opencode-anthropic-fix 0.1.32 → 0.2.1

package/lib/mimicry/request-body.mjs

@@ -0,0 +1,682 @@
+import {
+  buildSystemPromptBlocks,
+  normalizeSystemTextBlocks,
+  compactToolDescription,
+  tailSystemBlock,
+  isTitleGeneratorSystemBlocks,
+} from "./system-prompt.mjs";
+import {
+  applyTtlThinkingStrip,
+  applySessionToolResultDedupe,
+  applyProactiveMicrocompact,
+  applyTrailingSummaryTrim,
+  applyToolResultDedupe,
+  applyStableToolOrdering,
+  applyToolSchemaDeferral,
+  applyAdaptiveThinkingZero,
+  estimatePromptTokensFromParsed,
+  parseNaturalLanguageBudget,
+  injectTokenBudgetBlock,
+} from "../token-economy/transforms.mjs";
+import { resolveCacheTtl, shouldPlaceToolBreakpoint } from "./cache.mjs";
+import {
+  isAdaptiveThinkingModel,
+  isOpus46Model,
+  isOpus47Model,
+  isOpus48Model,
+  normalizeThinkingBlock,
+  CLAUDE_3_MODEL_RE,
+} from "./models.mjs";
+import {
+  resolveMaxTokens,
+  extractFirstUserMessageText,
+  buildRequestMetadata,
+  stripSlashCommandMessages,
+  repairOrphanedToolUseBlocks,
+} from "./request-helpers.mjs";
+import { isFalsyEnv } from "../env.mjs";
+import { sessionMetrics } from "../session-metrics.mjs";
+
+// Core tool names (CC PascalCase) that are always eager-loaded.
+export const CORE_TOOL_NAMES = new Set([
+  "Bash",
+  "Read",
+  "Glob",
+  "Grep",
+  "Edit",
+  "Write",
+  "WebFetch",
+  "TodoWrite",
+  "Skill",
+  "Task",
+  "Compress",
+]);
+
+/**
+ * Transform the request body: system prompt sanitization and tool prefixing.
+ * Preserves behaviors E1-E7.
+ *
+ * @param {string | undefined} body
+ * @param {{enabled: boolean, claudeCliVersion: string, promptCompactionMode: 'minimal' | 'off', provider?: string}} 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
+ * @returns {string | undefined}
+ */
+export function transformRequestBody(body, signature, runtime, betaHeader, config) {
+  if (!body || typeof body !== "string") return body;
+
+  try {
+    const parsed = JSON.parse(body);
+    // Diagnostic: log the incoming model + thinking/speed shape so model-routing
+    // issues (e.g. an unrecognized fast variant, or "No available account"
+    // symptoms tied to a specific model) are observable with /anthropic set debug on.
+    // NOTE: transformRequestBody is a top-level function outside the plugin
+    // closure where debugLog() lives, so we mirror debugLog's behaviour directly
+    // (console.error + same prefix, gated on config.debug).
+    if (config?.debug) {
+      console.error("[opencode-anthropic-auth]", "transformRequestBody: incoming model", {
+        model: parsed.model,
+        adaptive: isAdaptiveThinkingModel(parsed.model || ""),
+        thinkingType: parsed.thinking?.type,
+        hasEffort: parsed.effort !== undefined || parsed.output_config?.effort !== undefined,
+        speed: parsed.speed,
+      });
+    }
+    // Output cap: resolve max_tokens before any other body transforms
+    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;
+    }
+    // Strip any incoming "betas" field — API rejects it as unknown
+    if (Object.prototype.hasOwnProperty.call(parsed, "betas")) {
+      delete parsed.betas;
+    }
+    // Normalize thinking block for adaptive (Opus 4.6 / Sonnet 4.6) vs manual (older models).
+    // Real CC always sends thinking:{type:"adaptive"} for adaptive models even if the
+    // upstream SDK didn't include it. Inject it when missing to match the fingerprint.
