@nijaru/pi-provider-compaction 0.0.6 → 0.0.7

package/README.md

@@ -1,6 +1,6 @@
 # pi-provider-compaction
 
-Experimental provider-native compaction for [Pi](https://github.com/earendil-works/pi), with a meaningful portable summary saved alongside each native checkpoint. Minimum runtime: Pi 0.87.0. Development checks target Pi 1.2.0.
+Experimental provider-native compaction for [Pi](https://github.com/earendil-works/pi), with a meaningful portable summary saved alongside each native checkpoint. Minimum runtime: Pi 0.87.0. Development checks target Pi 1.3.0.
 
 **This experimental design is not production-validated. Do not enable it globally yet.** Real runner/adapter fixtures and a Codex live smoke test pass; broader backend and deployment validation remains incomplete.
 
@@ -73,7 +73,7 @@ pi --extension ./dist/index.js
 
 The manifest loads the built `dist/index.js`, including a bundled stateless Responses serializer validator; source or host dependency changes require a rebuild. `bun run check` rebuilds first and tests the artifact against Pi's actual host serializer, including function-call history replayed as grammar tools. Keep global installation disabled while testing. Publishing and global activation are separate approval steps.
 
-Tests include real Pi 1.2.0 runner/adapter fixtures for all three protocols, repeated prefix compaction, transformed suffix preservation, unchanged fallback, grammar tools, strict defaults, tool IDs, images and prompt updates. Set `PI_COMPACTOR_FIXTURE_PATH` to a `pi-compactor` entrypoint to check SDK flag precedence in both extension orders. The optional installed-fast-mode fixture (`PI_FAST_MODE_FIXTURE_PATH`) checks actual `pi-fast-mode` composition across `/reload`; it creates an isolated fast-mode configuration.
+Tests include real Pi 1.3.0 runner/adapter fixtures for all three protocols, repeated prefix compaction, transformed suffix preservation, unchanged fallback, grammar tools, strict defaults, tool IDs, images, prompt updates, cross-model thinking and mid-conversation effort changes. Set `PI_COMPACTOR_FIXTURE_PATH` to a `pi-compactor` entrypoint to check SDK flag precedence in both extension orders. The optional installed-fast-mode fixture (`PI_FAST_MODE_FIXTURE_PATH`) checks actual `pi-fast-mode` composition across `/reload`; it creates an isolated fast-mode configuration.
 
 Authenticated protocol smoke tests are separate and send three small requests:
 

