@raquezha/noheadroom 0.3.0 → 0.3.2

package/dist/bridge.js

@@ -32,6 +32,8 @@ export function applyCompressionResult(originalMessages, mappings, compressedMes
     }
     const nextMessages = structuredClone(originalMessages);
     let appliedMessages = 0;
+    let appliedTokensBefore = 0;
+    let appliedTokensAfter = 0;
     for (let index = 0; index < mappings.length; index++) {
         const mapping = mappings[index];
         const compressed = compressedMessages[index];
@@ -52,12 +54,26 @@ export function applyCompressionResult(originalMessages, mappings, compressedMes
         if (!hasContent(target) || !replaceTextContent(target, nextText)) {
             return { ok: false, reason: "target-content-unreplaceable" };
         }
+        appliedTokensBefore += estimateTokens(mapping.originalText);
+        appliedTokensAfter += estimateTokens(nextText);
         appliedMessages++;
     }
     if (appliedMessages === 0) {
         return { ok: false, reason: "no-applicable-message-changed" };
     }
-    return { ok: true, messages: nextMessages, appliedMessages };
+    const appliedTokensSaved = Math.max(0, appliedTokensBefore - appliedTokensAfter);
+    if (appliedTokensSaved === 0) {
+        return { ok: false, reason: "no-estimated-token-savings" };
+    }
+    return {
+        ok: true,
+        messages: nextMessages,
+        appliedMessages,
+        appliedTokensBefore,
+        appliedTokensAfter,
+        appliedTokensSaved,
+        appliedCompressionRatio: appliedTokensAfter / appliedTokensBefore,
+    };
 }
 export function convertMessage(message) {
     if (!message.role || !STANDARD_COMPRESSIBLE_ROLES.has(message.role))
@@ -221,6 +237,10 @@ function readStringProperty(value, key) {
 function isRecord(value) {
     return typeof value === "object" && value !== null;
 }
+function estimateTokens(text) {
+    // ponytail: cheap local estimate; exact tokenizer would add a heavy dependency for a footer number.
+    return text.length === 0 ? 0 : Math.max(1, Math.ceil(text.length / 4));
+}
 function naturalizeHeadroomMarkers(text) {
     // Translate Headroom's "[X items compressed to Y. Retrieve more: hash=...]" 
     // into a Pi-native instruction to use offset/limit instead of hallucinating a retrieve command.

package/dist/index.d.ts

@@ -12,6 +12,8 @@ interface HeadroomRuntimeState {
     lastInputFingerprint: string | null;
     lastOutputFingerprint: string | null;
     lastGuardSkipCandidateFingerprint: string | null;
+    seenCandidateContentFingerprints?: Set<string>;
+    seenCandidateContentOrder?: string[];
     lastCompressionTime: number;
     stats: HeadroomStats;
 }

