pi-mega-compact 0.20.30 → 0.20.32

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

@@ -50,8 +50,8 @@ export const SETTINGS = [
             boolFlag("MEGACOMPACT_MEMORY_GRAPH", "Memory Graph", "Dashboard-oriented memory graph traversal", true),
             boolFlag("MEGACOMPACT_HYDE", "HyDE", "Generate hypothetical answer via LLM, embed it, RRF-fuse", true, true),
             boolFlag("MEGACOMPACT_NEW_UI", "New Dashboard UI", "Tailwind + shadcn visual design (sidebar, glass panels)", true),
-            boolDirect("MEGACOMPACT_MESSAGE_SEPARATION", "Message Separation (P2)", "PLAN_V2: split conversation thread from tool results to grow stable cache prefix", false),
-            boolDirect("MEGACOMPACT_CACHE_STRIPING", "Cache Striping (P3)", "PLAN_V2: order stable context by stability score so durable chunks lead the prompt", false),
+            boolDirect("MEGACOMPACT_MESSAGE_SEPARATION", "Message Separation (P2)", "PLAN_V2: split conversation thread from tool results to grow stable cache prefix", true),
+            boolDirect("MEGACOMPACT_CACHE_STRIPING", "Cache Striping (P3)", "PLAN_V2: order stable context by stability score so durable chunks lead the prompt", true),
         ],
     },
     {

package/dist/extensions/mega-config-types.js

@@ -0,0 +1 @@
+export {};

package/dist/extensions/mega-config.js

@@ -167,8 +167,11 @@ export function loadConfig() {
         recallMaxTokens: envFlag("MEGACOMPACT_RECALL_MAX_TOKENS", 1500),
         windowDedupe: envBool("MEGACOMPACT_WINDOW_DEDUPE", true),
         recallTailInject: envBool("MEGACOMPACT_RECALL_TAIL_INJECT", true),
-        messageSeparation: envBool("MEGACOMPACT_MESSAGE_SEPARATION", false),
-        cacheStriping: envBool("MEGACOMPACT_CACHE_STRIPING", false),
+        // PC-A: positive sprint flag, default ON. =0 byte-identical to the
+        // pre-change OFF state (single gate lives at the call site in tailResult.ts).
+        messageSeparation: envBool("MEGACOMPACT_MESSAGE_SEPARATION", true),
+        // positive sprint flag: default ON, =0 byte-identical to OFF
+        cacheStriping: envBool("MEGACOMPACT_CACHE_STRIPING", true),
         tuiWidget: envBool("MEGACOMPACT_TUI_WIDGET", true),
         ragQueryReformulation: envBool("MEGACOMPACT_QUERY_REFORMULATION", false),
         ragTieredRouter: envBool("MEGACOMPACT_TIERED_ROUTER", false),

