@nijaru/pi-provider-compaction 0.0.5 → 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.
 
@@ -38,7 +38,7 @@ Request-only suffix additions, deletions, redactions and reordering can survive
 
 Version 3 stores the portable summary, native output, usage, route identity and validated prefix metadata in one compaction entry. Both generation requests contribute to usage accounting. Raw credentials, complete auth headers and request snapshots are never persisted.
 
-Native eligibility is deliberately **session-local and short-lived**. Restart, reload, resume, fork, tree navigation, model changes, covered-history edits, and relevant prompt/tool/route changes revoke it. A credential refresh may conservatively revoke it too. Persisted native metadata alone is not authority to replay hidden content under a new runtime's privacy policy.
+Native eligibility is deliberately **session-local and short-lived**. Restart, reload, SDK disposal, resume, fork, tree navigation, model changes, covered-history edits, and relevant prompt/tool/route changes revoke it. A credential refresh may conservatively revoke it too. Persisted native metadata alone is not authority to replay hidden content under a new runtime's privacy policy.
 
 Dynamic privacy policies that can change what previously covered history is allowed to reveal are **not transparently supported**. A cooperating policy owner must emit `pi-provider-compaction:invalidate` before such a change; otherwise keep this extension disabled. Matching visible summary text cannot establish that hidden native state obeys a new policy.
 
@@ -60,7 +60,7 @@ An explicit generic model takes precedence, matching `pi-compactor`:
 2. trusted `.pi/compaction-policy.json`;
 3. agent `compaction-policy.json`.
 
-An empty `models` list is not an override.
+An empty `models` list is not an override. When `pi-compactor` is loaded, its effective selection takes precedence, including flags injected by an SDK host rather than passed on the command line.
 
 ## Local development
 
@@ -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. The optional installed-fast-mode fixture (`PI_FAST_MODE_FIXTURE_PATH`) also checks actual `pi-fast-mode` composition across `/reload`; run it with an isolated `PI_CODING_AGENT_DIR` containing an active, persistent 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

@@ -1,6 +1,6 @@
 // index.ts
 import { createHash, createHmac, randomBytes, randomUUID } from "node:crypto";
-import { normalizeContext as normalizeContext2 } from "@earendil-works/pi-ai";
+import { normalizeContext as normalizeContext2, registerSessionResourceCleanup } from "@earendil-works/pi-ai";
 import { compact, convertToLlm } from "@earendil-works/pi-coding-agent";
 
 // policy.ts
@@ -119,18 +119,21 @@ function readCompactionModelFlagFromArgv(argv = process.argv.slice(2)) {
   return value;
 }
 function resolveCompactionModelPolicy(pi, ctx, argv = process.argv.slice(2)) {
+  let effective;
+  pi.events.emit("pi-compactor:resolve-model-policy", {
+    context: ctx,
+    respond: (hasSelectors) => {
+      if (typeof hasSelectors === "boolean")
+        effective = hasSelectors;
+    }
+  });
+  if (effective !== undefined)
+    return effective ? { hasSelectors: true, source: "compactor" } : { hasSelectors: false };
   const argvRaw = readCompactionModelFlagFromArgv(argv);
   if (typeof argvRaw === "string" && argvRaw.trim()) {
     const selector = normalizeSelector(argvRaw);
     return selector ? { hasSelectors: true, source: "flag" } : { hasSelectors: false };
   }
-  try {
-    const rawFlag = pi.getFlag(COMPACTION_MODEL_FLAG);
-    if (typeof rawFlag === "string" && rawFlag.trim()) {
-      const selector = normalizeSelector(rawFlag);
-      return selector ? { hasSelectors: true, source: "flag" } : { hasSelectors: false };
-    }
-  } catch {}
   const agentDir = activeAgentDir(ctx);
   const candidates = [];
   if (isProjectPolicyTrusted(ctx, agentDir)) {
@@ -147,8 +150,14 @@ function resolveCompactionModelPolicy(pi, ctx, argv = process.argv.slice(2)) {
 }
 
