opencode-anthropic-fix 0.1.12 → 0.1.14

package/index.mjs

@@ -3721,7 +3721,7 @@ export async function AnthropicAuthPlugin({ client, project, directory, worktree
                           cacheBreakState.lastAlertTurn !== sessionMetrics.turns
                         ) {
                           const drop = cacheBreakState.prevCacheRead - cacheRead;
-                          let alertMsg = `Cache break detected (u2212${drop.toLocaleString()} tokens)`;
+                          let alertMsg = `Cache break detected (−${drop.toLocaleString()} tokens)`;
 
                           // Identify changed sources if we have pending hashes
                           if (cacheBreakState._pendingHashes && cacheBreakState.sourceHashes.size > 0) {
@@ -4327,6 +4327,30 @@ function extractCacheSourceHashes(bodyStr, parsedBody = undefined) {
         }
       }
     }
+
+    // Hash messages prefix (everything except the last message) so we can
+    // detect when the conversation history prefix changes byte-for-byte
+    // between turns — a common cause of prompt-cache invalidation that
+    // system_prompt/tool hashes alone don't explain.
+    if (Array.isArray(parsed.messages) && parsed.messages.length > 1) {
+      const prefix = parsed.messages.slice(0, -1);
+      // Strip cache_control markers before hashing — they're legitimately
+      // re-stamped each turn and shouldn't trigger a false positive.
+      const normalized = prefix.map((m) => {
+        if (!Array.isArray(m.content)) return m;
+        return {
+          ...m,
+          content: m.content.map((b) => {
+            if (b && typeof b === "object" && b.cache_control) {
+              const { cache_control: _cc, ...rest } = b;
+              return rest;
+            }
+            return b;
+          }),
+        };
+      });
+      hashes.set("messages_prefix", hashCacheSource(JSON.stringify(normalized)));
+    }
   } catch {
     // Ignore parse errors
   }
@@ -7424,7 +7448,17 @@ function transformRequestBody(body, signature, runtime, betaHeader, config) {
     const tes = runtime?.tokenEconomySession;
     const isMainRole = runtime?.requestRole === "main" || runtime?.requestRole == null;
 
-    if (isMainRole && Array.isArray(parsed.messages) && tes) {
+    // `conservative` (default ON) disables all history-rewriting and tool-array
+    // transforms. These optimizations shrink each request body but cause the
+    // prompt-cache prefix to change turn-to-turn, invalidating the 1h cache
+    // and forcing a fresh cache_write each turn — which costs 2x base input
+    // tokens. For long opencode sessions, cache reuse dominates; flip to
+    // `false` only if you have measurements showing otherwise. Adaptive
+    // thinking zero-out remains active (affects only the thinking budget,
+    // not cached content).
+    const conservative = te.conservative !== false;
+
+    if (!conservative && isMainRole && Array.isArray(parsed.messages) && tes) {
       // (1) TTL-based thinking strip
       if (te.ttl_thinking_strip !== false) {
         const ttlMs = signature?.cachePolicy?.ttl === "5m" ? 5 * 60_000 : 60 * 60_000;
@@ -7474,7 +7508,7 @@ function transformRequestBody(body, signature, runtime, betaHeader, config) {
     }
 
     // === Token economy: tool-array transforms (stable ordering, deferral) ===
-    if (isMainRole && Array.isArray(parsed.tools)) {
+    if (!conservative && isMainRole && Array.isArray(parsed.tools)) {
       if (te.stable_tool_ordering !== false) {
         parsed.tools = applyStableToolOrdering(parsed.tools);
       }

package/lib/config.mjs

@@ -201,6 +201,17 @@ export const DEFAULT_CONFIG = {
      *  classified as "main-thread" (see classifyRequestRole). */
     context_hint: false,
 
+    /** Conservative mode (default ON). Disables every history-rewriting and
+     *  tool-array transform below that can invalidate the prompt-cache prefix
+     *  turn-to-turn. Rationale: every rewrite risks breaking the 1h cache,
+     *  forcing a fresh cache_write (2x base input cost). For long opencode
+     *  sessions, cache reuse dominates per-turn cost — the optimizations below
+     *  save kB in the request body but cost much more in cache misses.
+     *  Affects: ttl_thinking_strip, proactive_microcompact, trailing_summary_trim,
+     *  tool_result_dedupe, stable_tool_ordering, deferred_tool_names.
+     *  Unaffected: adaptive_thinking_zero_simple (only changes thinking budget). */
+    conservative: true,
+
     /** TTL-based thinking strip (CC parity). When the time since the last
      *  thinking strip exceeds the cache TTL, drop all `thinking` /
      *  `redacted_thinking` blocks from prior assistant messages. Mirrors CC's
@@ -644,6 +655,7 @@ function validateConfig(raw) {
       redact_thinking:
         typeof te.redact_thinking === "boolean" ? te.redact_thinking : DEFAULT_CONFIG.token_economy.redact_thinking,
       context_hint: typeof te.context_hint === "boolean" ? te.context_hint : DEFAULT_CONFIG.token_economy.context_hint,
+      conservative: typeof te.conservative === "boolean" ? te.conservative : DEFAULT_CONFIG.token_economy.conservative,
       ttl_thinking_strip:
         typeof te.ttl_thinking_strip === "boolean"
           ? te.ttl_thinking_strip

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-anthropic-fix",
-  "version": "0.1.12",
+  "version": "0.1.14",
   "license": "GPL-3.0-or-later",
   "main": "./index.mjs",
   "files": [