package/dist/extensions/mega-events/separated-prompt.js

@@ -12,8 +12,11 @@
  * 3 (thread: user/assistant turns) -> 4 (tool results at tail).
  *
  * Feature-gated by MEGACOMPACT_CACHE_STRIPING (default OFF).
- * MEGACOMPACT_MESSAGE_SEPARATION must also be ON for any layering to occur.
  * Flag-OFF = byte-identical to pre-sprint — returns messages unchanged.
+ *
+ * PC-A: buildSeparatedPrompt is PURE — the MEGACOMPACT_MESSAGE_SEPARATION gate
+ * lives at the single call site (tailResult.ts, config.messageSeparation),
+ * not inside this function.
  */
 import { openStore } from "../../src/store/sqlite/utils.js";
 import { getStateDir } from "../../src/store.js";
@@ -29,14 +32,11 @@ const TOPIC_SHIFT_THRESHOLD = 0.7;
  * Layer order: 0 (system) -> 1 (summary) -> 3 (thread: user/assistant turns) ->
  * 4 (tool results at tail).
  *
- * Feature gate: MEGACOMPACT_MESSAGE_SEPARATION env var.  When OFF (default),
- * returns `messages` unchanged (byte-identical to pre-sprint).
+ * PURE function: the MEGACOMPACT_MESSAGE_SEPARATION gate moved to the single
+ * call site (tailResult.ts, config.messageSeparation) — this never reads env.
+ * When there is nothing to reorder, returns `messages` unchanged (byte-identical).
  */
 export function buildSeparatedPrompt(messages, _opts) {
-    const flag = process.env.MEGACOMPACT_MESSAGE_SEPARATION;
-    if (flag === "0" || flag === "false" || flag === undefined || flag === "") {
-        return messages;
-    }
     // pi's AgentMessage union has no "system" role — the system prompt lives in
     // AgentState.systemPrompt, separate from this array. The cache-relevant,
     // low-risk transformation is moving volatile tool results/executions to the
@@ -169,18 +169,12 @@ export function refreshStripeAssignments(stateDir, epochId, limit = DEFAULT_STRI
  * Build a cache-optimized prompt with 5 layers:
  *   0 (system) -> 1 (summary) -> 2 (cache stripes) -> 3 (thread) -> 4 (tool)
  *
- * Feature gate: MEGACOMPACT_CACHE_STRIPING (default OFF). When OFF, delegates
- * to buildSeparatedPrompt (byte-identical to Phase 2 behavior).
- *
- * When ON but MEGACOMPACT_MESSAGE_SEPARATION is OFF, also delegates to
- * buildSeparatedPrompt (which returns messages unchanged).
+ * Positive sprint flag driven by config.cacheStriping at the call site
+ * (tailResult.ts); this function is pure and never reads process.env.
+ * With flag ON but no stripe rows for the epoch, it returns the base
+ * separated prompt unchanged (byte-identical to buildSeparatedPrompt).
  */
 export function buildCacheOptimizedPrompt(messages, opts) {
-    const flag = process.env.MEGACOMPACT_CACHE_STRIPING;
-    if (flag === "0" || flag === "false" || flag === undefined || flag === "") {
-        // Flag OFF: delegate to buildSeparatedPrompt (byte-identical).
-        return buildSeparatedPrompt(messages, opts);
-    }
     // Build the base 4-layer structure first.
     const base = buildSeparatedPrompt(messages, opts);
     // If base equals messages, separation is OFF — return unchanged.

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

@@ -126,13 +126,13 @@ export const SETTINGS: ReadonlyArray<SettingGroup> = [
 				"MEGACOMPACT_MESSAGE_SEPARATION",
 				"Message Separation (P2)",
 				"PLAN_V2: split conversation thread from tool results to grow stable cache prefix",
-				false,
+				true,
 			),
 			boolDirect(
 				"MEGACOMPACT_CACHE_STRIPING",
 				"Cache Striping (P3)",
 				"PLAN_V2: order stable context by stability score so durable chunks lead the prompt",
-				false,
+				true,
 			),
 		],
 	},