package/dist/index.js

@@ -8,6 +8,10 @@ const STATUS_KEY = "headroom";
 const SUBCOMMANDS = ["status", "on", "off", "health", "stats", "mode"];
 const MODES = ["normal", "quiet", "silent"];
 const NOTRACE_TELEMETRY_CHANNEL = "notrace.telemetry.extension";
+// Bounded FIFO, not config: this is a safety valve for duplicate-compression loops.
+// 512 hashes covers roughly 256 compressed candidates because we store original + compressed text hashes.
+// That is far beyond the reread-loop case while keeping memory bounded for long sessions.
+const MAX_SEEN_CANDIDATE_CONTENT_FINGERPRINTS = 512;
 const COMPRESSION_ENTRY_TYPE = "noheadroom.compression";
 export default function headroomExtension(pi) {
     registerCompressionRenderer(pi);
@@ -73,6 +77,8 @@ function createRuntime(pi) {
         lastInputFingerprint: null,
         lastOutputFingerprint: null,
         lastGuardSkipCandidateFingerprint: null,
+        seenCandidateContentFingerprints: new Set(),
+        seenCandidateContentOrder: [],
         lastCompressionTime: 0,
         stats: { attempts: 0, applied: 0, guardSkips: 0, tokensSaved: 0 },
     };
@@ -201,10 +207,17 @@ function registerCompressionRenderer(pi) {
 async function handleContextCompression(runtime, event, ctx) {
     if (runtime.state.processing)
         return undefined;
+    const bypassUncompressed = () => {
+        if (runtime.state.stats.last) {
+            runtime.state.stats.last = undefined;
+            runtime.refreshStatus(ctx);
+        }
+        return undefined;
+    };
     // Throttle: max 1 compression attempt per 3 seconds to kill infinite loops
     const now = Date.now();
     if (now - runtime.state.lastCompressionTime < 3000) {
-        return undefined;
+        return bypassUncompressed();
     }
     // Content-based guards to prevent infinite recursion
     const inputFingerprint = generateFingerprint(event.messages);
@@ -214,22 +227,26 @@ async function handleContextCompression(runtime, event, ctx) {
     }
     // 2. If this input matches our previous input, it didn't change. Stop.
     if (runtime.state.lastInputFingerprint === inputFingerprint) {
-        return undefined;
+        return bypassUncompressed();
     }
     if (shouldSkipBeforePayload(runtime, ctx))
-        return undefined;
+        return bypassUncompressed();
     const payload = buildCompressionPayload(event.messages, runtime.config.minMessageChars);
     if (payload.candidateCount === 0)
-        return undefined;
+        return bypassUncompressed();
     // 3. If eligible candidates haven't changed since the last skip/no-savings result,
     // don't spend another proxy call just because surrounding conversation changed.
     const candidateFingerprint = generateCandidateFingerprint(event.messages, payload);
     if (runtime.state.lastGuardSkipCandidateFingerprint === candidateFingerprint) {
-        return undefined;
+        return bypassUncompressed();
+    }
+    if (allCandidateContentSeen(runtime.state, payload)) {
+        return bypassUncompressed();
     }
+    ignoreSeenCandidateContent(runtime.state, payload);
     if (runtime.state.proxyOnline !== true) {
         void ensureProxyInBackground(runtime, ctx);
-        return undefined;
+        return bypassUncompressed();
     }
     runtime.state.processing = true;
     runtime.state.lastCompressionTime = now;
@@ -242,10 +259,7 @@ async function handleContextCompression(runtime, event, ctx) {
             // until the actual compressible candidate material changes.
             runtime.state.lastInputFingerprint = inputFingerprint;
             runtime.state.lastGuardSkipCandidateFingerprint = candidateFingerprint;
-            // ponytail: clear stats.last so renderFooterStatus shows idle instead of stale savings
-            runtime.state.stats.last = undefined;
-            runtime.refreshStatus(ctx);
-            return undefined;
+            return bypassUncompressed();
         }
         const applied = applyCompressionResult(event.messages, payload.mappings, result.messages, {
             minMessageChars: runtime.config.minMessageChars,
@@ -258,27 +272,75 @@ async function handleContextCompression(runtime, event, ctx) {
             recordGuardSkip(runtime.state.stats, applied.reason);
             emitNotraceTelemetry(runtime);
             announceGuardSkip(ctx, applied.reason, result, runtime.config.mode);
-            runtime.refreshStatus(ctx);
-            return undefined;
+            return bypassUncompressed();
         }
         // Store fingerprints to break the feedback loop
         runtime.state.lastInputFingerprint = inputFingerprint;
         runtime.state.lastOutputFingerprint = generateFingerprint(applied.messages);
         runtime.state.lastGuardSkipCandidateFingerprint = null;
-        recordAppliedCompression(runtime.state.stats, result, applied.appliedMessages);
+        recordSeenCandidateContent(runtime.state, payload, applied.messages);
+        const appliedResult = {
+            ...result,
+            tokensBefore: applied.appliedTokensBefore,
+            tokensAfter: applied.appliedTokensAfter,
+            tokensSaved: applied.appliedTokensSaved,
+            compressionRatio: applied.appliedCompressionRatio,
+        };
+        recordAppliedCompression(runtime.state.stats, appliedResult, applied.appliedMessages);
