@matthewfl/pi-contemplator 0.1.16 → 0.1.18

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@matthewfl/pi-contemplator",
-  "version": "0.1.16",
+  "version": "0.1.18",
   "description": "A Pi extension that keeps long-running agentic sessions on track with background memory, contemplation, and structural review.",
   "type": "module",
   "license": "MIT",

package/src/agents/observer/agent.ts

@@ -74,6 +74,8 @@ const RecordObservationsSchema = Type.Object({
 
 type RecordObservationsArgs = Static<typeof RecordObservationsSchema>;
 
+export const OBSERVER_MAX_LENGTH_ATTEMPTS = 4;
+
 /** A terminal provider/agent-loop failure that must not advance observation coverage. */
 export class ObserverStreamError extends Error {
 	readonly stopReason: string;
@@ -218,7 +220,7 @@ IMPORTANT: Now call record_observations to record the useful new observations fr
 	const loop = args.agentLoop ?? agentLoop;
 	const history: AgentMessage[] = [];
 	let terminalFailure: { stopReason: string; errorMessage?: string } | undefined;
-	let lengthRetryAttempted = false;
+	let lengthAttempts = 0;
 	const runInvocation = async (prompt: Message, afterLength = false): Promise<void> => {
 		const context: AgentContext = {
 			systemPrompt: OBSERVER_SYSTEM,
@@ -265,8 +267,8 @@ IMPORTANT: Now call record_observations to record the useful new observations fr
 	};
 
 	await runInvocation(initialPrompt);
-	if (accumulated.size === 0 && terminalFailure?.stopReason === "length") {
-		lengthRetryAttempted = true;
+	if (terminalFailure?.stopReason === "length") lengthAttempts = 1;
+	while (accumulated.size === 0 && terminalFailure?.stopReason === "length" && lengthAttempts < OBSERVER_MAX_LENGTH_ATTEMPTS) {
 		// A provider can impose a lower output ceiling than the advertised model
 		// maximum. agentLoop stops on `length` when no tool call was completed; it
 		// does not automatically send a continuation request. Preserve the partial
@@ -275,7 +277,8 @@ IMPORTANT: Now call record_observations to record the useful new observations fr
 		// assistant text at the LLM boundary because some provider templates strip
 		// historical reasoning; encrypted thinking retains its opaque structure.
 		// Then append a short tool-focused instruction and reduce reasoning to
-		// minimal. A second length stop fails forward at the bounded-chunk level.
+		// minimal. The bounded attempt limit prevents pathological chunks from
+		// consuming background tokens forever.
 		terminalFailure = undefined;
 		const retryPrompt: Message = {
 			role: "user",
@@ -283,6 +286,10 @@ IMPORTANT: Now call record_observations to record the useful new observations fr
 			timestamp: Date.now(),
 		};
 		await runInvocation(retryPrompt, true);
+		// runInvocation mutates this through its async stream callbacks; TypeScript
+		// cannot infer that mutation after the explicit reset above.
+		const retryFailure = terminalFailure as { stopReason: string; errorMessage?: string } | undefined;
+		if (retryFailure?.stopReason === "length") lengthAttempts++;
 	}
 	if (accumulated.size === 0 && !doneCalled && !terminalFailure && rejectedTotal === 0) {
 		const reminder: Message = {
@@ -298,8 +305,8 @@ IMPORTANT: Now call record_observations to record the useful new observations fr
 	// zero-observation stop is also a valid empty result after the reminder;
 	// actual stream failures, truncation, and malformed records still throw.
 	if (accumulated.size === 0 && terminalFailure) {
-		const detail = terminalFailure.stopReason === "length" && lengthRetryAttempted
-			? `provider reached the output limit twice without recording an observation (effective max output request: ${baseConfig.maxTokens} tokens)`
+		const detail = terminalFailure.stopReason === "length" && lengthAttempts > 0
+			? `provider reached the output limit ${lengthAttempts} times without recording an observation (effective max output request: ${baseConfig.maxTokens} tokens)`
 			: terminalFailure.errorMessage;
 		throw new ObserverStreamError(terminalFailure.stopReason, detail);
 	}

package/src/tools/compact-context.ts

@@ -20,12 +20,11 @@ export function createCompactContextTool(runtime: Runtime) {
 		promptGuidelines: [
 			"Use compact_context sparingly when either substantial additional work remains and there is not enough context left to complete it, or accumulated past context has become noisy, stale, or distracting enough that you are struggling to focus on the current task or reason about it reliably.",
 			"Do not use compact_context routinely, for short tasks, or merely because the conversation is long; use it for genuine context-capacity pressure or context degradation that is interfering with the work.",
-			"Call compact_context by itself, provide short_continuation_prompt with concrete instructions for the next agent step, and stop the current turn; pi-contemplator will compact the context and automatically resume with those instructions.",
+			"Call compact_context by itself and provide short_continuation_prompt with concrete instructions for the next agent step. The tool ends the current turn automatically; pi-contemplator will compact the context and resume with those instructions, so do not add a separate response after the tool call.",
 		],
 		parameters: Type.Object({
 			short_continuation_prompt: Type.String({
 				minLength: 1,
-				maxLength: 1_000,
 				pattern: "\\S",
 				description: "Short, concrete instructions to your post-compaction self describing the next action and any critical immediate constraint. Do not summarize the whole conversation.",
 			}),
@@ -33,14 +32,14 @@ export function createCompactContextTool(runtime: Runtime) {
 		async execute(_toolCallId, params) {
 			if (runtime.compactInFlight) {
 				return {
-					content: [{ type: "text" as const, text: "Context compaction is already in progress. Stop this turn and wait for automatic resume." }],
+					content: [{ type: "text" as const, text: "Context compaction is already in progress. This turn is ending automatically; wait for automatic resume." }],
 					details: { status: "in_progress" } as CompactContextDetails,
 					terminate: true,
 				};
 			}
 			if (runtime.compactRequested) {
 				return {
-					content: [{ type: "text" as const, text: "Context compaction is already scheduled. Stop this turn and wait for automatic resume." }],
+					content: [{ type: "text" as const, text: "Context compaction is already scheduled. This turn is ending automatically; wait for automatic resume." }],
 					details: { status: "already_pending" } as CompactContextDetails,
 					terminate: true,
 				};
@@ -49,7 +48,7 @@ export function createCompactContextTool(runtime: Runtime) {
 			runtime.compactRequested = true;
 			runtime.compactContinuationPrompt = params.short_continuation_prompt.trim();
 			return {
-				content: [{ type: "text" as const, text: "Context compaction scheduled. Stop this turn; the task will resume automatically after compaction." }],
+				content: [{ type: "text" as const, text: "Context compaction scheduled. This turn is ending automatically, and the task will resume after compaction." }],
 				details: { status: "scheduled" } as CompactContextDetails,
 				terminate: true,
 			};