opencode-anthropic-fix 0.2.0 → 0.3.0-beta.0

package/lib/mimicry/context-hint-threshold.mjs

@@ -0,0 +1,127 @@
+/**
+ * Claude Code 2.1.195 context-hint token accounting.
+ *
+ * The native bundle calls these values Pac, gao, Acp, aNn, and Ecp,
+ * respectively. Its internal transcript checks `type === "assistant"`; the
+ * plugin receives the equivalent wire shape as `role === "assistant"`.
+ *
+ * Genuine Claude Code also filters tool uses through `Hcp.has(block.name)`.
+ * In Claude Code 2.1.195, `Hcp` contains Read, Bash, PowerShell, Grep, Glob,
+ * WebSearch, WebFetch, Edit, and Write.
+ *
+ * @typedef {{ type?: string, id?: string, name?: string, tool_use_id?: string, content?: unknown, text?: unknown }} ContextBlock
+ * @typedef {{ role?: string, content?: unknown }} ContextMessage
+ */
+
+export const KEPT_TOOL_RESULT_GROUPS = 5;
+export const CONTEXT_HINT_TOKEN_THRESHOLD = 20_000;
+export const MEDIA_BLOCK_TOKENS = 2_000;
+export const CLEARED_MARKER = "[Old tool result content cleared]";
+export const PERSISTED_PREFIX = "<persisted-output>";
+export const CONTEXT_HINT_TOOL_NAMES = new Set([
+  "Read",
+  "Bash",
+  "PowerShell",
+  "Grep",
+  "Glob",
+  "WebSearch",
+  "WebFetch",
+  "Edit",
+  "Write",
+]);
+
+/**
+ * Genuine `If` estimator. This intentionally uses Math.round, unlike the
+ * plugin's other token estimators.
+ *
+ * @param {unknown} value
+ * @returns {number}
+ */
+export function estimateTextTokens(value) {
+  if (typeof value !== "string") return 0;
+  return Math.round(value.length / 4);
+}
+
+/**
+ * Genuine `Tcp` tool-result estimator.
+ *
+ * @param {ContextBlock} block
+ * @returns {number}
+ */
+export function estimateToolResultTokens(block) {
+  if (!block.content) return 0;
+  if (typeof block.content === "string") return estimateTextTokens(block.content);
+  if (!Array.isArray(block.content)) return 0;
+
+  return block.content.reduce((sum, item) => {
+    if (item?.type === "text") return sum + estimateTextTokens(item.text);
+    if (item?.type === "image" || item?.type === "document") return sum + MEDIA_BLOCK_TOKENS;
+    return sum;
+  }, 0);
+}
+
+/**
+ * Genuine `wcp` cleared-result check.
+ *
+ * @param {unknown} content
+ * @returns {boolean}
+ */
+export function isAlreadyCleared(content) {
+  return typeof content === "string" && (content === CLEARED_MARKER || content.startsWith(PERSISTED_PREFIX));
+}
+
+/**
+ * Genuine `vcp` collection order, adapted from the internal transcript's
+ * `type` discriminator to the wire message's `role` discriminator. Tool names
+ * are matched case-sensitively against Claude Code 2.1.195's native `Hcp` set.
+ *
+ * @param {ContextMessage[]} messages
+ * @returns {string[]}
+ */
+export function collectClearableToolUseIds(messages) {
+  const ids = [];
+
+  for (const message of messages) {
+    if (message?.role !== "assistant" || !Array.isArray(message.content)) continue;
+    for (const block of message.content) {
+      if (block?.type === "tool_use" && CONTEXT_HINT_TOOL_NAMES.has(block.name)) ids.push(block.id);
+    }
+  }
+
+  return ids;
+}
+
+/**
+ * Genuine `hao(messages, Pac)` saved-token calculation.
+ *
+ * @param {ContextMessage[]} messages
+ * @returns {number}
+ */
+export function computeContextHintTokensSaved(messages) {
+  const ids = collectClearableToolUseIds(messages);
+  const keep = Math.max(1, KEPT_TOOL_RESULT_GROUPS);
+  const keepSet = new Set(ids.slice(-keep));
+  const clearSet = new Set(ids.filter((id) => !keepSet.has(id)));
+  let tokensSaved = 0;
+
+  if (clearSet.size > 0) {
+    for (const message of messages) {
+      if (message?.role !== "user" || !Array.isArray(message.content)) continue;
+      for (const block of message.content) {
+        if (block?.type === "tool_result" && clearSet.has(block.tool_use_id) && !isAlreadyCleared(block.content)) {
+          tokensSaved += estimateToolResultTokens(block);
+        }
+      }
+    }
+  }
+
+  return tokensSaved;
+}
+
+/**
+ * @param {ContextMessage[]} messages
+ * @returns {boolean}
+ */
+export function shouldEmitContextHintBody(messages) {
+  return computeContextHintTokensSaved(messages) >= CONTEXT_HINT_TOKEN_THRESHOLD;
+}