package/dist/index.js

@@ -637,19 +637,13 @@ function transformMessages(messages, model, normalizeToolCallId) {
       const isSameModel = assistantMsg.provider === model.provider && assistantMsg.api === model.api && assistantMsg.model === model.id;
       const transformedContent = assistantMsg.content.flatMap((block) => {
         if (block.type === "thinking") {
-          if (block.redacted) {
-            return isSameModel ? block : [];
-          }
-          if (isSameModel && block.thinkingSignature)
+          if (!isSameModel)
+            return [];
+          if (block.redacted || block.thinkingSignature)
             return block;
           if (!block.thinking || block.thinking.trim() === "")
             return [];
-          if (isSameModel)
-            return block;
-          return {
-            type: "text",
-            text: block.thinking
-          };
+          return block;
         }
         if (block.type === "text") {
           if (isSameModel)
@@ -840,6 +834,12 @@ function convertResponsesMessages(model, context, allowedToolCallProviders, opti
     });
   };
   const includeInitialSystemMessage = options?.includeSystemPrompt ?? true;
+  const pushConfigurationUpdate = (effort) => {
+    messages.push({
+      type: "configuration_update",
+      reasoning: { effort }
+    });
+  };
   const compat = model.compat;
   const instructionRole = model.reasoning && compat?.supportsDeveloperRole !== false ? "developer" : "system";
   let msgIndex = 0;
@@ -946,6 +946,9 @@ function convertResponsesMessages(model, context, allowedToolCallProviders, opti
       }
       if (output.length === 0)
         continue;
+      if (options?.configurationUpdateEffort !== undefined && isSameModel && msg.providerThinkingLevel !== undefined) {
+        pushConfigurationUpdate(msg.providerThinkingLevel);
+      }
       messages.push(...output);
     } else if (msg.role === "toolResult") {
       const [callId] = msg.toolCallId.split("|");
@@ -967,6 +970,8 @@ function convertResponsesMessages(model, context, allowedToolCallProviders, opti
     if (!isLeadingSystemMessage)
       msgIndex++;
   }
+  if (options?.configurationUpdateEffort !== undefined)
+    pushConfigurationUpdate(options.configurationUpdateEffort);
   return messages;
 }
 function convertResponsesTools(tools, options) {
@@ -1463,7 +1468,7 @@ var SNAPSHOT_TTL_MS = 5 * 60000;
 function equal(left, right) {
   return JSON.stringify(left) === JSON.stringify(right);
 }
-function serializeInput(model, context, declarations = context) {
+function serializeInput(model, context, declarations = context, configurationUpdateEffort) {
   const compat = model.compat;
   const flag = (name) => compat?.[name] === true;
   const codex = model.api === "openai-codex-responses";
@@ -1471,6 +1476,7 @@ function serializeInput(model, context, declarations = context) {
   const tools = resolveTranscript2(declarations, flag("supportsMidConvoSystemMessages"));
   return convertResponsesMessages(model, normalized, model.api === "azure-openai-responses" ? AZURE_TOOL_CALL_PROVIDERS : TOOL_CALL_PROVIDERS, {
     includeSystemPrompt: !codex,
+    configurationUpdateEffort,
     grammarToolInputProperties: createGrammarToolInputProperties(getDeclaredTools(tools.messages), flag("supportsOpenAIGrammarTools")),
     supportsMidConvoSystemMessages: flag("supportsMidConvoSystemMessages"),
     supportsAdditionalTools: flag("supportsAdditionalTools"),
@@ -1495,12 +1501,15 @@ function capturePreparedRequest(model, context, canonical, payload, reject = ()
     reject("snapshot exceeds size limit");
     return;
   }
-  const expected = serializeInput(model, context);
+  const last = payload.input.at(-1);
+  const compat = model.compat;
+  const configurationUpdateEffort = model.api === "openai-responses" && compat?.supportsMidConvoEffort === true && last?.type === "configuration_update" && isObject(last.reasoning) && typeof last.reasoning.effort === "string" ? last.reasoning.effort : undefined;
+  const expected = serializeInput(model, context, context, configurationUpdateEffort);
   if (!equal(expected, payload.input)) {
     reject("prepared input differs from serializer");
     return;
   }
-  return { context: structuredClone(context), canonical: structuredClone(canonical), payload: structuredClone(payload), portableInput: expected, createdAt: Date.now() };
+  return { context: structuredClone(context), canonical: structuredClone(canonical), payload: structuredClone(payload), portableInput: expected, configurationUpdateEffort, createdAt: Date.now() };
 }
 function promptCount(input) {
   let count = 0;
@@ -1508,6 +1517,10 @@ function promptCount(input) {
     count++;
   return count;
 }
+function serializePrefix(model, snapshot, prefix) {
+  const input = serializeInput(model, prefix, snapshot.context, snapshot.configurationUpdateEffort);
+  return snapshot.configurationUpdateEffort === undefined ? input : input.slice(0, -1);
+}
 function closedTools(context) {
   const pending = new Set;
   for (const message of context.messages) {
@@ -1551,7 +1564,7 @@ function selectCoveredPrefix(model, snapshot, prefix, reject = () => {}) {
     reject("projected prefix differs from request context");
     return;
   }
-  const expected = serializeInput(model, mapped, snapshot.context);
+  const expected = serializePrefix(model, snapshot, mapped);
   if (!equal(snapshot.portableInput.slice(0, expected.length), expected)) {
     reject("serialized prefix does not match prepared input");
     return;
@@ -1584,7 +1597,7 @@ function summarySlot(model, snapshot, summaryPrefix) {
   const mapped = mappedPrefix(snapshot, summaryPrefix);
   if (!mapped)
     return;
-  const prefix = serializeInput(model, mapped, snapshot.context);
+  const prefix = serializePrefix(model, snapshot, mapped);
   const index = promptCount(prefix);
   if (prefix.length !== index + 1 || !equal(snapshot.portableInput.slice(0, prefix.length), prefix))
     return;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nijaru/pi-provider-compaction",
-  "version": "0.0.6",
+  "version": "0.0.7",
   "type": "module",
   "description": "Provider-native context compaction adapters for Pi",
   "keywords": [
@@ -44,9 +44,9 @@
     "LICENSE"
   ],
   "devDependencies": {
-    "@earendil-works/pi-ai": "1.2.0",
-    "@earendil-works/pi-agent-core": "1.2.0",
-    "@earendil-works/pi-coding-agent": "1.2.0",
+    "@earendil-works/pi-ai": "1.3.0",
+    "@earendil-works/pi-agent-core": "1.3.0",
+    "@earendil-works/pi-coding-agent": "1.3.0",
     "@types/bun": "latest",
     "oxlint": "^1.81.0",
     "typescript": "^6.0.3"

package/replay.ts

@@ -14,7 +14,7 @@ export function equal(left: unknown, right: unknown): boolean {
 }
 
