@magpie-community/opencode-workbuddy-auth 0.2.13 → 0.2.15

package/README.md

@@ -286,7 +286,11 @@ that names no level goes at the model's own default from the config
 (`reasoning.defaultEffort`, "high" for DeepSeek-V4.1-Flash), as WorkBuddy's
 client sends it: WorkBuddy doesn't think when a chat names none, so
 OpenCode's default variant got answers without thinking. An effort the
-client names, `none` too, goes as it came.
+client names, `none` too, goes as it came. A chat whose `max_tokens` is
+1024 or less gets no default (since 0.2.15, #643): a title or a
+classification asked at 64 tokens spent all 64 on reasoning and answered
+nothing. Anthropic's API takes no thinking under a 1024-token budget, so
+no client means to think in one that small.
 
 One value differs from the config, measured with real requests:
 DeepSeek-V4.1-Flash (and its paid `-sg` id) lists `none`, though the
@@ -399,6 +403,14 @@ What is measured, and what is easy to read wrongly:
   429 whose named reset is still ahead makes the window full — `used: 100`,
   `resetsAt` the reset the vendor named — whatever the rows add up to. The
   sum carries the window only until the first refusal.
+- **The sum is shown, not routed on.** Until the vendor refuses, the card is
+  set aside (`aside: true`): magpie draws it beside the account and never
+  counts it as the account's allowance, so an estimate near its cap does not
+  stop the model. The refusal's card does (`refused: true`), and magpie's
+  gateway rests this model alone on this account until the reset the
+  refusal names. Before 0.2.14 the sum stopped the model too, and a paid
+  China account read 99% of a cap it had never been refused at, with 2,457
+  credits left (plugins#641).
 - **The card names both ids**, `deepseek-v4.1-flash` and
   `deepseek-v4.1-flash-sg`: magpie matches a window's models by exact id
   (only an `@size` suffix folds into the bare id), so one alone would leave
@@ -435,10 +447,10 @@ exhausted, and the next request opens a fresh one. The China build
 (`copilot.tencent.com`) carries the same window: two Free-plan accounts
 refused on 2026-11-02 at 200,389,043 and 200,485,923 tokens, so its cap is
 measured at about 2e8 — and its refusals name the reset in Chinese
-(`将在 … UTC+8 重置`), which reads as an edge the same way. Its card stops
-at a **soft cap of 1.9e8**, 5% short of the measured one, so routing rests
-the model before the vendor refuses it; the headroom absorbs the account's
-spending on other machines, which this ledger cannot see.
+(`将在 … UTC+8 重置`), which reads as an edge the same way. Its card reads
+against 2e8. In 0.2.12 and 0.2.13 it read against a soft cap of 1.9e8, so
+that routing rested the model before the vendor refused it; with the sum set
+aside that headroom only stopped the model early, so it is gone.
 
 ## When free models work but a credit model does not
 

package/index.mjs

@@ -2276,6 +2276,17 @@ const compact = (n) => (Number.isInteger(n) ? String(n) : n.toFixed(2).replace(/
 //     full whatever the sum says. That is the one exact signal there is,
 //     and routing must not pass over it either.
 //
+// So the sum is an estimate, and it is shown, not routed on: its card is set
+// aside (aside: true), which magpie draws beside the account and never counts
+// as its allowance. Only a refusal's card stops the model. The estimate used
+// to stop it as well, and it read a paid China account at 99% of a cap it had
+// never been refused at, with 2,457 credits left (XiaoFeiyu9, plugins#641):
+// the cap is two accounts' measurement, the ledger is one machine's, and
+// neither is the vendor's word for this account. Nothing is lost by waiting
+// for that word: magpie's gateway reads the refusal's reset ("将在 … 重置",
+// "reset at …") and its "switch to the other models", and rests this model
+// alone, on this account alone, until the time the vendor named.
+//
 // The ledger is magpie's own, in magpie's config folder (appdir.Config()):
 // the folder the plugin host runs in, which is the portable data folder
 // when magpie runs portable, $XDG_CONFIG_HOME/magpie when that is set, and
@@ -2287,12 +2298,12 @@ const compact = (n) => (Number.isInteger(n) ? String(n) : n.toFixed(2).replace(/
 // caps: the same phone number can be signed in to both, and their ledgers
 // stay apart — one build's spending must not fill the other's window.
 //
-// The China cap is a soft one. Measured at about 2e8 (two accounts refusing
-// at 200.39M and 200.49M), the card stops the model at 1.9e8: 5% of headroom
-// for the account's spending on other machines (which this ledger cannot
-// see), so routing rests the model before the vendor's refusal does. A
-// refusal in force still reads the window full whatever the sum says.
-const FREE_WINDOW_CN = { provider: "workbuddy", cap: 1.9e8, short: "190M" }
+// The China cap is the measured one, 2e8 (two accounts refused at 200.39M
+// and 200.49M). It was a soft 1.9e8 in 0.2.12 and 0.2.13, so that routing
+// rested the model before the vendor refused it; since the sum is set aside
+// (freeWindowOf's card), the cap only scales what the card says, and the
+// headroom bought nothing but a model stopped 10M tokens early (plugins#641).
+const FREE_WINDOW_CN = { provider: "workbuddy", cap: 2e8, short: "200M" }
 const FREE_WINDOW_AI = { provider: "workbuddy-ai", cap: 1e8, short: "100M" }
 const FREE_WINDOW_SPAN = 24 * 3600 * 1000
 // The lead: how far back of an edge the read goes on. A chain read from a row
@@ -2438,7 +2449,9 @@ function freeWindowOf(rows, account, now = Date.now(), win = FREE_WINDOW_AI) {
     resetsAt: new Date(resetsAt).toISOString(),
     span: FREE_WINDOW_SPAN / 1000,
     models: FREE_WINDOW_MODELS.slice(),
-    ...(refused ? { refused: true } : {}),
+    // the ledger's sum is an estimate, shown and set aside; the vendor's
+    // refusal is the one that stops the model (see the window above)
+    ...(refused ? { refused: true } : { aside: true }),
   })
 
