feat(scheduler): add cache_ttl "never" sentinel for always-warm lanes

Lanes kept warm by external keepwarm mechanisms (prewarm proxies,
dedicated cache-keep tools) re-warm the provider prompt cache
out-of-band, so MC's idle>TTL heuristic false-positives: both TTL
consumers (scheduler idle-execute and the ttl_idle m[0] fold) initiate
a rebuild believed free that is actually a full paid cache-write
(measured 450-560K tokens on large sessions).

cache_ttl: "never" (string or per-model value, case-insensitive)
disables both consumers: parseCacheTtl returns Infinity, so
elapsed>ttl / elapsed>=ttl never fire. Rust scheduler mirrors the
sentinel with u64::MAX (predicates already use saturating_sub).
Status surfaces render "never expires (always-warm lane)" instead of
a bogus countdown: /ctx-status, the TUI sidebar (JSON-safe
cacheNeverExpires flag on StatusDetail), and Pi's status dialog
(which previously parsed the TTL with a private fallback parser).
hardCacheExpired extracted to a pure computeHardCacheExpired helper
with direct coverage of the never-chain.

Tradeoff documented in CONFIGURATION.md: on a genuinely cold start
the free-fold window is not detected on such lanes; mutations then
apply at the execute threshold.

Author: Tehan

CONFIGURATION.md

@@ -88,6 +88,13 @@ Per-model overrides for mixed-model workflows:
 
 Supported formats: `"30s"`, `"5m"`, `"1h"`.
 
+**`"never"` sentinel:** set `cache_ttl` to `"never"` to disable the idle-TTL heuristic entirely. Both consumers —
+the scheduler (which converts defer passes to execute after TTL expiry) and the HARD-fold trigger (which folds
+m[1] into m[0] on a "free" prefix rebuild) — no longer act on idle time. Use this on lanes kept warm by an
+external keepwarm mechanism (prewarm proxies, dedicated cache-keep tools) where the idle heuristic false-positives
+and causes paid full-prefix cache-writes. After a genuine cold start (e.g. the keepwarm process died), MC
+won't detect the free-fold window on that lane — mutations then apply only at the execute threshold.
+
 Higher-tier models with longer cache windows benefit from a longer TTL. Setting it too low wastes cache hits. Setting it too high delays reduction on long sessions.
 
 ---

assets/magic-context.schema.json

@@ -1075,7 +1075,7 @@
     },
     "cache_ttl": {
       "default": "5m",
-      "description": "Cache TTL: string (e.g. \"5m\") or per-model object ({ default: \"5m\", \"model-id\": \"10m\" })",
+      "description": "Cache TTL: string (e.g. \"5m\", \"1h\", \"30s\") or per-model object ({ default: \"5m\", \"model-id\": \"10m\" }). Set to \"never\" for lanes kept warm by an external keepwarm proxy — disables the idle-TTL heuristic so MC never initiates a rebuild based on elapsed time.",
       "anyOf": [
         {
           "type": "string"

