npm - @bitkyc08/opencodex - Versions diffs - 2.57.0 → 2.59.0 - Mend

@bitkyc08/opencodex 2.57.0 → 2.59.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (241) hide show

package/README.md +28 -10
package/gui/dist/assets/index-C5IebErG.js +136 -0
package/gui/dist/assets/{index-C5-RdDmD.css → index-OESInAjC.css} +1 -1
package/gui/dist/index.html +2 -2
package/gui/dist/provider-icons/crusoe.svg +1 -0
package/gui/dist/provider-icons/opper.svg +3 -0
package/package.json +2 -2
package/src/adapters/base.ts +11 -1
package/src/adapters/codebuddy/scaffold-guard.ts +5 -4
package/src/adapters/command-code.ts +13 -4
package/src/adapters/cursor/catalog.ts +11 -0
package/src/adapters/cursor/cursor-errors.ts +15 -0
package/src/adapters/cursor/discovery.ts +65 -1
package/src/adapters/cursor/effort-map.ts +16 -2
package/src/adapters/cursor/envelope-echo.ts +55 -2
package/src/adapters/cursor/live-transport.ts +5 -1
package/src/adapters/cursor/message-mapper.ts +3 -2
package/src/adapters/cursor/protobuf-events.ts +110 -11
package/src/adapters/cursor/protobuf-request.ts +27 -6
package/src/adapters/cursor/request-builder.ts +14 -3
package/src/adapters/cursor/text-toolcall.ts +230 -0
package/src/adapters/cursor/thread-continuity.ts +141 -0
package/src/adapters/cursor/tool-guidance.ts +5 -4
package/src/adapters/cursor/types.ts +5 -0
package/src/adapters/cursor.ts +97 -6
package/src/adapters/devin/cloud-direct/chat.ts +11 -2
package/src/adapters/devin/cloud-direct/index.ts +7 -0
package/src/adapters/devin/cloud-direct/stated-reset-retry.ts +103 -0
package/src/adapters/devin.ts +75 -13
package/src/adapters/google-antigravity-wire.ts +29 -2
package/src/adapters/google-http.ts +45 -13
package/src/adapters/google.ts +23 -4
package/src/adapters/mimo-free.ts +32 -17
package/src/adapters/ollama-native.ts +42 -8
package/src/adapters/openai-chat/response-events.ts +61 -0
package/src/adapters/openai-chat.ts +5 -10
package/src/adapters/openai-responses/passthrough.ts +40 -5
package/src/adapters/openai-responses/request-strips.ts +43 -0
package/src/adapters/openai-responses/tool-output-recovery.ts +75 -0
package/src/adapters/openai-responses/tool-schema.ts +19 -7
package/src/adapters/physical-send.ts +50 -0
package/src/adapters/responses-tool-schema.ts +76 -46
package/src/adapters/run-turn-queue.ts +17 -4
package/src/bridge/response-json.ts +2 -2
package/src/bridge/sse.ts +166 -25
package/src/claude/context-windows.ts +22 -0
package/src/claude/outbound.ts +46 -5
package/src/cli/account-api.ts +4 -3
package/src/cli/account-extended.ts +22 -2
package/src/cli/account-orca-import.ts +63 -0
package/src/cli/account.ts +32 -4
package/src/cli/capabilities.ts +40 -0
package/src/cli/claude.ts +29 -1
package/src/cli/codex-cli-update.ts +97 -2
package/src/cli/config-command.ts +35 -18
package/src/cli/dispatch.ts +71 -4
package/src/cli/doctor.ts +197 -2
package/src/cli/help.ts +4 -1
package/src/cli/index.ts +132 -22
package/src/cli/models-runtime.ts +33 -4
package/src/cli/registry.ts +11 -1
package/src/cli/runtime-api.ts +44 -0
package/src/cli/start-args.ts +94 -0
package/src/cli/system-command.ts +72 -1
package/src/cli/uninstall-client-state.ts +12 -0
package/src/client/machine-api.ts +4 -3
package/src/client/machine-listener.ts +14 -1
package/src/clients/config-export/constants.ts +2 -3
package/src/clients/config-export.ts +5 -5
package/src/codex/account-store.ts +81 -5
package/src/codex/auth-api/pool-quota-probe.ts +14 -3
package/src/codex/auth-api/routes.ts +17 -2
package/src/codex/auth-context.ts +58 -20
package/src/codex/catalog/build-entries.ts +25 -4
package/src/codex/catalog/derive-entry.ts +8 -1
package/src/codex/catalog/effort.ts +10 -6
package/src/codex/catalog/gather-capture.ts +1 -0
package/src/codex/catalog/model-hints.ts +37 -5
package/src/codex/catalog/parsing.ts +83 -5
package/src/codex/catalog/reserve-warn.ts +96 -0
package/src/codex/catalog/retained-sync.ts +19 -0
package/src/codex/catalog/routed-gather.ts +42 -3
package/src/codex/cli-installation-identity.ts +210 -0
package/src/codex/cli-installation-targets.ts +158 -0
package/src/codex/convergence.ts +5 -0
package/src/codex/desktop-switches.ts +145 -0
package/src/codex/history-job.ts +5 -1
package/src/codex/history-provider.ts +37 -5
package/src/codex/history-state-open.ts +105 -0
package/src/codex/history-worker.ts +14 -1
package/src/codex/inject/config-toml.ts +44 -2
package/src/codex/inject/remove.ts +145 -7
package/src/codex/inject/restore.ts +204 -32
package/src/codex/inject.ts +6 -9
package/src/codex/lineage.ts +83 -32
package/src/codex/loopback-target.ts +40 -0
package/src/codex/main-account-hard-lock.ts +2 -1
package/src/codex/main-account.ts +10 -3
package/src/codex/main-device-reauth.ts +17 -9
package/src/codex/model-entitlements.ts +60 -1
package/src/codex/native-profile-startup.ts +64 -20
package/src/codex/observed-model-denials.ts +137 -0
package/src/codex/orca-auth-source.ts +94 -0
package/src/codex/orca-import.ts +219 -0
package/src/codex/prompt-text-probe.ts +282 -12
package/src/codex/quota-401-recovery.ts +12 -0
package/src/codex/quota-types.ts +65 -0
package/src/codex/quota.ts +24 -19
package/src/codex/routing/cooldown-math.ts +8 -47
package/src/codex/routing/pin-drain.ts +57 -0
package/src/codex/routing.ts +13 -15
package/src/codex/subagent-model-fallback.ts +94 -0
package/src/codex/windows-installation-files.ts +224 -0
package/src/combos/failover.ts +122 -5
package/src/config/atomic-write.ts +83 -8
package/src/config/diagnostics.ts +21 -0
package/src/config/load-degrade.ts +15 -0
package/src/config/pending-teardown.ts +8 -0
package/src/config/process-state.ts +36 -3
package/src/config/provider-relative-send-path.ts +16 -0
package/src/config/proxy-env.ts +23 -5
package/src/config/schema/config-schema.ts +23 -0
package/src/config/schema/leaf-validators.ts +65 -17