         emitNotraceTelemetry(runtime);
-        announceAppliedCompression(runtime.pi, ctx, result, applied.appliedMessages, runtime.config.mode);
+        announceAppliedCompression(runtime.pi, ctx, appliedResult, applied.appliedMessages, runtime.config.mode);
         runtime.refreshStatus(ctx);
         return { messages: applied.messages };
     }
     catch (error) {
         recordCompressionError(runtime, ctx, error);
-        return undefined;
+        return bypassUncompressed();
     }
     finally {
         runtime.state.processing = false;
     }
 }
+function allCandidateContentSeen(state, payload) {
+    const seen = state.seenCandidateContentFingerprints;
+    if (!seen?.size)
+        return false;
+    const hashes = payload.mappings.filter((mapping) => mapping.applyTo).map((mapping) => stableHash(mapping.originalText));
+    return hashes.length > 0 && hashes.every((hash) => seen.has(hash));
+}
+function ignoreSeenCandidateContent(state, payload) {
+    const seen = state.seenCandidateContentFingerprints;
+    if (!seen?.size)
+        return;
+    for (const mapping of payload.mappings) {
+        if (mapping.applyTo && seen.has(stableHash(mapping.originalText)))
+            mapping.applyTo = null;
+    }
+}
+function recordSeenCandidateContent(state, payload, appliedMessages) {
+    state.seenCandidateContentFingerprints ??= new Set();
+    state.seenCandidateContentOrder ??= [];
+    payload.mappings.forEach((mapping) => {
+        if (!mapping.applyTo)
+            return;
+        addSeenCandidateHash(state, stableHash(mapping.originalText));
+        const applied = convertMessage(appliedMessages[mapping.sourceIndex]);
+        if (applied)
+            addSeenCandidateHash(state, stableHash(extractOpenAIText(applied)));
+    });
+}
+function addSeenCandidateHash(state, hash) {
+    state.seenCandidateContentFingerprints ??= new Set();
+    state.seenCandidateContentOrder ??= [];
+    if (state.seenCandidateContentFingerprints.has(hash))
+        return;
+    state.seenCandidateContentFingerprints.add(hash);
+    state.seenCandidateContentOrder.push(hash);
+    while (state.seenCandidateContentOrder.length > MAX_SEEN_CANDIDATE_CONTENT_FINGERPRINTS) {
+        const oldest = state.seenCandidateContentOrder.shift();
+        if (oldest)
+            state.seenCandidateContentFingerprints.delete(oldest);
+    }
+}
 function generateCandidateFingerprint(messages, payload) {
     const units = payload.mappings
         .filter((mapping) => mapping.applyTo)
@@ -349,7 +411,7 @@ function recordAppliedCompression(stats, result, appliedMessages) {
 function summarizeTelemetry(state) {
     if (state.stats.last) {
         const pct = Math.round((1 - state.stats.last.compressionRatio) * 100);
-        return `compressed ${state.stats.last.tokensBefore.toLocaleString()} to ${state.stats.last.tokensAfter.toLocaleString()} tokens; saved ${state.stats.last.tokensSaved.toLocaleString()} tokens across ${state.stats.last.appliedMessages} tool results (-${pct}%)`;
+        return `last compression estimated about ${state.stats.last.tokensSaved.toLocaleString()} tokens saved (-${pct}%) across ${state.stats.last.appliedMessages} tool results; local session estimate about ${state.stats.tokensSaved.toLocaleString()} tokens saved`;
     }
     if (!state.enabled)
         return "headroom loaded but disabled for this session";
@@ -401,7 +463,7 @@ function announceAppliedCompression(pi, ctx, result, appliedMessages, mode = "no
     if (mode === "quiet" || mode === "silent")
         return;
     const pct = Math.round((1 - result.compressionRatio) * 100);
-    const summary = `compressed ${result.tokensBefore.toLocaleString()} → ${result.tokensAfter.toLocaleString()} tokens (-${pct}%, saved ${result.tokensSaved.toLocaleString()}, messages ${appliedMessages})`;
+    const summary = `compressed ${appliedMessages} tool results; estimated ~${result.tokensSaved.toLocaleString()} tokens saved (-${pct}%)`;
     const line = `noheadroom: ${summary}`;
     ctx.ui.notify(line, "info");
     // Note: we no longer use pi.appendEntry or pi.sendMessage here because
@@ -519,10 +581,7 @@ function renderFooterStatus(ctx, config, state) {
         return paint("dim", "○ Headroom not running");
     if (state.proxyOnline === null && !state.stats.last)
         return paint("dim", "○ Headroom idle");
-    if (!state.stats.last)
-        return paint("success", "✓") + paint("dim", " Headroom");
-    const pct = Math.round((1 - state.stats.last.compressionRatio) * 100);
-    return (paint("success", "✓") + paint("dim", ` Headroom -${pct}% (${state.stats.last.tokensSaved.toLocaleString()} saved)`));
+    return paint("success", "✓") + paint("dim", " Headroom active");
 }
 async function showProxyStats(ctx, client, config) {
     if (isRemoteBlocked(config)) {
@@ -554,7 +613,8 @@ function renderStatus(config, state) {
         `  Attempts:     ${stats.attempts}`,
         `  Applied:      ${stats.applied}`,
         `  Guard skips:  ${stats.guardSkips}`,
-        `  Tokens saved: ${stats.tokensSaved.toLocaleString()}`,
+        `  Estimated tokens saved: ~${stats.tokensSaved.toLocaleString()}`,
+        `  Note: local estimate; proxy /stats does not track Pi extension /v1/compress calls.`,
     ];
     if (stats.last) {
         const pct = Math.round((1 - stats.last.compressionRatio) * 100);