crates/mc-module/src/scheduler.rs

@@ -288,6 +288,9 @@ struct CompiledPattern {
 /// Parse a cache idle TTL string into milliseconds.
 pub fn parse_cache_ttl(ttl: &str) -> Result<u64, CacheTtlParseError> {
     let normalized = ttl.trim();
+    if normalized.eq_ignore_ascii_case("never") {
+        return Ok(u64::MAX);
+    }
     let (number, multiplier) =
         if !normalized.is_empty() && normalized.chars().all(|c| c.is_ascii_digit()) {
             (normalized, 1.0)
@@ -1217,4 +1220,38 @@ mod tests {
             PassClass::EmergencyForce
         );
     }
+
+    #[test]
+    fn parse_cache_ttl_never_returns_u64_max() {
+        assert_eq!(parse_cache_ttl("never"), Ok(u64::MAX));
+        assert_eq!(parse_cache_ttl("NEVER"), Ok(u64::MAX));
+        assert_eq!(parse_cache_ttl(" never "), Ok(u64::MAX));
+        assert_eq!(parse_cache_ttl("Never"), Ok(u64::MAX));
+        assert_eq!(parse_cache_ttl("5m"), Ok(300_000));
+        assert_eq!(parse_cache_ttl("bad-format"), Err(CacheTtlParseError));
+    }
+
+    #[test]
+    fn never_ttl_predicates_are_always_false() {
+        // Scheduler: elapsed > u64::MAX is never true.
+        assert!(!ttl_execute_fired(u64::MAX, 0, u64::MAX));
+        assert!(!ttl_execute_fired(1_000_000, 0, u64::MAX));
+        // Hard: elapsed >= u64::MAX is never true.
+        assert!(!ttl_hard_expired(u64::MAX, 1, u64::MAX));
+        assert!(!ttl_hard_expired(1_000_000, 1, u64::MAX));
+    }
+
+    #[test]
+    fn never_ttl_scheduler_stays_deferred() {
+        let mut inputs = base_inputs();
+        // 10-day-old last response, well past any normal TTL
+        inputs.session.last_response_time_ms = 1_000;
+        inputs.session.cache_ttl = "never".to_string();
+        inputs.now_ms = 1_000 + 10 * 24 * 60 * 60 * 1000;
+        // Below threshold, so TTL is the only path to execute
+        inputs.usage.percentage = 50.0;
+        let outcome = decide(&inputs);
+        assert!(!outcome.idle_ttl_fired);
+        assert_eq!(outcome.pass, PassDecision::Defer);
+    }
 }

packages/docs/src/content/docs/reference/configuration.md

@@ -46,7 +46,7 @@ When and how aggressively Magic Context manages the session's context window. Pe
 
 | Key | Type | Default | Description |
 |---|---|---|---|
-| `cache_ttl` | string \\| map<string, string> | `"5m"` | Cache TTL: string (e.g. "5m") or per-model object ({ default: "5m", "model-id": "10m" }) |
+| `cache_ttl` | string \\| map<string, string> | `"5m"` | Cache TTL: string (e.g. "5m", "1h", "30s") or per-model object ({ default: "5m", "model-id": "10m" }). Set to "never" for lanes kept warm by an external keepwarm proxy — disables the idle-TTL heuristic so MC never initiates a rebuild based on elapsed time. |
 | `execute_threshold_percentage` | number (20–80) \\| map<string, number (20–80)> | `65` | Context percentage that forces queued operations to execute. Number or per-model object ({ default: 65, "provider/model": 45 }). Values above 80 are rejected because the runtime caps at 80% for cache safety (MAX_EXECUTE_THRESHOLD). Default: DEFAULT_EXECUTE_THRESHOLD_PERCENTAGE |
 | `execute_threshold_tokens` | object | — | Absolute token thresholds per model. When matched, overrides execute_threshold_percentage for that model. Accepts `default` for all models or per-model keys. Values above 80% × context_limit are clamped with a warning log. Min 5_000, max 2_000_000. |
 | `execute_threshold_tokens.default` | number (5000–2000000) | — |  |