 // provider.ts
+var STREAM_OBSERVER = Symbol.for("@nijaru/pi-provider-compaction/stream-observer");
 function optionsWithObserver(model, context, options, observe) {
-  const handler = observe(model, context, options ?? {});
+  let handler;
+  try {
+    handler = observe(model, context, options ?? {});
+  } catch {
+    return options;
+  }
   if (!handler)
     return options;
   return {
@@ -191,15 +200,22 @@ function installObserver(pi, ctx, observe) {
   const id = model.provider;
   const native = registry.getRegisteredNativeProvider(id);
   if (native) {
+    const isInstalled = () => registry.getRegisteredNativeProvider(id) === wrapper;
     const guarded = (...args) => {
+      if (!isInstalled())
+        return;
       const handler = observe(...args);
-      return handler ? (payload) => registry.getRegisteredNativeProvider(id) === wrapper ? handler(payload) : payload : undefined;
+      return handler ? (payload) => isInstalled() ? handler(payload) : payload : undefined;
     };
     const wrapper = decorateProvider(native, guarded);
     pi.registerProvider(wrapper);
-    return () => {
-      if (registry.getRegisteredNativeProvider(id) === wrapper)
-        pi.registerProvider(native);
+    return {
+      provider: id,
+      isInstalled,
+      remove() {
+        if (isInstalled())
+          registry.registerProvider(native);
+      }
     };
   }
   const original = registry.getRegisteredProviderConfig(id);
@@ -208,27 +224,24 @@ function installObserver(pi, ctx, observe) {
   if (original.api !== model.api)
     return;
   const delegate = original.streamSimple.bind(original);
+  const isInstalled = () => registry.getRegisteredProviderConfig(id)?.streamSimple === wrapper && !registry.getRegisteredNativeProvider(id);
   const guarded = (...args) => {
+    if (!isInstalled())
+      return;
     const handler = observe(...args);
-    return handler ? (payload) => registry.getRegisteredProviderConfig(id)?.streamSimple === wrapper ? handler(payload) : payload : undefined;
+    return handler ? (payload) => isInstalled() ? handler(payload) : payload : undefined;
   };
   const wrapper = (requestModel, context, options) => delegate(requestModel, context, optionsWithObserver(requestModel, context, options, guarded));
-  pi.registerProvider(id, { ...original, api: original?.api ?? model.api, streamSimple: wrapper });
-  return () => {
-    const current = registry.getRegisteredProviderConfig(id);
-    if (current?.streamSimple !== wrapper || registry.getRegisteredNativeProvider(id))
+  const remove = () => {
+    if (!isInstalled())
       return;
-    const restored = { ...current };
-    if (original?.streamSimple)
-      restored.streamSimple = original.streamSimple;
-    else
-      delete restored.streamSimple;
-    if (!original?.api && restored.api === model.api)
-      delete restored.api;
-    pi.unregisterProvider(id);
-    if (Object.keys(restored).length)
-      pi.registerProvider(id, restored);
+    const restored = { ...registry.getRegisteredProviderConfig(id), streamSimple: original.streamSimple };
+    registry.unregisterProvider(id);
+    registry.registerProvider(id, restored);
   };
+  Object.defineProperty(wrapper, STREAM_OBSERVER, { value: { delegate: original.streamSimple, release: remove } });
+  pi.registerProvider(id, { ...original, streamSimple: wrapper });
+  return { provider: id, isInstalled, remove };
 }
 
