@sahiljassal/opencode-anthropic-auth 2.1.0 → 2.2.0

package/README.md

@@ -5,7 +5,7 @@
 >
 > Use your best judgment and don't abuse your subscription.
 
-Fork of [ex-machina-co/opencode-anthropic-auth](https://github.com/ex-machina-co/opencode-anthropic-auth).
+Fork of [ex-machina-co/opencode-anthropic-auth](https://github.com/ex-machina-co/opencode-anthropic-auth), with caching improvements inspired by [cortexkit/anthropic-auth](https://github.com/cortexkit/anthropic-auth).
 
 An [OpenCode](https://github.com/anomalyco/opencode) plugin that provides Anthropic OAuth authentication, enabling Claude Pro/Max users to use their subscription directly with OpenCode.
 
@@ -27,12 +27,22 @@ Add to your OpenCode config (`~/.config/opencode/opencode.json`):
 
 ## Prompt Caching
 
-This fork applies **hybrid 1-hour ephemeral prompt caching** on every request:
+This fork applies **hybrid 1-hour ephemeral prompt caching** on every request, placing up to 4 breakpoints strategically:
 
-- Strips any existing `cache_control` blocks from the request
-- Anchors a `1h` ephemeral cache on the last system block (after identity) and the first two user messages
+| Breakpoint | Behaviour |
+|---|---|
+| **System anchor** | Last system block after the identity block (skipped when bridge occupies the slot) |
+| **messages[0]** | Magic-context split: anchors block[0] + block[1] when stable prefix and volatile delta are merged; otherwise anchors the last cacheable block |
+| **messages[1] / bridge** | Last cacheable block of messages[1]; replaced by a bridge anchor when the block count between two user anchors (counting ALL block types across every role) exceeds Anthropic's 20-block lookback window |
+| **Rolling latest** | Most recent user message beyond index 1, keeping cache hot across long sessions |
 
-This reduces token usage and latency on repeated requests by reusing cached prompt prefixes.
+Additional behaviours:
+
+- **System tail coalescing** — plugin-added system blocks beyond the primary prompt are merged into one block before placing the system anchor, preventing cache busts when block layout changes between requests
+- **Trailing assistant strip** — assistant messages at the tail of the request are removed before forwarding (OAuth rejects assistant prefill)
+- **Thinking block guard** — `thinking` and `redacted_thinking` blocks are excluded from cache anchor placement; messages containing only thinking blocks receive no `cache_control` (avoids Anthropic 400)
+- **SSE retryable errors** — transient server errors (`api_error`, `overloaded_error`, `server_error`) emitted inside HTTP 200 streams are detected and thrown as connection-reset errors so OpenCode auto-retries
+- **Buffered stream rewriting** — tool name stripping buffers partial `"name"` tokens across chunk boundaries to avoid corruption
 
 ## Configuration
 

package/dist/constants.d.ts

@@ -8,10 +8,12 @@ export declare const TOKEN_URL = "https://platform.claude.com/v1/oauth/token";
 export declare const OAUTH_SCOPES: string[];
 export declare const TOOL_PREFIX = "mcp_";
 /**
- * Anthropic's sliding-window lookback for cache breakpoints.
- * If the distance (in content blocks) between the previous user-role
- * message anchor and the latest one exceeds this threshold, a bridge
- * anchor is needed so the earlier slots are still within the window.
+ * Anthropic's cache lookback window size. Each explicit breakpoint scans at
+ * most this many content blocks backward (counting the breakpoint block as
+ * position 1, across all roles and types — text, thinking, tool_use,
+ * tool_result, …). When the estimated block count between two user-role anchors
+ * would exceed this threshold, a bridge anchor is inserted to keep the older
+ * breakpoint inside the window and the cached prefix reachable.
  */
 export declare const ANTHROPIC_CACHE_LOOKBACK_BLOCKS = 20;
 export declare const REQUIRED_BETAS: string[];
@@ -19,9 +21,9 @@ export declare const OPENCODE_IDENTITY_PREFIX = "You are OpenCode";
 export declare const CLAUDE_CODE_IDENTITY = "You are a Claude agent, built on Anthropic's Claude Agent SDK.";
 export declare const CCH_SALT = "59cf53e54c78";
 export declare const CCH_POSITIONS: number[];
-export declare const CLAUDE_CODE_VERSION = "2.1.87";
-export declare const CLAUDE_CODE_ENTRYPOINT = "sdk-cli";
-export declare const USER_AGENT = "claude-cli/2.1.87 (external, cli)";
+export declare const CLAUDE_CODE_VERSION = "2.1.177";
+export declare const CLAUDE_CODE_ENTRYPOINT = "cli";
+export declare const USER_AGENT = "claude-cli/2.1.177 (external, cli)";
 /**
  * Anchors that identify paragraphs to remove from the system prompt.
  * Any paragraph (text between blank lines) containing one of these