packages/pi-plugin/src/dialogs/status-dialog.ts

@@ -12,6 +12,7 @@ import {
 } from "@earendil-works/pi-tui";
 import { getCompartments } from "@magic-context/core/features/magic-context/compartment-storage";
 import { getMemoryCount } from "@magic-context/core/features/magic-context/memory/storage-memory";
+import { parseCacheTtl } from "@magic-context/core/features/magic-context/scheduler";
 import type { ContextDatabase } from "@magic-context/core/features/magic-context/storage";
 import { getOrCreateSessionMeta } from "@magic-context/core/features/magic-context/storage-meta";
 import { getSessionWorkMetrics } from "@magic-context/core/features/magic-context/storage-meta-persisted";
@@ -322,7 +323,9 @@ function renderInner(
 		} · ${
 			s.cacheExpired
 				? theme.fg("warning", "expired")
-				: `${Math.round(s.cacheRemainingMs / 1000)}s remaining`
+				: s.cacheRemainingMs === Number.POSITIVE_INFINITY
+					? "never expires (always-warm lane)"
+					: `${Math.round(s.cacheRemainingMs / 1000)}s remaining`
 		}`,
 	);
 	lines.push("");
@@ -519,11 +522,20 @@ export function buildPiStatusDetail(
 		},
 	);
 	const cacheTtl = meta.cacheTtl || "5m";
-	const cacheTtlMs = parseTtlString(cacheTtl);
+	let cacheTtlMs: number;
+	try {
+		cacheTtlMs = parseCacheTtl(cacheTtl);
+	} catch {
+		cacheTtlMs = 5 * 60 * 1000;
+	}
+	const neverExpires = cacheTtlMs === Number.POSITIVE_INFINITY;
 	const elapsed =
 		meta.lastResponseTime > 0 ? Date.now() - meta.lastResponseTime : 0;
-	const cacheRemainingMs =
-		meta.lastResponseTime > 0 ? Math.max(0, cacheTtlMs - elapsed) : cacheTtlMs;
+	const cacheRemainingMs = neverExpires
+		? Number.POSITIVE_INFINITY
+		: meta.lastResponseTime > 0
+			? Math.max(0, cacheTtlMs - elapsed)
+			: cacheTtlMs;
 	const cacheExpired = meta.lastResponseTime > 0 && cacheRemainingMs === 0;
 	const historyBlockTokens = compartmentTokens + factTokens;
 	const historyBudgetPercentage = deps.historyBudgetPercentage ?? 0.15;
@@ -748,22 +760,6 @@ function readPendingOpsCount(db: ContextDatabase, sessionId: string): number {
 	}
 }
 
-function parseTtlString(ttl: string): number {
-	const match = ttl.match(/^(\d+)(s|m|h)$/);
-	if (!match) return 5 * 60 * 1000;
-	const val = Number.parseInt(match[1] ?? "5", 10);
-	switch (match[2]) {
-		case "s":
-			return val * 1000;
-		case "m":
-			return val * 60 * 1000;
-		case "h":
-			return val * 3600 * 1000;
-		default:
-			return 5 * 60 * 1000;
-	}
-}
-
 function safeRead<T>(fn: () => T, fallback: T): T {
 	try {
 		return fn();

packages/plugin/src/config/schema/magic-context.ts

@@ -596,7 +596,7 @@ export const MagicContextConfigSchema = z
             .union([z.string(), z.object({ default: z.string() }).catchall(z.string())])
             .default("5m")
             .describe(
-                'Cache TTL: string (e.g. "5m") or per-model object ({ default: "5m", "model-id": "10m" })',
+                'Cache TTL: string (e.g. "5m", "1h", "30s") or per-model object ({ default: "5m", "model-id": "10m" }). Set to "never" for lanes kept warm by an external keepwarm proxy — disables the idle-TTL heuristic so MC never initiates a rebuild based on elapsed time.',
             ),
         toast_duration_ms: z
             .number()

packages/plugin/src/features/magic-context/scheduler.test.ts

@@ -119,6 +119,36 @@ describe("createScheduler", () => {
 
         expect(decision).toBe("defer");
     });
+
+    it("never executes from TTL idle when cacheTtl is 'never'", () => {
+        const scheduler = createScheduler({ executeThresholdPercentage: 65 });
+        // Last response was 10 days ago — well past any normal TTL.
+        const tenDaysAgo = BASE_TIME - 10 * 24 * 60 * 60 * 1000;
+        const sessionMeta = createSessionMeta({
+            cacheTtl: "never",
+            lastResponseTime: tenDaysAgo,
+        });
+        // Percentage is below threshold, so the only path to "execute" is TTL expiry.
+        const contextUsage = createContextUsage(50);
+
+        const decision = scheduler.shouldExecute(sessionMeta, contextUsage, BASE_TIME);
+
+        expect(decision).toBe("defer");
+    });
+
+    it("still executes on threshold when cacheTtl is 'never'", () => {
+        const scheduler = createScheduler({ executeThresholdPercentage: 65 });
+        const sessionMeta = createSessionMeta({
+            cacheTtl: "never",
+            lastResponseTime: BASE_TIME - 10_000,
+        });
+        // At threshold — should still execute.
+        const contextUsage = createContextUsage(65);
+
+        const decision = scheduler.shouldExecute(sessionMeta, contextUsage, BASE_TIME);
+
+        expect(decision).toBe("execute");
+    });
 });
 
 describe("parseCacheTtl", () => {
@@ -138,6 +168,13 @@ describe("parseCacheTtl", () => {
         expect(milliseconds).toBe(300_000);
     });
 