+    if (Object.prototype.hasOwnProperty.call(parsed, "thinking")) {
+      parsed.thinking = normalizeThinkingBlock(parsed.thinking, parsed.model || "");
+    } else if (parsed.model && isAdaptiveThinkingModel(parsed.model)) {
+      parsed.thinking = { type: "adaptive" };
+    }
+
+    // Fingerprint fix: real Claude Code v2.1.87+ nests the effort control inside
+    // `output_config.effort` (via Lyz() in cli.js). opencode's provider transform
+    // for variant=max on Opus 4.6 / Sonnet 4.6 sets `effort` at the top level,
+    // which causes Anthropic's server to fingerprint the body as non-CC and bill
+    // it as pay-as-you-go — surfacing as "You're out of extra usage" even on a
+    // valid Max plan. Move it into output_config when we're talking to an
+    // adaptive-thinking model so the wire shape matches real CC.
+    if (typeof parsed.effort === "string" && parsed.model && isAdaptiveThinkingModel(parsed.model)) {
+      if (!parsed.output_config || typeof parsed.output_config !== "object") {
+        parsed.output_config = {};
+      }
+      if (!("effort" in parsed.output_config)) {
+        parsed.output_config.effort = parsed.effort;
+      }
+      delete parsed.effort;
+    } else if (Object.prototype.hasOwnProperty.call(parsed, "effort")) {
+      // Non-adaptive models never carry a top-level effort in real CC — strip it
+      // to avoid polluting the fingerprint for models like Haiku.
+      delete parsed.effort;
+    }
+
+    // Claude Code v2.1.117+: default effort for Pro/Max on adaptive models is
+    // "high" (was "medium"). If the host omits effort entirely, inject the
+    // default so the wire shape matches real CC.
+    if (parsed.model && isAdaptiveThinkingModel(parsed.model)) {
+      if (!parsed.output_config || typeof parsed.output_config !== "object") {
+        parsed.output_config = {};
+      }
+      if (!("effort" in parsed.output_config)) {
+        parsed.output_config.effort = "high";
+      }
+    }
+
+    // Claude Code temperature rule: when extended thinking is active (any type),
+    // temperature must be omitted (undefined). Otherwise default to 1.
+    const thinkingActive =
+      parsed.thinking &&
+      typeof parsed.thinking === "object" &&
+      (parsed.thinking.type === "adaptive" || parsed.thinking.type === "enabled");
+    if (thinkingActive) {
+      // Anthropic API rejects temperature when thinking is enabled
+      delete parsed.temperature;
+
+      // Claude Code v2.1.84: inject the context_management body field ONLY when the
+      // user has explicitly opted in via token_economy.context_management. Note: as of
+      // v2.1.195 the context-management *beta header* is default-ON for first-party
+      // non-claude-3 models (see buildAnthropicBetaHeader), but the body field stays
+      // opt-in. A top-level context_management field WITHOUT the beta is rejected with
+      // "context_management: Extra inputs are not permitted"; the beta being present
+      // without the field is fine. Gating the field on explicit opt-in keeps the field
+      // ⊆ beta invariant, so the 400 never fires.
+      if (
+        config?.token_economy?.context_management &&
+        !CLAUDE_3_MODEL_RE.test(parsed.model || "") &&
+        !parsed.context_management
+      ) {
+        parsed.context_management = {
+          edits: [{ type: "clear_thinking_20251015", keep: "all" }],
+        };
+      }
+    } else {
+      // Claude Code always uses temperature: 1 for non-thinking requests (RE doc §5.2, never 0)
+      parsed.temperature = 1;
+    }
+
+    // Strip leaked /anthropic slash command messages from conversation history.
+    // OpenCode may include command text and sendCommandMessage output as regular
+    // user messages even when output.noReply = true was set. Filter them out
+    // so the agent never sees /anthropic commands in its context.