package/lib/mimicry/headers.mjs

@@ -7,11 +7,6 @@ export const CLAUDE_CODE_BETA_FLAG = "claude-code-20250219";
 export const _EFFORT_BETA_FLAG = "effort-2025-11-24";
 export const FAST_MODE_BETA_FLAG = "fast-mode-2026-02-01";
 export const TOKEN_COUNTING_BETA_FLAG = "token-counting-2024-11-01";
-export const BEDROCK_UNSUPPORTED_BETAS = new Set([
-  "interleaved-thinking-2025-05-14",
-  "context-1m-2025-08-07",
-  "tool-search-tool-2025-10-19",
-]);
 export const HOST_SDK_BETAS_BLOCKLIST = new Set([
   "fine-grained-tool-streaming-2025-05-14",
   "structured-outputs-2025-11-13",
@@ -69,24 +64,6 @@ export function supportsWebSearch(model) {
   return /claude|sonnet|opus|haiku|gpt|gemini/i.test(model);
 }
 
-export function detectProvider(requestUrl) {
-  // Match Claude Code provider precedence first (env-driven), then URL fallback.
-  if (isTruthyEnv(process.env.CLAUDE_CODE_USE_BEDROCK)) return "bedrock";
-  if (isTruthyEnv(process.env.CLAUDE_CODE_USE_FOUNDRY)) return "foundry";
-  if (isTruthyEnv(process.env.CLAUDE_CODE_USE_ANTHROPIC_AWS)) return "anthropicAws";
-  if (isTruthyEnv(process.env.CLAUDE_CODE_USE_MANTLE)) return "mantle";
-  if (isTruthyEnv(process.env.CLAUDE_CODE_USE_VERTEX)) return "vertex";
-
-  if (!requestUrl) return "anthropic";
-  const host = requestUrl.hostname.toLowerCase();
-  if (host.includes("mantle")) return "mantle";
-  if (host.includes("anthropicaws")) return "anthropicAws";
-  if (host.includes("bedrock") || host.includes("amazonaws.com")) return "bedrock";
-  if (host.includes("aiplatform") || host.includes("vertex")) return "vertex";
-  if (host.includes("foundry") || host.includes("azure")) return "foundry";
-  return "anthropic";
-}
-
 export function parseRequestBodyMetadata(body, parsedBody) {
   const parsed =
     parsedBody ||
@@ -110,29 +87,75 @@ export function parseRequestBodyMetadata(body, parsedBody) {
   return { model, tools, messages, hasFileReferences };
 }
 
+/**
+ * Walk every object that may carry a stainless helper marker: a tool object, a
+ * message object, and each block nested under `.content`. Shared by the header
+ * builder and the body stripper on purpose — what is READ to build
+ * `x-stainless-helper` and what is REMOVED from the body must be the exact same
+ * traversal, or the two would drift and a marker would leak to the wire.
+ *
+ * @param {unknown[]} tools
+ * @param {unknown[]} messages
+ * @param {(carrier: Record<string, unknown>) => void} visit
+ */
+function walkStainlessHelperCarriers(tools, messages, visit) {
+  const walk = (value) => {
+    if (!value || typeof value !== "object") return;
+
+    visit(value);
+
+    if (Array.isArray(value.content)) {
+      for (const contentBlock of value.content) {
+        walk(contentBlock);
+      }
+    }
+  };
+
+  if (Array.isArray(tools)) for (const tool of tools) walk(tool);
+  if (Array.isArray(messages)) for (const message of messages) walk(message);
+}
+
 export function buildStainlessHelperHeader(tools, messages) {
   const helpers = new Set();
 