 // replay.ts
@@ -624,19 +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)
@@ -827,6 +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;
@@ -933,6 +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("|");
@@ -954,6 +970,8 @@ function convertResponsesMessages(model, context, allowedToolCallProviders, opti
     if (!isLeadingSystemMessage)
       msgIndex++;
   }
+  if (options?.configurationUpdateEffort !== undefined)
+    pushConfigurationUpdate(options.configurationUpdateEffort);
   return messages;
 }
 function convertResponsesTools(tools, options) {
@@ -1450,7 +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";
@@ -1458,6 +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"),
@@ -1482,12 +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;
@@ -1495,6 +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) {
@@ -1538,7 +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;
@@ -1571,7 +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;
@@ -1692,8 +1718,8 @@ function pi_provider_compaction_default(pi) {
   let snapshot;
   let active;
   let pending;
-  let uninstall;
-  let installedProvider;
+  let observer;
+  let unregisterCleanup;
   let generation = 0;
   let captureStatus = "no request observed";
   let compactionStatus = "not attempted in this runtime";
@@ -1707,7 +1733,7 @@ function pi_provider_compaction_default(pi) {
     description: "Show content-free native compaction eligibility and last attempt",
     handler: async (_args, context) => {
       context.ui.notify(`Provider compaction
-Observer: ${installedProvider ? "installed" : "unavailable for current registration"}
+Observer: ${observer?.isInstalled() ? "installed" : "unavailable for current registration"}
 Capture: ${captureStatus}
 Last attempt: ${compactionStatus}`, "info");
     }
@@ -1720,12 +1746,30 @@ Last attempt: ${compactionStatus}`, "info");
     pending = undefined;
     captureStatus = "invalidated by lifecycle/history/policy change";
   };
+  const retire = () => {
+    invalidate();
+    const installed = observer;
+    observer = undefined;
+    ctx = undefined;
+    unregisterCleanup?.();
+    unregisterCleanup = undefined;
+    installed?.remove();
+  };
   const generic = (context) => resolveCompactionModelPolicy(pi, context).hasSelectors;
   const route = (model, options) => routeDigest({ model, apiKey: options.apiKey, headers: options.headers, env: options.env });
   const unchanged = (branch, length, hash) => branch.length >= length && digest(branch.slice(0, length)) === hash && !branch.slice(length).some((entry) => entry.type === "context_edit");
   const observe = (model, context, options) => {
     const owner = ctx;
-    if (!owner || options.sessionId !== owner.sessionManager.getSessionId())
+    if (!owner)
+      return;
+    let manager;
+    try {
+      manager = owner.sessionManager;
+    } catch {
+      retire();
+      return;
+    }
+    if (options.sessionId !== manager.getSessionId())
       return;
     if (!supportsNativeCompaction(model)) {
       captureRejected("unsupported API");
@@ -1744,10 +1788,19 @@ Last attempt: ${compactionStatus}`, "info");
       return;
     }
     const ticket = generation;
-    const branch = owner.sessionManager.getBranch();
+    const branch = manager.getBranch();
     const branchHash = digest(branch);
-    const canonical = transcript(owner.sessionManager.buildSessionProjection().messages);
-    const current = () => generation === ticket && ctx === owner && !options.signal?.aborted && supportedRouteOptions(model, options) && digest(owner.sessionManager.getBranch()) === branchHash;
+    const canonical = transcript(manager.buildSessionProjection().messages);
+    const current = () => {
+      if (generation !== ticket || ctx !== owner || options.signal?.aborted || !supportedRouteOptions(model, options))
+        return false;
+      try {
+        return digest(owner.sessionManager.getBranch()) === branchHash;
+      } catch {
+        retire();
+        return false;
+      }
+    };
     return async (payload) => {
       if (!current())
         return payload;
@@ -1811,21 +1864,26 @@ Last attempt: ${compactionStatus}`, "info");
   };
   const install = (context) => {
     ctx = context;
-    if (installedProvider === context.model?.provider)
+    if (observer?.provider === context.model?.provider && observer?.isInstalled())
       return;
-    uninstall?.();
-    uninstall = undefined;
-    installedProvider = undefined;
+    if (observer) {
+      invalidate();
+      observer.remove();
+      observer = undefined;
+    }
     if (!supportsNativeCompaction(context.model))
       return;
-    uninstall = installObserver(pi, context, observe);
-    if (uninstall)
-      installedProvider = context.model.provider;
-    captureStatus = uninstall ? "awaiting prepared request" : "unsupported provider registration";
+    observer = installObserver(pi, context, observe);
+    captureStatus = observer ? "awaiting prepared request" : "unsupported provider registration";
   };
   pi.on("session_start", (_event, context) => {
-    invalidate();
+    retire();
     install(context);
+    const sessionId = context.sessionManager.getSessionId();
+    unregisterCleanup = registerSessionResourceCleanup((disposedId) => {
+      if (disposedId === undefined || disposedId === sessionId)
+        retire();
+    });
     const latest = latestCompaction(context.sessionManager.getBranch());
     if (latest?.summary === NATIVE_COMPACTION_SUMMARY)
       context.ui.notify("Legacy native compaction has no portable summary. Native replay is disabled; recover from earlier history or start a new session.", "warning");
@@ -1840,13 +1898,7 @@ Last attempt: ${compactionStatus}`, "info");
   pi.on("session_tree", () => {
     invalidate();
   });
-  pi.on("session_shutdown", () => {
-    invalidate();
-    uninstall?.();
-    uninstall = undefined;
-    installedProvider = undefined;
-    ctx = undefined;
-  });
+  pi.on("session_shutdown", retire);
   const removeInvalidation = pi.events.on("pi-provider-compaction:invalidate", invalidate);
   pi.on("session_shutdown", () => {
     if (typeof removeInvalidation === "function")

