mohdel 0.91.0 → 0.93.0

package/js/core/events.js

@@ -64,6 +64,13 @@
  *   `'insufficientOutputBudget' | 'cancelled' | ...` additive union.
  * @property {ToolCall[]} [toolCalls]
  *   Present when `status === 'tool_use'`.
+ * @property {number} [maxInterFrameMs]
+ *   Longest gap (ms) between adapter events during the call —
+ *   from `startedAt` to the first frame, between consecutive
+ *   frames, and from the last frame to the terminal. Direct signal
+ *   for calibrating downstream read timeouts: a 15-min call that
+ *   streams deltas every 30s is safe; a 5-min call with zero
+ *   intermediate frames is dangerous.
  */
 
 /**
@@ -83,6 +90,12 @@
  * @property {string} id
  * @property {string} name
  * @property {object} arguments
+ * @property {string} [thoughtSignature]
+ *   Provider-specific opaque blob carried through tool-call
+ *   round-trips to preserve thinking state continuity. Set by the
+ *   Gemini adapter when the model emits one; absent for other
+ *   providers. Callers replaying tool results should pass the
+ *   ToolCall back unchanged so the adapter can re-attach it.
  */
 
 export const EVENT_TYPES = Object.freeze(['delta', 'done', 'error'])

package/js/session/run.js

@@ -141,9 +141,33 @@ export async function * run (envelope, {
   }
 
   let sawTerminal = false
+  // Track whether the adapter emitted any `delta` events during the
+  // call. Recorded as `mohdel.stream` on the span at finalize time —
+  // a non-streaming terminal (adapter fell back to wait-for-done, or
+  // the provider SDK collapsed stream events) is a hazard for any
+  // host-side read-timeout watchdog and shows up in traces before
+  // it turns into an abort downstream.
+  let sawDelta = false
+  // Track the longest gap between adapter events within this call —
+  // from `startedAt` to the first frame, between consecutive frames,
+  // and from the last frame to the terminal. A direct signal for any
+  // host-side read-timeout watchdog: a 15-min call that streams
+  // deltas every 30s is safe; a 5-min call with zero intermediate
+  // frames is dangerous. Surfaced on `AnswerResult.maxInterFrameMs`
+  // and the span's `mohdel.max_inter_frame_ms` attribute so hosts
+  // can aggregate it for adaptive-timeout calibration.
+  let lastFrameAt = startedAt
+  let maxInterFrameMs = 0
   try {
     for await (const ev of adapter(envelope, { signal, log, span })) {
-      if (ev.type === 'done') {
+      const now = Date.now()
+      const gap = now - lastFrameAt
+      if (gap > maxInterFrameMs) maxInterFrameMs = gap
+      lastFrameAt = now
+
+      if (ev.type === 'delta') {
+        sawDelta = true
+      } else if (ev.type === 'done') {
         sawTerminal = true
         // A cancelled terminal is the caller's action, not evidence
         // of provider recovery — don't wipe an accumulated failure
@@ -160,7 +184,11 @@ export async function * run (envelope, {
             (ev.result.thinkingTokens || 0)
           if (total > 0) limiter.recordTokens(bucketKey, total)
         }
-        finalizeSpanOk(span, ev.result)
+        // Surface on AnswerResult so hosts that pass the whole
+        // result upstream pick it up without needing a separate
+        // wire field.
+        if (ev.result) ev.result.maxInterFrameMs = maxInterFrameMs
+        finalizeSpanOk(span, ev.result, sawDelta, maxInterFrameMs)
         log.debug(summarizeDone(ev.result, startedAt), '[mohdel:answer] done')
       } else if (ev.type === 'error') {
         sawTerminal = true
@@ -168,7 +196,8 @@ export async function * run (envelope, {
         log.warn({
           err: ev.error,
           provider: envelope.provider,
-          totalMs: Date.now() - startedAt
+          totalMs: Date.now() - startedAt,
+          maxInterFrameMs
         }, '[mohdel:answer] failed')