package/src/generated/compatibility-version.json +337 -201
package/src/generated/model-metadata.ts +1 -1
package/src/lib/bounded-body.ts +4 -2
package/src/lib/bounded-subprocess.ts +62 -10
package/src/lib/destination-policy.ts +48 -6
package/src/lib/errors.ts +3 -15
package/src/lib/local-destinations.ts +32 -5
package/src/lib/provider-outbound.ts +3 -3
package/src/lib/proxy-env.ts +70 -3
package/src/lib/request-execution-budget.ts +11 -3
package/src/lib/response-body-inactivity.ts +193 -0
package/src/lib/retry-delay.ts +69 -0
package/src/lib/socks5-fetch.ts +631 -0
package/src/lib/spend-reservation-ledger.ts +115 -9
package/src/lib/windows-secret-acl.ts +151 -15
package/src/lib/windows-user-principal.ts +5 -1
package/src/lib/workflow-budget.ts +145 -8
package/src/oauth/account-quota-rank.ts +72 -15
package/src/oauth/generic-account-failover.ts +40 -27
package/src/oauth/orcarouter.ts +15 -2
package/src/oauth/store.ts +8 -0
package/src/providers/codex-capacity.ts +9 -0
package/src/providers/derive.ts +6 -0
package/src/providers/devin-provider-merge-migration.ts +33 -12
package/src/providers/free-directory.ts +20 -2
package/src/providers/key-failover.ts +261 -7
package/src/providers/model-discovery.ts +19 -7
package/src/providers/model-rename-migration.ts +1 -0
package/src/providers/openai-sidecar.ts +4 -0
package/src/providers/opencode-go-transport.ts +14 -5
package/src/providers/quota/report-cache.ts +3 -0
package/src/providers/registry/entries-core.ts +11 -0
package/src/providers/registry/entries-extended.ts +146 -28
package/src/providers/registry/model-seeds.ts +136 -29
package/src/providers/registry/types.ts +9 -0
package/src/responses/apply-patch-envelope.ts +44 -11
package/src/responses/bridge-search-replay-cache.ts +152 -0
package/src/responses/code-mode-helper-compat.ts +26 -16
package/src/responses/custom-tool-compat.ts +1 -1
package/src/responses/hosted-tool-policy.ts +85 -2
package/src/responses/schema.ts +9 -2
package/src/responses/spill-store.ts +17 -0
package/src/responses/state/body-policy.ts +25 -0
package/src/responses/state/spill-queue.ts +8 -6
package/src/responses/state.ts +3 -22
package/src/router.ts +4 -0
package/src/server/auth-cors.ts +27 -0
package/src/server/chat-completions.ts +9 -4
package/src/server/chat-native-sse.ts +26 -9
package/src/server/chat-native.ts +10 -4
package/src/server/claude-messages.ts +24 -2
package/src/server/gui-static.ts +36 -2
package/src/server/inbound-body-admission.ts +187 -0
package/src/server/index/websocket-handler.ts +48 -1
package/src/server/index.ts +15 -19
package/src/server/management/api-access.ts +3 -4
package/src/server/management/config-routes.ts +57 -10
package/src/server/management/provider-capability-config.ts +35 -7
package/src/server/management/provider-routes.ts +70 -18
package/src/server/models-capabilities.ts +24 -3
package/src/server/proxy-liveness.ts +97 -2
package/src/server/relay.ts +17 -24
package/src/server/request-log.ts +25 -1
package/src/server/responses/adapter-continuation.ts +71 -27
package/src/server/responses/adapter-delivery.ts +39 -8
package/src/server/responses/adapter-dispatch.ts +52 -24
package/src/server/responses/codex-ws-exchange.ts +65 -4
package/src/server/responses/combo-stream-preflight.ts +68 -5
package/src/server/responses/compact.ts +60 -11
package/src/server/responses/core-codex-account.ts +83 -22
package/src/server/responses/core-combo.ts +26 -0
package/src/server/responses/core-normalize.ts +12 -5
package/src/server/responses/core-options.ts +3 -0
package/src/server/responses/fetch-helpers.ts +72 -3
package/src/server/responses/native-injection-protocol.ts +42 -0
package/src/server/responses/native-injection-replay.ts +105 -0
package/src/server/responses/native-injection.ts +242 -0
package/src/server/responses/native-response-control.ts +56 -0
package/src/server/responses/native-response-json.ts +14 -0
package/src/server/responses/native-response-output.ts +37 -0
package/src/server/responses/native-steering-log.ts +44 -0
package/src/server/responses/native-steering-policy.ts +49 -0
package/src/server/responses/native-steering-replay.ts +126 -0
package/src/server/responses/native-steering-settings.ts +76 -0
package/src/server/responses/native-steering.ts +400 -0
package/src/server/responses/native-tool-results.ts +130 -0
package/src/server/responses/passthrough-delivery.ts +21 -1
package/src/server/responses/passthrough-dispatch.ts +146 -49
package/src/server/responses/passthrough-execution.ts +11 -1
package/src/server/responses/request-prepare.ts +70 -0
package/src/server/responses/request-send-budget.ts +84 -7
package/src/server/responses/request-sidecar-auth.ts +16 -8
package/src/server/responses/request-spend.ts +38 -9
package/src/server/responses/request-transport.ts +13 -10
package/src/server/responses/run-turn-execution.ts +20 -5
package/src/server/responses/sidecar-execution.ts +2 -0
package/src/server/responses/ws-upstream.ts +23 -2
package/src/server/responses-custom-tool-repair.ts +2 -2
package/src/server/sse-frame-buffer.ts +12 -10
package/src/server/sse-payload-rewrite.ts +36 -9
package/src/server/stop-teardown.ts +8 -1
package/src/server/system-env-shell.ts +5 -1
package/src/server/system-env.ts +7 -1
package/src/server/workflow-refusal.ts +56 -2
package/src/server/ws-bridge.ts +16 -1
package/src/service/cli.ts +29 -7
package/src/service/guards.ts +10 -0
package/src/service/health.ts +43 -0
package/src/service/state.ts +7 -2
package/src/types/accounts.ts +4 -0
package/src/types/config.ts +104 -3
package/src/types/provider.ts +32 -0
package/src/types/request.ts +7 -1
package/src/types/wire.ts +9 -1
package/src/usage/expected-prices.ts +28 -0
package/src/usage/log.ts +87 -4
package/src/web-search/passthrough-bridge.ts +39 -5
package/gui/dist/assets/index-Cz7CLdif.js +0 -128