package/extensions/mega-config-types.ts

@@ -0,0 +1,181 @@
+/**
+ * mega-config-types.ts — MegaConfig type-only declarations for the extension.
+ *
+ * Split out of mega-config.ts (delegate-shell) so the runtime barrel stays under
+ * the extensions/ soft line limit. Type-only file: no runtime code, no logic.
+ * `import type` keeps the only cross-reference (CompactTier) erased at build,
+ * so there is no runtime import cycle.
+ */
+import type { CompactTier } from "./mega-config.js";
+
+/**
+ * Resolved, frozen-at-load config. `tier` is the base compaction PRESET chosen
+ * by env (low/medium/high/ultra/mega) — it sets the threshold token budget and
+ * is NOT changed at runtime (the /mega-tier command was removed in S24). The
+ * *displayed* tier the user sees in the toolbar/dashboard is the LIVE pressure
+ * band (see MegaRuntime.pressureBand), which climbs low→mega as context fills.
+ */
+export interface MegaConfig {
+	tier: CompactTier | "custom";
+	/**
+	 * Compaction threshold as a fraction of the model context window (e.g. 0.70
+	 * for "high"). null for `custom` (explicit MEGACOMPACT_THRESHOLD_TOKENS, which
+	 * stays an ABSOLUTE token count, never percent-scaled).
+	 */
+	tierPct: number | null;
+	thresholdTokens: number;
+	stateDir: string;
+	fastGatePct: number;
+	anchorUserMessages: number;
+	preserveRecent: number;
+	/** High-pressure floor for preserveRecent — when context is near the limit
+	 *  we compact deeper, but never below this (keeps recent turns for coherence). */
+	preserveRecentMin: number;
+	/** D.1: minimum context-growth percentage delta before re-compacting instead of
+	 *  replaying the cached trim (Default 50). Set via MEGACOMPACT_RECOMPACT_PCT_DELTA.*/
+	recompactPctDelta: number;
+	auto: boolean;
+	autoInline: boolean;
+	autoInlineK: number;
+	/** S28: auto-continue the agent after a max-output-token length stop by
+	 *  reusing the existing S16 resume-nudge. Default true. Off = silent (the
+	 *  prior behavior). PREVENT-PI-003: restart via user-role sendUserMessage. */
+	autoContinueLengthStop: boolean;
+	/** S38: max retries for transient errors (5xx/429/network/max-output-token
+	 *  text that is NOT a length stopReason — S28 owns those). Default 5.
+	 *  `0` disables all transient retries (reverts to S28-only). */
+	autoRetryTransientMax: number;
+	/** S38: max retries for permanent errors (auth/config/malformed). Default 1.
+	 *  `0` disables permanent-error retries. */
+	autoRetryPermanentMax: number;
+	/** S38.5: strict race-guard — 30s cooldown + deferred ctx.compact() re-check
+	 *  (closes the first-race-in-burst window). Default true. `false` reverts to
+	 *  the v0.7.4 synchronous 10s-cooldown behavior. */
+	raceGuardStrict: boolean;
+	/** S38.6: max consecutive errors before circuit-breaker trips (stops retrying).
+	 *  Default 10. When `errorRetryCount` exceeds this across multiple turns,
+	 *  the extension stops retrying until a successful turn resets it. */
+	maxConsecutiveErrors: number;
+	/** S38.7: hard-stop switch — when true, ALL error retries are disabled.
+	 *  Default false. Set via env to force S28-only behavior (length-stop continues
+	 *  only). */
+	errorRetryHardStop: boolean;
+	/** R1 (retry redesign): base unit (ms) for errorRetryBackoffMs(count) pacing.
+	 *  The schedule is base, 2*base, 4*base, 6*base (cap) — so the default 5000
+	 *  yields 5s/10s/20s/30s. errorRetryUntil is now GATING (previously it was
+	 *  documented as non-gating); a nudge cannot fire before errorRetryUntil
+	 *  elapses. */
+	errorRetryBackoffMs: number;
+	/** R2: session-global cap on total S38 nudges across ALL bursts. Hitting it
+	 *  is terminal for the session — the extension stops nudging entirely,
+	 *  independent of the per-burst max and the circuit breaker. Default 3.
+	 *  `0` disables (reverts to per-burst + circuit-breaker only). */
+	errorRetrySessionMax: number;
+	/** R3: consecutive identical error-text count at which a 'transient'
+	 *  classification is upgraded to 'poisoned-context' (the stateful repeat
+	 *  signal). Default 3. Raise to make the upgrade less aggressive. */
+	poisonedContextRepeatThreshold: number;
+	/** R10: consecutive transient errors at which a calm "provider outage"
+	 *  advisory is sent to the user (distinct from the poisoned /clear advise).
+	 *  Default 3. `0` disables the advisory entirely. */
+	providerOutageAdviseThreshold: number;
+	/** R13: when true (default), poisoned-context and provider-outage advisories
+	 *  are dashboard-only (events tab + log) — no user-visible message injection.
+	 *  When false, the legacy sendUserMessage path runs (byte-identical pre-R13). */
+	advisoryChannel: boolean;
+	/** S29: override the auto-compact fire point for tiered configs, as a
+	 *  fraction of the context window (e.g. 0.85). null = inherit the tier's
+	 *  tierPct (default; preserves existing fire points). The context-handler
+	 *  gate fires on context % (reliable), not token count (under-reported),
+	 *  so it catches the overshoot that causes max-output-token truncation.
+	 *  `custom` (tierPct null) ignores this — it keeps the absolute token gate. */
+	autoPctTrigger: number | null;
+	dedupSim: number;
