reasonix 0.11.3 → 0.12.6

package/dist/cli/{prompt-YRY4HPMZ.js → prompt-HNDDXDRH.js}

@@ -2,9 +2,9 @@
 import {
   CODE_SYSTEM_PROMPT,
   codeSystemPrompt
-} from "./chunk-JDVY4JDU.js";
+} from "./chunk-PKPWI33U.js";
 export {
   CODE_SYSTEM_PROMPT,
   codeSystemPrompt
 };
-//# sourceMappingURL=prompt-YRY4HPMZ.js.map
+//# sourceMappingURL=prompt-HNDDXDRH.js.map

package/dist/index.d.ts

@@ -521,6 +521,46 @@ declare class VolatileScratch {
     reset(): void;
 }
 
+/**
+ * Predicate the breaker consults to decide whether a call mutates state.
+ * Mutating calls clear the recent-args buffer: re-reading a file after
+ * `edit_file` shouldn't count as "saw the same args before" — the file
+ * legitimately changed. Wire this from the caller using whatever source
+ * of truth is appropriate (e.g. the ToolRegistry's `readOnly` /
+ * `readOnlyCheck` flags). When undefined, every call is tracked the
+ * old way — preserves the original behavior for callers that don't
+ * thread a registry through.
+ */
+type IsMutating = (call: ToolCall) => boolean;
+/**
+ * Call-storm breaker.
+ *
+ * Detects (tool, args) tuples repeating within a sliding window and suppresses
+ * the offending call. Surfaces a synthetic tool_result advising the model to
+ * change strategy on its next turn.
+ *
+ * Buffer entries are tagged read-only vs mutating. When a mutating call
+ * runs, the breaker drops prior read-only entries — a re-read of the
+ * same path after `edit_file` is fresh, not a repeat. Mutating calls
+ * still count among themselves, so a model looping on identical
+ * `edit_file` invocations still trips on the threshold.
+ *
+ * Without an `isMutating` predicate everything is tracked the same way
+ * (back-compat for callers that don't thread a registry through).
+ */
+declare class StormBreaker {
+    private readonly windowSize;
+    private readonly threshold;
+    private readonly isMutating;
+    private readonly recent;
+    constructor(windowSize?: number, threshold?: number, isMutating?: IsMutating);
+    inspect(call: ToolCall): {
+        suppress: boolean;
+        reason?: string;
+    };
+    reset(): void;
+}
+
 /**
  * Schema flattening for DeepSeek tool calls.
  *
@@ -577,25 +617,6 @@ interface ScavengeResult {
 }
 declare function scavengeToolCalls(reasoningContent: string | null | undefined, opts: ScavengeOptions): ScavengeResult;
 