package/dist/constants.js

@@ -15,23 +15,27 @@ export const OAUTH_SCOPES = [
 ];
 export const TOOL_PREFIX = 'mcp_';
 /**
- * Anthropic's sliding-window lookback for cache breakpoints.
- * If the distance (in content blocks) between the previous user-role
- * message anchor and the latest one exceeds this threshold, a bridge
- * anchor is needed so the earlier slots are still within the window.
+ * Anthropic's cache lookback window size. Each explicit breakpoint scans at
+ * most this many content blocks backward (counting the breakpoint block as
+ * position 1, across all roles and types — text, thinking, tool_use,
+ * tool_result, …). When the estimated block count between two user-role anchors
+ * would exceed this threshold, a bridge anchor is inserted to keep the older
+ * breakpoint inside the window and the cached prefix reachable.
  */
 export const ANTHROPIC_CACHE_LOOKBACK_BLOCKS = 20;
 export const REQUIRED_BETAS = [
     'oauth-2025-04-20',
+    'claude-code-20250219',
     'interleaved-thinking-2025-05-14',
+    'fine-grained-tool-streaming-2025-05-14',
 ];
 export const OPENCODE_IDENTITY_PREFIX = 'You are OpenCode';
 export const CLAUDE_CODE_IDENTITY = "You are a Claude agent, built on Anthropic's Claude Agent SDK.";
 export const CCH_SALT = '59cf53e54c78';
 export const CCH_POSITIONS = [4, 7, 20];
-export const CLAUDE_CODE_VERSION = '2.1.87';
-export const CLAUDE_CODE_ENTRYPOINT = 'sdk-cli';
-export const USER_AGENT = 'claude-cli/2.1.87 (external, cli)';
+export const CLAUDE_CODE_VERSION = '2.1.177';
+export const CLAUDE_CODE_ENTRYPOINT = 'cli';
+export const USER_AGENT = 'claude-cli/2.1.177 (external, cli)';
 /**
  * Anchors that identify paragraphs to remove from the system prompt.
  * Any paragraph (text between blank lines) containing one of these

package/dist/transform.d.ts

@@ -9,7 +9,7 @@ export declare function mergeHeaders(input: FetchInput, init?: RequestInit): Hea
  */
 export declare function mergeBetaHeaders(headers: Headers): string;
 /**
- * Set OAuth-required headers on the request: authorization, beta, user-agent.
+ * Set OAuth-required headers: authorization, beta, user-agent.
  * Removes x-api-key since we're using OAuth.
  */
 export declare function setOAuthHeaders(headers: Headers, accessToken: string): Headers;
@@ -31,7 +31,6 @@ export declare function isInsecure(): boolean;
  * Rewrite the request URL to add ?beta=true for /v1/messages requests.
  * When ANTHROPIC_BASE_URL is set, overrides the origin (protocol + host)
  * for all API requests flowing through the fetch wrapper.
- * Returns the modified input and URL (if applicable).
  */
 export declare function rewriteUrl(input: FetchInput): {
     input: FetchInput;
@@ -62,13 +61,21 @@ type SystemBlock = {
  * Handles all Anthropic API system formats: undefined, string, or array of text blocks.
  */
 export declare function prependClaudeCodeIdentity(system: unknown): SystemBlock[];
+export declare function rewriteRequestBody(body: string): string;
 /**
- * Rewrite the full request body: sanitize system prompt, prefix tool names,
- * and apply hybrid 1h prompt caching.
+ * Error thrown when Anthropic emits a retryable server-side error inside
+ * an HTTP 200 stream. OpenCode recognises ECONNRESET + anthropic-sse syscall
+ * and applies its normal auto-retry flow instead of surfacing an unknown error.
  */
-export declare function rewriteRequestBody(body: string): string;
+export type RetryableAnthropicStreamError = Error & {
+    code: 'ECONNRESET';
+    syscall: 'anthropic-sse';
+    providerErrorType?: string;
+};
 /**
  * Create a streaming response that strips the tool prefix from tool names.
+ * Detects retryable Anthropic server errors inside HTTP 200 streams and
+ * throws a connection-reset-style error so OpenCode can auto-retry.
  */
 export declare function createStrippedStream(response: Response): Response;
 export {};