package/gui/dist/index.html CHANGED Viewed

@@ -16,8 +16,8 @@
         } catch (e) {}
       })();
     </script>
-    <script type="module" crossorigin src="/assets/index-Cz7CLdif.js"></script>
-    <link rel="stylesheet" crossorigin href="/assets/index-C5-RdDmD.css">
+    <script type="module" crossorigin src="/assets/index-C5IebErG.js"></script>
+    <link rel="stylesheet" crossorigin href="/assets/index-OESInAjC.css">
   </head>
   <body>
     <div id="root"></div>

package/gui/dist/provider-icons/crusoe.svg ADDED Viewed

@@ -0,0 +1 @@

+ <svg height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Crusoe</title><path d="M12 0L4.583 6.583c-3.23 2.869-3.23 7.965 0 10.834L12 24l7.417-6.583c3.23-2.869 3.23-7.965 0-10.834L12 0z" fill="url(#lobe-icons-crusoe-_R_0_)"></path><defs><linearGradient gradientUnits="userSpaceOnUse" id="lobe-icons-crusoe-_R_0_" x1="18.919" x2="4.853" y1="5.595" y2="18.301"><stop stop-color="#F4BF45"></stop><stop offset=".35" stop-color="#E48047"></stop><stop offset=".69" stop-color="#C73361"></stop><stop offset="1" stop-color="#A42F5F"></stop></linearGradient></defs></svg>

package/gui/dist/provider-icons/opper.svg ADDED Viewed

@@ -0,0 +1,3 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 315 315" fill="#000000">
+  <path fill-rule="evenodd" clip-rule="evenodd" d="M159.78 315C71.53 315 0 244.49 0 157.5C0 -18.9499 159.78 0.650075 159.78 0.650075C159.78 87.2201 88.36 157.4 0.2 157.5C149.8 157.64 159.78 315 159.78 315ZM160.52 217.98C160.52 217.98 156.94 161.65 105.04 157.52C120.6 157.34 160.52 151.54 160.52 96.5601C160.52 151.54 200.44 157.34 216 157.52C164.1 161.63 160.52 217.98 160.52 217.98Z"/>
+</svg>

package/package.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "@bitkyc08/opencodex",
-  "version": "2.57.0",
+  "version": "2.59.0",
   "description": "Universal provider proxy for OpenAI Codex & Claude Code — use any LLM with Codex CLI/App/SDK and Claude Code",
   "type": "module",
   "main": "./bin/package-main.mjs",
@@ -86,7 +86,7 @@
   "overrides": {
     "@hono/node-server": "2.1.0",
     "fast-uri": "^3.1.7",
-    "hono": "4.13.1",
+    "hono": "4.13.8",
     "ip-address": "^10.4.0",
     "qs": "^6.16.0"
   },

package/src/adapters/base.ts CHANGED Viewed

@@ -1,7 +1,7 @@
 import type { AdapterEvent, OcxParsedRequest } from "../types";
 import type { TranslatorBudget } from "../lib/translator-budget";
 import type { RequestExecutionBudget } from "../lib/request-execution-budget";
-import type { AttemptRecoveryKind } from "../usage/log";
+import type { AttemptRecoveryKind, AttemptRecoveryWithheld } from "../usage/log";
 import type { AdapterTierMetadata } from "../providers/fastwire";
 /** Metadata about the caller's incoming request, for auth-forwarding adapters. */
@@ -168,6 +168,16 @@ export interface AdapterFetchContext {
    * to `sendCount` and no regression could assert a count for them (#4546).
    */
   onPhysicalSend?: (send: { ordinal: number; recovery?: AttemptRecoveryKind }) => void;
+  /**
+   * Observes a recovery this adapter was ready to make and did not, because the send budget
+   * refused the dispatch.
+   *
+   * Separate from `onPhysicalSend` because nothing was sent: folding it in would inflate
+   * `sendCount`, the one number that means "requests this proxy actually made". Without it a
+   * log with one send cannot distinguish "no recovery was eligible" from "one was and the
+   * budget withheld it", and those need opposite follow-ups (#5044).
+   */
+  onRecoveryWithheld?: (withheld: { reason: AttemptRecoveryWithheld }) => void;
 }
 /**

