pi-mega-compact 0.21.6 → 0.21.8

package/README.md

@@ -1,9 +1,7 @@
 # pi-mega-compact
 
 > **⚠️ LTS — patches only (2026-08-13).** This extension is maintained for bug fixes only. New feature development has moved to **[radcode](https://github.com/TheArchitectit/radcode)**, a Rust pi.dev replacement that has ported mega-compact's compaction/recall/dedup/RAPTOR stack. See [`docs/LTS.md`](docs/LTS.md) and [`docs/SUCCESSION.md`](docs/SUCCESSION.md).
-> 
-Pending public release
-> 
+
 A local-first context compressor for the [pi coding agent](https://github.com/earendil-works/pi). Keeps long sessions running without overflowing the context window. Local by default — no cloud, no API calls, no telemetry. Bring your own localhost embedder (Ollama, ONNX, TEI) for better semantic matches, or opt in to a remote endpoint if you need to.
 
 ## Features

package/dist/extensions/dashboard-server/routes-rag-settings-helpers.js

@@ -132,6 +132,7 @@ export const SETTINGS = [
         settings: [
             num("MEGACOMPACT_THRESHOLD_PCT", "Compaction Threshold", "Fraction of the actual model context window at which compaction fires — 0.80 fires at 80% used (leaves 20% free). Applies to any model size; a per-model Model Thresholds row overrides it", 0.8, 0.1, 0.95),
             num("MEGACOMPACT_THRASH_REARM_PCT", "Thrash Re-arm %", "After an ineffective compaction (live window did not shrink), refuse to re-fire until the live window grows by this fraction of the effective threshold. Default 0.10 (10%)", 0.1, 0.01, 0.5),
+            boolDirect("MEGACOMPACT_OUTPUT_ERROR_COMPACT", "Output-Error Compact", "When a model response is truncated mid-output (stopReason: 'length'), trip a one-shot forced compaction to free input headroom. Closes the small-context deadlock where the model truncates below the input threshold.", true),
         ],
     },
     {
@@ -185,6 +186,8 @@ export const EXCLUDED_SETTINGS = [
     "MEGACOMPACT_RAPTOR_LEVEL_WEIGHTS",
     // Internal calibration, not user-facing.
     "MEGACOMPACT_FTS5_MAX_BM25",
+    // Boot-fallback context window — display-seed only, never a firing gate.
+    "MEGACOMPACT_DEFAULT_CONTEXT_WINDOW",
     // COS-FP-A: per-content-type L2 override landing slots. Null/unset by
     // default and NOT wired into the live L2 decision this sprint — the
     // top-level MEGACOMPACT_L2_THRESHOLD is the single runtime firing point. A

package/dist/extensions/mega-events/context-handler.js

@@ -87,7 +87,21 @@ export function registerContextHandler(pi, runtime, config) {
             /* non-fatal */
         }
         runtime.lastCtxPercent = pct ?? null;
-        runtime.lastCtxWindow = usage?.contextWindow ?? 0;
+        // Resolve the model window used by the token gate + live-trim tail-cap.
+        // Prefer the provider-reported usage window (authoritative when present);
+        // fall back to the captured model snapshot's contextWindow (populated from
+        // models.json / pi's model registry) when the provider does not report it
+        // via getContextUsage(). plexus (OpenAI-compatible) omits contextWindow in
+        // usage, so without this fallback lastCtxWindow is 0 for those providers —
+        // which silently disables the live-trim tail-cap (guarded on ctxWindow>0)
+        // and the token-gate window math, so mega-compact never reserves output
+        // headroom and a 32k model's own output overflows the window each turn.
+        // Mirrors the gate's existing pct fallback (gateCheck.ts S27).
+        const reportedWindow = usage?.contextWindow ?? 0;
+        runtime.lastCtxWindow =
+            reportedWindow > 0
+                ? reportedWindow
+                : (runtime.currentModel?.contextWindow ?? 0);
         runtime.snapshot(ctx);
         if (!config.auto) {
             const tailed = tailResult();
@@ -99,23 +113,6 @@ export function registerContextHandler(pi, runtime, config) {
         // S27 DB-mirror + VC1B ledger append. Runs BEFORE the fast-gate so every
         // message is captured, even if we don't compact this turn. Non-fatal.
         appendMirrorAndLedger(runtime, config, messages);
-        // Phase H.1: output-error catch noop guard. When forceCompactNextGate is
-        // armed (Phase H: model truncated its output → lengthStop.ts set the flag
-        // at turn_end), check piCompactWouldNoop BEFORE evaluateGate consumes the
-        // flag. If pi would noop (session too small to summarize — the transcript
-        // is under pi's keepRecentTokens floor), clear the flag + return the tail
-        // result silently instead of forcing a proceed that will throw "Nothing
-        // to compact (session too small)" to the user. The session is genuinely
-        // too small to compact — skipping is correct (the model retries on the
-        // next turn regardless). Gated by config.outputErrorCompact (default ON;
-        // OFF = byte-identical pre-H.1).
-        if (config.outputErrorCompact &&
-            runtime.rt.forceCompactNextGate &&
-            piCompactWouldNoop(ctx)) {
-            runtime.rt.forceCompactNextGate = false;
-            runtime.diagCtxOutputErrorTrip++;
-            return tailResult() ?? undefined;
-        }
         // S29 FAST GATE: drive the auto-trigger off the context percent (see
         // gateCheck.ts). Returns a tailed view when the gate does not pass.
         const gate = evaluateGate(runtime, config, { pct, currentTokens, tailResult });