package/index.ts

@@ -1,9 +1,9 @@
 import { createHash, createHmac, randomBytes, randomUUID } from "node:crypto";
 import type { AgentMessage } from "@earendil-works/pi-agent-core";
-import { normalizeContext, type Model, type Provider, type ProviderHeaders, type ProviderStreamOptions, type SimpleStreamOptions, type TranscriptContext, type Usage } from "@earendil-works/pi-ai";
+import { normalizeContext, registerSessionResourceCleanup, type Model, type Provider, type ProviderHeaders, type ProviderStreamOptions, type SimpleStreamOptions, type TranscriptContext, type Usage } from "@earendil-works/pi-ai";
 import { compact, convertToLlm, type ExtensionAPI, type ExtensionContext, type SessionEntry } from "@earendil-works/pi-coding-agent";
 import { resolveCompactionModelPolicy } from "./policy";
-import { installObserver, type ObserveRequest } from "./provider";
+import { installObserver, type InstalledObserver, type ObserveRequest } from "./provider";
 import { capturePreparedRequest, equal, requestPolicy, selectCoveredPrefix, substituteSummary, summarySlot, type PreparedRequest } from "./replay";
 import { isObject, requestProviderCompaction, resolveNativeProtocol, supportsResponsesApi, validateCompactedOutput, type NativeProtocol, type ResponseItem } from "./protocol";
 
@@ -152,8 +152,8 @@ export default function (pi: ExtensionAPI): void {
 	let snapshot: Captured | undefined;
 	let active: Active | undefined;
 	let pending: { details: NativeCompactionDetails; capture: Captured; summary: string } | undefined;
-	let uninstall: (() => void) | undefined;
-	let installedProvider: string | undefined;
+	let observer: InstalledObserver | undefined;
+	let unregisterCleanup: (() => void) | undefined;
 	let generation = 0;
 	let captureStatus = "no request observed";
 	let compactionStatus = "not attempted in this runtime";
@@ -162,17 +162,33 @@ export default function (pi: ExtensionAPI): void {
 	pi.registerCommand("provider-compaction-status", {
 		description: "Show content-free native compaction eligibility and last attempt",
 		handler: async (_args, context) => {
-			context.ui.notify(`Provider compaction\nObserver: ${installedProvider ? "installed" : "unavailable for current registration"}\nCapture: ${captureStatus}\nLast attempt: ${compactionStatus}`, "info");
+			context.ui.notify(`Provider compaction\nObserver: ${observer?.isInstalled() ? "installed" : "unavailable for current registration"}\nCapture: ${captureStatus}\nLast attempt: ${compactionStatus}`, "info");
 		},
 	});
 	const invalidate = () => { generation++; runtimeId = randomUUID(); snapshot = undefined; active = undefined; pending = undefined; captureStatus = "invalidated by lifecycle/history/policy change"; };
+	const retire = () => {
+		invalidate();
+		const installed = observer;
+		observer = undefined; ctx = undefined;
+		unregisterCleanup?.(); unregisterCleanup = undefined;
+		installed?.remove();
+	};
 	const generic = (context: ExtensionContext) => resolveCompactionModelPolicy(pi, context).hasSelectors;
 	const route = (model: Model<any>, options: ProviderStreamOptions | SimpleStreamOptions) => routeDigest({ model, apiKey: options.apiKey, headers: options.headers, env: options.env });
 	const unchanged = (branch: readonly SessionEntry[], length: number, hash: string) => branch.length >= length && digest(branch.slice(0, length)) === hash && !branch.slice(length).some((entry) => entry.type === "context_edit");
 