+	/** RAPTOR hierarchical recall enabled (Fix D). Drives both live recall and
+	 *  the durable-trim summary source (root summary). */
+	raptorEnabled: boolean;
+	/** Legacy v0.4.28 behavior: auto-trigger calls ctx.compact() (which STOPS
+	 *  the agent). Default false — the S16 redesign uses the live context-event
+	 *  trim + pi native auto-compaction instead (compact and continue). Kept for
+	 *  one release as rollback. */
+	legacyDurableTrim: boolean;
+	/** S27: durable raw-transcript DB mirror (MEGACOMPACT_DB_MIRROR). When on,
+	 *  raw message bytes + checkpoint-epoch bookkeeping are appended to the
+	 *  SQLite store so a compacted window can be rehydrated locally instead of
+	 *  from the pi runtime transcript. Default OFF — additive, no behavior
+	 *  change until flipped on. legacyDurableTrim takes precedence (the legacy
+	 *  v0.4.28 ctx.compact() path does not emit the S27 mirror hook). */
+	dbMirror: boolean;
+	/** S49: isolated per-turn store (turns.db). Default ON. OFF = legacy main-db
+	 *  turn path (S48 behavior — byte-identical). Mirrors TurnsConfig.TURNS_DB_ENABLED. */
+	turnsDbEnabled: boolean;
+	/** S51: auto-categorizing wiki (k-means + TF-IDF over real embeddings). Default ON.
+	 *  Mirrors TurnsConfig.AUTO_WIKI_ENABLED. Rebuild fires every Nth compaction. */
+	autoWikiEnabled: boolean;
+	/** Cross-repo recall enabled (S17). Resume + /mega-recall --cross-repo can
+	 *  pull checkpoints from OTHER repos via the PGlite HNSW index. Default true. */
+	crossRepoEnabled: boolean;
+	/** Stricter cosine floor for cross-repo hits (S17). Default 0.90 (trigram) /
+	 *  tighter than same-repo so only genuinely-relevant cross-repo context is
+	 *  injected. */
+	crossRepoCosine: number;
+	/** Memory-RAG auto-review enabled (S20). Every memoryReviewInterval turns the
+	 *  conversation is auto-reviewed into durable add/replace/remove memories. */
+	memoryAutoReview: boolean;
+	/** Turn cadence for the auto-review scan (S20). Default 10. */
+	memoryReviewInterval: number;
+	/** Token ceiling for the re-injected recall block (Fix C). Recall stops
+	 *  adding checkpoints once the block would exceed this — bounds read-path
+	 *  token cost so it can never net-inflate the window. */
+	recallMaxTokens: number;
+	/** Inline-dedupe recalled checkpoints against the live window (Fix C): drop
+	 *  a hit whose summary is ≥ dedupSim similar to a live message — "dedupe on
+	 *  inline/read" so we never re-inject context already resident. */
+	windowDedupe: boolean;
+	/** S53: Recall Tail Injection — inject staged recall block as a user message at
+	 *  the tail of the view when auto is OFF AND no trim action is needed. Default ON
+	 *  (true). When false, restores the pre-sprint systemPrompt prepend behavior. */
+	recallTailInject: boolean;
+	/** A1 PLAN_V2 Phase 2: Message Separation — isolate user/assistant turns
+	 *  from volatile tool results so the prompt-cache prefix stays stable.
+	 *  PC-A: positive sprint flag, now default ON; flag-OFF (=0) is byte-identical
+	 *  to the pre-change OFF state. The single gate lives at the call site
+	 *  (tailResult.ts, config.messageSeparation), not inside buildSeparatedPrompt. */
+	messageSeparation: boolean;
+	/** P3: Cache-aware striping (PLAN_V2 Phase 3). Inserts stability-ordered
+	 *  cache stripes between summaries and thread. Default OFF. */
+	cacheStriping: boolean;
+	debug: boolean;
+	/** Master reconciliation: TUI shutdown widget (MEGACOMPACT_TUI_WIDGET=0 to disable). */
+	tuiWidget: boolean;
+	/** S57 B1: Query reformulation via embedding-neighbor keyword expansion. */
+	ragQueryReformulation: boolean;
+	/** S57 B2: Tiered recall router (L0 cache -> L1 FTS5 -> L2). */
+	ragTieredRouter: boolean;
+	/** S57 B3: Recall quality metrics (precision/recall scoring + logging). */
+	ragRecallMetrics: boolean;
+	/** S57 B4: Memory graph traversal (dashboard-oriented). */
+	ragMemoryGraph: boolean;
+	/** D1: Seed initial wiki topic model from live turns when no context_chunks exist yet. */
+	wikiSeedFromTurns: boolean;
+	/** D3 Source A: Include structural turn nodes in the memory graph (metadata only, no content). */
+	memoryGraphSeedTurns: boolean;
+	/** D3 Source B: Include raw_transcript content nodes in the memory graph (requires dbMirror). */
+	memoryGraphSeedTurnContent: boolean;
+	/** D3 Source C: Include memory review nodes in the memory graph. */
+	memoryGraphSeedMemories: boolean;
+	/** D3 edges: Stricter cosine floor for cross-type edges (e.g. turn↔checkpoint). */
+	memoryGraphCrossTypeThreshold: number;
+	/** D3 edges: Cosine floor for within-type semantic edges. */
+	memoryGraphWithinTypeThreshold: number;
+	/** v0.12: Context health monitoring (drift + output quality + cache poison). Default ON. */
+	contextHealth: boolean;
+	/** Sub-flag: drift detection (topic drift + error escalation + prefix instability). */
+	contextHealthDrift: boolean;
+	/** Sub-flag: output quality analysis (repetition, coherence, token salad). */
+	contextHealthOutputQuality: boolean;
+	/** Sub-flag: tri-layer KV cache poison validation. */
+	contextHealthCachePoison: boolean;
+	/** v0.12: KV cache poison mitigation — inject prefix break on mismatch. Default OFF. */
+	contextHealthMitigate: boolean;
+}