-  const collect = (value) => {
-    if (!value || typeof value !== "object") return;
-
+  walkStainlessHelperCarriers(tools, messages, (value) => {
     for (const key of STAINLESS_HELPER_KEYS) {
       if (typeof value[key] === "string" && value[key]) {
         helpers.add(value[key]);
       }
     }
+  });
 
-    if (Array.isArray(value.content)) {
-      for (const contentBlock of value.content) {
-        collect(contentBlock);
+  return Array.from(helpers).join(", ");
+}
+
+/**
+ * Remove the stainless helper markers from the request body, in place.
+ *
+ * The marker is an INTERNAL signal whose only purpose is to compute the
+ * `x-stainless-helper` header. It is not part of the Anthropic wire contract:
+ * the API has never known these keys, and the shared wire-compat package
+ * rejects the whole request with `INVALID_INPUT` when one reaches it. Call this
+ * AFTER `buildStainlessHelperHeader` so the header is still derived from the
+ * untouched body, and before the body is handed to the package or to `fetch`.
+ *
+ * @param {unknown[]} tools
+ * @param {unknown[]} messages
+ * @returns {number} how many marker keys were removed
+ */
+export function stripStainlessHelperMarkers(tools, messages) {
+  let removed = 0;
+
+  walkStainlessHelperCarriers(tools, messages, (value) => {
+    for (const key of STAINLESS_HELPER_KEYS) {
+      if (key in value) {
+        delete value[key];
+        removed += 1;
       }
     }
-  };
+  });
 
-  for (const tool of tools) collect(tool);
-  for (const message of messages) collect(message);
-
-  return Array.from(helpers).join(", ");
+  return removed;
 }
 
 export function isEffortCapableModel(model) {
@@ -145,7 +168,6 @@ export function buildAnthropicBetaHeader(
   incomingBeta,
   signatureEnabled,
   model,
-  provider,
   customBetas,
   strategy,
   requestPath,
@@ -177,8 +199,6 @@ export function buildAnthropicBetaHeader(
     return mergedBetas.join(",");
   }
 
-  const nonInteractive = isNonInteractiveMode();
-  const haiku = isHaikuModel(model);
   const isRoundRobin = strategy === "round-robin";
   const te = tokenEconomy || {};
 
@@ -228,17 +248,16 @@ export function buildAnthropicBetaHeader(
 
   // === CONDITIONAL BETAS (model/context-dependent) ===
 
-  // First-party provider predicate (OAuth path). Mirrors CC's ZO(provider) which
-  // excludes bedrock/vertex/mantle. context-management and effort are first-party
-  // only here, so they must never be pushed on bedrock/vertex.
-  const isFirstPartyProvider = provider !== "vertex" && provider !== "bedrock" && provider !== "mantle";
+  // The plugin is first-party OAuth only; Bedrock/Vertex/Foundry/Mantle support
+  // was removed, so the former isFirstPartyProvider predicate is always true and
+  // the gates below no longer test it.
 
   // v2.1.195: context-management-2025-06-27 is default-ON for first-party non-claude-3
   // models (incl. Haiku 4.5), mirroring CC's n0d(model) eligibility path. Earlier
   // analyses read only the separate USE_API_CONTEXT_MANAGEMENT env term (hardcoded
   // && false) and recorded it as off; 2.1.195 confirms it ships by default for modern
   // first-party models. Opt out via token_economy.context_management = false.
-  if (te.context_management !== false && isFirstPartyProvider && !CLAUDE_3_MODEL_RE.test(model)) {
+  if (te.context_management !== false && !CLAUDE_3_MODEL_RE.test(model)) {
     betas.push("context-management-2025-06-27");
   }
 