package/src/adapters/codebuddy/scaffold-guard.ts CHANGED Viewed

@@ -5,10 +5,10 @@ export const CODEBUDDY_SCAFFOLD_ERROR_CODE = "vendor_scaffold_detected";
 // The observed control protocol uses FULLWIDTH VERTICAL LINE (U+FF5C). Detection stays
 // deliberately narrower than the marker spelling: a calls control line must be followed by an
-// invoke line for a functions.* tool. That distinguishes an agent scaffold from prose quoting or
+// invoke line with a non-empty tool name. That distinguishes an agent scaffold from prose quoting or
 // discussing one tag.
 const DSML_CALLS_LINE = "<｜｜dsml｜｜ calls>";
-const DSML_INVOKE_PREFIX = "<｜｜dsml｜｜ invoke name=\"functions.";
+const DSML_INVOKE_PREFIX = "<｜｜dsml｜｜ invoke name=\"";
 export interface CodeBuddyScaffoldFilterResult {
   /** Bytes released from a suffix withheld by an earlier event on this channel. */
@@ -39,7 +39,7 @@ function prefixAtEnd(text: string, at: number, expected: string): boolean {
  * Control tags are recognized only at column zero and outside fenced Markdown. Inline code,
  * quoted strings, blockquotes, indented source, and prose all add syntax before the tag and are
  * therefore forwarded unchanged. A calls line alone is harmless; refusal requires the observed
- * two-line calls-plus-functions-invoke grammar.
+ * two-line calls-plus-named-invoke grammar.
  */
 function scan(
   text: string,
@@ -77,7 +77,8 @@ function scan(
           if (invokeAt >= 0) {
             const invokeRest = text.slice(invokeAt).toLowerCase();
-            if (invokeRest.startsWith(DSML_INVOKE_PREFIX)) {
+            const invokeNameStart = invokeRest[DSML_INVOKE_PREFIX.length];
+            if (invokeRest.startsWith(DSML_INVOKE_PREFIX) && invokeNameStart && !/[\s"]/.test(invokeNameStart)) {
               return { safe: text.slice(0, index), held: "", fail: true, fence, lineStart };
             }
             if (invokeRest.length === 0 || DSML_INVOKE_PREFIX.startsWith(invokeRest)) {

package/src/adapters/command-code.ts CHANGED Viewed

@@ -13,6 +13,8 @@ import { commandCodeReasoningEfforts, refreshCommandCodeReasoningEfforts } from
 import { identifyRoutedModel } from "./identity";
 import { buildNonOpenAIToolCatalogNudgeForTools } from "./tool-catalog-nudge";
 import { parseDataUrl } from "./image";
+import { createAdapterPhysicalSend } from "./physical-send";
+import { SendBudgetExhaustedError } from "../lib/upstream-retry";
 // Retain the short ids emitted by the first local integration. New requests use the live catalog's
 // provider-native IDs directly; this map is compatibility-only and is not a model fallback list.
@@ -469,7 +471,7 @@ async function fetchCommandCode(request: AdapterRequest, ctx: AdapterFetchContex
   const timer = setTimeout(() => timeout.abort(new DOMException("Timeout elapsed", "TimeoutError")), ctx?.timeoutMs ?? 200_000);
   const callerSignal = ctx?.abortSignal ?? new AbortController().signal;
   try {
-    return await (ctx?.executor ?? executor)(request.url, {
+    return await executor(request.url, {
       method: request.method,
       headers: request.headers,
       body: request.body,
@@ -556,7 +558,8 @@ export function createCommandCodeAdapter(provider: OcxProviderConfig): ProviderA
       };
     },
     async fetchResponse(request: AdapterRequest, ctx?: AdapterFetchContext): Promise<Response> {
-      const response = await fetchCommandCode(request, ctx, executor);
+      const send = createAdapterPhysicalSend(ctx, executor);
+      const response = await send({ url: request.url, dispatch: physical => fetchCommandCode(request, ctx, physical) });
       if (response.ok) return response;
       const currentEffort = (() => {
         try { return (JSON.parse(request.body) as { params?: { reasoning_effort?: unknown } }).params?.reasoning_effort; } catch { return undefined; }
@@ -577,8 +580,14 @@ export function createCommandCodeAdapter(provider: OcxProviderConfig): ProviderA
       if (!refreshed || refreshed.includes(currentEffort)) return response;
       const retry = requestWithoutReasoningEffort(request);
       if (!retry) return response;
-      try { void response.body?.cancel(); } catch { /* already closed */ }
-      return fetchCommandCode(retry, ctx, executor);
+      try {
+        return await send({ url: retry.url, sendClass: "repair", recovery: "reasoning-effort-downgrade",
+          beforeDispatch: () => { try { void response.body?.cancel().catch(() => {}); } catch { /* already closed */ } },
+          dispatch: physical => fetchCommandCode(retry, ctx, physical) });
+      } catch (error) {
+        if (error instanceof SendBudgetExhaustedError) return response;
+        throw error;
+      }
     },
     async *parseStream(response: Response, budget: TranslatorBudget): AsyncGenerator<AdapterEvent> {
       let sawFinish = false;

package/src/adapters/cursor/catalog.ts CHANGED Viewed

@@ -60,6 +60,8 @@ const CONTEXT_500K = 500 * K;
 const CONTEXT_1M = 1_000 * K;
 /** Gemini publishes the exact power-of-two window, not a rounded 1M. */
 const CONTEXT_GEMINI = 1_048_576;
+/** Meta publishes 1,048,576 for both Muse Spark 1.3 tiers (dev.meta.ai/docs/models). */
+const CONTEXT_MUSE = 1_048_576;
 const FULL = ["low", "medium", "high", "xhigh", "max"] as const;
 const T = "thinking-then-effort" as const;
@@ -211,6 +213,15 @@ export const CURSOR_CAPABILITIES: Record<string, CursorCapability> = {
     defaultVariant: "regular",
     variants: { regular: { levels: ["low", "medium", "high"] } },
   },
+  // Seeded from the live GetUsableModels roster attached to #4820, which advertises six
+  // muse-spark-1.3 effort variants. The ladder stops at xhigh on purpose: see the matching
+  // effort-map entry for why Cursor advertising `-max` is not evidence that it runs.
+  "muse-spark-1.3": {
+    displayName: "Muse Spark 1.3",
+    window: CONTEXT_MUSE,
+    defaultVariant: "regular",
+    variants: { regular: { levels: ["minimal", "low", "medium", "high", "xhigh"] } },
+  },
   "kimi-k3": {
     displayName: "Kimi K3",
     window: CONTEXT_1M,