         endSpanError(span, new Error(ev.error?.message || 'adapter error'))
       }
@@ -177,11 +206,12 @@ export async function * run (envelope, {
   } catch (e) {
     if (signal?.aborted && !sawTerminal) {
       const fallback = cancelledFallback()
-      finalizeSpanOk(span, fallback.result)
+      if (fallback.result) fallback.result.maxInterFrameMs = maxInterFrameMs
+      finalizeSpanOk(span, fallback.result, sawDelta, maxInterFrameMs)
       yield fallback
       return
     }
-    log.warn({ err: e, provider: envelope.provider }, '[mohdel:answer] adapter threw')
+    log.warn({ err: e, provider: envelope.provider, maxInterFrameMs }, '[mohdel:answer] adapter threw')
     endSpanError(span, e)
     yield errorEvent(messageOf(e), 'SESSION_ADAPTER_THREW')
     return
@@ -190,11 +220,12 @@ export async function * run (envelope, {
   if (!sawTerminal) {
     if (signal?.aborted) {
       const fallback = cancelledFallback()
-      finalizeSpanOk(span, fallback.result)
+      if (fallback.result) fallback.result.maxInterFrameMs = maxInterFrameMs
+      finalizeSpanOk(span, fallback.result, sawDelta, maxInterFrameMs)
       yield fallback
     } else {
       const err = 'adapter returned without a terminal event'
-      log.error({ provider: envelope.provider }, '[mohdel:answer] no terminal event')
+      log.error({ provider: envelope.provider, maxInterFrameMs }, '[mohdel:answer] no terminal event')
       endSpanError(span, new Error(err))
       yield errorEvent(err, 'SESSION_ADAPTER_NO_TERMINAL')
     }
@@ -288,14 +319,25 @@ function scopedLogger (logger, envelope, span) {
 /**
  * @param {any} span
  * @param {any} result
+ * @param {boolean} sawDelta  Whether the adapter emitted at least one
+ *   `delta` event during the call. Surfaces as `mohdel.stream` on the
+ *   span so traces can flag calls that completed without streaming —
+ *   a non-streaming long-running call is invisible to upstream idle
+ *   watchdogs until the terminal lands.
+ * @param {number} maxInterFrameMs  Longest gap between adapter events
+ *   during the call (including pre-first-frame and post-last-frame).
+ *   Surfaces as `mohdel.max_inter_frame_ms` so downstream timeout
+ *   calibration has a direct signal independent of total elapsed.
  */
-function finalizeSpanOk (span, result) {
+function finalizeSpanOk (span, result, sawDelta = false, maxInterFrameMs = 0) {
   /** @type {Record<string, any>} */
   const attrs = {
     'gen_ai.usage.input_tokens': result?.inputTokens || 0,
     'gen_ai.usage.output_tokens': result?.outputTokens || 0,
     'mohdel.thinking_tokens': result?.thinkingTokens || 0,
-    'mohdel.status': result?.status || 'unknown'
+    'mohdel.status': result?.status || 'unknown',
+    'mohdel.stream': !!sawDelta,
+    'mohdel.max_inter_frame_ms': maxInterFrameMs
   }
   if (result?.cost != null) attrs['mohdel.cost'] = result.cost
   if (result?.warning) attrs['mohdel.warning'] = result.warning
@@ -323,7 +365,8 @@ function summarizeDone (result, startedAt) {
     think: result?.thinkingTokens || 0,
     cost: result?.cost,
     warning: result?.warning,
-    totalMs: Date.now() - startedAt
+    totalMs: Date.now() - startedAt,
+    maxInterFrameMs: result?.maxInterFrameMs
   }
 }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mohdel",
-  "version": "0.91.0",
+  "version": "0.93.0",
   "license": "MIT",
   "author": {
     "name": "Christophe Le Bars",
@@ -87,10 +87,10 @@
     "@opentelemetry/exporter-trace-otlp-grpc": "^0.215.0",
     "@opentelemetry/sdk-node": "^0.215.0",
     "chalk": "^5.4.0",
-    "mohdel-thin-gate-linux-x64-gnu": "0.91.0"
+    "mohdel-thin-gate-linux-x64-gnu": "0.93.0"
   },
   "dependencies": {
-    "@anthropic-ai/sdk": "^0.90.0",
+    "@anthropic-ai/sdk": "^0.91.0",
     "@cerebras/cerebras_cloud_sdk": "^1.61.1",
     "@google/genai": "^1.50.1",
     "@opentelemetry/api": "^1.9.1",