package/dist/transform.js

@@ -1,17 +1,8 @@
 import { ANTHROPIC_CACHE_LOOKBACK_BLOCKS, CLAUDE_CODE_IDENTITY, OPENCODE_IDENTITY_PREFIX, PARAGRAPH_REMOVAL_ANCHORS, REQUIRED_BETAS, TEXT_REPLACEMENTS, TOOL_PREFIX, USER_AGENT, } from "./constants.js";
-/**
- * Prefix a tool name with TOOL_PREFIX and uppercase the first character.
- * Claude Code uses PascalCase tool names (e.g. mcp_Bash, mcp_Read);
- * lowercase names (mcp_bash, mcp_read) are flagged as non-Claude-Code clients.
- */
 function prefixName(name) {
     return `${TOOL_PREFIX}${name.charAt(0).toUpperCase()}${name.slice(1)}`;
 }
-/**
- * Reverse prefixName: strip TOOL_PREFIX and restore the original leading case.
- */
 function unprefixName(name) {
-    // StructuredOutput is still used as StructuredOutput
     if (name === 'StructuredOutput') {
         return name;
     }
@@ -65,7 +56,7 @@ export function mergeBetaHeaders(headers) {
     return [...new Set([...REQUIRED_BETAS, ...incomingBetasList])].join(',');
 }
 /**
- * Set OAuth-required headers on the request: authorization, beta, user-agent.
+ * Set OAuth-required headers: authorization, beta, user-agent.
  * Removes x-api-key since we're using OAuth.
  */
 export function setOAuthHeaders(headers, accessToken) {
@@ -142,7 +133,6 @@ function resolveBaseUrl() {
  * Rewrite the request URL to add ?beta=true for /v1/messages requests.
  * When ANTHROPIC_BASE_URL is set, overrides the origin (protocol + host)
  * for all API requests flowing through the fetch wrapper.
- * Returns the modified input and URL (if applicable).
  */
 export function rewriteUrl(input) {
     let requestUrl = null;
@@ -192,14 +182,11 @@ export function rewriteUrl(input) {
  * somewhere in the paragraph, the removal works.
  */
 export function sanitizeSystemText(text) {
-    // Split into paragraphs (separated by one or more blank lines)
     const paragraphs = text.split(/\n\n+/);
     const filtered = paragraphs.filter((paragraph) => {
         if (paragraph.includes(OPENCODE_IDENTITY_PREFIX)) {
-            // If the paragraph contains the identity, drop it entirely
             return false;
         }
-        // Remove paragraphs containing any removal anchor
         for (const anchor of PARAGRAPH_REMOVAL_ANCHORS) {
             if (paragraph.includes(anchor))
                 return false;
@@ -207,7 +194,6 @@ export function sanitizeSystemText(text) {
         return true;
     });
     let result = filtered.join('\n\n');
-    // Apply inline text replacements
     for (const rule of TEXT_REPLACEMENTS) {
         result = result.replace(rule.match, rule.replacement);
     }
@@ -217,9 +203,6 @@ function isRecord(value) {
     return value != null && typeof value === 'object' && !Array.isArray(value);
 }
 const CACHE_1H = { type: 'ephemeral', ttl: '1h' };
-// ---------------------------------------------------------------------------
-// Cache control primitives
-// ---------------------------------------------------------------------------
 function removeCacheControl(value) {
     if (!isRecord(value))
         return;
@@ -248,21 +231,23 @@ function setWireCacheControl(value) {
     value.cache_control = { ...CACHE_1H };
     return true;
 }
-// ---------------------------------------------------------------------------
-// Content-block helpers
-// ---------------------------------------------------------------------------
 /**
  * Returns true for content block types that accept cache_control.
- * Anthropic rejects cache_control on thinking / redacted_thinking blocks.
+ * Anthropic rejects cache_control on thinking / redacted_thinking blocks
+ * and silently skips empty text blocks without caching them.
  */
 function isCacheableContentBlock(block) {
     if (!isRecord(block))
         return false;
-    return block.type !== 'thinking' && block.type !== 'redacted_thinking';
+    if (block.type === 'thinking' || block.type === 'redacted_thinking')
+        return false;
+    if (block.type === 'text' && !String(block.text ?? '').trim())
+        return false;
+    return true;
 }
 /**
  * Normalises message content to an array of blocks, then filters to only
- * cacheable types.  Returns undefined when there is nothing to anchor.
+ * cacheable types. Returns undefined when there is nothing to anchor.
  */