 	const observe: ObserveRequest = (model, context, options) => {
 		const owner = ctx;
-		if (!owner || options.sessionId !== owner.sessionManager.getSessionId()) return;
+		if (!owner) return;
+		let manager: ExtensionContext["sessionManager"];
+		try { manager = owner.sessionManager; }
+		catch {
+			// Bare SDK dispose() invalidates contexts without session_shutdown.
+			// Retire the owned registration before dispatching another session.
+			retire();
+			return;
+		}
+		if (options.sessionId !== manager.getSessionId()) return;
 		if (!supportsNativeCompaction(model)) { captureRejected("unsupported API"); return; }
 		if (!options.onPayload) { captureRejected("no payload callback"); return; }
 		if (generic(owner)) { captureRejected("generic compaction model takes precedence"); return; }
@@ -180,10 +196,14 @@ export default function (pi: ExtensionAPI): void {
 		// fallbacks. Those must not silently select a route absent from our evidence.
 		if (!supportedRouteOptions(model, options)) { captureRejected("unsupported route overrides"); return; }
 		const ticket = generation;
-		const branch = owner.sessionManager.getBranch();
+		const branch = manager.getBranch();
 		const branchHash = digest(branch);
-		const canonical = transcript(owner.sessionManager.buildSessionProjection().messages);
-		const current = () => generation === ticket && ctx === owner && !options.signal?.aborted && supportedRouteOptions(model, options) && digest(owner.sessionManager.getBranch()) === branchHash;
+		const canonical = transcript(manager.buildSessionProjection().messages);
+		const current = () => {
+			if (generation !== ticket || ctx !== owner || options.signal?.aborted || !supportedRouteOptions(model, options)) return false;
+			try { return digest(owner.sessionManager.getBranch()) === branchHash; }
+			catch { retire(); return false; }
+		};
 		return async (payload) => {
 			if (!current()) return payload;
 			snapshot = undefined;
@@ -234,22 +254,27 @@ export default function (pi: ExtensionAPI): void {
 
 	const install = (context: ExtensionContext) => {
 		ctx = context;
-		if (installedProvider === context.model?.provider) return;
-		uninstall?.(); uninstall = undefined; installedProvider = undefined;
+		if (observer?.provider === context.model?.provider && observer?.isInstalled()) return;
+		if (observer) { invalidate(); observer.remove(); observer = undefined; }
 		if (!supportsNativeCompaction(context.model)) return;
-		uninstall = installObserver(pi, context, observe);
-		if (uninstall) installedProvider = context.model.provider;
-		captureStatus = uninstall ? "awaiting prepared request" : "unsupported provider registration";
+		observer = installObserver(pi, context, observe);
+		captureStatus = observer ? "awaiting prepared request" : "unsupported provider registration";
 	};
 	pi.on("session_start", (_event, context) => {
-		invalidate(); install(context);
+		retire(); install(context);
+		const sessionId = context.sessionManager.getSessionId();
+		// AgentSession.dispose() cleans session resources even without emitting
+		// session_shutdown. The captured registry keeps owned cleanup usable.
+		unregisterCleanup = registerSessionResourceCleanup((disposedId) => {
+			if (disposedId === undefined || disposedId === sessionId) retire();
+		});
 		const latest = latestCompaction(context.sessionManager.getBranch());
 		if (latest?.summary === NATIVE_COMPACTION_SUMMARY) context.ui.notify("Legacy native compaction has no portable summary. Native replay is disabled; recover from earlier history or start a new session.", "warning");
 	});
 	pi.on("before_agent_start", (_event, context) => { install(context); });
 	pi.on("model_select", (_event, context) => { invalidate(); install(context); });
 	pi.on("session_tree", () => { invalidate(); });
-	pi.on("session_shutdown", () => { invalidate(); uninstall?.(); uninstall = undefined; installedProvider = undefined; ctx = undefined; });
+	pi.on("session_shutdown", retire);
 	// Privacy-policy owners must invalidate before changing a policy that can affect
 	// hidden covered history. Arbitrary dynamic privacy compositions are unsupported.
 	const removeInvalidation = pi.events.on("pi-provider-compaction:invalidate", invalidate);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nijaru/pi-provider-compaction",
-  "version": "0.0.5",
+  "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/policy.ts