@@ -246,11 +265,12 @@ export function buildAnthropicBetaHeader(
   // (Opus 4.5/4.6/4.7/4.8, Sonnet 4.6), mirroring CC's Kw(model). Excluded for
   // claude-3-*, opus-4-0, opus-4-1, sonnet-4-0, sonnet-4-5, haiku-4-5. First-party
   // only; opt out via token_economy.effort = false.
-  if (te.effort !== false && isFirstPartyProvider && isEffortCapableModel(model)) {
+  if (te.effort !== false && isEffortCapableModel(model)) {
     betas.push(_EFFORT_BETA_FLAG);
   }
 
-  // v2.1.150: structured-outputs gated behind opt-in (CC has behind tengu_tool_pear flag).
+  // structured-outputs is gated by the caller supplying an output format.
+  // `tengu_tool_pear` instead gates `tool.strict = true` on the tool-schema path.
   if (te.structured_outputs && supportsStructuredOutputs(model)) {
     betas.push("structured-outputs-2025-12-15");
   }
@@ -268,21 +288,13 @@ export function buildAnthropicBetaHeader(
     betas.push("advisor-tool-2026-03-01");
   }
 
-  // context-hint-2026-04-09 — introduced in CC v2.1.110. Paired with the body
-  // field `context_hint: { enabled: true }` (injected in transformRequestBody).
-  // Real CC gates it on includeFirstPartyBetas + querySource startsWith
-  // "repl_main_thread" (i.e. interactive sessions on 1P Anthropic only).
-  // Server responds with hints suggesting which messages to forget for
-  // token-efficient context compaction. Sticky: disabled permanently on
-  // 400/409/529 errors referencing the hint. Users can opt out via
-  // token_economy.context_hint = false.
-  // Mimicry note: real CC gates context-hint on querySource.startsWith("repl_main_thread").
-  // We infer main-thread from body shape via classifyRequestRole and pass it via
-  // tokenEconomy.__requestRole. Treat absent marker as "main" for backward compat.
-  const _isMainThread = te.__requestRole == null || te.__requestRole === "main";
-  if (isFirstPartyProvider && !CLAUDE_3_MODEL_RE.test(model) && te.context_hint !== false && _isMainThread) {
-    betas.push("context-hint-2026-04-09");
-  }
+  // context-hint-2026-04-09 is NOT emitted, on any path, under any knob.
+  // The profile derived from the real Claude Code 2.1.195 binary sends neither
+  // this beta nor its paired `context_hint` body field. Emitting either makes
+  // the request distinguishable from the genuine client — the exact opposite of
+  // what this plugin exists to do. `token_economy.context_hint` is therefore
+  // deprecated rather than honoured: validateConfig warns on explicit opt-in so
+  // the switch is not a silent no-op. See lib/config.mjs.
 
   // Files API — scoped to file endpoints/references
   if (isFilesEndpoint || hasFileReferences) {
@@ -355,11 +367,6 @@ export function buildAnthropicBetaHeader(
     mergedBetas = mergedBetas.filter((beta) => !EXPERIMENTAL_BETA_FLAGS.has(beta));
   }
 
-  // Remove betas unsupported by Bedrock
-  if (provider === "bedrock") {
-    mergedBetas = mergedBetas.filter((beta) => !BEDROCK_UNSUPPORTED_BETAS.has(beta));
-  }
-
   return mergedBetas.join(",");
 }
 
@@ -415,7 +422,6 @@ export function buildRequestHeaders(
   // Preserve all incoming beta headers while ensuring OAuth requirements
   const incomingBeta = requestHeaders.get("anthropic-beta") || "";
   const { model, tools, messages, hasFileReferences } = parseRequestBodyMetadata(requestBody);
-  const provider = detectProvider(requestUrl);
   // Detect fast mode structurally from the already-transformed body string.
   // transformRequestBody runs before buildRequestHeaders (see index.mjs:2986 vs 3039),
   // so speed:"fast" is already present when fast mode fired. Same detection idiom as the
@@ -428,7 +434,6 @@ export function buildRequestHeaders(
     incomingBeta,
     signature.enabled,
     model,
-    provider,
     signature.customBetas,
     signature.strategy,
     requestUrl?.pathname,