 function getCacheableContentBlocks(message) {
     if (!isRecord(message))
@@ -272,7 +257,6 @@ function getCacheableContentBlocks(message) {
         blocks = message.content;
     }
     else if (typeof message.content === 'string') {
-        // Normalise inline string to block array in place so downstream sees array
         const normalised = [{ type: 'text', text: message.content }];
         message.content = normalised;
         blocks = normalised;
@@ -283,38 +267,21 @@ function getCacheableContentBlocks(message) {
     const cacheable = blocks.filter(isCacheableContentBlock);
     return cacheable.length > 0 ? cacheable : undefined;
 }
-/**
- * Total cacheable content-block count for a message (used for lookback math).
- */
 function messageContentBlockCount(message) {
     return getCacheableContentBlocks(message)?.length ?? 0;
 }
-// ---------------------------------------------------------------------------
-// Message-anchor setters
-// ---------------------------------------------------------------------------
-/**
- * Anchor the last cacheable block of a message.
- * Returns false (and sets nothing) when there are no cacheable blocks.
- */
 function setMessageCacheAnchor(message) {
     const blocks = getCacheableContentBlocks(message);
     if (!blocks)
         return false;
     return setWireCacheControl(blocks[blocks.length - 1]);
 }
-/**
- * Anchor the FIRST cacheable block of a message (for magic-context split).
- */
 function setFirstMessageCacheAnchor(message) {
     const blocks = getCacheableContentBlocks(message);
     if (!blocks)
         return false;
     return setWireCacheControl(blocks[0]);
 }
-/**
- * Anchor the SECOND cacheable block of a message (for magic-context split).
- * Returns false when fewer than two cacheable blocks exist.
- */
 function setSecondMessageCacheAnchor(message) {
     const blocks = getCacheableContentBlocks(message);
     if (!blocks || blocks.length < 2)
@@ -322,53 +289,106 @@ function setSecondMessageCacheAnchor(message) {
     return setWireCacheControl(blocks[1]);
 }
 /**
- * Walk all messages and collect user-role (or tool_result) anchor positions.
- * Returns the `latest` position (index > 1) and a `bridge` position placed
- * whenever the cumulative block distance from bridge→latest exceeds
- * ANTHROPIC_CACHE_LOOKBACK_BLOCKS, ensuring both are within the sliding window.
+ * Total positional block count for a message regardless of role or type.
+ * Anthropic's cache lookback window counts every content block (text,
+ * thinking, tool_use, tool_result, …), so distance math must use raw counts,
+ * not the cacheable-only filter used to validate anchor placement.
+ */
+function rawBlockCount(message) {
+    if (!isRecord(message))
+        return 0;
+    const { content } = message;
+    if (Array.isArray(content))
+        return content.length;
+    if (typeof content === 'string')
+        return 1;
+    return 0;
+}
+/**
+ * Collect user-role anchor indices that hold at least one cacheable block,
+ * then pick the `latest` (index > 1) and a `bridge`. The bridge is the
+ * furthest-back valid user anchor where the block count between it and
+ * `latest` — counting ALL content blocks across user and assistant turns,
+ * excluding the anchor's own blocks — does not exceed
+ * ANTHROPIC_CACHE_LOOKBACK_BLOCKS. This keeps the older breakpoint inside
+ * Anthropic's sliding lookback window so the cached prefix stays reachable.
  */
 function selectHybridMessageAnchors(messages) {
-    // Collect positions of user-role messages that have cacheable content
-    const userPositions = messages
-        .map((msg, index) => {
-        if (!isRecord(msg))
-            return null;
-        if (msg.role !== 'user')
-            return null;
-        const blockCount = messageContentBlockCount(msg);
-        if (blockCount === 0)
-            return null;
-        return { index, blockCount };
-    })
-        .filter((p) => p !== null);
-    // We only care about positions beyond index 1 (0 and 1 are always anchored)
-    const rollingPositions = userPositions.filter((p) => p.index > 1);
-    if (rollingPositions.length === 0) {
+    // Collect indices of user messages beyond index 1 that have ≥1 cacheable block.
+    // Indices 0 and 1 are handled as fixed slots in applyHybridCache1h, not rolling.