@@ -1,31 +1,9 @@
 /**
- * Generic-compaction-model precedence resolution shared with pi-compactor.
- *
- * When pi-compactor's generic compaction model is configured — via the
- * `compaction-model` flag or a `compaction-policy.json` file — the explicit
- * generic model takes precedence over provider-native compaction: this
- * extension must not request a native compaction and must not replay a
- * previously persisted native window.
- *
- * Pi scopes `getFlag` to flags the calling extension registered, and Pi
- * rejects two extensions registering the same flag name. This extension
- * therefore never registers `compaction-model` (owned by pi-compactor) and
- * reads the shared CLI value from `process.argv` instead. The policy-file
- * fallback mirrors pi-compactor exactly: project policy requires project
- * trust; the agent-dir policy does not; a valid file with an empty `models`
- * list disables the generic model. This module must stay in sync with
- * `readModelSelectors` in pi-compactor's policy.ts (see
- * tests/policy.test.ts, which always checks the safety contracts and also
- * compares both implementations when the sibling checkout is present).
- *
- * Flag-source note (verified against Pi 0.85.1): extension flag values are
- * populated solely from CLI unknown flags (parseArgs in cli/args.js feeds
- * extensionFlagValues in main.js, applied by applyExtensionFlagValues in
- * agent-session-services.js). There is no config-file flag source, so reading
- * process.argv with Pi's exact parsing rules observes every flag value the
- * host can deliver. A programmatic SDK host injecting extensionFlagValues
- * without argv would not be seen here; flag scoping provides no shared read
- * API for that path.
+ * An explicit generic model suppresses both native acquisition and replay.
+ * Ask pi-compactor for its effective policy first: only the flag's owner can
+ * read SDK-injected extensionFlagValues. The event reply is synchronous.
+ * Without a companion reply, argv and policy files retain standalone support;
+ * their trust, bounds, and empty-policy rules mirror pi-compactor's resolver.
  */
 
 import {
@@ -49,7 +27,7 @@ type PolicyContext = Pick<ExtensionContext, "cwd" | "isProjectTrusted"> & {
 	sessionManager?: Pick<ExtensionContext["sessionManager"], "getSessionDir">;
 };
 
-export type CompactionModelSource = "flag" | "project-policy" | "agent-policy";
+export type CompactionModelSource = "compactor" | "flag" | "project-policy" | "agent-policy";
 
 export interface CompactionModelPolicy {
 	/** True when a generic compaction model takes precedence over native compaction. */
@@ -172,27 +150,24 @@ export function readCompactionModelFlagFromArgv(argv: readonly string[] = proces
 
 /** Resolve whether pi-compactor's generic compaction model takes precedence. */
 export function resolveCompactionModelPolicy(
-	pi: Pick<ExtensionAPI, "getFlag">,
+	pi: { events: Pick<ExtensionAPI["events"], "emit"> },
 	ctx: PolicyContext,
 	argv: readonly string[] = process.argv.slice(2),
 ): CompactionModelPolicy {
+	let effective: boolean | undefined;
+	pi.events.emit("pi-compactor:resolve-model-policy", {
+		context: ctx,
+		respond: (hasSelectors: boolean) => { if (typeof hasSelectors === "boolean") effective = hasSelectors; },
+	});
+	if (effective !== undefined) return effective ? { hasSelectors: true, source: "compactor" } : { hasSelectors: false };
+
 	const argvRaw = readCompactionModelFlagFromArgv(argv);
 	if (typeof argvRaw === "string" && argvRaw.trim()) {
 		const selector = normalizeSelector(argvRaw);
 		return selector ? { hasSelectors: true, source: "flag" } : { hasSelectors: false };
 	}
-	// A bare `--compaction-model` (true) or whitespace value falls through like
-	// pi-compactor: Pi reports "requires a value" and the policy files still apply.
+	// A bare or whitespace flag leaves policy files eligible, like pi-compactor.
 	// An oversized selector above already returned hasSelectors:false.
-	try {
-		const rawFlag = pi.getFlag(COMPACTION_MODEL_FLAG);
-		if (typeof rawFlag === "string" && rawFlag.trim()) {
-			const selector = normalizeSelector(rawFlag);
-			return selector ? { hasSelectors: true, source: "flag" } : { hasSelectors: false };
-		}
-	} catch {
-		// getFlag can throw on a stale context; fall through to policy files.
-	}
 
 	const agentDir = activeAgentDir(ctx);
 	const candidates: Array<{ path: string; source: CompactionModelSource }> = [];

package/provider.ts

@@ -3,8 +3,20 @@ import type { ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-a
 
 export type ObserveRequest = (model: Model<any>, context: TranscriptContext, options: StreamOptions) => ((payload: unknown) => unknown | Promise<unknown>) | undefined;
 