package/extensions/dashboard-server/routes-rag-settings-helpers.ts

@@ -294,6 +294,12 @@ export const SETTINGS: ReadonlyArray<SettingGroup> = [
 				0.01,
 				0.5,
 			),
+			boolDirect(
+				"MEGACOMPACT_OUTPUT_ERROR_COMPACT",
+				"Output-Error Compact",
+				"When a model response is truncated mid-output (stopReason: 'length'), trip a one-shot forced compaction to free input headroom. Closes the small-context deadlock where the model truncates below the input threshold.",
+				true,
+			),
 		],
 	},
 	{
@@ -376,6 +382,8 @@ export const EXCLUDED_SETTINGS: readonly string[] = [
 	"MEGACOMPACT_RAPTOR_LEVEL_WEIGHTS",
 	// Internal calibration, not user-facing.
 	"MEGACOMPACT_FTS5_MAX_BM25",
+	// Boot-fallback context window — display-seed only, never a firing gate.
+	"MEGACOMPACT_DEFAULT_CONTEXT_WINDOW",
 	// COS-FP-A: per-content-type L2 override landing slots. Null/unset by
 	// default and NOT wired into the live L2 decision this sprint — the
 	// top-level MEGACOMPACT_L2_THRESHOLD is the single runtime firing point. A

package/extensions/mega-events/context-handler.ts

@@ -116,7 +116,21 @@ export function registerContextHandler(
 			/* non-fatal */
 		}
 		runtime.lastCtxPercent = pct ?? null;
-		runtime.lastCtxWindow = usage?.contextWindow ?? 0;
+		// Resolve the model window used by the token gate + live-trim tail-cap.
+		// Prefer the provider-reported usage window (authoritative when present);
+		// fall back to the captured model snapshot's contextWindow (populated from
+		// models.json / pi's model registry) when the provider does not report it
+		// via getContextUsage(). plexus (OpenAI-compatible) omits contextWindow in
+		// usage, so without this fallback lastCtxWindow is 0 for those providers —
+		// which silently disables the live-trim tail-cap (guarded on ctxWindow>0)
+		// and the token-gate window math, so mega-compact never reserves output
+		// headroom and a 32k model's own output overflows the window each turn.
+		// Mirrors the gate's existing pct fallback (gateCheck.ts S27).
+		const reportedWindow = usage?.contextWindow ?? 0;
+		runtime.lastCtxWindow =
+			reportedWindow > 0
+				? reportedWindow
+				: (runtime.currentModel?.contextWindow ?? 0);
 		runtime.snapshot(ctx);
 		if (!config.auto) {
 			const tailed = tailResult();
@@ -130,26 +144,6 @@ export function registerContextHandler(
 		// message is captured, even if we don't compact this turn. Non-fatal.
 		appendMirrorAndLedger(runtime, config, messages);
 
-		// Phase H.1: output-error catch noop guard. When forceCompactNextGate is
-		// armed (Phase H: model truncated its output → lengthStop.ts set the flag
-		// at turn_end), check piCompactWouldNoop BEFORE evaluateGate consumes the
-		// flag. If pi would noop (session too small to summarize — the transcript
-		// is under pi's keepRecentTokens floor), clear the flag + return the tail
-		// result silently instead of forcing a proceed that will throw "Nothing
-		// to compact (session too small)" to the user. The session is genuinely
-		// too small to compact — skipping is correct (the model retries on the
-		// next turn regardless). Gated by config.outputErrorCompact (default ON;
-		// OFF = byte-identical pre-H.1).
-		if (
-			config.outputErrorCompact &&
-			runtime.rt.forceCompactNextGate &&
-			piCompactWouldNoop(ctx)
-		) {
-			runtime.rt.forceCompactNextGate = false;
-			runtime.diagCtxOutputErrorTrip++;
-			return tailResult() ?? undefined;
-		}
-
 		// S29 FAST GATE: drive the auto-trigger off the context percent (see
 		// gateCheck.ts). Returns a tailed view when the gate does not pass.
 		const gate = evaluateGate(runtime, config, { pct, currentTokens, tailResult });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-mega-compact",
-  "version": "0.21.6",
+  "version": "0.21.8",
   "description": "Layered, local, vector-backed context compressor for pi — supersede/collapse/cluster compaction with deduped inline recall.",
   "type": "module",
   "license": "BSD-3-Clause",