+    const rollingIndices = [];
+    messages.forEach((msg, index) => {
+        if (index > 1 &&
+            isRecord(msg) &&
+            msg.role === 'user' &&
+            messageContentBlockCount(msg) > 0) {
+            rollingIndices.push(index);
+        }
+    });
+    if (rollingIndices.length === 0) {
         return { latest: undefined, bridge: undefined };
     }
-    // Non-null: rollingPositions is non-empty (early return guards above)
     // biome-ignore lint/style/noNonNullAssertion: guarded by length check above
-    const latest = rollingPositions[rollingPositions.length - 1];
-    // Find a bridge: walk back from latest until cumulative blocks exceed window
+    const latestIndex = rollingIndices[rollingIndices.length - 1];
+    const rollingSet = new Set(rollingIndices);
     let bridge;
-    let cumulativeBlocks = latest.blockCount;
-    for (let i = rollingPositions.length - 2; i >= 0; i--) {
-        // biome-ignore lint/style/noNonNullAssertion: index is within bounds
-        cumulativeBlocks += rollingPositions[i].blockCount;
-        if (cumulativeBlocks > ANTHROPIC_CACHE_LOOKBACK_BLOCKS) {
-            bridge = rollingPositions[i];
-            break;
+    let cumulativeBlocks = 0;
+    for (let i = latestIndex - 1; i >= 0; i--) {
+        if (rollingSet.has(i)) {
+            if (cumulativeBlocks > ANTHROPIC_CACHE_LOOKBACK_BLOCKS)
+                break;
+            bridge = i;
         }
+        cumulativeBlocks += rawBlockCount(messages[i]);
     }
-    return { latest, bridge };
+    return { latest: latestIndex, bridge };
+}
+function systemBlockText(block) {
+    return isRecord(block) && typeof block.text === 'string' ? block.text : '';
+}
+/**
+ * Merge all plugin-added system instruction blocks (those after the primary
+ * OpenCode/system prompt block) into a single block before placing the hybrid
+ * system cache anchor.
+ *
+ * OpenCode normally emits these as one merged block, but some hooks can cause
+ * them to arrive split across multiple blocks. Without coalescing, byte-
+ * identical system text flips between merged/split layouts and moves the
+ * cache_control breakpoint — busting the cache every turn.
+ *
+ * Block layout after prependClaudeCodeIdentity:
+ *   [billing-header?] [identity] [primary system prompt] [plugin blocks…]
+ * We preserve everything up to and including the primary prompt block and
+ * merge all remaining plugin blocks into one.
+ */
+function coalesceHybridSystemTail(parsed) {
+    if (!Array.isArray(parsed.system))
+        return;
+    const system = parsed.system;
+    let prefixCount = 0;
+    if (systemBlockText(system[prefixCount]).startsWith('x-anthropic-billing-header:')) {
+        prefixCount++;
+    }
+    if (systemBlockText(system[prefixCount]) === CLAUDE_CODE_IDENTITY) {
+        prefixCount++;
+    }
+    const tailStart = prefixCount + 1;
+    if (tailStart >= system.length - 1)
+        return;
+    const firstTail = system[tailStart];
+    if (!isRecord(firstTail))
+        return;
+    const mergedText = system.slice(tailStart).map(systemBlockText).join('\n');
+    system.splice(tailStart, system.length - tailStart, {
+        ...firstTail,
+        type: 'text',
+        text: mergedText,
+    });
 }
-// ---------------------------------------------------------------------------
-// System-anchor setter
-// ---------------------------------------------------------------------------
 /**
  * Place a cache anchor on the last system block that follows the
- * CLAUDE_CODE_IDENTITY block.  When there are no system blocks, or all
- * system blocks precede the identity, nothing is anchored.
+ * CLAUDE_CODE_IDENTITY block. When there are no system blocks after the
+ * identity, nothing is anchored.
  */
 function setHybridSystemAnchor(parsed) {
     if (!Array.isArray(parsed.system))
@@ -379,11 +399,8 @@ function setHybridSystemAnchor(parsed) {
         .filter(isRecord);
     setWireCacheControl(afterIdentity[afterIdentity.length - 1]);
 }
-// ---------------------------------------------------------------------------
-// Trailing-assistant strip (Tier 2)
-// ---------------------------------------------------------------------------
 /**