+    if (Array.isArray(parsed.messages)) {
+      parsed.messages = stripSlashCommandMessages(parsed.messages);
+    }
+
+    // === Token economy: layered message/history compaction ===
+    // Only applies to main-thread requests (subagents/title-gen stay untouched).
+    // Strategies stack in order: TTL thinking strip → proactive microcompact →
+    // trailing-summary trim → tool_result dedupe. Each is independently gated.
+    const te = config?.token_economy || {};
+    const tes = runtime?.tokenEconomySession;
+    const isMainRole = runtime?.requestRole === "main" || runtime?.requestRole == null;
+
+    // `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;
+        const res = applyTtlThinkingStrip(parsed.messages, {
+          lastClearMs: tes.lastThinkingStripMs,
+          ttlMs,
+        });
+        if (res.changed) {
+          parsed.messages = res.messages;
+          tes.lastThinkingStripMs = res.ranStripAt;
+          tes.thinkingStripped += res.cleared;
+        }
+      }
+
+      // (1b) Session-wide reproducible-tool result dedupe (Phase C C3, opt-in)
+      // Pure over message history; runs before microcompact so dedup'd stubs
+      // are visible to downstream size estimation. Gated by
+      // token_economy_strategies.tool_result_dedupe_session_wide (default off).
+      if (config?.token_economy_strategies?.tool_result_dedupe_session_wide === true) {
+        const res = applySessionToolResultDedupe(parsed.messages);
+        if (res.changed) parsed.messages = res.messages;
+      }
+
+      // (2) Proactive microcompact (client-side, pre-422)
+      if (te.proactive_microcompact !== false) {
+        const estimated = estimatePromptTokensFromParsed(parsed);
+        const cw = 200_000; // conservative — 1M models still benefit
+        const res = applyProactiveMicrocompact(parsed.messages, {
+          estimatedTokens: estimated,
+          contextWindow: cw,
+          percent: te.microcompact_percent ?? 70,
+          keepRecent: te.microcompact_keep_recent ?? 8,
+        });
+        if (res.changed) {
+          parsed.messages = res.messages;
+          tes.lastMicrocompactMs = Date.now();
+          tes.toolResultsCompacted += res.cleared;
+        }
+      }
+
+      // (3) Trailing-summary trim (opt-in)
+      if (te.trailing_summary_trim === true) {
+        const res = applyTrailingSummaryTrim(parsed.messages);
+        if (res.changed) parsed.messages = res.messages;
+      }
+
+      // (4) Cross-turn tool_result dedupe (opt-in)
+      if (te.tool_result_dedupe === true) {
+        const SAFE_READ_TOOLS = new Set(["Read", "Grep", "Glob", "LS", "BashOutput"]);
+        const res = applyToolResultDedupe(parsed.messages, {
+          seen: tes.seenContentHashes,
+          safeTools: SAFE_READ_TOOLS,
+        });
+        if (res.changed) parsed.messages = res.messages;
+      }
+    }
+
+    // === Token economy: tool-array transforms (stable ordering, deferral) ===
+    if (!conservative && isMainRole && Array.isArray(parsed.tools)) {
+      if (te.stable_tool_ordering !== false) {
+        parsed.tools = applyStableToolOrdering(parsed.tools);
+      }
+      if (Array.isArray(te.deferred_tool_names) && te.deferred_tool_names.length > 0) {
+        // "Invoked" means any assistant message in the convo has used the tool.
