@magpie-community/opencode-cursor-auth 0.2.5 → 0.2.7

package/README.md

@@ -63,7 +63,11 @@ plugin's `fetch` answers them on Cursor's agent API:
   so subagents under one session are conversations of their own. With
   nothing naming the session, each Run has a new one.
 - Text, thinking (`reasoning_content`) and token usage come back streamed
-  or not. Cursor's `input_tokens` counts the cached prompt too: the cache
+  or not. Replayed assistant `reasoning_content` (or `reasoning`) is kept as
+  AI SDK reasoning parts in the next Run's transcript, including when a
+  caller splits parallel tool calls across assistant messages. It stays
+  separate from reply text; no thinking signature is invented.
+  Cursor's `input_tokens` counts the cached prompt too: the cache
   read and written is taken out of it, and given as `cached_tokens` and
   `cache_write_tokens`, so it is counted once; its reasoning tokens are
   `reasoning_tokens`. A step that calls tools gets no usage in its Run
@@ -131,6 +135,13 @@ Accounts with a Grok Bot allowance also show its usage and reset time;
 active trials are labelled separately. A spent Bot allowance doesn't stop
 Cursor requests.
 
+An account that may spend on-demand once its included usage is gone (a
+team or personal spend limit, as `cursor-agent`'s usage view shows it) also
+shows On-demand: its spend, and its limit when it has one. Cursor keeps
+serving such an account past the included usage, so magpie doesn't count
+it used up when Cursor Models or Other Models reach 100%; it does when the
+on-demand spend reaches its limit.
+
 Sign in to the Cursor account linked in Grok Bot. A linked SuperGrok
 subscription grants Bot access without a paid Cursor plan; see
 [Grok Bot plans and billing](https://cursor.com/help/grok-bot/plans).

package/index.mjs

@@ -380,16 +380,67 @@ function firstParty(model) {
   return ["grok-4.5", "grok-4.6", "grok-4.7"].some((b) => model === b || model.startsWith(b + "-"))
 }
 
+// onDemandOf is what Cursor still serves once the included usage is spent,
+// as cursor-agent's own usage view tells it (its On-demand line): from the
+// period's spendLimitUsage (cents) and, where that sets no limit of the
+// user's, GetHardLimit's reply (dollars; undefined when it couldn't be
+// read). kind is "fixed" (up to limitDollars), "unlimited", "disabled" or
+// "unavailable" (not known).
+function onDemandOf(data, hard) {
+  const s = data?.spendLimitUsage
+  const usedDollars = Number(s?.individualUsed ?? 0) / 100
+  const own =
+    s?.individualLimit === undefined
+      ? undefined
+      : Number(s.individualLimit) > 0
+        ? { kind: "fixed", usedDollars, limitDollars: Number(s.individualLimit) / 100 }
+        : { kind: "disabled", usedDollars }
+  if (s?.limitType === "team") {
+    if (own) return own
+    if (hard) return hard.noUsageBasedAllowed || !(Number(hard.hardLimit) > 0) ? { kind: "disabled", usedDollars } : { kind: "unlimited", usedDollars }
+    return Number(s.pooledLimit ?? 0) > 0 ? { kind: "unlimited", usedDollars } : { kind: "unavailable", usedDollars }
+  }
+  if (!hard) return own ?? { kind: "unavailable", usedDollars }
+  if (hard.noUsageBasedAllowed) return { kind: "disabled", usedDollars }
+  const limit = Number(hard.hardLimit ?? 0)
+  if (limit >= 2147483647) return { kind: "unlimited", usedDollars }
+  return limit > 0 ? { kind: "fixed", usedDollars, limitDollars: limit } : { kind: "disabled", usedDollars }
+}
+
+// hardLimit is GetHardLimit's reply, undefined when it couldn't be read.
+async function hardLimit(tok) {
+  try {
+    const res = await fetch(API + "/aiserver.v1.DashboardService/GetHardLimit", {
+      method: "POST",
+      headers: { Authorization: `Bearer ${tok}`, "Content-Type": "application/json", "Connect-Protocol-Version": "1" },
+      body: "{}",
+      signal: AbortSignal.timeout(15_000),
+    })
+    return res.ok ? ((await res.json()) ?? undefined) : undefined
+  } catch {
+    return undefined
+  }
+}
+
+const dollars = (n) => "$" + n.toFixed(2)
+
 // usage is the account's windows this billing period; an enterprise plan
 // reports spend instead, and gets none. ids are the models it can be
-// asked for, each counted by the pool Cursor bills it to.
+// asked for, each counted by the pool Cursor bills it to. An account that
+// may spend on-demand once its included usage is gone (fottencity on
+// Discord: a team that pays for more) is still served when the pools are
+// spent, so they are shown aside, and its On-demand spend is what can run
+// out: at its limit, or never when it has none.
 async function usage(tok, ids) {
-  const res = await fetch(API + "/aiserver.v1.DashboardService/GetCurrentPeriodUsage", {
-    method: "POST",
-    headers: { Authorization: `Bearer ${tok}`, "Content-Type": "application/json", "Connect-Protocol-Version": "1" },
-    body: "{}",
-    signal: AbortSignal.timeout(15_000),
-  })
+  const [res, hard] = await Promise.all([
+    fetch(API + "/aiserver.v1.DashboardService/GetCurrentPeriodUsage", {
+      method: "POST",
+      headers: { Authorization: `Bearer ${tok}`, "Content-Type": "application/json", "Connect-Protocol-Version": "1" },
+      body: "{}",
+      signal: AbortSignal.timeout(15_000),
+    }),
+    hardLimit(tok),
+  ])
   if (!res.ok) return { error: statusText(res.status), windows: [] }
   const data = await res.json()
   const u = data?.planUsage
@@ -412,11 +463,20 @@ async function usage(tok, ids) {
   const asked = ids.flatMap((id) => [id, id.replace(/@[^@]*$/, "")])
   const pool = [...new Set(["auto", ...asked, ...(data.autoBucketModels ?? [])])].filter(inPool)
   const num = (v) => (typeof v === "number" ? v : 0)
+  const od = onDemandOf(data, hard)
+  const pooled = od.kind === "fixed" || od.kind === "unlimited" ? { aside: true } : {}
+  const onDemand =
+    od.kind === "fixed"
+      ? [{ name: "On-demand", used: Math.max(0, Math.min(100, (100 * od.usedDollars) / od.limitDollars)), display: `${dollars(od.usedDollars)} / ${dollars(od.limitDollars)}`, ...resets }]
+      : od.kind === "unlimited"
+        ? [{ name: "On-demand", used: 0, display: dollars(od.usedDollars), ...resets, aside: true }]
+        : []
   // the two pools fit the line; the total goes in its tooltip
   return {
     windows: [
-      { name: "Cursor Models", used: num(u.autoPercentUsed), ...resets, models: pool },
-      { name: "Other Models", used: num(u.apiPercentUsed), ...resets, notModels: pool },
+      { name: "Cursor Models", used: num(u.autoPercentUsed), ...resets, models: pool, ...pooled },
+      { name: "Other Models", used: num(u.apiPercentUsed), ...resets, notModels: pool, ...pooled },
+      ...onDemand,
       { name: "Total", used: num(u.totalPercentUsed), ...resets, aside: true },
     ],
   }
@@ -1021,6 +1081,10 @@ function statusOf(status, code, msg) {
 const textOf = (c) =>
   typeof c === "string" ? c : Array.isArray(c) ? c.filter((p) => p?.type === "text").map((p) => p.text ?? "").join("") : ""
 