- * Remove trailing assistant-role messages.  OAuth endpoints reject requests
+ * Remove trailing assistant-role messages. OAuth endpoints reject requests
  * that end with an assistant turn (assistant prefill is not supported).
  */
 function stripTrailingAssistantMessages(parsed) {
@@ -395,51 +412,60 @@ function stripTrailingAssistantMessages(parsed) {
         parsed.messages.pop();
     }
 }
-// ---------------------------------------------------------------------------
-// Main hybrid cache logic
-// ---------------------------------------------------------------------------
+/**
+ * Returns true when messages[0] carries a merged stable-prefix layout
+ * (≥2 cacheable blocks). In that case anchoring the last block would bust
+ * the cache every turn because the tail is volatile; instead we anchor
+ * block[0] and block[1] (the two stable-prefix blocks).
+ */
+function isMagicContextLayout(blocks) {
+    return blocks.length >= 2;
+}
 /**
  * Apply hybrid 1h prompt-caching breakpoints to parsed request body.
  *
- * Breakpoint slots (Anthropic supports max 4 per request):
- *   1. Last system block after the identity block  (skipped when bridge used)
- *   2. messages[0] — first cacheable block         (magic-context: slot 2a)
- *   3. messages[0] — second cacheable block        (magic-context split only)
- *      OR messages[1] last cacheable block         (normal path)
- *      OR bridge user message                      (when bridge detected)
- *   4. Latest user/tool-result message (index > 1) (when present)
+ * Anthropic supports max 4 cache breakpoints per request. Slot allocation:
+ *
+ *   Slot 1 — system anchor: last block after identity.
+ *             Skipped when a bridge is used (bridge takes that slot).
+ *
+ *   Slots 2+3 — messages[0]:
+ *     • Normal layout (1 cacheable block): slot 2 = last block of msg[0];
+ *       slot 3 = bridge message OR messages[1] (no bridge).
+ *     • Magic-context layout (≥2 cacheable blocks): slot 2 = block[0],
+ *       slot 3 = block[1]. messages[1] is skipped (msg[0] uses both slots).
+ *
+ *   Slot 4 — rolling latest: last user/tool-result message beyond index 1.
+ *
+ * Bridge and magic-context are independent: when both apply simultaneously,
+ * the slot budget is: system(skipped) + msg0-block0 + bridge + latest = 4.
+ * The bridge anchor is ALWAYS placed when detected, regardless of msg0 layout.
  */
 function applyHybridCache1h(parsed) {
     removeAllCacheControls(parsed);
+    coalesceHybridSystemTail(parsed);
     const messages = Array.isArray(parsed.messages) ? parsed.messages : [];
     const { latest, bridge } = selectHybridMessageAnchors(messages);
-    // --- Slot 1: system anchor (skip when bridge will occupy a slot) ---
-    if (!bridge) {
+    if (bridge === undefined) {
         setHybridSystemAnchor(parsed);
     }
-    // --- Slots 2 & 3: messages[0] ---
     const msg0 = messages[0];
     const msg0Blocks = getCacheableContentBlocks(msg0);
-    if (msg0Blocks && msg0Blocks.length >= 2) {
-        // Magic-context split: stable prefix is in block[0] and block[1];
-        // anchoring last block would bust cache every turn.
+    if (msg0Blocks && isMagicContextLayout(msg0Blocks)) {
         setFirstMessageCacheAnchor(msg0);
         setSecondMessageCacheAnchor(msg0);
     }
     else {
         setMessageCacheAnchor(msg0);
-        // --- Slot 3 (normal): messages[1] or bridge ---
-        if (bridge) {
-            setHybridSystemAnchor(parsed); // system anchor reclaimed for bridge support
-            setMessageCacheAnchor(messages[bridge.index]);
-        }
-        else {
+        if (bridge === undefined) {
             setMessageCacheAnchor(messages[1]);
         }
     }
-    // --- Slot 4: rolling latest user anchor ---
-    if (latest) {
-        setMessageCacheAnchor(messages[latest.index]);
+    if (bridge !== undefined) {
+        setMessageCacheAnchor(messages[bridge]);
+    }
+    if (latest !== undefined) {
+        setMessageCacheAnchor(messages[latest]);
     }
 }
 /**
@@ -481,16 +507,11 @@ export function prependClaudeCodeIdentity(system) {
         }
         return { type: 'text', text: String(item) };
     });
-    // Idempotency: don't double-prepend if first block already has the identity
     if (sanitized[0]?.text === CLAUDE_CODE_IDENTITY) {