package/extensions/mega-config.ts

@@ -10,6 +10,7 @@
 import { STATE_DIR_DEFAULT } from "../src/config.js";
 import { join } from "node:path";
 import { execSync } from "node:child_process"; // guardrails-allow PREVENT-PI-004: read-only `git rev-parse` to scope the store per-repo
+import type { MegaConfig } from "./mega-config-types.js";
 
 /**
  * Named compaction tiers. A tier sets the token threshold at which the
@@ -40,176 +41,9 @@ export const TIER_PCT: Record<CompactTier, number> = {
 	mega: 0.75,
 };
 
-/**
- * Resolved, frozen-at-load config. `tier` is the base compaction PRESET chosen
- * by env (low/medium/high/ultra/mega) — it sets the threshold token budget and
- * is NOT changed at runtime (the /mega-tier command was removed in S24). The
- * *displayed* tier the user sees in the toolbar/dashboard is the LIVE pressure
- * band (see MegaRuntime.pressureBand), which climbs low→mega as context fills.
- */
-export interface MegaConfig {
-	tier: CompactTier | "custom";
-	/**
-	 * Compaction threshold as a fraction of the model context window (e.g. 0.70
-	 * for "high"). null for `custom` (explicit MEGACOMPACT_THRESHOLD_TOKENS, which
-	 * stays an ABSOLUTE token count, never percent-scaled).
-	 */
-	tierPct: number | null;
-	thresholdTokens: number;
-	stateDir: string;
-	fastGatePct: number;
-	anchorUserMessages: number;
-	preserveRecent: number;
-	/** High-pressure floor for preserveRecent — when context is near the limit
-	 *  we compact deeper, but never below this (keeps recent turns for coherence). */
-	preserveRecentMin: number;
-	/** D.1: minimum context-growth percentage delta before re-compacting instead of
-	 *  replaying the cached trim (Default 50). Set via MEGACOMPACT_RECOMPACT_PCT_DELTA.*/
-	recompactPctDelta: number;
-	auto: boolean;
-	autoInline: boolean;
-	autoInlineK: number;
-	/** S28: auto-continue the agent after a max-output-token length stop by
-	 *  reusing the existing S16 resume-nudge. Default true. Off = silent (the
-	 *  prior behavior). PREVENT-PI-003: restart via user-role sendUserMessage. */
-	autoContinueLengthStop: boolean;
-	/** S38: max retries for transient errors (5xx/429/network/max-output-token
-	 *  text that is NOT a length stopReason — S28 owns those). Default 5.
-	 *  `0` disables all transient retries (reverts to S28-only). */
-	autoRetryTransientMax: number;
-	/** S38: max retries for permanent errors (auth/config/malformed). Default 1.
-	 *  `0` disables permanent-error retries. */
-	autoRetryPermanentMax: number;
-	/** S38.5: strict race-guard — 30s cooldown + deferred ctx.compact() re-check
-	 *  (closes the first-race-in-burst window). Default true. `false` reverts to
-	 *  the v0.7.4 synchronous 10s-cooldown behavior. */
-	raceGuardStrict: boolean;
-	/** S38.6: max consecutive errors before circuit-breaker trips (stops retrying).
-	 *  Default 10. When `errorRetryCount` exceeds this across multiple turns,
-	 *  the extension stops retrying until a successful turn resets it. */
-	maxConsecutiveErrors: number;
-	/** S38.7: hard-stop switch — when true, ALL error retries are disabled.
-	 *  Default false. Set via env to force S28-only behavior (length-stop continues
-	 *  only). */
-	errorRetryHardStop: boolean;
-	/** R1 (retry redesign): base unit (ms) for errorRetryBackoffMs(count) pacing.
-	 *  The schedule is base, 2*base, 4*base, 6*base (cap) — so the default 5000
-	 *  yields 5s/10s/20s/30s. errorRetryUntil is now GATING (previously it was
-	 *  documented as non-gating); a nudge cannot fire before errorRetryUntil
-	 *  elapses. */
-	errorRetryBackoffMs: number;
-	/** R2: session-global cap on total S38 nudges across ALL bursts. Hitting it
-	 *  is terminal for the session — the extension stops nudging entirely,
-	 *  independent of the per-burst max and the circuit breaker. Default 3.
-	 *  `0` disables (reverts to per-burst + circuit-breaker only). */