package/lib/mimicry/request-body.mjs

@@ -57,7 +57,7 @@ export const CORE_TOOL_NAMES = new Set([
  * Preserves behaviors E1-E7.
  *
  * @param {string | undefined} body
- * @param {{enabled: boolean, claudeCliVersion: string, promptCompactionMode: 'minimal' | 'off', provider?: string}} signature
+ * @param {{enabled: boolean, claudeCliVersion: string, promptCompactionMode: 'minimal' | 'off', provider?: string, useAdapter?: boolean}} signature
  * @param {{persistentUserId: string, sessionId: string, accountId: string, turns?: number, usedTools?: Set<string>}} runtime
  * @param {string} [betaHeader] - Pre-computed anthropic-beta header value to inject into the body.
  * @param {import('../config.mjs').AnthropicAuthConfig} [config] - Plugin configuration for output cap
@@ -66,8 +66,6 @@ export const CORE_TOOL_NAMES = new Set([
 export function transformRequestBody(body, signature, runtime, betaHeader, config) {
   if (!body || typeof body !== "string") return body;
 
-  const TOOL_PREFIX = "mcp_";
-
   try {
     const parsed = JSON.parse(body);
     // Diagnostic: log the incoming model + thinking/speed shape so model-routing
@@ -89,17 +87,10 @@ export function transformRequestBody(body, signature, runtime, betaHeader, confi
     if (config?.output_cap?.enabled) {
       parsed.max_tokens = resolveMaxTokens(parsed, config);
     }
-    // Bedrock requires betas in the body as "anthropic_beta" (underscore) since it
-    // doesn't forward custom HTTP headers. First-party API rejects "betas" in body
-    // with "Extra inputs are not permitted" — betas are header-only for first-party.
-    if (signature.enabled && betaHeader && signature.provider === "bedrock") {
-      const betaArray = betaHeader
-        .split(",")
-        .map((b) => b.trim())
-        .filter(Boolean)
-        .filter((b) => b !== "oauth-2025-04-20");
-      parsed.anthropic_beta = betaArray;
-    }
+    // The body-level `anthropic_beta` array existed only for Bedrock, which does
+    // not forward custom HTTP headers. Multi-provider support was removed, and
+    // the first-party API rejects betas in the body with "Extra inputs are not
+    // permitted" — they are header-only here.
     // Strip any incoming "betas" field — API rejects it as unknown
     if (Object.prototype.hasOwnProperty.call(parsed, "betas")) {
       delete parsed.betas;
@@ -332,7 +323,7 @@ export function transformRequestBody(body, signature, runtime, betaHeader, confi
       requestRole: runtime?.requestRole,
       leanNonMain: config?.token_economy?.lean_system_non_main === true,
       // Simple-system-prompt mode for Opus 4.7+ (CC v2.1.133+ parity, gated by
-      // GrowthBook `tengu_vellum_lantern` in real CC). Plugin gate: model
+      // GrowthBook `tengu_velvet_cascade` in real CC). Plugin gate: model
       // eligibility + opt-in flag. See buildSystemPromptBlocks for what it
       // strips (anti-verbosity boilerplate only; identity/billing untouched).
       simpleSystemPrompt: config?.token_economy?.simple_system_prompt === true,
@@ -340,6 +331,11 @@ export function transformRequestBody(body, signature, runtime, betaHeader, confi
       // Mirrors real CC's `--workload <tag>` CLI flag (process-scoped, used by
       // SDK daemon callers spawning cron subprocesses). Empty string -> omit.
       workload: typeof config?.signature_emulation?.workload === "string" ? config.signature_emulation.workload : "",
+      // Adapter path only: suppress the canonical billing + identity blocks
+      // because the shared package prepends its own. Every other stage of the
+      // system pipeline (anti-verbosity, subagent CC-prefix, title-generator
+      // swap, dedupe, compaction, tailing, token-budget block) still runs.
+      suppressCanonicalBlocks: signature.useAdapter === true,
     };
     // Sanitize system prompt and optionally inject Claude Code identity/billing blocks.
     parsed.system = buildSystemPromptBlocks(normalizeSystemTextBlocks(parsed.system), signatureWithModel);
@@ -379,7 +375,13 @@ export function transformRequestBody(body, signature, runtime, betaHeader, confi
       }
     }
 
