chore: pi family 0.85.1 -> 0.87.1 (finishTurn, transcript contexts, typebox 1.3.27)

Mechanical fallout only; semantic deltas flagged in the wave report:
- shouldStopAfterTurn -> finishTurn ({action:'end'} wrap; loop order verified
  equivalent for Wave C stranding)
- transcript rework: seeded leading SystemMessage filtered at both
  state.messages read seams (ui-stream settleWith, pi-run slice); prompt
  and tool declarations now ride the transcript
- streamOnce sites normalizeContext() (Context -> TranscriptContext)
- AgentToolResult details / ToolCall.arguments JsonValue narrowing (casts)
- AUTH_PROVIDER_KEYS regenerated (meta added); glm-5.3-flash price pin
  follows the 0.87.1 catalog (0.15/0.5/0.05)
- vanish probe: 0/12 vs 2/6 baseline; live smoke green on PORT=3004

src/agent/subagents/pdf-reader.ts

@@ -1,4 +1,5 @@
 import type { AssistantMessage, Message, ToolResultMessage, Usage, Tool } from "@earendil-works/pi-ai";
+import { normalizeContext } from "@earendil-works/pi-ai";
 import { Type } from "typebox";
 import { piEngineFor, type PiEngine, type ThinkingLevel } from "../models.ts";
 import type { PdfResult, PdfToolInput } from "../../core/contracts.ts";
@@ -392,7 +393,10 @@ async function streamOnce(
 	reasoning: ThinkingLevel | undefined,
 	abortSignal?: AbortSignal,
 ): Promise<AssistantMessage> {
-	const stream = engine.streamFn(engine.model, { systemPrompt, messages, tools }, {
+	// 0.87.1: provider-facing contexts are TranscriptContexts — the prompt
+	// and tool declarations fold into a leading system message.
+	// normalizeContext is the library's only Context→TranscriptContext entry.
+	const stream = engine.streamFn(engine.model, normalizeContext({ systemPrompt, messages, tools }), {
 		reasoning: reasoning && reasoning !== "off" ? reasoning : undefined,
 		apiKey: process.env.OPENROUTER_API_KEY || undefined,
 		signal: abortSignal,

src/agent/subagents/pi-run.ts

@@ -65,9 +65,13 @@ export async function runSubagent(tools: AgentTool<any, any>[], opts: SubagentRu
 			messages: opts.messages ?? [],
 		},
 		streamFn: engine.streamFn,
-		shouldStopAfterTurn: (context) => {
+		// 0.87.1: shouldStopAfterTurn → finishTurn (same call point; "end"
+		// exits before the steering poll, exactly the old boolean-true).
+		finishTurn: (turn) => {
 			turns++;
-			return turns >= opts.maxTurns || opts.shouldStop?.(context.message, context.toolResults) === true;
+			return turns >= opts.maxTurns || opts.shouldStop?.(turn.message, turn.toolResults) === true
+				? { action: "end" }
+				: undefined;
 		},
 	});
 
@@ -119,7 +123,12 @@ export async function runSubagent(tools: AgentTool<any, any>[], opts: SubagentRu
 	}
 	if (errorMessage !== undefined) throw new Error(errorMessage);
 
-	const newMessages = agent.state.messages.slice(before);
+	// 0.87.1: drop the seeded leading SystemMessage (transcript rework —
+	// see ui-stream's settleWith) before slicing past the input count, or
+	// the seed displaces the slice and the last INPUT message leaks in.
+	const newMessages = agent.state.messages
+		.filter((m) => (m as { role?: string }).role !== "system")
+		.slice(before);
 	const assistants = newMessages.filter((m) => (m as { role?: string }).role === "assistant") as AssistantMessage[];
 	const last = assistants[assistants.length - 1];
 	const text = last

src/agent/subagents/visual-maker.ts

@@ -1,4 +1,5 @@
 import type { AssistantMessage, Message, Tool, ToolResultMessage, Usage } from "@earendil-works/pi-ai";
+import { normalizeContext } from "@earendil-works/pi-ai";
 import { Type } from "typebox";
 import { piEngineFor, type PiEngine, type ThinkingLevel } from "../models.ts";
 import type { VisualResult } from "../../core/contracts.ts";
@@ -307,7 +308,10 @@ async function streamOnce(
 	reasoning: ThinkingLevel | undefined,
 	abortSignal?: AbortSignal,
 ): Promise<AssistantMessage> {
-	const stream = engine.streamFn(engine.model, { systemPrompt, messages, tools }, {
+	// 0.87.1: provider-facing contexts are TranscriptContexts — the prompt
+	// and tool declarations fold into a leading system message.
+	// normalizeContext is the library's only Context→TranscriptContext entry.
+	const stream = engine.streamFn(engine.model, normalizeContext({ systemPrompt, messages, tools }), {
 		reasoning: reasoning && reasoning !== "off" ? reasoning : undefined,
 		apiKey: process.env.OPENROUTER_API_KEY || undefined,
 		signal: abortSignal,