         return sanitized;
     }
     return [identityBlock, ...sanitized];
 }
-/**
- * Rewrite the full request body: sanitize system prompt, prefix tool names,
- * and apply hybrid 1h prompt caching.
- */
 export function rewriteRequestBody(body) {
     try {
         const parsed = JSON.parse(body);
@@ -503,8 +524,135 @@ export function rewriteRequestBody(body) {
         return body;
     }
 }
+function findSseBoundary(value) {
+    const lf = value.indexOf('\n\n');
+    const crlf = value.indexOf('\r\n\r\n');
+    if (lf === -1)
+        return crlf === -1 ? null : { index: crlf, length: 4 };
+    if (crlf === -1 || lf < crlf)
+        return { index: lf, length: 2 };
+    return { index: crlf, length: 4 };
+}
+function asDiagnosticRecord(value) {
+    return value != null && typeof value === 'object' && !Array.isArray(value)
+        ? value
+        : undefined;
+}
+function stringField(record, key) {
+    const v = record?.[key];
+    return typeof v === 'string' ? v : undefined;
+}
+function isRetryableAnthropicStreamError(errorType, message) {
+    const t = errorType?.toLowerCase();
+    const m = message.toLowerCase();
+    return (t === 'api_error' ||
+        t === 'overloaded_error' ||
+        t === 'server_error' ||
+        t === 'internal_server_error' ||
+        m.includes('internal server error') ||
+        m.includes('server overloaded'));
+}
+function retryableAnthropicStreamError(errorType, message) {
+    const detail = errorType ? `${errorType}: ${message}` : message;
+    const err = new Error(`Anthropic stream error: ${detail}`);
+    err.code = 'ECONNRESET';
+    err.syscall = 'anthropic-sse';
+    if (errorType)
+        err.providerErrorType = errorType;
+    return err;
+}
+function retryableAnthropicStreamErrorFromRawEvent(rawEvent) {
+    if (!rawEvent.includes('error'))
+        return null;
+    let eventName;
+    const dataLines = [];
+    for (const line of rawEvent.split(/\r?\n/)) {
+        if (line.startsWith('event:')) {
+            eventName = line.slice('event:'.length).trim();
+        }
+        else if (line.startsWith('data:')) {
+            const v = line.slice('data:'.length);
+            dataLines.push(v.startsWith(' ') ? v.slice(1) : v);
+        }
+    }
+    const dataText = dataLines.join('\n');
+    if (!dataText || dataText === '[DONE]')
+        return null;
+    let parsed;
+    try {
+        parsed = JSON.parse(dataText);
+    }
+    catch {
+        return null;
+    }
+    const data = asDiagnosticRecord(parsed);
+    if (eventName !== 'error' && stringField(data, 'type') !== 'error')
+        return null;
+    const errorObj = asDiagnosticRecord(data?.error);
+    const errorType = stringField(errorObj, 'type') ?? stringField(errorObj, 'code') ?? undefined;
+    const message = stringField(errorObj, 'message') ??
+        stringField(data, 'message') ??
+        errorType ??
+        'Anthropic stream error';
+    if (!isRetryableAnthropicStreamError(errorType, message))
+        return null;
+    return retryableAnthropicStreamError(errorType, message);
+}
+function createSseErrorState() {
+    return { pending: '' };
+}
+function updateSseErrorState(state, text) {
+    if (!text)
+        return null;
+    state.pending += text;
+    while (true) {
+        const boundary = findSseBoundary(state.pending);
+        if (!boundary)
+            break;
+        const rawEvent = state.pending.slice(0, boundary.index);
+        state.pending = state.pending.slice(boundary.index + boundary.length);
+        const err = retryableAnthropicStreamErrorFromRawEvent(rawEvent);
+        if (err)
+            return err;
+    }
+    return null;
+}
+/**
+ * Rewrite the tool prefix from the safe portion of a text buffer.
+ * Holds back any suffix that could be the start of a partial `"name"` marker
+ * spanning a chunk boundary. Pass flush=true on stream end to emit everything.
+ */
+function splitToolPrefixRewriteBuffer(buffer, flush = false) {
+    if (flush)
+        return { ready: stripToolPrefix(buffer), pending: '' };
+    let keepFrom = buffer.length;
+    const marker = '"name"';
+    const partialStart = Math.max(0, buffer.length - marker.length + 1);
+    for (let i = partialStart; i < buffer.length; i++) {
+        if (marker.startsWith(buffer.slice(i))) {