package/src/adapters/cursor/cursor-errors.ts CHANGED Viewed

@@ -46,6 +46,21 @@ export class CursorStreamTruncatedError extends Error {
   }
 }
+export const CURSOR_INCOMPLETE_TOOL_CALL_MESSAGE_PREFIX =
+  "Cursor stream ended with incomplete tool call(s):";
+/**
+ * True when Cursor ended the stream with a client tool still open. The adapter fail-closes
+ * the current turn (no partial `tool_call_start`) and remints the conversation afterwards
+ * so the next turn does not resume a session left waiting for `mcpResult`.
+ */
+export function isCursorIncompleteToolCallMessage(value: unknown): boolean {
+  const message = typeof value === "string" ? value : errorMessage(value);
+  const lower = message.toLowerCase();
+  return lower.includes(CURSOR_INCOMPLETE_TOOL_CALL_MESSAGE_PREFIX.toLowerCase())
+    || lower.includes("tool call(s) left incomplete");
+}
 /**
  * A cancel-shaped stream failure that WE did not request. `cancelCursorRun` is the only place
  * that cancels our own stream, and it sets `expectedClose` first, so a cancel arriving without it

package/src/adapters/cursor/discovery.ts CHANGED Viewed

@@ -24,8 +24,54 @@ const CONTEXT_272K = 272_000;
 const CONTEXT_262K = 262_144;
 const CONTEXT_256K = 256_000;
 const CONTEXT_200K = 200_000;
+export const CURSOR_OBSERVED_CONTEXT_WINDOW_MAX_ENTRIES = 2_048;
-export function inferCursorContextWindow(modelId: string): number {
+/**
+ * Process-local ceilings from `ConversationTokenDetails.maxTokens` on live
+ * checkpoints. Each observation belongs to the Cursor identity scope that
+ * produced it; plan-gated accounts sharing one proxy must not overwrite each
+ * other's overflow prior (senpi `cursor-context-limit`).
+ */
+const observedCursorContextWindows = new Map<string, number>();
+interface CursorContextWindowOptions {
+  identityScope?: string;
+  observed?: number;
+}
+function normalizeObservedWindowKey(modelId: string, identityScope?: string): string {
+  return `${identityScope?.trim() || "local"}\0${modelId.trim().toLowerCase()}`;
+}
+export function recordObservedCursorContextWindow(
+  modelId: string,
+  maxTokens: number | undefined,
+  options: Pick<CursorContextWindowOptions, "identityScope"> = {},
+): void {
+  if (!modelId.trim()) return;
+  if (typeof maxTokens !== "number" || !Number.isFinite(maxTokens) || maxTokens <= 0) return;
+  const key = normalizeObservedWindowKey(modelId, options.identityScope);
+  observedCursorContextWindows.delete(key);
+  observedCursorContextWindows.set(key, Math.floor(maxTokens));
+  while (observedCursorContextWindows.size > CURSOR_OBSERVED_CONTEXT_WINDOW_MAX_ENTRIES) {
+    const oldest = observedCursorContextWindows.keys().next().value;
+    if (oldest === undefined) break;
+    observedCursorContextWindows.delete(oldest);
+  }
+}
+export function observedCursorContextWindow(
+  modelId: string,
+  options: Pick<CursorContextWindowOptions, "identityScope"> = {},
+): number | undefined {
+  return observedCursorContextWindows.get(normalizeObservedWindowKey(modelId, options.identityScope));
+}
+export function resetObservedCursorContextWindowsForTests(): void {
+  observedCursorContextWindows.clear();
+}
+function inferCursorContextWindowHeuristic(modelId: string): number {
   const id = modelId.trim().toLowerCase();
   if (id.includes("1m")) return CONTEXT_1M;
   if (id.startsWith("gemini-")) return CONTEXT_1M;
@@ -40,6 +86,24 @@ export function inferCursorContextWindow(modelId: string): number {
   return CURSOR_DEFAULT_CONTEXT_WINDOW;
 }
+/**
+ * Infer a conservative context window for a Cursor model id.
+ *
+ * A positive explicit observation wins, then an identity-scoped process-local
+ * checkpoint `maxTokens`, then the id heuristic. Cursor's `AvailableModelsResponse`
+ * does not currently include per-model context window metadata.
+ */
+export function inferCursorContextWindow(
+  modelId: string,
+  options: CursorContextWindowOptions = {},
+): number {
+  const { observed } = options;
+  if (typeof observed === "number" && Number.isFinite(observed) && observed > 0) {
+    return Math.floor(observed);
+  }
+  return observedCursorContextWindow(modelId, options) ?? inferCursorContextWindowHeuristic(modelId);
+}
 function normalizeInputModalities(input: string[] | undefined): string[] {
   const values = (input ?? [...CURSOR_DEFAULT_INPUT_MODALITIES])
     .map(item => item.trim())

package/src/adapters/cursor/effort-map.ts CHANGED Viewed

@@ -38,8 +38,10 @@ const CURSOR_MODEL_EFFORT_TIERS: Record<string, readonly string[]> = {
   "claude-opus-5-fast": ["low", "medium", "high"],
   "claude-sonnet-5": ["low", "medium", "high", "xhigh", "max"],
   "glm-5.2": ["high", "max"],
-  // 260825 live GetUsableModels. gemini-3.6-flash is the only Cursor model exposing `minimal`;
-  // listing it here is also what admits the suffix into CANONICAL_EFFORT_SUFFIXES below.
+  // 260825 live GetUsableModels. gemini-3.6-flash was the first Cursor model exposing
+  // `minimal`; listing a rung here is also what admits the suffix into
+  // CANONICAL_EFFORT_SUFFIXES below. muse-spark-1.3 now carries it too, so `minimal` no
+  // longer depends on this single row.
   "gemini-3.6-flash": ["minimal", "low", "medium", "high"],
   "gemini-3.7-flash": ["low", "medium", "high"],
   // 260903 preemptive: gemini-3.8-flash seeded ahead of Cursor's lineup update, the same way
@@ -91,6 +93,18 @@ const CURSOR_MODEL_EFFORT_TIERS: Record<string, readonly string[]> = {
   "gpt-5.6-sol": ["low", "medium", "high", "xhigh", "max"],
   "gpt-5.6-terra": ["low", "medium", "high", "xhigh", "max"],
   "gpt-5.6-luna": ["low", "medium", "high", "xhigh", "max"],
+  // 260916 live GetUsableModels (#4820) advertises muse-spark-1.3 at minimal, low, medium,
+  // high, xhigh AND max. The seed stops at xhigh deliberately.
+  //
+  // Meta publishes minimal..xhigh for Muse Spark and lists no `max` at all
+  // (dev.meta.ai/docs/reasoning), and an independent OpenCode Zen probe of
+  // muse-spark-1.3-contributor-free rejected max with `unknown variant` — both already
+  // recorded on META_MUSE_REASONING_EFFORTS in src/providers/registry/model-seeds.ts.
+  // Cursor advertising a wire id is not evidence that Run accepts it; that is exactly the
+  // advertised-but-not-callable shape CURSOR_KNOWN_UNCALLABLE_MODEL_IDS was created for.
+  // Publishing the rung anyway would invent a capability on two sources' contrary evidence.
+  // Add `max` here once a Cursor Run at max is observed to succeed.
+  "muse-spark-1.3": ["minimal", "low", "medium", "high", "xhigh"],
 };
 /** All effort suffixes accepted when matching live Cursor model ids to configured base ids. */