package/src/lib/catalog/anthropic.js

@@ -0,0 +1,44 @@
+const BASE_URL = 'https://api.anthropic.com/v1'
+const API_VERSION = '2023-06-01'
+
+const fetchModels = async ({ apiKey }) => {
+  const all = []
+  let afterId = null
+  while (true) {
+    const url = new URL(`${BASE_URL}/models`)
+    url.searchParams.set('limit', '100')
+    if (afterId) url.searchParams.set('after_id', afterId)
+    const res = await fetch(url, {
+      headers: {
+        'x-api-key': apiKey,
+        'anthropic-version': API_VERSION
+      }
+    })
+    if (!res.ok) throw new Error(`${res.status} ${res.statusText} fetching ${url}`)
+    const body = await res.json()
+    const page = Array.isArray(body?.data) ? body.data : []
+    all.push(...page)
+    if (!body?.has_more || !page.length) break
+    afterId = page[page.length - 1].id
+  }
+  return all
+}
+
+export default (sdkConfig) => {
+  let cache = null
+  const load = async () => {
+    if (!cache) cache = await fetchModels(sdkConfig)
+    return cache
+  }
+  return {
+    listModels: async () => (await load()).map(m => ({ id: m.id, label: m.display_name || m.id })),
+    getModelInfo: async (id) => {
+      const m = (await load()).find(x => x.id === id)
+      if (!m) return null
+      const info = { model: m.id }
+      if (m.display_name) info.label = m.display_name
+      if (m.created_at) info.createdAt = m.created_at
+      return info
+    }
+  }
+}

package/src/lib/catalog/cerebras.js

@@ -0,0 +1,26 @@
+const BASE_URL = 'https://api.cerebras.ai/v1'
+
+const fetchModels = async ({ apiKey }) => {
+  const res = await fetch(`${BASE_URL}/models`, { headers: { Authorization: `Bearer ${apiKey}` } })
+  if (!res.ok) throw new Error(`${res.status} ${res.statusText} fetching cerebras models`)
+  const body = await res.json()
+  return Array.isArray(body?.data) ? body.data : []
+}
+
+export default (sdkConfig) => {
+  let cache = null
+  const load = async () => {
+    if (!cache) cache = await fetchModels(sdkConfig)
+    return cache
+  }
+  return {
+    listModels: async () => (await load()).map(m => ({ id: m.id, label: m.id })),
+    getModelInfo: async (id) => {
+      const m = (await load()).find(x => x.id === id)
+      if (!m) return null
+      const info = { model: m.id }
+      if (typeof m.created === 'number') info.created = m.created
+      return info
+    }
+  }
+}

package/src/lib/catalog/fireworks.js

@@ -0,0 +1,38 @@
+// Fireworks' /inference/v1/models is openai-compatible for listing.
+// Upstream IDs carry a `accounts/fireworks/models/` prefix that the
+// runtime adapter adds back automatically — strip it here so curated
+// entries stay short.
+const DEFAULT_BASE_URL = 'https://api.fireworks.ai/inference/v1'
+const FW_PREFIX = 'accounts/fireworks/models/'
+
+const shortId = (id) => id?.startsWith(FW_PREFIX) ? id.slice(FW_PREFIX.length) : id
+
+const fetchModels = async ({ apiKey, baseURL }) => {
+  const url = `${(baseURL || DEFAULT_BASE_URL).replace(/\/$/, '')}/models`
+  const res = await fetch(url, { headers: { Authorization: `Bearer ${apiKey}` } })
+  if (!res.ok) throw new Error(`${res.status} ${res.statusText} fetching ${url}`)
+  const body = await res.json()
+  return Array.isArray(body?.data) ? body.data : []
+}
+
+export default (sdkConfig) => {
+  let cache = null
+  const load = async () => {
+    if (!cache) cache = await fetchModels(sdkConfig)
+    return cache
+  }
+  return {
+    listModels: async () => (await load()).map(m => {
+      const id = shortId(m.id)
+      return { id, label: id }
+    }),
+    getModelInfo: async (id) => {
+      const m = (await load()).find(x => shortId(x.id) === id)
+      if (!m) return null
+      const info = { model: shortId(m.id) }
+      if (typeof m.context_length === 'number') info.contextTokenLimit = m.context_length
+      if (typeof m.created === 'number') info.created = m.created
+      return info
+    }
+  }
+}