src/agent/teacher-agent.ts

@@ -122,7 +122,8 @@ export interface AgentBuildOptions {
  *
  * Question results also carry terminate: true (tool-execute's
  * questionResult), so single-question batches stop natively at the pi
- * loop's own batch check; this shouldStopAfterTurn hook is the
+ * loop's own batch check; this finishTurn hook (0.87.1 rename of
+ * shouldStopAfterTurn) is the
  * belt-and-suspenders HALF for the mixed batch (a question + other calls
  * in one message) that terminate-only semantics would let run on — the
  * learner cannot answer while a turn is in flight (per-chat 409), so a
@@ -164,7 +165,12 @@ function baseAgentOptions(kind: AgentKind, opts: AgentBuildOptions, tools: Agent
 			messages: opts.messages ?? [],
 		},
 		streamFn: engine.streamFn,
-		shouldStopAfterTurn: stopper,
+		// 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
+		// Wave C registerLiveAgent-before-runPiTurn design depends on.
+		// undefined preserves normal scheduling (the old boolean false).
+		finishTurn: (turn) => (stopper(turn) ? { action: "end" } : undefined),
 		// Wave C (2026-09-26, owner correction same day): markers handed to a
 		// live run drain from pi's STEERING queue — the lesson loop is one
 		// continuous teach→quiz→teach run, so delivery must inject at turn

src/core/auth-keys.ts

@@ -95,6 +95,7 @@ export const AUTH_PROVIDER_KEYS: readonly AuthProviderInfo[] = [
 	{ id: "groq", envVar: "GROQ_API_KEY", label: "Groq" },
 	{ id: "huggingface", envVar: "HF_TOKEN", label: "Hugging Face" },
 	{ id: "kimi-coding", envVar: "KIMI_API_KEY", label: "Kimi For Coding" },
+	{ id: "meta", envVar: "META_API_KEY", label: "Meta" },
 	{ id: "minimax", envVar: "MINIMAX_API_KEY", label: "MiniMax" },
 	{ id: "minimax-cn", envVar: "MINIMAX_CN_API_KEY", label: "MiniMax CN" },
 	{ id: "mistral", envVar: "MISTRAL_API_KEY", label: "Mistral" },

src/core/native-messages.ts

@@ -117,7 +117,9 @@ export function newAnswerResult(
 		toolCallId,
 		toolName,
 		content: [{ type: "text", text: JSON.stringify(answer) }],
-		details: answer,
+		// 0.87.1 narrowed ToolResultMessage.details to JsonRepresentation —
+		// our payloads are JSON-safe answer structures, so the cast is honest.
+		details: answer as ToolResultMessage["details"],
 		isError: false,
 		timestamp: Date.now(),
 	};

src/server/pi/ui-stream.ts

@@ -115,8 +115,18 @@ export function runPiTurn(spec: PiTurnSpec): PiTurnHandle {
 	// The settle seam: the record write is the agent's OWN final state
 	// (native history — signatures, usage, provider stamps all real).
 	// Copy the array: the agent may outlive this call's read.
+	// 0.87.1 (transcript rework): the Agent seeds a leading SystemMessage
+	// from initialState.systemPrompt (prompt + tool declarations) — it is
+	// synthesized per turn and must NEVER persist (the record would bloat
+	// with tool declarations AND replay would see messages[0] === system,
+	// suppressing the fresh per-turn re-seed). Our app mints no system
+	// messages of its own (the prompt is composed per turn), so dropping
+	// the role wholesale is exact; revisit if prompt-updates ever land.
 	const settleWith = () => {
-		onFinish([...agent.state.messages], sawAbort);
+		onFinish(
+			[...agent.state.messages].filter((m) => (m as { role?: string }).role !== "system"),
+			sawAbort,
+		);
 	};
 
 	const handle = (event: AgentEvent) => {