-    if (signature.enabled) {
+    // Adapter path only: the shared package composes `metadata.user_id` itself
+    // from the runtime correlation triple, and the two env features become
+    // `metadataOverrides` on the transport instead (see adapter-input.mjs).
+    // Writing metadata here as well would collide with the package's own
+    // composition. CONDITIONAL, never unconditional: the legacy path still
+    // needs buildRequestMetadata.
+    if (signature.enabled && signature.useAdapter !== true) {
       const currentMetadata =
         parsed.metadata && typeof parsed.metadata === "object" && !Array.isArray(parsed.metadata)
           ? parsed.metadata
@@ -587,14 +589,11 @@ export function transformRequestBody(body, signature, runtime, betaHeader, confi
       }
     }
 
-    // Context hint (CC v2.1.110+): pair the context-hint-2026-04-09 beta with
-    // the body field context_hint:{enabled:true}. Extracted from CC bundle
-    // buildRequestParams(): `{betaHeader, body:{context_hint:{enabled:!0}}}`.
-    // When active, the server responds with optional hints suggesting which
-    // messages to drop for token-efficient context compaction.
-    if (betaHeader && betaHeader.includes("context-hint-2026-04-09") && !parsed.context_hint) {
-      parsed.context_hint = { enabled: true };
-    }
+    // The `context_hint` body field is never injected, on either path. The
+    // profile derived from the real Claude Code 2.1.195 binary does not send it
+    // (nor the paired context-hint-2026-04-09 beta), so emitting it would be a
+    // fingerprint. `token_economy.context_hint` is deprecated in lib/config.mjs,
+    // which warns on explicit opt-in rather than silently doing nothing.
 
     // Fast mode: inject speed parameter for fast-mode-eligible models.
     // Per Anthropic fast-mode docs, `speed: "fast"` (beta fast-mode-2026-02-01) is

package/lib/mimicry/system-prompt.mjs

@@ -6,6 +6,8 @@ const BILLING_HASH_SALT = "59cf53e54c78";
 const BILLING_HASH_INDICES = [4, 7, 20];
 
 export const CLAUDE_CODE_IDENTITY_STRING = "You are Claude Code, Anthropic's official CLI for Claude.";
+/** Marker that identifies the canonical billing block inside a system array. */
+export const BILLING_HEADER_PREFIX = "x-anthropic-billing-header:";
 const KNOWN_IDENTITY_STRINGS = new Set([
   CLAUDE_CODE_IDENTITY_STRING,
   "You are Claude Code, Anthropic's official CLI for Claude, running within the Claude Agent SDK.",
@@ -25,11 +27,12 @@ const COMPACT_TITLE_GENERATOR_SYSTEM_PROMPT = [
 
 /**
  * Anti-verbosity system prompt text.
- * Extracted from CC v2.1.100 (gated on quiet_salted_ember A/B test for Opus 4.6).
+ * Text verified against the CC v2.1.195 binary; the heading corrects a stale
+ * v2.1.100 value.
  * Significantly reduces output token count by instructing the model to be concise.
  */
 const ANTI_VERBOSITY_SYSTEM_PROMPT = [
-  "# Communication style",
+  "# Text output (does not apply to tool calls)",
   "Assume users can't see most tool calls or thinking — only your text output. Before your first tool call, state in one sentence what you're about to do. While working, give short updates at key moments: when you find something, when you change direction, or when you hit a blocker. Brief is good — silent is not. One sentence per update is almost always enough.",
   "",
   "Don't narrate your internal deliberation. User-facing text should be relevant communication to the user, not a running commentary on your thought process. State results and decisions directly, and focus user-facing text on relevant updates for the user.",
@@ -43,13 +46,6 @@ const ANTI_VERBOSITY_SYSTEM_PROMPT = [
   "In code: default to writing no comments. Never write multi-paragraph docstrings or multi-line comment blocks — one short line max. Don't create planning, decision, or analysis documents unless the user asks for them — work from conversation context, not intermediate files.",
 ].join("\n");
 