package/src/lib/catalog/gemini.js

@@ -0,0 +1,46 @@
+const BASE_URL = 'https://generativelanguage.googleapis.com/v1beta'
+
+const fetchModels = async ({ apiKey }) => {
+  const all = []
+  let pageToken = null
+  while (true) {
+    const url = new URL(`${BASE_URL}/models`)
+    url.searchParams.set('key', apiKey)
+    url.searchParams.set('pageSize', '1000')
+    if (pageToken) url.searchParams.set('pageToken', pageToken)
+    const res = await fetch(url)
+    if (!res.ok) throw new Error(`${res.status} ${res.statusText} fetching ${url.pathname}`)
+    const body = await res.json()
+    const page = Array.isArray(body?.models) ? body.models : []
+    all.push(...page)
+    if (!body?.nextPageToken) break
+    pageToken = body.nextPageToken
+  }
+  return all
+}
+
+const shortId = (name) => name.startsWith('models/') ? name.slice('models/'.length) : name
+
+export default (sdkConfig) => {
+  let cache = null
+  const load = async () => {
+    if (!cache) cache = await fetchModels(sdkConfig)
+    return cache
+  }
+  return {
+    listModels: async () => (await load())
+      .filter(m => (m.supportedGenerationMethods || []).includes('generateContent'))
+      .map(m => ({ id: shortId(m.name), label: m.displayName || shortId(m.name) })),
+    getModelInfo: async (id) => {
+      const m = (await load()).find(x => shortId(x.name) === id)
+      if (!m) return null
+      const info = { model: shortId(m.name) }
+      if (m.displayName) info.label = m.displayName
+      if (m.description) info.description = m.description
+      if (typeof m.inputTokenLimit === 'number') info.contextTokenLimit = m.inputTokenLimit
+      if (typeof m.outputTokenLimit === 'number') info.outputTokenLimit = m.outputTokenLimit
+      if (m.version) info.version = m.version
+      return info
+    }
+  }
+}

package/src/lib/catalog/groq.js

@@ -0,0 +1,30 @@
+const BASE_URL = 'https://api.groq.com/openai/v1'
+
+const fetchModels = async ({ apiKey }) => {
+  const res = await fetch(`${BASE_URL}/models`, { headers: { Authorization: `Bearer ${apiKey}` } })
+  if (!res.ok) throw new Error(`${res.status} ${res.statusText} fetching groq models`)
+  const body = await res.json()
+  return Array.isArray(body?.data) ? body.data : []
+}
+
+export default (sdkConfig) => {
+  let cache = null
+  const load = async () => {
+    if (!cache) cache = await fetchModels(sdkConfig)
+    return cache
+  }
+  return {
+    listModels: async () => (await load())
+      .filter(m => m.active !== false)
+      .map(m => ({ id: m.id, label: m.id })),
+    getModelInfo: async (id) => {
+      const m = (await load()).find(x => x.id === id)
+      if (!m) return null
+      const info = { model: m.id }
+      if (typeof m.context_window === 'number') info.contextTokenLimit = m.context_window
+      if (typeof m.max_completion_tokens === 'number') info.outputTokenLimit = m.max_completion_tokens
+      if (typeof m.created === 'number') info.created = m.created
+      return info
+    }
+  }
+}

package/src/lib/catalog/openai.js

@@ -0,0 +1,34 @@
+// Catalog fetcher for any OpenAI-compatible /v1/models endpoint
+// (openai, deepseek, mistral, novita, xai). Inference path lives in
+// js/session/adapters/openai.js — this module is CLI-only.
+
+const DEFAULT_BASE_URL = 'https://api.openai.com/v1'
+
+const fetchModels = async ({ apiKey, baseURL }) => {
+  const url = `${(baseURL || DEFAULT_BASE_URL).replace(/\/$/, '')}/models`
+  const res = await fetch(url, { headers: { Authorization: `Bearer ${apiKey}` } })
+  if (!res.ok) throw new Error(`${res.status} ${res.statusText} fetching ${url}`)
+  const body = await res.json()
+  return Array.isArray(body?.data) ? body.data : []
+}
+
+export default (sdkConfig) => {
+  let cache = null
+  const load = async () => {
+    if (!cache) cache = await fetchModels(sdkConfig)
+    return cache
+  }
+  return {
+    listModels: async () => (await load()).map(m => ({ id: m.id, label: m.id })),
+    getModelInfo: async (id) => {
+      const m = (await load()).find(x => x.id === id)
+      if (!m) return null
+      const info = { model: m.id }
+      if (typeof m.context_window === 'number') info.contextTokenLimit = m.context_window
+      if (typeof m.max_context_length === 'number') info.contextTokenLimit = m.max_context_length
+      if (typeof m.max_completion_tokens === 'number') info.outputTokenLimit = m.max_completion_tokens
+      if (typeof m.created === 'number') info.created = m.created
+      return info
+    }
+  }
+}