+    it("returns Infinity for 'never' regardless of casing and whitespace", () => {
+        expect(parseCacheTtl("never")).toBe(Number.POSITIVE_INFINITY);
+        expect(parseCacheTtl("NEVER")).toBe(Number.POSITIVE_INFINITY);
+        expect(parseCacheTtl(" never ")).toBe(Number.POSITIVE_INFINITY);
+        expect(parseCacheTtl("Never")).toBe(Number.POSITIVE_INFINITY);
+    });
+
     it("throws on invalid ttl format", () => {
         expect(() => parseCacheTtl("bad-format")).toThrow();
     });

packages/plugin/src/features/magic-context/scheduler.ts

@@ -30,6 +30,12 @@ interface SchedulerConfig {
 export function parseCacheTtl(ttl: string): number {
     const normalizedTtl = ttl.trim();
 
+    // "never" sentinel: lanes kept warm by external keepwarm proxies — the idle
+    // heuristic is disabled, so MC never initiates a rebuild based on elapsed time.
+    if (normalizedTtl.toLowerCase() === "never") {
+        return Number.POSITIVE_INFINITY;
+    }
+
     // Intentional: bare numeric strings are treated as milliseconds. The setup CLI writes
     // "5m" or "59m" so users don't encounter bare numbers through normal config paths.
     if (NUMERIC_PATTERN.test(normalizedTtl)) {

packages/plugin/src/hooks/magic-context/execute-status.test.ts

@@ -69,4 +69,21 @@ describe("executeStatus", () => {
         expect(status).not.toContain("[clamped:");
         db.close();
     });
+
+    test("renders 'never expires' for cacheTtl 'never'", () => {
+        const db = new Database(":memory:");
+        initializeDatabase(db);
+        getOrCreateSessionMeta(db, SESSION_ID);
+        db.prepare("UPDATE session_meta SET cache_ttl = 'never' WHERE session_id = ?").run(
+            SESSION_ID,
+        );
+
+        const status = executeStatus(db, SESSION_ID, 20);
+
+        expect(status).toContain("- Configured: never");
+        expect(status).toContain("- Remaining: never expires (always-warm lane)");
+        expect(status).toContain("never expires");
+        expect(status).not.toContain("Infinity");
+        db.close();
+    });
 });

packages/plugin/src/hooks/magic-context/execute-status.ts

@@ -127,8 +127,8 @@ export function executeStatus(
             "### Cache TTL",
             `- Configured: ${meta.cacheTtl}`,
             `- Last response: ${meta.lastResponseTime > 0 ? `${Math.round(elapsed / 1000)}s ago` : "never"}`,
-            `- Remaining: ${cacheExpired ? "expired" : `${Math.round(remainingMs / 1000)}s`}`,
-            `- Queue will auto-execute: ${cacheExpired ? "yes (cache expired)" : `when TTL expires or context >= ${executeThresholdPercentage}%`}`,
+            `- Remaining: ${cacheExpired ? "expired" : ttlMs === Number.POSITIVE_INFINITY ? "never expires (always-warm lane)" : `${Math.round(remainingMs / 1000)}s`}`,
+            `- Queue will auto-execute: ${cacheExpired ? "yes (cache expired)" : ttlMs === Number.POSITIVE_INFINITY ? `when context >= ${executeThresholdPercentage}%` : `when TTL expires or context >= ${executeThresholdPercentage}%`}`,
             "",
             "### Execute Threshold",
             `- Execute threshold: ${formatExecuteThreshold(thresholdDetail, displayContextLimit)}`,