-/**
- * Numeric length anchors text.
- * Extracted from CC v2.1.100. Hard word-count limits for output.
- */
-const NUMERIC_LENGTH_ANCHORS_PROMPT =
-  "Length limits: keep text between tool calls to ≤25 words. Keep final responses to ≤100 words unless the task requires more detail.";
-
 // Max system prompt length that passes CC billing validation.
 // The server pattern-matches the system prompt against the real CC prompt.
 // Opencode's customizations after ~5800 chars diverge and trigger extra usage billing.
@@ -98,7 +94,7 @@ export function computeBillingCacheHash(firstUserMessage, version) {
 
 /**
  * Models eligible for the "simple system prompt" mode that real CC ships in
- * v2.1.133+ under GrowthBook flag `tengu_vellum_lantern` (or forced via
+ * v2.1.133+ under GrowthBook flag `tengu_velvet_cascade` (or forced via
  * `CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=1`).
  *
  * Real CC's `sR9` matches: `claude-opus-4-7`, any model with `-eap` (early
@@ -112,7 +108,7 @@ export function isSimpleSystemPromptEligible(model) {
   if (!model) return false;
   if (isOpus47Model(model) || isOpus48Model(model)) return true;
   // -eap suffix variant, e.g. "claude-opus-4-7-eap" or "claude-opus-4-7-eap[1m]"
-  if (/-eap(?:\[|$|-)/i.test(model)) return true;
+  if (/-eap($|\[)/i.test(model)) return true;
   return false;
 }
 
@@ -126,12 +122,11 @@ export function isSimpleSystemPromptEligible(model) {
  *
  * @param {string} version - CLI version (e.g., "2.1.97")
  * @param {string} [firstUserMessage] - First user message text for fingerprint computation
- * @param {string} [provider] - API provider ("anthropic" | "bedrock" | "vertex" | "foundry" | "anthropicAws" | "mantle")
  * @param {string} [workloadOverride] - Explicit workload tag from config (`signature_emulation.workload`).
  *   Takes precedence over `CLAUDE_CODE_WORKLOAD` env var. Mirrors real CC's `--workload <tag>` flag.
  * @returns {string}
  */
-export function buildAnthropicBillingHeader(version, firstUserMessage, provider, workloadOverride) {
+export function buildAnthropicBillingHeader(version, firstUserMessage, workloadOverride) {
   if (isFalsyEnv(process.env.CLAUDE_CODE_ATTRIBUTION_HEADER)) return "";
   // Real CC sends cc_entrypoint=cli (confirmed via proxy capture).
   const entrypoint = process.env.CLAUDE_CODE_ENTRYPOINT || "cli";
@@ -144,15 +139,15 @@ export function buildAnthropicBillingHeader(version, firstUserMessage, provider,
   const ccVersion = `${version}.${fingerprint}`;
   // cch: v2.1.97 sends static "cch=00000" — xxHash64 attestation was removed.
   // The server uses the PRESENCE of cch=00000 as a CC identification signal.
-  const cchDisabled = provider === "bedrock" || provider === "anthropicAws" || provider === "mantle";
-  const cchPart = cchDisabled ? "" : " cch=00000;";
+  // The plugin is first-party only, so cch is always emitted; the former
+  // bedrock/anthropicAws/mantle suppression went away with multi-provider support.
+  const cchPart = " cch=00000;";
   // Build workload part (upstream concatenates directly, no regex replace).
   // Config wins over env so `signature_emulation.workload` is portable across hosts.
   let workloadPart = "";
   const workload = workloadOverride || process.env.CLAUDE_CODE_WORKLOAD;
-  if (workload && !cchDisabled) {
+  if (workload) {
     // QA fix M5: sanitize workload value to prevent header injection.
-    // Real CC skips cc_workload on bedrock/anthropicAws/mantle (same gate as cch).
     const safeWorkload = String(workload).replace(/[;\s\r\n]/g, "_");
     if (safeWorkload) workloadPart = ` cc_workload=${safeWorkload};`;
   }
@@ -550,7 +545,7 @@ export function buildSystemPromptBlocks(system, signature) {
   // Anti-verbosity injection (CC v2.1.100 quiet_salted_ember equivalent).
   // Applies to Opus 4.6 / 4.7 for non-title-generator requests.
   //
-  // Simple-system-prompt gate (CC v2.1.133+ `tengu_vellum_lantern` equivalent):
+  // Simple-system-prompt gate (CC v2.1.133+ `tengu_velvet_cascade` equivalent):
   // when `signature.simpleSystemPrompt` is true AND the model is eligible
   // (Opus 4.7, -eap variants), skip the anti-verbosity boilerplate. Real CC's
   // simple-prompt mode strips this kind of post-hoc behavioral hand-holding on
@@ -570,9 +565,6 @@ export function buildSystemPromptBlocks(system, signature) {
     if (avConfig?.enabled !== false) {
       sanitized.push({ type: "text", text: ANTI_VERBOSITY_SYSTEM_PROMPT });
     }
-    if (avConfig?.length_anchors !== false) {
-      sanitized.push({ type: "text", text: NUMERIC_LENGTH_ANCHORS_PROMPT });
-    }
   }
 