package/src/lib/catalog/openrouter.js

@@ -0,0 +1,48 @@
+const BASE_URL = 'https://openrouter.ai/api/v1'
+
+const fetchModels = async ({ apiKey }) => {
+  const res = await fetch(`${BASE_URL}/models`, { headers: { Authorization: `Bearer ${apiKey}` } })
+  if (!res.ok) throw new Error(`${res.status} ${res.statusText} fetching openrouter models`)
+  const body = await res.json()
+  return Array.isArray(body?.data) ? body.data : []
+}
+
+// OpenRouter prices are decimal USD per token — convert to per-million-tokens.
+const toPerMillion = (raw) => {
+  const n = typeof raw === 'string' ? parseFloat(raw) : raw
+  if (!Number.isFinite(n)) return undefined
+  return n * 1_000_000
+}
+
+export default (sdkConfig) => {
+  let cache = null
+  const load = async () => {
+    if (!cache) cache = await fetchModels(sdkConfig)
+    return cache
+  }
+  return {
+    listModels: async () => (await load()).map(m => ({ id: m.id, label: m.name || m.id })),
+    getModelInfo: async (id) => {
+      const m = (await load()).find(x => x.id === id)
+      if (!m) return null
+      const info = { model: m.id }
+      if (m.name) info.label = m.name
+      if (m.description) info.description = m.description
+      if (typeof m.context_length === 'number') info.contextTokenLimit = m.context_length
+      const topCtx = m.top_provider?.context_length
+      if (!info.contextTokenLimit && typeof topCtx === 'number') info.contextTokenLimit = topCtx
+      const maxOut = m.top_provider?.max_completion_tokens
+      if (typeof maxOut === 'number') info.outputTokenLimit = maxOut
+      const inP = toPerMillion(m.pricing?.prompt)
+      const outP = toPerMillion(m.pricing?.completion)
+      if (inP !== undefined) info.inputPrice = inP
+      if (outP !== undefined) info.outputPrice = outP
+      const modalities = m.architecture?.input_modalities
+      if (Array.isArray(modalities) && modalities.length) {
+        info.inputFormat = modalities.filter(x => typeof x === 'string')
+      }
+      if (typeof m.created === 'number') info.created = m.created
+      return info
+    }
+  }
+}

package/src/lib/select.js

@@ -37,11 +37,9 @@ export const initializeAPIs = async () => {
       // Create configuration
       const sdkConfig = config.createConfiguration(apiKey)
 
-      // Import the SDK module dynamically
-      const sdkPath = `./sdk/${config.sdk}.js`
+      const sdkPath = `./catalog/${config.sdk}.js`
       const { default: API } = await import(sdkPath)
 
-      // Initialize the provider with the configuration, no specs
       api[name] = API(sdkConfig, {}, silent)
       providersWithKeys.push(name)
     } catch (err) {
@@ -195,10 +193,22 @@ export const promptMissingFields = async (entry, curatedKey) => {
     } else {
       const creatorVal = await clack.select({
         message: `Creator for ${curatedKey}:`,
-        options: creatorsList.map(c => ({ value: c, label: c }))
+        options: [
+          ...creatorsList.map(c => ({ value: c, label: c })),
+          { value: '__other', label: 'Other… (enter a name)' }
+        ]
       })
       if (clack.isCancel(creatorVal)) return entry
-      entry.creator = creatorVal
+      if (creatorVal === '__other') {
+        const custom = await clack.text({
+          message: `New creator name for ${curatedKey}:`,
+          validate: (v) => v?.trim() ? undefined : 'Creator is required'
+        })
+        if (clack.isCancel(custom)) return entry
+        entry.creator = custom.trim()
+      } else {
+        entry.creator = creatorVal
+      }
     }
   }