package/dist/types.d.ts

@@ -96,6 +96,10 @@ export type ApplyCompressionResult = {
     ok: true;
     messages: AgentMessage[];
     appliedMessages: number;
+    appliedTokensBefore: number;
+    appliedTokensAfter: number;
+    appliedTokensSaved: number;
+    appliedCompressionRatio: number;
 } | {
     ok: false;
     reason: string;

package/extensions/bridge.ts

@@ -61,6 +61,8 @@ export function applyCompressionResult(
 
 	const nextMessages = structuredClone(originalMessages) as AgentMessage[];
 	let appliedMessages = 0;
+	let appliedTokensBefore = 0;
+	let appliedTokensAfter = 0;
 
 	for (let index = 0; index < mappings.length; index++) {
 		const mapping = mappings[index];
@@ -84,14 +86,28 @@ export function applyCompressionResult(
 		if (!hasContent(target) || !replaceTextContent(target, nextText)) {
 			return { ok: false, reason: "target-content-unreplaceable" };
 		}
+		appliedTokensBefore += estimateTokens(mapping.originalText);
+		appliedTokensAfter += estimateTokens(nextText);
 		appliedMessages++;
 	}
 
 	if (appliedMessages === 0) {
 		return { ok: false, reason: "no-applicable-message-changed" };
 	}
+	const appliedTokensSaved = Math.max(0, appliedTokensBefore - appliedTokensAfter);
+	if (appliedTokensSaved === 0) {
+		return { ok: false, reason: "no-estimated-token-savings" };
+	}
 
-	return { ok: true, messages: nextMessages, appliedMessages };
+	return {
+		ok: true,
+		messages: nextMessages,
+		appliedMessages,
+		appliedTokensBefore,
+		appliedTokensAfter,
+		appliedTokensSaved,
+		appliedCompressionRatio: appliedTokensAfter / appliedTokensBefore,
+	};
 }
 
 export function convertMessage(message: AnyMessage): OpenAIMessage | undefined {
@@ -267,6 +283,11 @@ function isRecord(value: unknown): value is Record<string, unknown> {
 	return typeof value === "object" && value !== null;
 }
 
+function estimateTokens(text: string): number {
+	// ponytail: cheap local estimate; exact tokenizer would add a heavy dependency for a footer number.
+	return text.length === 0 ? 0 : Math.max(1, Math.ceil(text.length / 4));
+}
+
 function naturalizeHeadroomMarkers(text: string): string {
 	// Translate Headroom's "[X items compressed to Y. Retrieve more: hash=...]" 
 	// into a Pi-native instruction to use offset/limit instead of hallucinating a retrieve command.

package/extensions/index.ts

@@ -16,6 +16,10 @@ const STATUS_KEY = "headroom";
 const SUBCOMMANDS = ["status", "on", "off", "health", "stats", "mode"] as const;
 const MODES = ["normal", "quiet", "silent"] as const;
 const NOTRACE_TELEMETRY_CHANNEL = "notrace.telemetry.extension";
+// Bounded FIFO, not config: this is a safety valve for duplicate-compression loops.
+// 512 hashes covers roughly 256 compressed candidates because we store original + compressed text hashes.
+// That is far beyond the reread-loop case while keeping memory bounded for long sessions.
+const MAX_SEEN_CANDIDATE_CONTENT_FINGERPRINTS = 512;
 
 type Subcommand = (typeof SUBCOMMANDS)[number];
 