   if (!signature.enabled) {
@@ -598,7 +590,6 @@ export function buildSystemPromptBlocks(system, signature) {
   const billingHeader = buildAnthropicBillingHeader(
     signature.claudeCliVersion,
     signature.firstUserMessage,
-    signature.provider,
     signature.workload,
   );
 
@@ -617,9 +608,21 @@ export function buildSystemPromptBlocks(system, signature) {
   // Run the real CC splitSysPromptPrefix algorithm to get blocks with cacheScope
   const scopedBlocks = splitSysPromptPrefix(sanitized, billingHeader || undefined, identityString, useBoundaryMode);
 
+  // Adapter path only: the shared package prepends its OWN billing block
+  // (index 0) and identity block (index 1), each with its own cache_control
+  // marker. Emitting ours too would duplicate both blocks on the wire.
+  //
+  // This is CONDITIONAL, never unconditional: the legacy path (any request the
+  // signature-emulation adapter declines — non-Anthropic provider, or emulation
+  // off) still builds the request through buildRequestHeaders and depends on
+  // these two blocks being present. Dropping them for every caller would break
+  // that path silently — no error, just a wrong request on the wire.
+  const emittedBlocks =
+    signature.suppressCanonicalBlocks === true ? dropCanonicalPrefixBlocks(scopedBlocks, identityString) : scopedBlocks;
+
   // Convert scoped blocks to wire format using getCacheControlForScope
   // (mirrors real CC buildSystemPromptBlocks → map + getCacheControl)
-  return scopedBlocks.map((block) => {
+  return emittedBlocks.map((block) => {
     const cc = getCacheControlForScope(block.cacheScope, effectiveCachePolicy);
     return {
       type: "text",
@@ -629,6 +632,34 @@ export function buildSystemPromptBlocks(system, signature) {
   });
 }
 
+/**
+ * Drop the two canonical prefix blocks the shared package prepends itself.
+ *
+ * Both are matched with the SAME predicates `splitSysPromptPrefix` uses to
+ * recognise them on input, and only the FIRST occurrence of each is removed, so
+ * a caller-supplied block that happens to repeat the identity string survives.
+ *
+ * @param {{text: string, cacheScope: string | null}[]} blocks
+ * @param {string} identityString
+ * @returns {{text: string, cacheScope: string | null}[]}
+ */
+function dropCanonicalPrefixBlocks(blocks, identityString) {
+  let billingDropped = false;
+  let identityDropped = false;
+  return blocks.filter((block) => {
+    const text = typeof block.text === "string" ? block.text : "";
+    if (!billingDropped && text.startsWith(BILLING_HEADER_PREFIX)) {
+      billingDropped = true;
+      return false;
+    }
+    if (!identityDropped && text === identityString) {
+      identityDropped = true;
+      return false;
+    }
+    return true;
+  });
+}
+
 export function getCachedCCPrompt() {
   return cachedCCPrompt;
 }