug: when set AND LH_PROVIDER_EVENTS=1, raw provider stream
+	 *  events append to data/debug/provider-events/<chatId>.jsonl (context
+	 *  "job:<role>"). Absent or env-off → no observer. */
+	chatId?: string;
 }
 
 export interface SubagentRunResult {
@@ -56,6 +61,8 @@ export function subagentEngine(role: ModelRole, override?: PiEngine): PiEngine {
 export async function runSubagent(tools: AgentTool<any, any>[], opts: SubagentRunOptions): Promise<SubagentRunResult> {
 	const engine = subagentEngine(opts.role, opts.engine);
 	let turns = 0;
+	// Debug recorder (owner, 2026-10-04): raw provider events per job run.
+	const providerEvents = opts.chatId ? providerEventObserver(opts.chatId, `job:${opts.role}`) : undefined;
 	const agent = new Agent({
 		initialState: {
 			systemPrompt: opts.systemPrompt,
@@ -65,6 +72,7 @@ export async function runSubagent(tools: AgentTool<any, any>[], opts: SubagentRu
 			messages: opts.messages ?? [],
 		},
 		streamFn: engine.streamFn,
+		...(providerEvents ? { onProviderStreamEvent: providerEvents } : {}),
 		// 0.87.1: shouldStopAfterTurn → finishTurn (same call point; "end"
 		// exits before the steering poll, exactly the old boolean-true).
 		finishTurn: (turn) => {

src/agent/subagents/researcher.ts

@@ -71,6 +71,8 @@ export interface ResearcherRunOptions {
 	abortSignal?: AbortSignal;
 	onUsage?: (usage: Usage) => void;
 	onToolStart?: (toolName: string, args: unknown) => void;
+	/** Debug: raw provider-event recording key (LH_PROVIDER_EVENTS=1). */
+	chatId?: string;
 }
 
 /**
@@ -88,6 +90,7 @@ export async function runResearcherAgent(prompt: string, opts: ResearcherRunOpti
 		abortSignal: opts.abortSignal,
 		onUsage: opts.onUsage,
 		onToolStart: opts.onToolStart,
+		chatId: opts.chatId,
 	});
 	return text;
 }

src/agent/subagents/runner.ts

@@ -106,6 +106,9 @@ interface LoopContext {
 	/** Fires when abortJob kills the run; abortSignal is passed into every
 	 * model call so the spend stops mid-call, not at the next loop check. */
 	abortSignal: AbortSignal;
+	/** The job's chat (debug: raw provider-event recording under
+	 * LH_PROVIDER_EVENTS=1 — forwarded into every subagent run). */
+	chatId?: string;
 	/** Fold one model call's usage into the job record (progress write). */
 	onUsage: (step: UsageStepLike) => void;
 }
@@ -127,6 +130,7 @@ export function createJobRunner(deps: JobRunnerDeps): JobRunner {
 	const usageTotals = new Map<string, JobUsage>();
 
 	function makeLoopContext(jobId: string, engine: PiEngine): LoopContext {
+		const chatId = jobStore.getJob(jobId)?.chatId;
 		// Provider-qualified stamp: the engine's actual model — the only
 		// thing that ran (BYOK: no config-pin fallback stamps exist).
 		const model = `${engine.model.provider}:${engine.model.id}`;
@@ -137,6 +141,7 @@ export function createJobRunner(deps: JobRunnerDeps): JobRunner {
 		registerJobAbort(jobId, controller);
 		return {
 			abortSignal: controller.signal,
+			chatId,
 			onUsage: (step) => {
 				const next = accumulateJobUsage(usageTotals.get(jobId), model, step);
 				usageTotals.set(jobId, next);
@@ -355,6 +360,7 @@ export function createJobRunner(deps: JobRunnerDeps): JobRunner {
 				engine: model,
 				search: deps.searchProvider,
 				abortSignal: loop.abortSignal,
+				chatId: loop.chatId,
 				// Money channel: each agent step is one model call — fold its
 				// usage into the record as it completes (progress write).
 				onUsage: (usage) => loop.onUsage(piUsageToStep(usage)),
@@ -403,6 +409,7 @@ export function createJobRunner(deps: JobRunnerDeps): JobRunner {
 				{
 					engine: model,
 					abortSignal: loop.abortSignal,
+					chatId: loop.chatId,
 					onUsage: (usage) => loop.onUsage(piUsageToStep(usage)),
 					onActivity: (line) => jobStore.updateJob(jobId, { activity: line }),
 				},
@@ -438,6 +445,7 @@ export function createJobRunner(deps: JobRunnerDeps): JobRunner {
 				{
 					engine: model,
 					abortSignal: loop.abortSignal,
+					chatId: loop.chatId,
 					onUsage: (usage) => loop.onUsage(piUsageToStep(usage)),
 					onActivity: (line) => jobStore.updateJob(jobId, { activity: line }),
 				},
@@ -476,6 +484,7 @@ export function createJobRunner(deps: JobRunnerDeps): JobRunner {
 			{
 				engine: model,
 				abortSignal: loop.abortSignal,
+				chatId: loop.chatId,
 				onUsage: (usage) => loop.onUsage(piUsageToStep(usage)),
 				onActivity: (line) => jobStore.updateJob(jobId, { activity: line }),
 			},