+            keepFrom = Math.min(keepFrom, i);
+            break;
+        }
+    }
+    const lastMarker = buffer.lastIndexOf(marker);
+    if (lastMarker !== -1) {
+        const tail = buffer.slice(lastMarker);
+        if (/^"name"\s*(?::\s*(?:"[^"]*)?)?$/.test(tail)) {
+            keepFrom = Math.min(keepFrom, lastMarker);
+        }
+    }
+    if (keepFrom < buffer.length) {
+        return {
+            ready: stripToolPrefix(buffer.slice(0, keepFrom)),
+            pending: buffer.slice(keepFrom),
+        };
+    }
+    return { ready: stripToolPrefix(buffer), pending: '' };
+}
 /**
  * Create a streaming response that strips the tool prefix from tool names.
+ * Detects retryable Anthropic server errors inside HTTP 200 streams and
+ * throws a connection-reset-style error so OpenCode can auto-retry.
  */
 export function createStrippedStream(response) {
     if (!response.body)
@@ -512,16 +660,71 @@ export function createStrippedStream(response) {
     const reader = response.body.getReader();
     const decoder = new TextDecoder();
     const encoder = new TextEncoder();
+    let pending = '';
+    let readerReleased = false;
+    const sseErrors = createSseErrorState();
+    const releaseReader = () => {
+        if (readerReleased)
+            return;
+        readerReleased = true;
+        reader.releaseLock();
+    };
     const stream = new ReadableStream({
         async pull(controller) {
-            const { done, value } = await reader.read();
-            if (done) {
-                controller.close();
-                return;
+            try {
+                const { done, value } = await reader.read();
+                if (done) {
+                    const finalDecoded = decoder.decode();
+                    let retryableError = updateSseErrorState(sseErrors, finalDecoded);
+                    if (!retryableError && sseErrors.pending) {
+                        retryableError = retryableAnthropicStreamErrorFromRawEvent(sseErrors.pending);
+                    }
+                    if (retryableError) {
+                        try {
+                            await reader.cancel();
+                        }
+                        catch {
+                            /* ignore cancel failure */
+                        }
+                        releaseReader();
+                        throw retryableError;
+                    }
+                    const { ready } = splitToolPrefixRewriteBuffer(`${pending}${finalDecoded}`, true);
+                    if (ready)
+                        controller.enqueue(encoder.encode(ready));
+                    releaseReader();
+                    controller.close();
+                    return;
+                }
+                const decoded = decoder.decode(value, { stream: true });
+                const retryableError = updateSseErrorState(sseErrors, decoded);
+                if (retryableError) {
+                    try {
+                        await reader.cancel();
+                    }
+                    catch {
+                        /* ignore cancel failure */
+                    }
+                    releaseReader();
+                    throw retryableError;
+                }
+                const { ready, pending: nextPending } = splitToolPrefixRewriteBuffer(pending + decoded);
+                pending = nextPending;
+                if (ready)
+                    controller.enqueue(encoder.encode(ready));
+            }
+            catch (error) {
+                releaseReader();
+                throw error;
+            }
+        },
+        async cancel(reason) {
+            try {
+                await reader.cancel(reason);
+            }
+            finally {
+                releaseReader();
             }
-            let text = decoder.decode(value, { stream: true });
-            text = stripToolPrefix(text);
-            controller.enqueue(encoder.encode(text));
         },
     });
     return new Response(stream, {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sahiljassal/opencode-anthropic-auth",
-  "version": "2.1.0",
+  "version": "2.2.0",
   "repository": {
     "type": "git",
     "url": "git+https://github.com/shljsl75891/opencode-anthropic-auth.git"