-/**
- * Call-storm breaker.
- *
- * Detects (tool, args) tuples repeating within a sliding window and suppresses
- * the offending call. Surfaces a synthetic tool_result advising the model to
- * change strategy on its next turn.
- */
-declare class StormBreaker {
-    private readonly windowSize;
-    private readonly threshold;
-    private readonly recent;
-    constructor(windowSize?: number, threshold?: number);
-    inspect(call: ToolCall): {
-        suppress: boolean;
-        reason?: string;
-    };
-    reset(): void;
-}
-
 /**
  * Pillar 3 — Tool-Call Repair pipeline.
  *
@@ -619,6 +640,14 @@ interface ToolCallRepairOptions {
     stormWindow?: number;
     stormThreshold?: number;
     maxScavenge?: number;
+    /**
+     * Optional predicate the storm breaker consults to identify state-
+     * changing calls — those clear the sliding window so a post-edit
+     * verify-read isn't mistaken for a repeat. Production callers wire
+     * this off the ToolRegistry's `readOnly` / `readOnlyCheck` flags;
+     * tests that don't supply it keep the original behavior.
+     */
+    isMutating?: IsMutating;
 }
 declare class ToolCallRepair {
     private readonly storm;
@@ -899,6 +928,12 @@ interface CacheFirstLoopOptions {
      * `max` for Reasonix (agent-class use per DeepSeek V4 docs).
      */
     reasoningEffort?: "high" | "max";
+    /**
+     * Master switch for auto-escalation paths. See ReconfigurableOptions
+     * — defaults to `true` (current behavior); the `flash` and `pro`
+     * presets pass `false` to lock the running session to one model.
+     */
+    autoEscalate?: boolean;
     /**
      * Session name. When set, the loop pre-loads the session's prior messages
      * into its log on construction, and appends every new log entry to
@@ -943,6 +978,15 @@ interface ReconfigurableOptions {
      * mid-session for cheaper, faster turns on simple tasks.
      */
     reasoningEffort?: "high" | "max";
+    /**
+     * Master switch for the auto-escalation paths — both the
+     * `<<<NEEDS_PRO>>>` marker scavenge and the failure-count threshold.
+     * `true` (default) preserves the original "flash baseline, jump to
+     * pro when struggling" behavior. `false` locks the active turn to
+     * whatever `model` is set to — used by the `flash` and `pro` presets
+     * which want a hard model commitment.
+     */
+    autoEscalate?: boolean;
 }
 declare class CacheFirstLoop {
     readonly client: DeepSeekClient;
@@ -961,6 +1005,13 @@ declare class CacheFirstLoop {
     branchOptions: BranchOptions;
     /** See ReconfigurableOptions — mutable so `/effort` can flip mid-session. */
     reasoningEffort: "high" | "max";
+    /**
+     * Auto-escalation toggle. `true` lets the loop self-promote to pro
+     * mid-turn (NEEDS_PRO marker / failure threshold); `false` keeps it
+     * pinned to `model`. Mutable so the dashboard's preset switcher can
+     * flip it live alongside `model`.
+     */
+    autoEscalate: boolean;
     sessionName: string | null;
     /**
      * Hook list, mutable so `/hooks reload` can swap it without
@@ -3532,8 +3583,21 @@ declare function codeSystemPrompt(rootDir: string, opts?: CodeSystemPromptOption
  * from `reasonix setup`: preset, MCP servers, session. This is what
  * makes `reasonix chat` with no flags "just work" after first-run.
  */
-/** One of the preset bundles (model + harvest + branch combo). */
-type PresetName = "fast" | "smart" | "max";
+/**
+ * Preset names — three model-commitment levels.
+ *   - `auto`  — flash baseline + auto-escalate to pro on hard turns
+ *               (NEEDS_PRO marker / failure-count threshold both fire).
+ *               Default. Closest match to the legacy `smart` preset.
+ *   - `flash` — flash always. No auto-escalation. `/pro` still works
+ *               for one-shot manual escalation. Cheapest predictable.
+ *   - `pro`   — pro always. No downgrade. ~3× cost vs flash at the
+ *               2026-04 discount rate; more outside the window.
+ *
+ * Legacy `fast | smart | max` names stay in the union for back-compat
+ * with existing `~/.reasonix/config.json` files; resolvePreset() maps
+ * them to the new semantics.
+ */
+type PresetName = "auto" | "flash" | "pro" | "fast" | "smart" | "max";
 /**
  * How `reasonix code` handles model-issued tool calls. Two axes folded
  * into one enum because users think about "how trusting am I right now?"
@@ -3812,6 +3876,16 @@ interface UsageBucket {
     cacheMissTokens: number;
     costUsd: number;
     claudeEquivUsd: number;
+    /**
+     * USD that DeepSeek's prompt cache shaved off the bill — sum of
+     * `cacheHitTokens × (missPrice − hitPrice)` per record. Recomputed
+     * from the current pricing table on every aggregate, not frozen at
+     * write time, so a price-cut announcement updates retroactively. The
+     * trade-off is mild inconsistency with `costUsd` (which IS frozen);
+     * acceptable because cache savings is a "what does this mechanism
+     * give me" narrative, not a billing record.
+     */
+    cacheSavingsUsd: number;
 }
 /** Cache hit ratio for a bucket — zero denominator returns 0. */
 declare function bucketCacheHitRatio(b: UsageBucket): number;