src/agent/subagents/visual-maker.ts

@@ -2,6 +2,7 @@ import type { AssistantMessage, Message, Tool, ToolResultMessage, Usage } from "
 import { normalizeContext } from "@earendil-works/pi-ai";
 import { Type } from "typebox";
 import { piEngineFor, type PiEngine, type ThinkingLevel } from "../models.ts";
+import { providerEventObserver, type ProviderEventObserver } from "../provider-events.ts";
 import type { VisualResult } from "../../core/contracts.ts";
 import { getHeadless, type HeadlessApi } from "./headless.ts";
 
@@ -44,6 +45,8 @@ export interface VisualMakerOptions {
 	/** Money channel: one call per loop iteration (streamSimple), folded
 	 * into JobRecord.usage. */
 	onUsage?: (usage: Usage) => void;
+	/** Debug: raw provider-event recording key (LH_PROVIDER_EVENTS=1). */
+	chatId?: string;
 }
 
 const INSTRUCTIONS = `You are a visual maker — a diagram author + renderer for precise, spatial figures. You
@@ -155,7 +158,15 @@ export async function runVisualMaker(input: VisualMakerInput, opts: VisualMakerO
 		for (let iteration = 0; iteration < maxIterations; iteration++) {
 			// One model call per iteration (streamSimple, no agent loop — this
 			// loop owns tool execution and the image-in-the-loop injection).
-			const result = await streamOnce(engine, INSTRUCTIONS, messages, toolDeclarations, reasoning, opts.abortSignal);
+			const result = await streamOnce(
+				engine,
+				INSTRUCTIONS,
+				messages,
+				toolDeclarations,
+				reasoning,
+				opts.abortSignal,
+				opts.chatId ? providerEventObserver(opts.chatId, "job:visualMaker") : undefined,
+			);
 			// Money channel: each iteration is one model call.
 			opts.onUsage?.(result.usage);
 			messages = [...messages, result];
@@ -307,6 +318,7 @@ async function streamOnce(
 	tools: Tool[],
 	reasoning: ThinkingLevel | undefined,
 	abortSignal?: AbortSignal,
+	onProviderStreamEvent?: ProviderEventObserver,
 ): Promise<AssistantMessage> {
 	// 0.87.1: provider-facing contexts are TranscriptContexts — the prompt
 	// and tool declarations fold into a leading system message.
@@ -315,6 +327,7 @@ async function streamOnce(
 		reasoning: reasoning && reasoning !== "off" ? reasoning : undefined,
 		apiKey: process.env.OPENROUTER_API_KEY || undefined,
 		signal: abortSignal,
+		onProviderStreamEvent: onProviderStreamEvent as never,
 	} as never);
 	for await (const event of stream) {
 		if (event.type === "done") return event.message;

src/agent/teacher-agent.ts

@@ -1,6 +1,7 @@
 import { Agent } from "@earendil-works/pi-agent-core";
 import type { AgentEvent, AgentMessage, AgentTool } from "@earendil-works/pi-agent-core";
-import { questionResult, ToolInputError } from "./tool-execute.ts";
+import { invalidInputResult, questionResult } from "./tool-execute.ts";
+import { providerEventObserver } from "./provider-events.ts";
 import type { Usage } from "@earendil-works/pi-ai";
 import { PATHS } from "../../config.ts";
 import { piEngineFor, type PiEngine } from "./models.ts";
@@ -111,6 +112,11 @@ export interface AgentBuildOptions {
 	 *  fresh per turn, honoring enabled overrides (switch on + file
 	 *  exists); absent → the cached pure-source composition. */
 	promptSwitches?: Record<string, boolean>;
+	/** Debug: when set AND LH_PROVIDER_EVENTS=1, raw provider stream events
+	 *  (pre-normalization) append to data/debug/provider-events/
+	 *  <chatId>.jsonl via providerEventObserver — the GLM-truncation /
+	 *  DeepSeek-id incident layer. Absent or env-off → no observer. */
+	chatId?: string;
 }
 