-	errorRetrySessionMax: number;
-	/** R3: consecutive identical error-text count at which a 'transient'
-	 *  classification is upgraded to 'poisoned-context' (the stateful repeat
-	 *  signal). Default 3. Raise to make the upgrade less aggressive. */
-	poisonedContextRepeatThreshold: number;
-	/** R10: consecutive transient errors at which a calm "provider outage"
-	 *  advisory is sent to the user (distinct from the poisoned /clear advise).
-	 *  Default 3. `0` disables the advisory entirely. */
-	providerOutageAdviseThreshold: number;
-	/** R13: when true (default), poisoned-context and provider-outage advisories
-	 *  are dashboard-only (events tab + log) — no user-visible message injection.
-	 *  When false, the legacy sendUserMessage path runs (byte-identical pre-R13). */
-	advisoryChannel: boolean;
-	/** S29: override the auto-compact fire point for tiered configs, as a
-	 *  fraction of the context window (e.g. 0.85). null = inherit the tier's
-	 *  tierPct (default; preserves existing fire points). The context-handler
-	 *  gate fires on context % (reliable), not token count (under-reported),
-	 *  so it catches the overshoot that causes max-output-token truncation.
-	 *  `custom` (tierPct null) ignores this — it keeps the absolute token gate. */
-	autoPctTrigger: number | null;
-	dedupSim: number;
-	/** RAPTOR hierarchical recall enabled (Fix D). Drives both live recall and
-	 *  the durable-trim summary source (root summary). */
-	raptorEnabled: boolean;
-	/** Legacy v0.4.28 behavior: auto-trigger calls ctx.compact() (which STOPS
-	 *  the agent). Default false — the S16 redesign uses the live context-event
-	 *  trim + pi native auto-compaction instead (compact and continue). Kept for
-	 *  one release as rollback. */
-	legacyDurableTrim: boolean;
-	/** S27: durable raw-transcript DB mirror (MEGACOMPACT_DB_MIRROR). When on,
-	 *  raw message bytes + checkpoint-epoch bookkeeping are appended to the
-	 *  SQLite store so a compacted window can be rehydrated locally instead of
-	 *  from the pi runtime transcript. Default OFF — additive, no behavior
-	 *  change until flipped on. legacyDurableTrim takes precedence (the legacy
-	 *  v0.4.28 ctx.compact() path does not emit the S27 mirror hook). */
-	dbMirror: boolean;
-	/** S49: isolated per-turn store (turns.db). Default ON. OFF = legacy main-db
-	 *  turn path (S48 behavior — byte-identical). Mirrors TurnsConfig.TURNS_DB_ENABLED. */
-	turnsDbEnabled: boolean;
-	/** S51: auto-categorizing wiki (k-means + TF-IDF over real embeddings). Default ON.
-	 *  Mirrors TurnsConfig.AUTO_WIKI_ENABLED. Rebuild fires every Nth compaction. */
-	autoWikiEnabled: boolean;
-	/** Cross-repo recall enabled (S17). Resume + /mega-recall --cross-repo can
-	 *  pull checkpoints from OTHER repos via the PGlite HNSW index. Default true. */
-	crossRepoEnabled: boolean;
-	/** Stricter cosine floor for cross-repo hits (S17). Default 0.90 (trigram) /
-	 *  tighter than same-repo so only genuinely-relevant cross-repo context is
-	 *  injected. */
-	crossRepoCosine: number;
-	/** Memory-RAG auto-review enabled (S20). Every memoryReviewInterval turns the
-	 *  conversation is auto-reviewed into durable add/replace/remove memories. */
-	memoryAutoReview: boolean;
-	/** Turn cadence for the auto-review scan (S20). Default 10. */
-	memoryReviewInterval: number;
-	/** Token ceiling for the re-injected recall block (Fix C). Recall stops
-	 *  adding checkpoints once the block would exceed this — bounds read-path
-	 *  token cost so it can never net-inflate the window. */
-	recallMaxTokens: number;
-	/** Inline-dedupe recalled checkpoints against the live window (Fix C): drop
-	 *  a hit whose summary is ≥ dedupSim similar to a live message — "dedupe on
-	 *  inline/read" so we never re-inject context already resident. */
-	windowDedupe: boolean;