package/src/adapters/cursor/envelope-echo.ts CHANGED Viewed

@@ -13,6 +13,58 @@
  */
 const ECHO_MARKERS = ["[Tool Result]", "[Tool Error]", "[tool_result]"] as const;
+function isEchoMarkerLine(line: string): boolean {
+  return (ECHO_MARKERS as readonly string[]).includes(line.replace(/^[ \t]+/, ""));
+}
+/**
+ * Drop echoed tool-result envelopes from assistant history before Cursor root replay.
+ *
+ * The prefix sniffer catches an echo that STARTS a turn, but grok-4.6 routinely writes a real
+ * sentence first and pastes the envelope after it. That text has already reached the client and
+ * is stored as assistant output, so replaying it verbatim re-primes the next turn with the very
+ * envelope the model is copying.
+ *
+ * Scope starts AT the marker line and runs to the next blank line, rather than to the end of
+ * the message. The echoed envelope has no terminator we can recognise — we build it as a marker
+ * line plus arbitrary result text (protobuf-request.ts), and the observed copies are not
+ * byte-exact, so matching against the replayed envelope is not available either. Truncating to
+ * the end of the message was the alternative, and it discards a genuine answer whenever the
+ * model resumes after the echo. A blank line is the one boundary the model reliably writes when
+ * it goes back to prose.
+ *
+ * The tradeoff is explicit: an echoed envelope whose pasted result itself contains a blank line
+ * leaves its remainder in replay. That is the safer direction to be wrong in — conversation
+ * remint, not this filter, is the primary defence against a poisoned conversation, and this only
+ * stops the transcript from feeding itself.
+ *
+ * Only whole-line markers count, so prose such as "the string [Tool Result] appeared" survives.
+ */
+export function stripAssistantEchoedToolEnvelope(text: string): string {
+  if (!text || !ECHO_MARKERS.some(marker => text.includes(marker))) return text;
+  const newline = text.includes("\r\n") ? "\r\n" : "\n";
+  const lines = text.split(/\r?\n/);
+  const kept: string[] = [];
+  let dropped = false;
+  let index = 0;
+  while (index < lines.length) {
+    const line = lines[index] ?? "";
+    if (!isEchoMarkerLine(line)) {
+      kept.push(line);
+      index += 1;
+      continue;
+    }
+    dropped = true;
+    index += 1;
+    // The envelope body is the contiguous non-blank run after the marker. The blank line that
+    // ends it is left in place, so surviving prose on either side stays separated.
+    while (index < lines.length && (lines[index] ?? "").trim() !== "") index += 1;
+  }
+  if (!dropped) return text;
+  return kept.join(newline).trimEnd();
+}
 const MAX_SNIFF_BYTES = 40;
 /** Mid-stream observer: max leading whitespace on a line before matching disarms. */
 const MAX_MIDSTREAM_LINE_INDENT = 128;
@@ -66,8 +118,9 @@ export interface MidstreamEchoFinding {
  * MIDDLE of an agent message — after legitimate leading text — one of them
  * carrying a whitespace-spliced call-id ("fc_x mar-y" instead of "fc_x-y").
  * Deltas at that point have already reached the client, so this observer
- * never throws and never withholds output: it records findings so the
- * adapter can emit a structured diagnostic at turn end. Only fixed marker
+ * never throws and never withholds output. It records findings so the adapter
+ * can emit a structured diagnostic and remint the conversation for the next
+ * turn at turn end. Only fixed marker
  * enums, numeric offsets, and corruption booleans are retained — never
  * content bytes.
  */

package/src/adapters/cursor/live-transport.ts CHANGED Viewed

@@ -20,6 +20,7 @@ import {
   createCursorContextUsageTracker,
   createCursorProtobufEventState,
   finalizeTurnEvents,
+  hasBufferedTextToolCalls,
   mapCursorProtobufServerMessage,
   mapSyntheticMcpExecToToolEvents,
   reportableContextTokens,
@@ -732,6 +733,8 @@ class LiveCursorTransport implements CursorTransport {
         syntheticStructuredEditToolNames,
         translatorBudget: this.translatorBudget,
         contextUsage,
+        wireModelId: request.modelId,
+        identityScope: request._cursorIdentityScope,
         ...(prepared.estimatedInputTokens !== undefined
           ? { estimatedInputTokens: prepared.estimatedInputTokens }
           : {}),
@@ -1265,6 +1268,7 @@ class LiveCursorTransport implements CursorTransport {
             state.openToolCalls.size > 0
             || this.sawAssistantText
             || hasPendingClientToolFinalization
+            || hasBufferedTextToolCalls(state)
           )
         ) {
           const terminal = hasPendingClientToolFinalization && state.openToolCalls.size === 0
@@ -1431,7 +1435,7 @@ class LiveCursorTransport implements CursorTransport {
           settler.settleFinish();
           return;
         }
-        if (this.framesReceived > 0 && this.sawAssistantText) {
+        if (this.framesReceived > 0 && (this.sawAssistantText || hasBufferedTextToolCalls(state))) {
           for (const event of finalizeTurnEvents(state)) push(event);
           releaseBacklogLease();
           settler.settleFinish();