 /**
@@ -158,6 +164,10 @@ function baseAgentOptions(kind: AgentKind, opts: AgentBuildOptions, tools: Agent
 	const systemPrompt = opts.promptSwitches
 		? composeInstructions(kind, opts.promptSwitches)
 		: composedInstructions[kind]!;
+	// Debug recorder (owner, 2026-10-04): raw provider events when
+	// LH_PROVIDER_EVENTS=1 and the caller handed a chatId; undefined is
+	// the normal no-observer path (tests, env-off).
+	const providerEvents = opts.chatId ? providerEventObserver(opts.chatId, kind) : undefined;
 	const agent = new Agent({
 		initialState: {
 			systemPrompt,
@@ -167,6 +177,7 @@ function baseAgentOptions(kind: AgentKind, opts: AgentBuildOptions, tools: Agent
 			messages: opts.messages ?? [],
 		},
 		streamFn: engine.streamFn,
+		...(providerEvents ? { onProviderStreamEvent: providerEvents } : {}),
 		// 0.87.1: shouldStopAfterTurn → finishTurn. Same call point (after
 		// tool results attach, before turn_end) and `{ action: "end" }` still
 		// exits BEFORE the steering poll — the parked-marker stranding our
@@ -232,9 +243,9 @@ export function createTeacherAgent(
 }
 
 /** The bare teach question set (tests, bare agents): quiz with the shared
- *  authoring refinement (quizInputProblems, run in execute — throws
- *  ToolInputError, which pi's loop converts to the error tool result the
- *  model retries on) but NO per-session nodeId check (no record to read),
+ *  authoring refinement (quizInputProblems, run in execute — returns an
+ *  invalidInputResult isError result the model retries on, pi 1.0
+ *  convention) but NO per-session nodeId check (no record to read),
  *  plus ask. */
 export function bareQuestionTools(): Record<string, AgentTool<any>> {
 	// A typed const (not an inline literal): the specific AgentTool typing
@@ -248,7 +259,7 @@ export function bareQuestionTools(): Record<string, AgentTool<any>> {
 		executionMode: "sequential",
 		execute: async (_toolCallId, input) => {
 			const problems = quizInputProblems(input);
-			if (problems.length > 0) throw new ToolInputError("quiz", problems);
+			if (problems.length > 0) return invalidInputResult("quiz", problems);
 			return questionResult();
 		},
 	};

src/server/chat-handler.ts

@@ -18,6 +18,7 @@ import {
 } from "./jobs.ts";
 import { piTurnResponse, runPiTurn } from "./pi/ui-stream.ts";
 import { collectToolCallIds } from "./pi/call-ids.ts";
+import { providerEventNote } from "../agent/provider-events.ts";
 import { effectiveSettings, stores } from "./stores.ts";
 import { searchAvailable, searchProvider } from "./search.ts";
 import { providerKeyPresent } from "./keys.ts";
@@ -321,6 +322,9 @@ export async function handleChatInput(
 		messages: asAgentMessages(merged),
 		onUsage: collectUsage,
 		toolToggles: effective.agent.tools,
+		// Debug (owner, 2026-10-04): teacher/prober raw provider events under
+		// LH_PROVIDER_EVENTS=1 (provider-events.ts); no-op otherwise.
+		chatId,
 		// The prompt-override switches (P2, 2026-09-19): same resolution
 		// point as the tool delta — the agent layer composes the system
 		// prompt fresh per turn, swapping in any enabled override segments.
@@ -378,6 +382,10 @@ export async function handleChatInput(
 		// coalescing). Settlement still cannot beat the flow below: it needs
 		// a streamed response, which never happens inside the constructor.
 		registerLiveAgent(chatId, constructed);
+		// Debug recorder run marker (owner, 2026-10-04): correlates the
+		// JSONL event stream with the record's turns; no-op unless
+		// LH_PROVIDER_EVENTS=1.
+		providerEventNote(chatId, "turn", `run start (base: ${merged.length} messages)`);
 		const handle = runPiTurn({
 			agent: constructed,
 			// THE LACAN FIX (2026-10-04): seed the uniquifier with every