+        const invoked = new Set();
+        if (Array.isArray(parsed.messages)) {
+          for (const m of parsed.messages) {
+            if (m.role !== "assistant" || !Array.isArray(m.content)) continue;
+            for (const b of m.content) {
+              if (b?.type === "tool_use" && typeof b.name === "string") invoked.add(b.name);
+            }
+          }
+        }
+        const res = applyToolSchemaDeferral(parsed.tools, {
+          deferred: new Set(te.deferred_tool_names),
+          invoked,
+        });
+        parsed.tools = res.tools;
+      }
+    }
+
+    // === Token economy: adaptive thinking zero-out for trivial follow-ups ===
+    if (isMainRole && te.adaptive_thinking_zero_simple !== false) {
+      applyAdaptiveThinkingZero(parsed);
+    }
+
+    // QA fix H2: avoid mutating the signature parameter; capture modelId locally
+    const modelId = parsed.model || "";
+    // Extract first user message text for billing hash computation (cch)
+    const firstUserMessage = extractFirstUserMessageText(parsed.messages);
+    // Resolve the prompt-cache TTL ONCE so system blocks, tools, and messages
+    // all share the SAME ttl. Anthropic processes cache_control blocks in the
+    // order tools → system → messages and rejects any ttl='1h' block that comes
+    // AFTER a ttl='5m' block. Before this fix the role-scoped 5m downgrade was
+    // applied only to tools/messages (resolveCacheTtl in the breakpoint loop)
+    // while the system blocks kept the configured 1h — so subagent requests
+    // (5m tools, then 1h system) tripped:
+    //   "system.1.cache_control.ttl: a ttl='1h' cache_control block must not
+    //    come after a ttl='5m' cache_control block".
+    // Threading it through buildSystemPromptBlocks keeps every cache_control ttl
+    // consistent (real CC derives the ttl uniformly from querySource).
+    const baseCachePolicy = signature.cachePolicy || { ttl: "1h", ttl_supported: true };
+    const cachingEnabledForTtl = baseCachePolicy.ttl !== "off" && baseCachePolicy.ttl_supported !== false;
+    const resolvedCacheTtl = cachingEnabledForTtl
+      ? resolveCacheTtl({
+          configuredTtl: baseCachePolicy.ttl || "1h",
+          roleScopedTtl: config?.token_economy?.role_scoped_cache_ttl !== false,
+          isMainForCache: runtime?.requestRole === "main" || runtime?.requestRole == null,
+          isSubagent: runtime?.isSubagent === true,
+          env: process.env,
+        })
+      : baseCachePolicy.ttl;
+    const signatureWithModel = {
+      ...signature,
+      modelId,
+      // Override cachePolicy.ttl with the role/subagent-resolved ttl so the
+      // system blocks match the tool/message breakpoint ttl (see comment above).
+      cachePolicy: { ...baseCachePolicy, ttl: resolvedCacheTtl },
+      firstUserMessage,
+      antiVerbosity: config?.anti_verbosity,
+      // Role-aware system-prompt leaning: for non-main-thread requests (title,
+      // small, empty shapes) strip billing identity + CC identity injection.
+      // Title-gen path is handled separately by isTitleGeneratorSystemBlocks().
+      // Default off — opt-in via `token_economy.lean_system_non_main: true`.
+      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
+      // 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,
+      // Workload tag for x-anthropic-billing-header `cc_workload=` segment.
+      // 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 : "",
+    };
+    // Sanitize system prompt and optionally inject Claude Code identity/billing blocks.
+    parsed.system = buildSystemPromptBlocks(normalizeSystemTextBlocks(parsed.system), signatureWithModel);
+
+    // Strategy 5 — System prompt tailing: after N turns, trim large system blocks
+    // to essential sections only. The model has internalized verbose instructions
+    // (shell strategy, package manager tables, delegation protocols) by this point.
+    // Preserves: first paragraph (identity/role), lines containing MUST/NEVER/CRITICAL/
+    // IMPORTANT, section headers, and short blocks. Drops verbose body paragraphs.