@@ -30,6 +34,8 @@ interface HeadroomRuntimeState {
 	lastInputFingerprint: string | null;
 	lastOutputFingerprint: string | null;
 	lastGuardSkipCandidateFingerprint: string | null;
+	seenCandidateContentFingerprints?: Set<string>;
+	seenCandidateContentOrder?: string[];
 	lastCompressionTime: number;
 	stats: HeadroomStats;
 }
@@ -127,6 +133,8 @@ function createRuntime(pi: ExtensionAPI): HeadroomRuntime {
 		lastInputFingerprint: null,
 		lastOutputFingerprint: null,
 		lastGuardSkipCandidateFingerprint: null,
+		seenCandidateContentFingerprints: new Set(),
+		seenCandidateContentOrder: [],
 		lastCompressionTime: 0,
 		stats: { attempts: 0, applied: 0, guardSkips: 0, tokensSaved: 0 },
 	};
@@ -262,10 +270,18 @@ async function handleContextCompression(
 ): Promise<{ messages?: AgentMessage[] } | undefined> {
 	if (runtime.state.processing) return undefined;
 
+	const bypassUncompressed = () => {
+		if (runtime.state.stats.last) {
+			runtime.state.stats.last = undefined;
+			runtime.refreshStatus(ctx);
+		}
+		return undefined;
+	};
+
 	// Throttle: max 1 compression attempt per 3 seconds to kill infinite loops
 	const now = Date.now();
 	if (now - runtime.state.lastCompressionTime < 3000) {
-		return undefined;
+		return bypassUncompressed();
 	}
 
 	// Content-based guards to prevent infinite recursion
@@ -278,22 +294,26 @@ async function handleContextCompression(
 
 	// 2. If this input matches our previous input, it didn't change. Stop.
 	if (runtime.state.lastInputFingerprint === inputFingerprint) {
-		return undefined;
+		return bypassUncompressed();
 	}
 
-	if (shouldSkipBeforePayload(runtime, ctx)) return undefined;
+	if (shouldSkipBeforePayload(runtime, ctx)) return bypassUncompressed();
 	const payload = buildCompressionPayload(event.messages, runtime.config.minMessageChars);
-	if (payload.candidateCount === 0) return undefined;
+	if (payload.candidateCount === 0) return bypassUncompressed();
 
 	// 3. If eligible candidates haven't changed since the last skip/no-savings result,
 	// don't spend another proxy call just because surrounding conversation changed.
 	const candidateFingerprint = generateCandidateFingerprint(event.messages, payload);
 	if (runtime.state.lastGuardSkipCandidateFingerprint === candidateFingerprint) {
-		return undefined;
+		return bypassUncompressed();
+	}
+	if (allCandidateContentSeen(runtime.state, payload)) {
+		return bypassUncompressed();
 	}
+	ignoreSeenCandidateContent(runtime.state, payload);
 	if (runtime.state.proxyOnline !== true) {
 		void ensureProxyInBackground(runtime, ctx);
-		return undefined;
+		return bypassUncompressed();
 	}
 
 	runtime.state.processing = true;
@@ -307,10 +327,7 @@ async function handleContextCompression(
 			// until the actual compressible candidate material changes.
 			runtime.state.lastInputFingerprint = inputFingerprint;
 			runtime.state.lastGuardSkipCandidateFingerprint = candidateFingerprint;
-			// ponytail: clear stats.last so renderFooterStatus shows idle instead of stale savings
-			runtime.state.stats.last = undefined;
-			runtime.refreshStatus(ctx);
-			return undefined;
+			return bypassUncompressed();
 		}
 
 		const applied = applyCompressionResult(event.messages, payload.mappings, result.messages, {
@@ -325,28 +342,77 @@ async function handleContextCompression(
 			recordGuardSkip(runtime.state.stats, applied.reason);
 			emitNotraceTelemetry(runtime);
 			announceGuardSkip(ctx, applied.reason, result, runtime.config.mode);
-			runtime.refreshStatus(ctx);
-			return undefined;
+			return bypassUncompressed();
 		}
 