package/src/adapters/cursor/message-mapper.ts CHANGED Viewed

@@ -46,7 +46,8 @@ export function mapCursorServerMessage(
       state.writeClient(cursorExecResult(message.requestId, message.execCase));
       return [];
     case "local_side_effect":
-      // Internal retry-safety signal only; keep the bridge alive without producing protocol output.
-      return [{ type: "heartbeat" }];
+      // Internal retry-safety signal only; keep the bridge alive without producing protocol output,
+      // while preventing server-level failover from replaying the completed local operation.
+      return [{ type: "heartbeat", replayUnsafe: true }];
   }
 }

package/src/adapters/cursor/protobuf-events.ts CHANGED Viewed

@@ -11,13 +11,19 @@ import {
   isCodexShellBridgeToolName,
   isCursorStructuredEditToolName,
   normalizeCursorWireName,
-  normalizeCursorTextToolMarkers,
   OCX_RESPONSES_TOOL_PROVIDER,
   resolveShellBridgeAliasKey,
   responsesToolNameFromCursorWire,
 } from "./tool-definitions";
+import {
+  drainCursorTextToolCalls,
+  type DrainedTextToolCall,
+  type SuppressedTextToolCallScan,
+} from "./text-toolcall";
+import { recordObservedCursorContextWindow } from "./discovery";
 import type { CursorServerMessage } from "./types";
 import type { TranslatorBudget } from "../../lib/translator-budget";
+import { CURSOR_INCOMPLETE_TOOL_CALL_MESSAGE_PREFIX } from "./cursor-errors";
 const DEFAULT_CONTEXT_USAGE_MAX_ENTRIES = 200;
 const DEFAULT_CONTEXT_USAGE_TTL_MS = 60 * 60 * 1_000;
@@ -176,6 +182,23 @@ export interface CursorProtobufEventState {
    */
   syntheticStructuredEditToolNames?: ReadonlySet<string>;
   translatorBudget?: TranslatorBudget;
+  /**
+   * Incomplete `[TOOL_CALL]…[ARGS]{` prefix held across `textDelta` frames so a
+   * marker split by the stream cannot leak into assistant text.
+   */
+  pendingTextToolCall?: string;
+  /** Constant-space scanner used after an incomplete textual marker exceeds its retained byte cap. */
+  suppressedTextToolCall?: SuppressedTextToolCallScan;
+  /** Parsed textual fallback calls held until turn finalization establishes that no real frame won. */
+  bufferedTextToolCalls?: DrainedTextToolCall[];
+  /** True once this turn carries any real client-tool frame, including an incomplete one. */
+  sawRealClientToolCall?: boolean;
+  /** Monotonic id suffix for tool calls promoted from text markers. */
+  textToolCallSeq?: number;
+  /** Wire model id used to record checkpoint `maxTokens` for the next turn. */
+  wireModelId?: string;
+  /** Normalized Cursor identity scope that owns the observed checkpoint ceiling. */
+  identityScope: string;
 }
@@ -215,6 +238,10 @@ export function createCursorProtobufEventState(options: {
    */
   estimatedInputTokens?: number;
   translatorBudget?: TranslatorBudget;
+  /** Wire model id for recording checkpoint `maxTokens` into the process-local window map. */
+  wireModelId?: string;
+  /** Cursor request identity scope; normalized identically to request-builder continuity. */
+  identityScope?: string;
 } = {}): CursorProtobufEventState {
   return {
     // Cursor provides no authoritative usage frame; token counts are heuristic estimates from
@@ -244,6 +271,8 @@ export function createCursorProtobufEventState(options: {
       && options.estimatedInputTokens > 0
       ? { estimatedInputTokens: options.estimatedInputTokens }
       : {}),
+    ...(options.wireModelId?.trim() ? { wireModelId: options.wireModelId.trim() } : {}),
+    identityScope: options.identityScope?.trim() || "local",
   };
 }
@@ -1048,6 +1077,7 @@ export function mapSyntheticMcpExecToToolEvents(
 ): CursorServerMessage[] {
   if (args.providerIdentifier !== OCX_RESPONSES_TOOL_PROVIDER) return [];
   if (options.state?.terminated) return [];
+  if (options.state) options.state.sawRealClientToolCall = true;
   if (options.allowEmptyArgs !== true && !hasMcpArgBytes(args)) return [];
   const cursorWireName = mcpWireNameFromArgs(args);
   if (!cursorWireName) return [{ type: "error", message: "Cursor requested a Responses tool without a tool name" }];
@@ -1117,6 +1147,11 @@ function recordToolCall(state: CursorProtobufEventState, callId: string, cursorW
   return [];
 }
+function recordRealToolCall(state: CursorProtobufEventState, callId: string, cursorWireName: string): CursorServerMessage[] {
+  state.sawRealClientToolCall = true;
+  return recordToolCall(state, callId, cursorWireName);
+}
 /**
  * Emit a completed client tool call as one atomic unit: `tool_call_start` (deferred from open time),
  * the full normalized arguments delta when present, then `tool_call_end`. The call must already be
@@ -1232,33 +1267,70 @@ export function mapCursorProtobufServerMessage(
   if (state.terminated) return [];
   if (serverMessage.message.case === "conversationCheckpointUpdate") {
-    const usedTokens = serverMessage.message.value.tokenDetails?.usedTokens ?? 0;
+    const tokenDetails = serverMessage.message.value.tokenDetails;
+    const usedTokens = tokenDetails?.usedTokens ?? 0;
     // `usedTokens` is the ABSOLUTE conversation context size, not a per-turn output delta. Track it
     // separately (monotonic max) and surface it as `done.usage.totalTokens`; folding it into
     // `outputTokens` (which also accumulates `tokenDelta`) double-counts in Codex. See contextTokens.
     observeContextTokens(state, usedTokens);
+    // First checkpoints often send maxTokens=0 (senpi). Only a positive ceiling
+    // replaces the id heuristic for the next turn's overflow vs 429 size prior.
+    if (state.wireModelId) {
+      recordObservedCursorContextWindow(state.wireModelId, tokenDetails?.maxTokens, {
+        identityScope: state.identityScope,
+      });
+    }
     return [];
   }
   if (serverMessage.message.case !== "interactionUpdate") return [];
   const update = serverMessage.message.value.message;
   switch (update.case) {
-    case "textDelta":
-      // #2305: fold Cursor display aliases inside textual pseudo tool-call markers back to
-      // the advertised wire name before any client sees the text. Real frames are already
-      // normalized structurally (mcpWireNameFromArgs above).