+    const tailThreshold = signature.systemPromptTailTurns ?? 6;
+    if (signature.systemPromptTailing === true && runtime.turns >= tailThreshold && Array.isArray(parsed.system)) {
+      const maxChars = signature.systemPromptTailMaxChars ?? 2000;
+      for (let i = 0; i < parsed.system.length; i++) {
+        const block = parsed.system[i];
+        if (block.type === "text" && block.text && block.text.length > maxChars * 2) {
+          block.text = tailSystemBlock(block.text, maxChars, tailThreshold);
+        }
+      }
+    }
+
+    // Token budget (A9): parse NL budget from last user message, inject status block
+    if (config?.token_budget?.enabled && Array.isArray(parsed.messages)) {
+      const budgetExpr = parseNaturalLanguageBudget(parsed.messages);
+      if (budgetExpr > 0) {
+        sessionMetrics.tokenBudget.limit = budgetExpr;
+      } else if (config.token_budget.default > 0 && sessionMetrics.tokenBudget.limit === 0) {
+        sessionMetrics.tokenBudget.limit = config.token_budget.default;
+      }
+      // If budget is active, inject status into system prompt
+      if (sessionMetrics.tokenBudget.limit > 0) {
+        const threshold = config.token_budget.completion_threshold ?? 0.9;
+        parsed.system = injectTokenBudgetBlock(parsed.system, sessionMetrics.tokenBudget, threshold);
+        // Soft stop: if we've exceeded the threshold, cap max_tokens to 1
+        if (sessionMetrics.tokenBudget.used >= sessionMetrics.tokenBudget.limit * threshold) {
+          parsed.max_tokens = 1;
+        }
+      }
+    }
+
+    if (signature.enabled) {
+      const currentMetadata =
+        parsed.metadata && typeof parsed.metadata === "object" && !Array.isArray(parsed.metadata)
+          ? parsed.metadata
+          : {};
+      parsed.metadata = {
+        ...currentMetadata,
+        ...buildRequestMetadata({
+          persistentUserId: runtime.persistentUserId,
+          accountId: runtime.accountId,
+          sessionId: runtime.sessionId,
+        }),
+      };
+    }
+
+    // Cache breakpoint optimization: add cache_control to the last content block
+    // of each user/assistant message for maximum prefix caching.
+    // Skip for round-robin strategy (cache defeated by account rotation).
+    // Skip for title generators / fire-and-forget queries: these are one-shot
+    // requests that don't benefit from caching and would pollute the cache pool.
+    const isTitleGen = isTitleGeneratorSystemBlocks(parsed.system || []);
+    if (
+      signature.enabled &&
+      signature.cachePolicy?.ttl !== "off" &&
+      signature.cachePolicy?.ttl_supported !== false &&
+      !isTitleGen
+    ) {
+      // Strip ALL incoming cache_control from tools and messages to prevent
+      // TTL ordering violations (host SDK may set ttl=5m which conflicts with
+      // our system prompt ttl=1h). Then add our own to the last user message
+      // (matching real CC behavior seen in proxy capture).
+      //
+      // Role-scoped TTL: real CC's REH(querySource) decides the `ttl:"1h"` field.
+      // Decompiled from CC 2.1.154 (function REH, offset 225174828), precedence:
+      //   1. FORCE_PROMPT_CACHING_5M  => 5m (highest priority)
+      //   2. ENABLE_PROMPT_CACHING_1H => 1h
+      //   3. no tengu support / overage => 5m
+      //   4. querySource ∈ allowlist [repl_main_thread*, sdk, auto_mode, ...] => 1h, else 5m
+      // We map step 4 via classifyRequestRole: main (interactive/auto_mode thread)
+      // → 1h; side-queries (title-gen, etc.) → 5m. The env vars are honored
+      // exactly as CC does, giving the user a manual override + mimicry fidelity.
+      // Reuse the request-wide resolved ttl (computed above where
+      // signatureWithModel is built) so tools/messages match the system blocks
+      // and never trip the 1h-after-5m ordering rule on subagent requests.