package/dist/index.js

@@ -1358,25 +1358,32 @@ function coerceToToolCall(candidateJson, allowedNames) {
 var StormBreaker = class {
   windowSize;
   threshold;
+  isMutating;
   recent = [];
-  constructor(windowSize = 6, threshold = 3) {
+  constructor(windowSize = 6, threshold = 3, isMutating) {
     this.windowSize = windowSize;
     this.threshold = threshold;
+    this.isMutating = isMutating;
   }
   inspect(call) {
-    const sig = signature(call);
-    if (!sig) return { suppress: false };
-    const count = this.recent.reduce(
-      (n, [name, args]) => name === sig[0] && args === sig[1] ? n + 1 : n,
-      0
-    );
+    const name = call.function?.name;
+    if (!name) return { suppress: false };
+    const args = call.function?.arguments ?? "";
+    const mutating = this.isMutating ? this.isMutating(call) : false;
+    const readOnly = !mutating;
+    if (mutating) {
+      for (let i = this.recent.length - 1; i >= 0; i--) {
+        if (this.recent[i].readOnly) this.recent.splice(i, 1);
+      }
+    }
+    const count = this.recent.reduce((n, e) => e.name === name && e.args === args ? n + 1 : n, 0);
     if (count >= this.threshold - 1) {
       return {
         suppress: true,
-        reason: `call-storm suppressed: ${sig[0]} called with identical args ${count + 1} times within window=${this.windowSize}`
+        reason: `call-storm suppressed: ${name} called with identical args ${count + 1} times within window=${this.windowSize}`
       };
     }
-    this.recent.push(sig);
+    this.recent.push({ name, args, readOnly });
     while (this.recent.length > this.windowSize) this.recent.shift();
     return { suppress: false };
   }
@@ -1384,11 +1391,6 @@ var StormBreaker = class {
     this.recent.length = 0;
   }
 };
-function signature(call) {
-  const name = call.function?.name;
-  if (!name) return null;
-  return [name, call.function?.arguments ?? ""];
-}
 
 // src/repair/truncation.ts
 function repairTruncatedJson(input) {
@@ -1466,7 +1468,7 @@ var ToolCallRepair = class {
   opts;
   constructor(opts) {
     this.opts = opts;
-    this.storm = new StormBreaker(opts.stormWindow ?? 6, opts.stormThreshold ?? 3);
+    this.storm = new StormBreaker(opts.stormWindow ?? 6, opts.stormThreshold ?? 3, opts.isMutating);
   }
   /**
    * Drop the StormBreaker's sliding window of recent (name, args)
@@ -1490,13 +1492,13 @@ var ToolCallRepair = class {
       allowedNames: this.opts.allowedToolNames,
       maxCalls: this.opts.maxScavenge ?? 4
     });
-    const seenSignatures = new Set(declaredCalls.map(signature2));
+    const seenSignatures = new Set(declaredCalls.map(signature));
     const merged = [...declaredCalls];
     for (const sc of scavenged.calls) {
-      if (!seenSignatures.has(signature2(sc))) {
+      if (!seenSignatures.has(signature(sc))) {
         merged.push(sc);
         report.scavenged++;
-        seenSignatures.add(signature2(sc));
+        seenSignatures.add(signature(sc));
       }
     }
     report.notes.push(...scavenged.notes);
@@ -1522,7 +1524,7 @@ var ToolCallRepair = class {
     return { calls: filtered, report };
   }
 };
-function signature2(call) {
+function signature(call) {
   return `${call.function?.name ?? ""}::${call.function?.arguments ?? ""}`;
 }
 
