mohdel 0.101.0 → 0.103.0

package/js/session/adapters/_chat_completions.js

@@ -330,14 +330,16 @@ function buildRequest (envelope, spec, config) {
 
   if (spec.thinkingEffortLevels) {
     const effort = envelope.outputEffort ?? spec.defaultThinkingEffort ?? 'low'
-    if (effort && effort !== 'none') {
+    if (effort && spec.thinkingEffortLevels[effort] != null) {
       const headroom = spec.thinkingEffortLevels[effort]
       if (args.max_tokens && typeof headroom === 'number') {
         args.max_tokens += headroom
       }
-      delete args.temperature
+      // When reasoning is disabled ('none') the model accepts
+      // temperature again — only strip it when reasoning is on.
+      if (effort !== 'none') delete args.temperature
       if (config.reasoningField === 'cerebras_zai' && /zai/i.test(bareOf(envelope.model))) {
-        args.disable_reasoning = false
+        args.disable_reasoning = (effort === 'none')
       } else {
         args.reasoning_effort = effort
       }

package/js/session/adapters/anthropic.js

@@ -203,12 +203,18 @@ export async function * anthropic (envelope, deps = {}) {
   }
 
   const end = String(process.hrtime.bigint())
-  // Estimate thinking tokens from streamed thinking_delta char count
-  // (Anthropic API doesn't report them separately). Cap at total
-  // output tokens reported by usage.
+  // Estimate thinking tokens. Primary path: count streamed thinking_delta
+  // chars (sonnet emits these). Fallback: gap between Anthropic's reported
+  // output_tokens and what actually streamed as visible output (text +
+  // tool input JSON) — catches redacted_thinking blocks (opus 4.7 default)
+  // that consume output tokens but emit no streaming deltas.
+  const streamedOutput = currentOutput()
+  const streamedOutputChars = streamedOutput.length +
+    [...toolBlocks.values()].reduce((s, b) => s + b.inputJson.length, 0)
+  const streamedOutputTokens = Math.ceil(streamedOutputChars / ANTHROPIC_THINKING_CHARS_PER_TOKEN)
   const estimatedThinkingTokens = thinkingChars > 0
     ? Math.min(Math.ceil(thinkingChars / ANTHROPIC_THINKING_CHARS_PER_TOKEN), outputTokens)
-    : 0
+    : Math.max(0, outputTokens - streamedOutputTokens)
   const messageOutputTokens = Math.max(0, outputTokens - estimatedThinkingTokens)
 
   /** @type {import('#core/events.js').DoneEvent} */
@@ -216,7 +222,7 @@ export async function * anthropic (envelope, deps = {}) {
     type: 'done',
     result: {
       status,
-      output: currentOutput() || null,
+      output: streamedOutput || null,
       inputTokens,
       outputTokens: messageOutputTokens,
       thinkingTokens: estimatedThinkingTokens,

package/js/session/adapters/openai.js

@@ -225,19 +225,18 @@ function buildRequest (envelope, input, instructions) {
 
   // Thinking: when the spec has `thinkingEffortLevels`, set
   // `reasoning.effort` and add the thinking-budget headroom on top
-  // of the user's `outputBudget`. `reasoning` is an OpenAI-only
-  // parameter — xAI reasoning is automatic, so add the headroom
-  // but skip the request field on xAI.
+  // of the user's `outputBudget`. Both OpenAI (gpt-5.x) and xAI
+  // (grok-4.3+) accept the same `reasoning: { effort }` shape on
+  // the Responses API, including the literal value 'none' to
+  // disable reasoning entirely.
   if (spec?.thinkingEffortLevels) {
     const effort = envelope.outputEffort ?? spec.defaultThinkingEffort ?? 'low'
-    if (effort && effort !== 'none') {
+    if (effort && spec.thinkingEffortLevels[effort] != null) {
       const headroom = spec.thinkingEffortLevels[effort]
       if (request.max_output_tokens && typeof headroom === 'number') {
         request.max_output_tokens += headroom
       }
-      if (provider === 'openai') {
-        request.reasoning = { effort }
-      }
+      request.reasoning = { effort }
     }
   }
 

package/js/session/run.js

@@ -264,12 +264,16 @@ export async function * run (envelope, {
 function normalizeModelEffort (envelope, resolveSpec) {
   const candidate = effortOf(envelope.model)
   if (!candidate) return { envelope }
-  if (envelope.outputEffort) return { envelope } // explicit wins
 
   const base = catalogKey(envelope.model)
   const baseSpec = resolveSpec(base)
   if (!baseSpec) return { envelope } // base not known — let full string fall through to not-found
 
+  // Explicit outputEffort wins; still strip the suffix so spans/logs see the canonical id.
+  if (envelope.outputEffort) {
+    return { envelope: { ...envelope, model: base } }
+  }
+
   if (!baseSpec.thinkingEffortLevels) {
     return {
       envelope,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mohdel",
-  "version": "0.101.0",
+  "version": "0.103.0",
   "license": "MIT",
   "author": {
     "name": "Christophe Le Bars",
@@ -84,19 +84,19 @@
   },
   "optionalDependencies": {
     "@clack/prompts": "^1.3.0",
-    "@opentelemetry/exporter-trace-otlp-grpc": "^0.216.0",
-    "@opentelemetry/sdk-node": "^0.216.0",
+    "@opentelemetry/exporter-trace-otlp-grpc": "^0.217.0",
+    "@opentelemetry/sdk-node": "^0.217.0",
     "chalk": "^5.4.0",
-    "mohdel-thin-gate-linux-x64-gnu": "0.101.0"
+    "mohdel-thin-gate-linux-x64-gnu": "0.103.0"
   },
   "dependencies": {
-    "@anthropic-ai/sdk": "^0.91.1",
+    "@anthropic-ai/sdk": "^0.95.1",
     "@cerebras/cerebras_cloud_sdk": "^1.61.1",
-    "@google/genai": "^1.51.0",
+    "@google/genai": "^2.0.0",
     "@opentelemetry/api": "^1.9.1",
     "env-paths": "^4.0.0",
     "groq-sdk": "^1.1.2",
-    "openai": "^6.35.0",
+    "openai": "^6.37.0",
     "undici": "^7.24.5"
   },
   "lint-staged": {
@@ -104,7 +104,7 @@
   },
   "devDependencies": {
     "gpt-tokenizer": "^3.4.0",
-    "lint-staged": "^16.4.0",
+    "lint-staged": "^17.0.3",
     "release-it": "^20.0.1",
     "standard": "^17.1.2",
     "vitest": "^4.1.5"