 		// Store fingerprints to break the feedback loop
 		runtime.state.lastInputFingerprint = inputFingerprint;
 		runtime.state.lastOutputFingerprint = generateFingerprint(applied.messages);
 		runtime.state.lastGuardSkipCandidateFingerprint = null;
-
-		recordAppliedCompression(runtime.state.stats, result, applied.appliedMessages);
+		recordSeenCandidateContent(runtime.state, payload, applied.messages);
+		const appliedResult = {
+			...result,
+			tokensBefore: applied.appliedTokensBefore,
+			tokensAfter: applied.appliedTokensAfter,
+			tokensSaved: applied.appliedTokensSaved,
+			compressionRatio: applied.appliedCompressionRatio,
+		};
+
+		recordAppliedCompression(runtime.state.stats, appliedResult, applied.appliedMessages);
 		emitNotraceTelemetry(runtime);
-		announceAppliedCompression(runtime.pi, ctx, result, applied.appliedMessages, runtime.config.mode);
+		announceAppliedCompression(runtime.pi, ctx, appliedResult, applied.appliedMessages, runtime.config.mode);
 		runtime.refreshStatus(ctx);
 		return { messages: applied.messages };
 	} catch (error) {
 		recordCompressionError(runtime, ctx, error);
-		return undefined;
+		return bypassUncompressed();
 	} finally {
 		runtime.state.processing = false;
 	}
 }
 
+function allCandidateContentSeen(state: HeadroomRuntimeState, payload: CompressionPayload): boolean {
+	const seen = state.seenCandidateContentFingerprints;
+	if (!seen?.size) return false;
+	const hashes = payload.mappings.filter((mapping) => mapping.applyTo).map((mapping) => stableHash(mapping.originalText));
+	return hashes.length > 0 && hashes.every((hash) => seen.has(hash));
+}
+
+function ignoreSeenCandidateContent(state: HeadroomRuntimeState, payload: CompressionPayload): void {
+	const seen = state.seenCandidateContentFingerprints;
+	if (!seen?.size) return;
+	for (const mapping of payload.mappings) {
+		if (mapping.applyTo && seen.has(stableHash(mapping.originalText))) mapping.applyTo = null;
+	}
+}
+
+function recordSeenCandidateContent(
+	state: HeadroomRuntimeState,
+	payload: CompressionPayload,
+	appliedMessages: AgentMessage[],
+): void {
+	state.seenCandidateContentFingerprints ??= new Set();
+	state.seenCandidateContentOrder ??= [];
+	payload.mappings.forEach((mapping) => {
+		if (!mapping.applyTo) return;
+		addSeenCandidateHash(state, stableHash(mapping.originalText));
+		const applied = convertMessage(appliedMessages[mapping.sourceIndex]);
+		if (applied) addSeenCandidateHash(state, stableHash(extractOpenAIText(applied)));
+	});
+}
+
+function addSeenCandidateHash(state: HeadroomRuntimeState, hash: string): void {
+	state.seenCandidateContentFingerprints ??= new Set();
+	state.seenCandidateContentOrder ??= [];
+	if (state.seenCandidateContentFingerprints.has(hash)) return;
+	state.seenCandidateContentFingerprints.add(hash);
+	state.seenCandidateContentOrder.push(hash);
+	while (state.seenCandidateContentOrder.length > MAX_SEEN_CANDIDATE_CONTENT_FINGERPRINTS) {
+		const oldest = state.seenCandidateContentOrder.shift();
+		if (oldest) state.seenCandidateContentFingerprints.delete(oldest);
+	}
+}
+
 function generateCandidateFingerprint(messages: AgentMessage[], payload: CompressionPayload): string {
 	const units = payload.mappings
 		.filter((mapping) => mapping.applyTo)
@@ -426,7 +492,7 @@ function recordAppliedCompression(stats: HeadroomStats, result: CompressResult,
 function summarizeTelemetry(state: HeadroomRuntimeState): string | null {
 	if (state.stats.last) {
 		const pct = Math.round((1 - state.stats.last.compressionRatio) * 100);
-		return `compressed ${state.stats.last.tokensBefore.toLocaleString()} to ${state.stats.last.tokensAfter.toLocaleString()} tokens; saved ${state.stats.last.tokensSaved.toLocaleString()} tokens across ${state.stats.last.appliedMessages} tool results (-${pct}%)`;
+		return `last compression estimated about ${state.stats.last.tokensSaved.toLocaleString()} tokens saved (-${pct}%) across ${state.stats.last.appliedMessages} tool results; local session estimate about ${state.stats.tokensSaved.toLocaleString()} tokens saved`;
 	}
 	if (!state.enabled) return "headroom loaded but disabled for this session";
 	if (state.stats.lastSkipReason) return `compression not applied; last guard skip: ${state.stats.lastSkipReason}`;
@@ -478,7 +544,7 @@ function announceAppliedCompression(
 ): void {
 	if (mode === "quiet" || mode === "silent") return;
 	const pct = Math.round((1 - result.compressionRatio) * 100);
-	const summary = `compressed ${result.tokensBefore.toLocaleString()} → ${result.tokensAfter.toLocaleString()} tokens (-${pct}%, saved ${result.tokensSaved.toLocaleString()}, messages ${appliedMessages})`;
+	const summary = `compressed ${appliedMessages} tool results; estimated ~${result.tokensSaved.toLocaleString()} tokens saved (-${pct}%)`;
 	const line = `noheadroom: ${summary}`;
 	ctx.ui.notify(line, "info");
 	// Note: we no longer use pi.appendEntry or pi.sendMessage here because
@@ -615,12 +681,7 @@ function renderFooterStatus(ctx: ExtensionContext, config: HeadroomConfig, state
 	if (state.proxyStarting) return paint("dim", "⏳ Headroom starting");
 	if (state.proxyOnline === false) return paint("dim", "○ Headroom not running");
 	if (state.proxyOnline === null && !state.stats.last) return paint("dim", "○ Headroom idle");
-	if (!state.stats.last) return paint("success", "✓") + paint("dim", " Headroom");
-
-	const pct = Math.round((1 - state.stats.last.compressionRatio) * 100);
-	return (
-		paint("success", "✓") + paint("dim", ` Headroom -${pct}% (${state.stats.last.tokensSaved.toLocaleString()} saved)`)
-	);
+	return paint("success", "✓") + paint("dim", " Headroom active");
 }
 