@@ -1661,6 +1663,12 @@ function outputCostUsd(model, usage) {
   if (!p) return 0;
   return usage.completionTokens * p.output / 1e6;
 }
+function cacheSavingsUsd(model, hitTokens) {
+  if (hitTokens <= 0) return 0;
+  const p = DEEPSEEK_PRICING[model];
+  if (!p) return 0;
+  return hitTokens * (p.inputCacheMiss - p.inputCacheHit) / 1e6;
+}
 function claudeEquivalentCost(usage) {
   return (usage.promptTokens * CLAUDE_SONNET_PRICING.input + usage.completionTokens * CLAUDE_SONNET_PRICING.output) / 1e6;
 }
@@ -1751,6 +1759,13 @@ var CacheFirstLoop = class {
   branchOptions;
   /** See ReconfigurableOptions — mutable so `/effort` can flip mid-session. */
   reasoningEffort;
+  /**
+   * Auto-escalation toggle. `true` lets the loop self-promote to pro
+   * mid-turn (NEEDS_PRO marker / failure threshold); `false` keeps it
+   * pinned to `model`. Mutable so the dashboard's preset switcher can
+   * flip it live alongside `model`.
+   */
+  autoEscalate = true;
   sessionName;
   /**
    * Hook list, mutable so `/hooks reload` can swap it without
@@ -1815,6 +1830,7 @@ var CacheFirstLoop = class {
     this.tools = opts.tools ?? new ToolRegistry();
     this.model = opts.model ?? "deepseek-v4-flash";
     this.reasoningEffort = opts.reasoningEffort ?? "max";
+    if (opts.autoEscalate !== void 0) this.autoEscalate = opts.autoEscalate;
     this.maxToolIters = opts.maxToolIters ?? 64;
     this.hooks = opts.hooks ?? [];
     this.hookCwd = opts.hookCwd ?? process.cwd();
@@ -1832,7 +1848,26 @@ var CacheFirstLoop = class {
     this._streamPreference = opts.stream ?? true;
     this.stream = this.branchEnabled ? false : this._streamPreference;
     const allowedNames = /* @__PURE__ */ new Set([...this.prefix.toolSpecs.map((s) => s.function.name)]);
-    this.repair = new ToolCallRepair({ allowedToolNames: allowedNames });
+    const registry = this.tools;
+    const isMutating = (call) => {
+      const name = call.function?.name;
+      if (!name) return false;
+      const def = registry.get(name);
+      if (!def) return false;
+      if (def.readOnlyCheck) {
+        let args = {};
+        try {
+          args = JSON.parse(call.function?.arguments ?? "{}") ?? {};
+        } catch {
+        }
+        try {
+          if (def.readOnlyCheck(args)) return false;
+        } catch {
+        }
+      }
+      return def.readOnly !== true;
+    };
+    this.repair = new ToolCallRepair({ allowedToolNames: allowedNames, isMutating });
     this.sessionName = opts.session ?? null;
     if (this.sessionName) {
       const prior = loadSessionMessages(this.sessionName);
@@ -2013,6 +2048,7 @@ var CacheFirstLoop = class {
     if (opts.model !== void 0) this.model = opts.model;
     if (opts.stream !== void 0) this._streamPreference = opts.stream;
     if (opts.reasoningEffort !== void 0) this.reasoningEffort = opts.reasoningEffort;
+    if (opts.autoEscalate !== void 0) this.autoEscalate = opts.autoEscalate;
     if (opts.branch !== void 0) {
       if (typeof opts.branch === "number") {
         this.branchOptions = { budget: opts.branch };
@@ -2128,7 +2164,7 @@ var CacheFirstLoop = class {
       if (repair.truncationsFixed > 0) bump("truncated", repair.truncationsFixed);
       if (repair.stormsBroken > 0) bump("storm-broken", repair.stormsBroken);
     }
-    if (bumped && !this._escalateThisTurn && this._turnFailureCount >= FAILURE_ESCALATION_THRESHOLD) {
+    if (bumped && !this._escalateThisTurn && this.autoEscalate && this._turnFailureCount >= FAILURE_ESCALATION_THRESHOLD) {
       this._escalateThisTurn = true;
       return true;
     }
@@ -2373,7 +2409,7 @@ var CacheFirstLoop = class {
           const callBuf = /* @__PURE__ */ new Map();
           const readyIndices = /* @__PURE__ */ new Set();
           const callModel = this.modelForCurrentCall();
-          const bufferForEscalation = callModel !== ESCALATION_MODEL;
+          const bufferForEscalation = this.autoEscalate && callModel !== ESCALATION_MODEL;
           let escalationBuf = "";
           let escalationBufFlushed = false;
           for await (const chunk of this.client.stream({
@@ -2485,7 +2521,7 @@ var CacheFirstLoop = class {
         };
         return;
       }
-      if (this.modelForCurrentCall() !== ESCALATION_MODEL && this.isEscalationRequest(assistantContent)) {
+      if (this.autoEscalate && this.modelForCurrentCall() !== ESCALATION_MODEL && this.isEscalationRequest(assistantContent)) {
         const { reason } = this.parseEscalationMarker(assistantContent);
         this._escalateThisTurn = true;
         const reasonSuffix = reason ? ` \u2014 ${reason}` : "";
@@ -8158,7 +8194,8 @@ function emptyBucket(label, since) {
     cacheHitTokens: 0,
     cacheMissTokens: 0,
     costUsd: 0,
-    claudeEquivUsd: 0
+    claudeEquivUsd: 0,
+    cacheSavingsUsd: 0
   };
 }
 function addToBucket(b, r) {
@@ -8169,6 +8206,7 @@ function addToBucket(b, r) {
   b.cacheMissTokens += r.cacheMissTokens;
   b.costUsd += r.costUsd;
   b.claudeEquivUsd += r.claudeEquivUsd;
+  b.cacheSavingsUsd += cacheSavingsUsd(r.model, r.cacheHitTokens);
 }
 function aggregateUsage(records, opts = {}) {
   const now = opts.now ?? Date.now();