+// Chat clients replay thinking beside content. Cursor's transcript blobs
+// use AI SDK reasoning parts; putting it in text would change its role.
+const reasoningOf = (m) => [m.reasoning_content, m.reasoning].find((s) => typeof s === "string" && s.trim()) ?? ""
+
 // joinSplitCalls joins an assistant message onto the assistant message
 // before it when that one made tool calls: Anthropic takes consecutive
 // assistant messages as one turn, and an agent can send a turn's parallel
@@ -1033,7 +1097,8 @@ function joinSplitCalls(msgs) {
     const prev = out.at(-1)
     if (m?.role === "assistant" && prev?.role === "assistant" && prev.tool_calls?.length) {
       const content = [textOf(prev.content), textOf(m.content)].filter(Boolean).join("\n\n")
-      out[out.length - 1] = { ...prev, content, tool_calls: [...prev.tool_calls, ...(m.tool_calls ?? [])] }
+      const reasoning_content = [reasoningOf(prev), reasoningOf(m)].filter(Boolean).join("\n\n")
+      out[out.length - 1] = { ...prev, content, reasoning_content, tool_calls: [...prev.tool_calls, ...(m.tool_calls ?? [])] }
       continue
     }
     out.push(m)
@@ -1138,9 +1203,11 @@ function conversation(chat, tools) {
     if (m.role === "assistant") {
       answer()
       const content = []
+      const reasoning = reasoningOf(m)
+      if (reasoning) content.push({ type: "reasoning", text: reasoning })
       const text = textOf(m.content)
       if (text) content.push({ type: "text", text })
-      size += text.length
+      size += reasoning.length + text.length
       for (const c of m.tool_calls ?? []) {
         if (c?.type && c.type !== "function") continue
         const args = parseArgs(c.function?.arguments)
@@ -2197,7 +2264,7 @@ export async function CursorAuthPlugin() {
         }
         const [u, plan, bot] = await Promise.all([usage(tok, Object.keys(provider?.models ?? {})), planOf(tok), botUsage(tok)])
         // Put Bot before Total for magpie's three menu-bar rings.
-        const windows = [...u.windows.filter((w) => !w.aside), ...bot, ...u.windows.filter((w) => w.aside)]
+        const windows = [...u.windows.filter((w) => w.name !== "Total"), ...bot, ...u.windows.filter((w) => w.name === "Total")]
         return { ...u, windows, ...(plan ? { plan } : {}), signIn: "kept" }
       },
     },
@@ -2240,4 +2307,4 @@ export async function CursorAuthPlugin() {
 }
 
 // for tests
-export const _internal = { resetH1: () => (h1Until = 0), onH1: () => Date.now() < h1Until, H2, H1, h2Blocked, open, openH1, STEP_WAIT, stepUsage, ledger, owed, flights, misses, unreadable, catalog, usable, request, modelsOf, tokensOf, maxRequired, conversationID, usageOf, sessionOf, SESSION, errorResponse, kept, poolBase, conversation, buildRun, decode, exec, fields, pb, pbValue, pbAny, frame, frames, failure, toolsOf, expiry }
+export const _internal = { onDemandOf,resetH1: () => (h1Until = 0), onH1: () => Date.now() < h1Until, H2, H1, h2Blocked, open, openH1, STEP_WAIT, stepUsage, ledger, owed, flights, misses, unreadable, catalog, usable, request, modelsOf, tokensOf, maxRequired, conversationID, usageOf, sessionOf, SESSION, errorResponse, kept, poolBase, conversation, buildRun, decode, exec, fields, pb, pbValue, pbAny, frame, frames, failure, toolsOf, expiry }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@magpie-community/opencode-cursor-auth",
-  "version": "0.2.5",
+  "version": "0.2.7",
   "description": "Cursor subscriptions (Pro, Pro+, Ultra, Teams) in OpenCode and magpie, through the API cursor-agent talks to",
   "type": "module",
   "main": "./index.mjs",