-	/** S53: Recall Tail Injection — inject staged recall block as a user message at
-	 *  the tail of the view when auto is OFF AND no trim action is needed. Default ON
-	 *  (true). When false, restores the pre-sprint systemPrompt prepend behavior. */
-	recallTailInject: boolean;
-	/** A1 PLAN_V2 Phase 2: Message Separation — isolate user/assistant turns
-	 *  from volatile tool results so the prompt-cache prefix stays stable.
-	 *  Default OFF (flag-OFF = byte-identical pre-sprint). */
-	messageSeparation: boolean;
-	/** P3: Cache-aware striping (PLAN_V2 Phase 3). Inserts stability-ordered
-	 *  cache stripes between summaries and thread. Default OFF. */
-	cacheStriping: boolean;
-	debug: boolean;
-	/** Master reconciliation: TUI shutdown widget (MEGACOMPACT_TUI_WIDGET=0 to disable). */
-	tuiWidget: boolean;
-	/** S57 B1: Query reformulation via embedding-neighbor keyword expansion. */
-	ragQueryReformulation: boolean;
-	/** S57 B2: Tiered recall router (L0 cache -> L1 FTS5 -> L2). */
-	ragTieredRouter: boolean;
-	/** S57 B3: Recall quality metrics (precision/recall scoring + logging). */
-	ragRecallMetrics: boolean;
-	/** S57 B4: Memory graph traversal (dashboard-oriented). */
-	ragMemoryGraph: boolean;
-	/** D1: Seed initial wiki topic model from live turns when no context_chunks exist yet. */
-	wikiSeedFromTurns: boolean;
-	/** D3 Source A: Include structural turn nodes in the memory graph (metadata only, no content). */
-	memoryGraphSeedTurns: boolean;
-	/** D3 Source B: Include raw_transcript content nodes in the memory graph (requires dbMirror). */
-	memoryGraphSeedTurnContent: boolean;
-	/** D3 Source C: Include memory review nodes in the memory graph. */
-	memoryGraphSeedMemories: boolean;
-	/** D3 edges: Stricter cosine floor for cross-type edges (e.g. turn↔checkpoint). */
-	memoryGraphCrossTypeThreshold: number;
-	/** D3 edges: Cosine floor for within-type semantic edges. */
-	memoryGraphWithinTypeThreshold: number;
-	/** v0.12: Context health monitoring (drift + output quality + cache poison). Default ON. */
-	contextHealth: boolean;
-	/** Sub-flag: drift detection (topic drift + error escalation + prefix instability). */
-	contextHealthDrift: boolean;
-	/** Sub-flag: output quality analysis (repetition, coherence, token salad). */
-	contextHealthOutputQuality: boolean;
-	/** Sub-flag: tri-layer KV cache poison validation. */
-	contextHealthCachePoison: boolean;
-	/** v0.12: KV cache poison mitigation — inject prefix break on mismatch. Default OFF. */
-	contextHealthMitigate: boolean;
-}
-
+// MegaConfig type moved to mega-config-types.ts (delegate-shell split, PC-A)
+// so this runtime config barrel stays under the 400-line soft limit.
+export type { MegaConfig } from "./mega-config-types.js";
 function envFlag(name: string, fallback: number): number {
 	const v = process.env[name];
 	if (v == null || v === "") return fallback;
@@ -369,8 +203,11 @@ export function loadConfig(): MegaConfig {
 		recallMaxTokens: envFlag("MEGACOMPACT_RECALL_MAX_TOKENS", 1500),
 		windowDedupe: envBool("MEGACOMPACT_WINDOW_DEDUPE", true),
 		recallTailInject: envBool("MEGACOMPACT_RECALL_TAIL_INJECT", true),
-		messageSeparation: envBool("MEGACOMPACT_MESSAGE_SEPARATION", false),
-			cacheStriping: envBool("MEGACOMPACT_CACHE_STRIPING", false),
+		// PC-A: positive sprint flag, default ON. =0 byte-identical to the
+		// pre-change OFF state (single gate lives at the call site in tailResult.ts).
+		messageSeparation: envBool("MEGACOMPACT_MESSAGE_SEPARATION", true),
+		// positive sprint flag: default ON, =0 byte-identical to OFF
+		cacheStriping: envBool("MEGACOMPACT_CACHE_STRIPING", true),
 		tuiWidget: envBool("MEGACOMPACT_TUI_WIDGET", true),
 		ragQueryReformulation: envBool("MEGACOMPACT_QUERY_REFORMULATION", false),
 		ragTieredRouter: envBool("MEGACOMPACT_TIERED_ROUTER", false),