+      const ccTtl = resolvedCacheTtl;
+      if (Array.isArray(parsed.tools) && parsed.tools.length > 0) {
+        for (const tool of parsed.tools) {
+          if (tool.cache_control) delete tool.cache_control;
+        }
+        // Adaptive breakpoint placement (opt-in, runtime-driven). Real CC always
+        // pins the breakpoint on the last tool because CC controls a STABLE tool
+        // array. opencode may reorder/add/remove tools between turns, which moves
+        // the "last tool" and invalidates the cached prefix after it. When the
+        // detector reports the tool boundary is thrashing while the system prompt
+        // is stable, we SKIP the volatile last-tool breakpoint and let the stable
+        // system-prompt breakpoint (added by buildSystemPromptBlocks) anchor the
+        // cache. Falls back to exact CC behavior when no stability data exists.
+        const placeToolBreakpoint = shouldPlaceToolBreakpoint(runtime?.cacheBoundaryStability);
+        if (placeToolBreakpoint) {
+          parsed.tools[parsed.tools.length - 1].cache_control = { type: "ephemeral", ttl: ccTtl };
+        }
+      }
+      if (Array.isArray(parsed.messages)) {
+        for (const msg of parsed.messages) {
+          if (Array.isArray(msg.content)) {
+            for (const block of msg.content) {
+              // CONTRACT GUARD: `thinking` / `redacted_thinking` blocks MUST be
+              // round-tripped byte-identical to the model's original response
+              // (Anthropic extended-thinking rules). ANY mutation — including
+              // `delete block.cache_control` — triggers the 400 error
+              // "thinking or redacted_thinking blocks in the latest assistant
+              // message cannot be modified". `cache_control` is not even a valid
+              // field on a thinking block, so we never strip nor add it here;
+              // we leave the block exactly as received. This is required for
+              // adaptive-thinking models (Opus 4.6/4.7/4.8, Sonnet 4.6) on
+              // tool-continuation turns. See docs/mimese-http-header-system-prompt.md.
+              if (!block || typeof block !== "object") continue;
+              if (block.type === "thinking" || block.type === "redacted_thinking") {
+                continue;
+              }
+              if (block.cache_control) delete block.cache_control;
+            }
+          }
+        }
+        // Add cache_control to last user message (real CC does this).
+        // (User messages never contain thinking blocks, so the guard above
+        // does not affect this breakpoint.)
+        for (let i = parsed.messages.length - 1; i >= 0; i--) {
+          const msg = parsed.messages[i];
+          if (msg.role !== "user" || !Array.isArray(msg.content) || msg.content.length === 0) continue;
+          const lastBlock = msg.content[msg.content.length - 1];
+          if (lastBlock && typeof lastBlock === "object") {
+            lastBlock.cache_control = { type: "ephemeral", ttl: ccTtl };
+          }
+          break;
+        }
+      }
+    }
+
+    // Tool name sanitization: Anthropic's server blocklists known non-CC tool names.
+    // opencode uses lowercase names while CC uses PascalCase. While only "todowrite"
+    // is currently confirmed blocklisted, we rename ALL core opencode tools to match
+    // CC's naming convention as a preventive measure against future blocklist additions.
+    const OC_TO_CC_TOOL_NAMES = {
+      bash: "Bash",
+      read: "Read",
+      glob: "Glob",
+      grep: "Grep",
+      edit: "Edit",
+      write: "Write",
+      webfetch: "WebFetch",
+      todowrite: "TodoWrite",
+      skill: "Skill",
+      task: "Task",
+      compress: "Compress",
+    };
+    if (Array.isArray(parsed.tools)) {
+      for (const tool of parsed.tools) {
+        if (tool.name && OC_TO_CC_TOOL_NAMES[tool.name]) {
+          tool.name = OC_TO_CC_TOOL_NAMES[tool.name];
+        }
+      }
+    }
+    // Also rename in tool_use blocks in messages (assistant responses referencing the tool)
+    if (Array.isArray(parsed.messages)) {
+      for (const msg of parsed.messages) {
+        if (!Array.isArray(msg.content)) continue;
+        for (const block of msg.content) {
+          if (block.type === "tool_use" && block.name && OC_TO_CC_TOOL_NAMES[block.name]) {
+            block.name = OC_TO_CC_TOOL_NAMES[block.name];
+          }
+        }
+      }
+    }
+    // Track which tools the model has used (from assistant tool_use blocks).