 async function showProxyStats(
@@ -660,7 +721,8 @@ function renderStatus(config: HeadroomConfig, state: HeadroomRuntimeState): stri
 		`  Attempts:     ${stats.attempts}`,
 		`  Applied:      ${stats.applied}`,
 		`  Guard skips:  ${stats.guardSkips}`,
-		`  Tokens saved: ${stats.tokensSaved.toLocaleString()}`,
+		`  Estimated tokens saved: ~${stats.tokensSaved.toLocaleString()}`,
+		`  Note: local estimate; proxy /stats does not track Pi extension /v1/compress calls.`,
 	];
 	if (stats.last) {
 		const pct = Math.round((1 - stats.last.compressionRatio) * 100);

package/extensions/types.ts

@@ -108,5 +108,13 @@ export interface ApplyCompressionOptions {
 }
 
 export type ApplyCompressionResult =
-	| { ok: true; messages: AgentMessage[]; appliedMessages: number }
+	| {
+			ok: true;
+			messages: AgentMessage[];
+			appliedMessages: number;
+			appliedTokensBefore: number;
+			appliedTokensAfter: number;
+			appliedTokensSaved: number;
+			appliedCompressionRatio: number;
+	  }
 	| { ok: false; reason: string };

package/package.json

@@ -1,11 +1,12 @@
 {
   "name": "@raquezha/noheadroom",
-  "version": "0.3.0",
+  "version": "0.3.2",
   "description": "Personal Headroom compression bridge for Pi with Docker-first defaults and Pi tool-name adaptation",
   "main": "dist/index.js",
   "type": "module",
   "scripts": {
     "build": "tsc",
+    "bench:loop": "node bench-loop-hardening.mjs",
     "prepare": "npm run build"
   },
   "peerDependencies": {