tests/provider-events.test.ts

@@ -0,0 +1,117 @@
+/**
+ * provider-events — the raw-wire recorder pins (owner, 2026-10-04).
+ *
+ * The recorder is env-gated debug tooling (LH_PROVIDER_EVENTS=1); these
+ * pins hold its mechanics: off = zero surface (no observer, no file),
+ * on = one JSONL per chat under data/debug/provider-events/, safe
+ * degradation on non-serializable adapter payloads, and the byte cap
+ * that stops a runaway turn from filling the disk.
+ */
+
+import test from "node:test";
+import assert from "node:assert/strict";
+import * as fs from "node:fs";
+import * as os from "node:os";
+import * as path from "node:path";
+import { PATHS } from "../config.ts";
+import {
+	providerEventNote,
+	providerEventObserver,
+	providerEventsCap,
+	providerEventsDir,
+	providerEventsEnabled,
+} from "../src/agent/provider-events.ts";
+
+const ENV = "LH_PROVIDER_EVENTS";
+const REAL_DIR = path.join(PATHS.debug, "provider-events");
+const REAL_CAP = 20 * 1024 * 1024;
+
+function tmpDir(): string {
+	return fs.mkdtempSync(path.join(os.tmpdir(), "lh-provider-events-"));
+}
+
+function readLines(dir: string, chatId: string): Array<Record<string, unknown>> {
+	const file = path.join(dir, `${chatId}.jsonl`);
+	return fs
+		.readFileSync(file, "utf-8")
+		.trim()
+		.split("\n")
+		.map((l) => JSON.parse(l) as Record<string, unknown>);
+}
+
+/** Common harness: redirect the dir, enable the env, restore both. */
+function withRecorder(t: { after: (fn: () => void) => void }): string {
+	const dir = tmpDir();
+	providerEventsDir.current = dir;
+	process.env[ENV] = "1";
+	t.after(() => {
+		delete process.env[ENV];
+		providerEventsDir.current = REAL_DIR;
+	});
+	return dir;
+}
+
+test("env off: no observer, no note, no file — zero surface", (t) => {
+	const dir = tmpDir();
+	providerEventsDir.current = dir;
+	delete process.env[ENV];
+	t.after(() => {
+		delete process.env[ENV];
+		providerEventsDir.current = REAL_DIR;
+	});
+
+	assert.equal(providerEventsEnabled(), false);
+	assert.equal(providerEventObserver("chat_x", "teacher"), undefined);
+	providerEventNote("chat_x", "teacher", "run start"); // must be a silent no-op
+	assert.equal(fs.existsSync(path.join(dir, "chat_x.jsonl")), false, "nothing written while off");
+});
+
+test("env on: observer writes pinned-shape JSONL lines", (t) => {
+	const dir = withRecorder(t);
+
+	const observe = providerEventObserver("chat_y", "teacher");
+	assert.ok(observe, "enabled → observer exists");
+	observe!({ type: "raw", delta: "hello" }, { provider: "openrouter", id: "z-ai/glm-5.3-flash" });
+	providerEventNote("chat_y", "teacher", "run end");
+
+	const lines = readLines(dir, "chat_y");
+	assert.equal(lines.length, 2);
+	assert.deepEqual(
+		{ ...lines[0], t: 0 },
+		{ t: 0, ctx: "teacher", provider: "openrouter", model: "z-ai/glm-5.3-flash", event: { type: "raw", delta: "hello" } },
+		"event line: {t, ctx, provider, model, event}",
+	);
+	assert.match(String(lines[1]!.note), /run end/);
+});
+
+test("non-serializable adapter payloads degrade to a type note, never throw", (t) => {
+	const dir = withRecorder(t);
+
+	const cyclic: Record<string, unknown> = { type: "chunk" };
+	cyclic.self = cyclic;
+	const observe = providerEventObserver("chat_c", "prober");
+	observe!(cyclic, { provider: "anthropic", id: "claude-fable-5" });
+
+	const [line] = readLines(dir, "chat_c");
+	assert.deepEqual(line!.event, { __nonSerializable: "object" });
+});
+
+test("the byte cap writes one marker and stops recording", (t) => {
+	const dir = withRecorder(t);
+	// Deterministic sizing: cap 250, each event line ≈ 85B base + 200B pad
+	// ≈ 285B → the FIRST line already crosses; every later line is dropped.
+	providerEventsCap.current = 250;
+	t.after(() => {
+		providerEventsCap.current = REAL_CAP;
+	});
+
+	const observe = providerEventObserver("chat_cap", "teacher");
+	assert.ok(observe);
+	observe!({ type: "a", pad: "x".repeat(200) }, { provider: "p", id: "m" }); // crosses → written + marker
+	observe!({ type: "b" }, { provider: "p", id: "m" }); // capped: dropped
+
+	const lines = readLines(dir, "chat_cap");
+	assert.equal(lines.length, 2, "one event + one cap marker");
+	assert.match(String(lines.at(-1)!.note), /cap 250 bytes reached/);
+	assert.equal(lines.some((l) => (l.event as { type?: string })?.type === "b"), false, "post-cap event dropped");
+});