+    // Names are already CC PascalCase after renaming above.
+    if (Array.isArray(parsed.messages)) {
+      for (const msg of parsed.messages) {
+        if (msg.role === "assistant" && Array.isArray(msg.content)) {
+          for (const block of msg.content) {
+            if (block.type === "tool_use" && block.name) {
+              runtime.usedTools.add(block.name);
+            }
+          }
+        }
+      }
+    }
+
+    // Strategy 4 — Adaptive tool set: after turn 3, defer non-core tools that
+    // the model hasn't used yet. This saves schema bytes on turns where tools
+    // are unlikely to be needed. Core tools stay eager. Resets on /clear.
+    if (
+      Array.isArray(parsed.tools) &&
+      signature.adaptiveToolSet !== false &&
+      runtime.turns >= 3 &&
+      parsed.model &&
+      !/claude-3-|haiku/i.test(parsed.model)
+    ) {
+      const used = runtime.usedTools;
+      for (const tool of parsed.tools) {
+        if (tool.name && !used.has(tool.name) && !CORE_TOOL_NAMES.has(tool.name)) {
+          tool.defer_loading = true;
+        }
+      }
+    }
+
+    // Tool description compaction: apply the same compaction logic used for
+    // system prompts (strip examples, collapse whitespace, dedup lines) to tool
+    // descriptions. The top 4 tools (Bash 10.6K, TodoWrite 9.7K, Task 5.5K,
+    // Compress 4.5K) account for 30KB. Compaction typically saves 30-50%.
+    if (Array.isArray(parsed.tools) && signature.toolDescriptionCompaction !== false) {
+      for (const tool of parsed.tools) {
+        if (tool.description && tool.description.length > 500) {
+          tool.description = compactToolDescription(tool.description);
+        }
+      }
+    }
+
+    // MCP tool deferral: mark non-core tools with defer_loading: true.
+    // CC defers all MCP tools by default — the API omits their full schemas from token
+    // counting, only sending the tool name. When the model needs a deferred tool, it uses
+    // tool_reference to load the schema on demand. This saves ~20KB per turn.
+    // Core tools (OC_TO_CC_TOOL_NAMES) are always eager-loaded.
+    if (
+      Array.isArray(parsed.tools) &&
+      signature.toolDeferral !== false &&
+      parsed.model &&
+      !/claude-3-|haiku/i.test(parsed.model)
+    ) {
+      const coreToolNames = new Set(Object.values(OC_TO_CC_TOOL_NAMES));
+      for (const tool of parsed.tools) {
+        if (tool.name && !coreToolNames.has(tool.name)) {
+          tool.defer_loading = true;
+        }
+      }
+    }
+
+    // Task budgets: when the task-budgets beta is active, preserve or inject output_config.
+    // The beta unlocks output_config.max_output_tokens for per-task budget control.
+    // Model-router compatibility: the beta header + output_config body are forwarded as-is.
+    if (betaHeader && betaHeader.includes("task-budgets-2026-03-13")) {
+      if (!parsed.output_config) {
+        // Default: set a reasonable per-task output budget for long-running agentic tasks.
+        // Claude Code tasks typically need generous output budgets.
+        parsed.output_config = { max_output_tokens: 16384 };
+      }
+    }
+
+    // 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 };
+    }
+
+    // Fast mode: inject speed parameter for fast-mode-eligible models.
+    // Per Anthropic fast-mode docs, `speed: "fast"` (beta fast-mode-2026-02-01) is
+    // supported on Opus 4.6, Opus 4.7, and Opus 4.8 (research preview). Opus 4.7
+    // is the /fast default in real Claude Code v2.1.142+. Sonnet is NOT eligible.