 /** Validator only. No serialized item from this function is transmitted. */
-export function serializeInput(model: Model<any>, context: TranscriptContext, declarations = context): ResponseItem[] {
+export function serializeInput(model: Model<any>, context: TranscriptContext, declarations = context, configurationUpdateEffort?: string): ResponseItem[] {
 	const compat = model.compat as Record<string, unknown> | undefined;
 	const flag = (name: string) => compat?.[name] === true;
 	const codex = model.api === "openai-codex-responses";
@@ -22,6 +22,7 @@ export function serializeInput(model: Model<any>, context: TranscriptContext, de
 	const tools = resolveTranscript(declarations, flag("supportsMidConvoSystemMessages"));
 	return convertResponsesMessages(model, normalized, model.api === "azure-openai-responses" ? AZURE_TOOL_CALL_PROVIDERS : TOOL_CALL_PROVIDERS, {
 		includeSystemPrompt: !codex,
+		configurationUpdateEffort,
 		grammarToolInputProperties: createGrammarToolInputProperties(getDeclaredTools(tools.messages), flag("supportsOpenAIGrammarTools")),
 		supportsMidConvoSystemMessages: flag("supportsMidConvoSystemMessages"),
 		supportsAdditionalTools: flag("supportsAdditionalTools"),
@@ -40,6 +41,8 @@ export interface PreparedRequest {
 	payload: JsonObject & { input: ResponseItem[] };
 	/** The portable wire input before our sole summary substitution. */
 	portableInput: ResponseItem[];
+	/** The final request-only effort update; historical efforts come from the transcript. */
+	configurationUpdateEffort?: string;
 	replacement?: { index: number; count: number };
 	createdAt: number;
 }
@@ -52,9 +55,16 @@ export function capturePreparedRequest(
 	// Reject stateful provider requests: coverage must be the complete visible input.
 	if (payload.previous_response_id || payload.conversation) { reject("stateful request"); return; }
 	if (Buffer.byteLength(JSON.stringify({ context, payload })) > MAX_SNAPSHOT_BYTES) { reject("snapshot exceeds size limit"); return; }
-	const expected = serializeInput(model, context);
+	// Read the actual prepared request's mode, not provider/auth guesses. Exact conversion
+	// below still validates every historical update, its position and its complete shape.
+	const last = payload.input.at(-1);
+	const compat = model.compat as Record<string, unknown> | undefined;
+	const configurationUpdateEffort = model.api === "openai-responses" && compat?.supportsMidConvoEffort === true &&
+		last?.type === "configuration_update" && isObject(last.reasoning) && typeof last.reasoning.effort === "string"
+		? last.reasoning.effort : undefined;
+	const expected = serializeInput(model, context, context, configurationUpdateEffort);
 	if (!equal(expected, payload.input)) { reject("prepared input differs from serializer"); return; }
-	return { context: structuredClone(context), canonical: structuredClone(canonical), payload: structuredClone(payload) as PreparedRequest["payload"], portableInput: expected, createdAt: Date.now() };
+	return { context: structuredClone(context), canonical: structuredClone(canonical), payload: structuredClone(payload) as PreparedRequest["payload"], portableInput: expected, configurationUpdateEffort, createdAt: Date.now() };
 }
 
 function promptCount(input: readonly ResponseItem[]): number {
@@ -63,6 +73,13 @@ function promptCount(input: readonly ResponseItem[]): number {
 	return count;
 }
 
+function serializePrefix(model: Model<any>, snapshot: PreparedRequest, prefix: TranscriptContext): ResponseItem[] {
+	const input = serializeInput(model, prefix, snapshot.context, snapshot.configurationUpdateEffort);
+	// The converter appends the next reply's effort even to a partial transcript.
+	// It is not historical coverage and must remain in the live suffix on replay.
+	return snapshot.configurationUpdateEffort === undefined ? input : input.slice(0, -1);
+}
+
 /** No open/synthetic tool exchange may cross the coverage boundary. */
 function closedTools(context: TranscriptContext): boolean {
 	const pending = new Set<string>();
@@ -94,7 +111,7 @@ export function selectCoveredPrefix(model: Model<any>, snapshot: PreparedRequest
 	if (!closedTools(prefix)) { reject("prefix has an incomplete tool exchange or failed assistant"); return; }
 	const mapped = mappedPrefix(snapshot, prefix);
 	if (!mapped) { reject("projected prefix differs from request context"); return; }
-	const expected = serializeInput(model, mapped, snapshot.context);
+	const expected = serializePrefix(model, snapshot, mapped);
 	if (!equal(snapshot.portableInput.slice(0, expected.length), expected)) { reject("serialized prefix does not match prepared input"); return; }
 	const start = promptCount(expected);
 	let end = expected.length;
@@ -121,7 +138,7 @@ export function substituteSummary(
 export function summarySlot(model: Model<any>, snapshot: PreparedRequest, summaryPrefix: TranscriptContext): number | undefined {
 	const mapped = mappedPrefix(snapshot, summaryPrefix);
 	if (!mapped) return;
-	const prefix = serializeInput(model, mapped, snapshot.context);
+	const prefix = serializePrefix(model, snapshot, mapped);
 	const index = promptCount(prefix);
 	if (prefix.length !== index + 1 || !equal(snapshot.portableInput.slice(0, prefix.length), prefix)) return;
 	return index;