billion-context-dsh 0.2.18 → 0.2.20

package/dist/region.d.ts

@@ -39,7 +39,19 @@ export interface AcpBlockLedgerEntry {
 }
 /** The open turn number, or null when the log ends between turns. */
 export declare function findOpenTurn(events: readonly SessionEvent[]): number | null;
-/** Reject a second concurrent compaction for the same session. */
+/**
+ * Reject a second concurrent compaction for the same session.
+ *
+ * Compaction is synchronous and a session is single-writer, so a
+ * `compaction/start` with NO matching `compaction/end` in the durable log can
+ * only be a stale leftover from a prior run that died mid-write (a hard kill,
+ * not a caught throw — every caught throw is paired with a compensating
+ * `compaction/end` in runCompactionTransaction). Such a leftover must NOT
+ * permanently block every later compress call: this treats it as stale,
+ * surfaces it once, and lets a new compaction proceed. The old "already
+ * active" throw only fired when a genuine concurrent compaction existed,
+ * which the synchronous single-writer premise makes impossible.
+ */
 export declare function assertNoActiveCompaction(events: readonly SessionEvent[]): void;
 /**
  * A requested range whose EVERY live message was already shadowed by one or

package/dist/window.d.ts

@@ -1,12 +1,17 @@
 /**
  * Auto context-window detection — resolve the model's real context window
- * from the host LLM runtime instead of trusting a hardcoded config default.
+ * from the host LLM runtime instead of trusting a hardcoded config default,
+ * plus the adapter's per-request output cap (the output reservation subtracted
+ * from it so pressure decisions run against the SUSTAINABLE input budget, not
+ * the raw window).
  *
  * `agent.ctx.llm` (the cordis `LlmRuntime` service) exposes
- * `resolveModelInfo(provider, model)` → `{ context: { contextWindow } }`, the
- * exact-route capacity the adapter learned from the provider API (pi-ai reads
- * `context_window`/`context_length` during discovery). Probing is a standalone
- * capability query — no request is sent.
+ * `resolveModelInfo(provider, model)` →
+ * `{ context: { contextWindow }, defaultMaxTokens }` — the exact-route
+ * capacity the adapter learned from the provider API (pi-ai reads
+ * `context_window`/`context_length` during discovery) plus the output cap it
+ * applies when callers omit one. Probing is a standalone capability query —
+ * no request is sent.
  * @module billion-context-dsh/window
  */
 import type { Agent } from '@deepseek-ai/dsh-agent';
@@ -34,6 +39,19 @@ export interface AcpWindow {
      * silently fell back to 128K and produced false emergency nudges).
      */
     readonly probeFailed?: boolean;
+    /**
+     * The model's TOTAL context window in tokens, before the output reservation
+     * was subtracted. Set only when `outputReserved` is set:
+     * `limit = rawLimit - outputReserved`.
+     */
+    readonly rawLimit?: number;
+    /**
+     * The adapter's per-request output cap (`defaultMaxTokens`) in tokens,
+     * subtracted from `rawLimit` to yield `limit` — the output reservation the
+     * provider guarantees at the end of the window on every request. Set only
+     * when the host discloses it and it is smaller than the raw window.
+     */
+    readonly outputReserved?: number;
 }
 /** Human label for an AcpWindow's source (used by /acp status). */
 export declare function windowSourceLabel(window: AcpWindow): string;
@@ -48,6 +66,26 @@ export declare function windowSourceLabel(window: AcpWindow): string;
  * Returns null when the host exposes no projection or disclosed no window.
  */
 export declare function projectedContextWindow(agent: Agent): number | null;
+/** The model window plus the adapter's per-request output cap, in one probe. */
+export interface ModelWindowProbe {
+    /** The model's total context window in tokens, when disclosed. */
+    readonly contextWindow: number | null;
+    /** The adapter's per-request output cap (`defaultMaxTokens`), when disclosed. */
+    readonly outputReservation: number | null;
+}
+/**
+ * Probe the model's real context window AND the adapter's per-request output
+ * cap in a single `resolveModelInfo` call. The cap is the output reservation
+ * the provider guarantees at the end of the window on every request —
+ * pressure decisions must run against the SUSTAINABLE input budget (window
+ * minus cap), not the raw window: a 96K window with a 16K cap carries at
+ * most 80K of input, so the raw denominator understates usage by cap/window
+ * (≈17% there — and far worse on short-window models, where the same cap is
+ * a quarter or more of the window). Returns nulls — never throws — when the
+ * host provides no llm service, discloses nothing, or the probe throws;
+ * callers keep the raw-window behavior in those cases.
+ */
+export declare function probeModelWindow(agent: Agent, provider: string, model: string): Promise<ModelWindowProbe>;
 /**
  * Probe the model's real context window. Returns null when the host provides
  * no llm service, the adapter discloses no window, or the probe throws —

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "billion-context-dsh",
-	"version": "0.2.18",
+	"version": "0.2.20",
 	"description": "Active Context Pruning (ACP) for the DeepSeek Harness — model-driven context management as a CompactionEngine backend.",
 	"keywords": [
 		"deepseek",