package/extensions/mega-events/separated-prompt.ts

@@ -12,8 +12,11 @@
  * 3 (thread: user/assistant turns) -> 4 (tool results at tail).
  *
  * Feature-gated by MEGACOMPACT_CACHE_STRIPING (default OFF).
- * MEGACOMPACT_MESSAGE_SEPARATION must also be ON for any layering to occur.
  * Flag-OFF = byte-identical to pre-sprint — returns messages unchanged.
+ *
+ * PC-A: buildSeparatedPrompt is PURE — the MEGACOMPACT_MESSAGE_SEPARATION gate
+ * lives at the single call site (tailResult.ts, config.messageSeparation),
+ * not inside this function.
  */
 
 import type { AgentMessage } from "@earendil-works/pi-agent-core";
@@ -61,18 +64,14 @@ const TOPIC_SHIFT_THRESHOLD = 0.7;
  * Layer order: 0 (system) -> 1 (summary) -> 3 (thread: user/assistant turns) ->
  * 4 (tool results at tail).
  *
- * Feature gate: MEGACOMPACT_MESSAGE_SEPARATION env var.  When OFF (default),
- * returns `messages` unchanged (byte-identical to pre-sprint).
+ * PURE function: the MEGACOMPACT_MESSAGE_SEPARATION gate moved to the single
+ * call site (tailResult.ts, config.messageSeparation) — this never reads env.
+ * When there is nothing to reorder, returns `messages` unchanged (byte-identical).
  */
 export function buildSeparatedPrompt(
   messages: AgentMessage[],
   _opts?: SeparatedPromptOptions,
 ): AgentMessage[] {
-  const flag = process.env.MEGACOMPACT_MESSAGE_SEPARATION;
-  if (flag === "0" || flag === "false" || flag === undefined || flag === "") {
-    return messages;
-  }
-
   // pi's AgentMessage union has no "system" role — the system prompt lives in
   // AgentState.systemPrompt, separate from this array. The cache-relevant,
   // low-risk transformation is moving volatile tool results/executions to the
@@ -237,22 +236,15 @@ export function refreshStripeAssignments(
  * Build a cache-optimized prompt with 5 layers:
  *   0 (system) -> 1 (summary) -> 2 (cache stripes) -> 3 (thread) -> 4 (tool)
  *
- * Feature gate: MEGACOMPACT_CACHE_STRIPING (default OFF). When OFF, delegates
- * to buildSeparatedPrompt (byte-identical to Phase 2 behavior).
- *
- * When ON but MEGACOMPACT_MESSAGE_SEPARATION is OFF, also delegates to
- * buildSeparatedPrompt (which returns messages unchanged).
+ * Positive sprint flag driven by config.cacheStriping at the call site
+ * (tailResult.ts); this function is pure and never reads process.env.
+ * With flag ON but no stripe rows for the epoch, it returns the base
+ * separated prompt unchanged (byte-identical to buildSeparatedPrompt).
  */
 export function buildCacheOptimizedPrompt(
   messages: AgentMessage[],
   opts?: SeparatedPromptOptions,
 ): AgentMessage[] {
-  const flag = process.env.MEGACOMPACT_CACHE_STRIPING;
-  if (flag === "0" || flag === "false" || flag === undefined || flag === "") {
-    // Flag OFF: delegate to buildSeparatedPrompt (byte-identical).
-    return buildSeparatedPrompt(messages, opts);
-  }
-
   // Build the base 4-layer structure first.
   const base = buildSeparatedPrompt(messages, opts);
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-mega-compact",
-  "version": "0.20.30",
+  "version": "0.20.32",
   "description": "Layered, local, vector-backed context compressor for pi — supersede/collapse/cluster compaction with deduped inline recall.",
   "type": "module",
   "license": "BSD-3-Clause",