-      return update.value.text ? [{ type: "text", text: normalizeCursorTextToolMarkers(update.value.text) }] : [];
+    case "textDelta": {
+      // Textual `[TOOL_CALL]name[ARGS]{…}` is not assistant prose. Leaving it in
+      // the text channel (even after #2305 renamed the display alias) leaks a
+      // synthetic protocol marker that later turns few-shot-mimic as inert text.
+      // Strip complete markers, buffer advertised fallbacks until finalize,
+      // and hold or suppress-scan an incomplete opener across deltas.
+      const chunk = update.value.text ?? "";
+      if (!chunk && !state.pendingTextToolCall && !state.suppressedTextToolCall) return [];
+      const drained = drainCursorTextToolCalls(
+        state.pendingTextToolCall ?? "",
+        chunk,
+        state.suppressedTextToolCall,
+      );
+      if (drained.pending) state.pendingTextToolCall = drained.pending;
+      else delete state.pendingTextToolCall;
+      if (drained.suppressed) state.suppressedTextToolCall = drained.suppressed;
+      else delete state.suppressedTextToolCall;
+      const out: CursorServerMessage[] = [];
+      if (drained.text) out.push({ type: "text", text: drained.text });
+      for (const call of drained.calls) {
+        const advertised = resolveAdvertisedClientToolName(state, call.name);
+        if (
+          state.sawRealClientToolCall
+          || !state.clientToolNames
+          || !advertised
+          || (state.bufferedTextToolCalls?.length ?? 0) >= state.maxClientToolCalls
+        ) continue;
+        (state.bufferedTextToolCalls ??= []).push({
+          name: advertised,
+          args: normalizeJsonText(call.args, advertised, state),
+        });
+      }
+      return out;
+    }
     case "thinkingDelta":
       return update.value.text ? [{ type: "thinking", thinking: update.value.text }] : [];
     case "toolCallStarted": {
       const name = mcpCursorWireName(update.value.toolCall);
       // Record the open call but defer the outward tool_call_start to completion (atomic emission).
-      return name ? recordToolCall(state, update.value.callId, name) : [];
+      return name ? recordRealToolCall(state, update.value.callId, name) : [];
     }
     case "partialToolCall": {
       const out: CursorServerMessage[] = [];
       const name = mcpCursorWireName(update.value.toolCall);
-      if (name) out.push(...recordToolCall(state, update.value.callId, name));
+      if (name) out.push(...recordRealToolCall(state, update.value.callId, name));
       if (out.some(event => event.type === "error")) return out;
       // Buffer cumulative args; do not emit a delta. Args are emitted once, normalized, at completion.
       if (state.openToolCalls.has(update.value.callId)) {
@@ -1274,6 +1346,7 @@ export function mapCursorProtobufServerMessage(
       const out: CursorServerMessage[] = [];
       if (state.completedToolCalls.has(update.value.callId)) return [];
       const name = mcpCursorWireName(update.value.toolCall);
+      if (name) state.sawRealClientToolCall = true;
       const args = mcpArgsFromToolCall(update.value.toolCall);
       const openBeforeStart = state.openToolCalls.get(update.value.callId);
       // Empty-arg completion handling:
@@ -1362,20 +1435,46 @@ export function resolvedTurnUsage(state: CursorProtobufEventState): OcxUsage {
  * with corrupt/empty arguments. Emit an explicit error instead of done (fail-closed).
  * Mirrors kiro-truncation.ts behavior.
  */
+/**
+ * True when this turn holds textual fallback tool calls that only turn finalization can emit.
+ *
+ * Cursor can close a stream with a clean Connect END_STREAM and no turnEnded frame. The transport
+ * finalizes that path only for a turn it can see is unfinished, and a turn whose entire visible
+ * text was a stripped marker looks empty from the outside. Without this the deferred fallback
+ * would be dropped exactly when the marker was the turn's only content.
+ */
+export function hasBufferedTextToolCalls(state: CursorProtobufEventState): boolean {
+  return (state.bufferedTextToolCalls?.length ?? 0) > 0;
+}
 export function finalizeTurnEvents(state: CursorProtobufEventState): CursorServerMessage[] {
   state.terminated = true;
+  delete state.pendingTextToolCall;
+  delete state.suppressedTextToolCall;
+  const bufferedTextToolCalls = state.bufferedTextToolCalls ?? [];
+  delete state.bufferedTextToolCalls;
   if (state.openToolCalls.size > 0) {
     const openCallIds = [...state.openToolCalls.keys()];
     const openIds = openCallIds.join(", ");
     // Clear so a second turnEnded (should not happen, but defensive) doesn't re-emit.
     for (const callId of openCallIds) state.translatorBudget?.closeCall(callId);
     state.openToolCalls.clear();
-    return [{ type: "error", message: `Cursor stream ended with incomplete tool call(s): ${openIds}. Arguments may be truncated; the call was not committed.` }];
+    return [{ type: "error", message: `${CURSOR_INCOMPLETE_TOOL_CALL_MESSAGE_PREFIX} ${openIds}. Arguments may be truncated; the call was not committed.` }];
+  }
+  const out: CursorServerMessage[] = [];
+  if (!state.sawRealClientToolCall) {
+    for (const call of bufferedTextToolCalls) {
+      state.textToolCallSeq = (state.textToolCallSeq ?? 0) + 1;
+      const callId = `textcall_${state.textToolCallSeq}`;
+      out.push(...recordToolCall(state, callId, call.name));
+      if (state.openToolCalls.has(callId)) out.push(...commitToolCall(state, callId, call.args));
+    }
   }
   // Surface the absolute context size (when Cursor reported a checkpoint) as both totalTokens and
   // the estimated input side of Codex's visible `input + output` counter. Codex status lines can
   // render the additive pair instead of total_tokens, so leaving inputTokens at 0 makes a 16k-context
   // first turn display as "9 used". Keep outputTokens as the per-turn delta and clamp the inferred
   // input to 0 in case Cursor reports a checkpoint smaller than the streamed output delta.
-  return [{ type: "done", usage: resolvedTurnUsage(state) }];
+  out.push({ type: "done", usage: resolvedTurnUsage(state) });
+  return out;
 }