+    // NOTE: switching speed invalidates system + message prompt caches (per
+    // Anthropic fast-mode docs), so only flip it deliberately.
+    // When the selected account's fast pool is cooling down (a prior fast 429),
+    // suppress speed:"fast" entirely so this turn runs at standard speed on the
+    // same account rather than re-hitting the exhausted fast pool.
+    const fastPoolAvailable = !signature.fastRateLimited;
+    const isFastModeEligibleModel = (m) =>
+      fastPoolAvailable && (isOpus46Model(m) || isOpus47Model(m) || isOpus48Model(m));
+    const fastModeEnabled = signature.fastMode && !isFalsyEnv(process.env.OPENCODE_ANTHROPIC_DISABLE_FAST_MODE);
+    let fastModeAutoApplied = false;
+    if (
+      !fastModeEnabled &&
+      te.fast_mode_auto === true &&
+      isMainRole &&
+      parsed.model &&
+      isFastModeEligibleModel(parsed.model) &&
+      Array.isArray(parsed.messages) &&
+      parsed.messages.length >= 2
+    ) {
+      // Simple exchange heuristic: last user message is short, no tool_result,
+      // no file references. Suggests a follow-up question that doesn't need
+      // deep reasoning.
+      const last = parsed.messages[parsed.messages.length - 1];
+      if (last && last.role === "user") {
+        let txt = "";
+        let hasToolResult = false;
+        if (typeof last.content === "string") txt = last.content;
+        else if (Array.isArray(last.content)) {
+          for (const b of last.content) {
+            if (b?.type === "tool_result") hasToolResult = true;
+            if (b?.type === "text" && typeof b.text === "string") txt += b.text;
+          }
+        }
+        if (!hasToolResult && txt.length < 240 && !/\bfile:|\.md\b|\.mjs\b|\.ts\b/i.test(txt)) {
+          fastModeAutoApplied = true;
+        }
+      }
+    }
+    if ((fastModeEnabled || fastModeAutoApplied) && parsed.model && isFastModeEligibleModel(parsed.model)) {
+      parsed.speed = "fast";
+    }
+
+    // Guard: repair orphaned tool_use blocks anywhere in the message array.
+    // The Anthropic API requires that every assistant message containing tool_use
+    // blocks is immediately followed by a user message with matching tool_result
+    // blocks. When OpenCode crashes/hangs mid-tool-execution, the conversation
+    // state may be saved with unpaired tool_use blocks. This causes:
+    //   "messages.N: `tool_use` ids were found without `tool_result` blocks
+    //    immediately after: toolu_XXXXX"
+    // We scan the full array and synthesize missing tool_result messages.
+    if (Array.isArray(parsed.messages) && parsed.messages.length > 0) {
+      parsed.messages = repairOrphanedToolUseBlocks(parsed.messages);
+
+      // Also ensure the array never ends with an assistant message (prefill guard).
+      const lastMsg = parsed.messages[parsed.messages.length - 1];
+      if (lastMsg && lastMsg.role === "assistant") {
+        const lastContent = Array.isArray(lastMsg.content) ? lastMsg.content : [];
+        const toolUseBlocks = lastContent.filter((b) => b.type === "tool_use");
+        if (toolUseBlocks.length > 0) {
+          parsed.messages.push({
+            role: "user",
+            content: toolUseBlocks.map((tu) => ({
+              type: "tool_result",
+              tool_use_id: tu.id,
+              content: "[Result unavailable — conversation was restructured]",
+            })),
+          });
+        } else {
+          parsed.messages.push({
+            role: "user",
+            content: [{ type: "text", text: "Continue." }],
+          });
+        }
+      }
+    }
+
+    return JSON.stringify(parsed);
+  } catch {
+    // ignore parse errors
+    return body;
+  }
+}