+export interface InstalledObserver {
+	provider: string;
+	isInstalled(): boolean;
+	remove(): void;
+}
+
+// Explicit cooperation with pi-fast-mode, not permission to replace arbitrary
+// streams: this observer delegates unchanged and can release only its own slot.
+export const STREAM_OBSERVER = Symbol.for("@nijaru/pi-provider-compaction/stream-observer");
+
 function optionsWithObserver<T extends StreamOptions>(model: Model<any>, context: TranscriptContext, options: T | undefined, observe: ObserveRequest): T | undefined {
-	const handler = observe(model, context, options ?? {});
+	let handler: ReturnType<ObserveRequest>;
+	try { handler = observe(model, context, options ?? {}); }
+	catch { return options; } // An expired or failing optional observer must not veto dispatch.
 	if (!handler) return options;
 	return {
 		...options,
@@ -40,21 +52,27 @@ export function decorateProvider(base: Provider, observe: ObserveRequest): Provi
 	* delegate their original callback (including /fast). Never register an effective
 	* composed provider as a native base: that loses model headers and freezes config.
 	*/
-export function installObserver(pi: ExtensionAPI, ctx: ExtensionContext, observe: ObserveRequest): (() => void) | undefined {
+export function installObserver(pi: ExtensionAPI, ctx: ExtensionContext, observe: ObserveRequest): InstalledObserver | undefined {
 	const model = ctx.model;
 	if (!model) return;
 	const registry = ctx.modelRegistry;
 	const id = model.provider;
 	const native = registry.getRegisteredNativeProvider(id);
 	if (native) {
+		const isInstalled = () => registry.getRegisteredNativeProvider(id) === wrapper;
 		const guarded: ObserveRequest = (...args) => {
+			if (!isInstalled()) return;
 			const handler = observe(...args);
-			return handler ? (payload) => registry.getRegisteredNativeProvider(id) === wrapper ? handler(payload) : payload : undefined;
+			return handler ? (payload) => isInstalled() ? handler(payload) : payload : undefined;
 		};
 		const wrapper = decorateProvider(native, guarded);
 		pi.registerProvider(wrapper);
-		return () => {
-			if (registry.getRegisteredNativeProvider(id) === wrapper) pi.registerProvider(native);
+		return {
+			provider: id, isInstalled,
+			remove() {
+				// Keep cleanup usable after bare SDK disposal invalidates the extension API.
+				if (isInstalled()) registry.registerProvider(native);
+			},
 		};
 	}
 	const original = registry.getRegisteredProviderConfig(id);
@@ -65,21 +83,21 @@ export function installObserver(pi: ExtensionAPI, ctx: ExtensionContext, observe
 	// An injected legacy streamSimple overlay would also intercept full stream()
 	// calls and change reasoning semantics. Only decorate an existing stream owner.
 	const delegate = original.streamSimple.bind(original);
+	const isInstalled = () => registry.getRegisteredProviderConfig(id)?.streamSimple === wrapper && !registry.getRegisteredNativeProvider(id);
 	const guarded: ObserveRequest = (...args) => {
+		if (!isInstalled()) return;
 		const handler = observe(...args);
-		return handler ? (payload) => registry.getRegisteredProviderConfig(id)?.streamSimple === wrapper ? handler(payload) : payload : undefined;
+		return handler ? (payload) => isInstalled() ? handler(payload) : payload : undefined;
 	};
 	const wrapper: Provider["streamSimple"] = (requestModel, context, options) => delegate(requestModel, context, optionsWithObserver(requestModel, context, options, guarded));
-	pi.registerProvider(id, { ...original, api: original?.api ?? model.api, streamSimple: wrapper });
-	return () => {
-		const current = registry.getRegisteredProviderConfig(id);
-		if (current?.streamSimple !== wrapper || registry.getRegisteredNativeProvider(id)) return;
+	const remove = () => {
+		if (!isInstalled()) return;
 		// registerProvider merges undefined properties: remove our layer before restore.
-		const restored = { ...current };
-		if (original?.streamSimple) restored.streamSimple = original.streamSimple;
-		else delete restored.streamSimple;
-		if (!original?.api && restored.api === model.api) delete restored.api;
-		pi.unregisterProvider(id);
-		if (Object.keys(restored).length) pi.registerProvider(id, restored);
+		const restored = { ...registry.getRegisteredProviderConfig(id)!, streamSimple: original.streamSimple };
+		registry.unregisterProvider(id);
+		registry.registerProvider(id, restored);
 	};
+	Object.defineProperty(wrapper, STREAM_OBSERVER, { value: { delegate: original.streamSimple, release: remove } });
+	pi.registerProvider(id, { ...original, streamSimple: wrapper });
+	return { provider: id, isInstalled, remove };
 }

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;