   // A refusal still in force is exact: the account is spending this window
@@ -3246,11 +3259,27 @@ function defaultEffortOf(d, efforts) {
   return typeof own === "string" && efforts.includes(own) ? own : undefined
 }
 
+// SMALL_BUDGET is the largest output budget a chat can name and still be
+// one no client meant to think in: Anthropic's API takes thinking only with
+// a budget of at least 1024 tokens and a max_tokens above it, so a request
+// of 1024 or less can't have asked for thinking there, and a client's own
+// short asks (a session title, a classification) name such budgets. A
+// default effort spends them on reasoning: itspig (#643) saw 230+ title
+// requests at max_tokens 64 on deepseek-v4-pro end out=64, reasoning=64,
+// no text. Measured 2026-11-03 on www.workbuddy.ai, deepseek-v4.1-flash-sg
+// at its default high: max_tokens 64 ended "length" with 64 reasoning
+// tokens and no text; without an effort it answered the title.
+const SMALL_BUDGET = 1024
+
 // withDefaultEffort gives a chat that names no effort the model's default
 // (defaultEffortOf), as WorkBuddy's client sends it; true when it did. An
-// effort the client named, "none" too, goes as it came.
+// effort the client named, "none" too, goes as it came. A chat whose
+// max_tokens is SMALL_BUDGET or less is given none: WorkBuddy doesn't
+// think without one, so its budget goes to the answer.
 function withDefaultEffort(chat, list) {
   if (!chat || typeof chat !== "object" || namesEffort(chat)) return false
+  const limit = chat.max_tokens ?? chat.max_completion_tokens
+  if (typeof limit === "number" && limit > 0 && limit <= SMALL_BUDGET) return false
   const m = list?.find((x) => x.id === chat.model)
   if (!m?.effort) return false
   chat.reasoning_effort = m.effort
@@ -3788,9 +3817,8 @@ function makePlugin(site) {
           // the free model's own window, which the credit meter above does
           // not show: what actually refuses a chat long before 160 credits
           // are spent. Both builds: WorkBuddy AI at 1e8 (ten days of
-          // refusals), the China build's refusals measured at about 2e8 on
-          // 2026-11-02, its card stopping at the 1.9e8 soft cap (see the
-          // window definitions above)
+          // refusals), the China build's at about 2e8 (2026-11-02). Set
+          // aside until the vendor refuses (see the window definitions above)
           {
             const account = a.accountId || a.user || a.uid || ""
             const win = site.ai ? FREE_WINDOW_AI : FREE_WINDOW_CN
@@ -3940,4 +3968,4 @@ export const WorkBuddyAuthPlugin = makePlugin(SITES.workbuddy)
 export const WorkBuddyAIAuthPlugin = makePlugin(SITES["workbuddy-ai"])
 
 // for tests
-export const _internal = { planCodes, idFor, OFF_TAKEN, replyAfterToolsOn, withDefaultEffort, defaultEffortOf, gzipped, FREE_ONLY, PRICE_TTL, keyAccount, remoteConfig, cycleOf, typeUnions, systemAsUser, systemKey, DEFAULT_SYSTEM, EDGE_READS, fitShape, replyAfterTools, AFTER_TOOLS_REFUSED, enterpriseUsageOf, accountName, notDesktop, textOnly, runsOutOf, packagesOf, endOf, enterpriseHeaders, enterpriseOf, liveModels, browserSignIn, withSystem, asReasoning, normalizeToolChoice, fitTools, fitted, renamed, unsaid, FILTER_SAYS, unflagged, usageOf, desktopHeld, explained, REFUSED_HINT, RESTRICTED_HINT, EXHAUSTED_HINT, REFUSED_DOC, commaPaused, commaPausedOn, opened, freeCredits, creditsOf, priceOf, fresh, SITES, renewing, renewed, SESSION, attend, scopeOf, turnKey, conversationKey, signatureOf, isMessages, withoutUnsigned, withoutSignedThinking, prepared, answeredBefore, nativeOf, modelOf, configModels, freeWindowOf, tokensShort, freeWindowRows, freeWindowLedger, readLedgerTail, pinned, pinning, rowTime, rowAccount, ofAccount, windowOffset, windowReset, FREE_WINDOW_MODELS, FREE_WINDOW_CN, FREE_WINDOW_AI, FREE_WINDOW_SPAN, FREE_WINDOW_GAP, FREE_WINDOW_LEAD, checkinOf, API_MESSAGES: NPM_MESSAGES, API_CHAT: NPM }
+export const _internal = { planCodes, idFor, OFF_TAKEN, replyAfterToolsOn, withDefaultEffort, SMALL_BUDGET, defaultEffortOf, gzipped, FREE_ONLY, PRICE_TTL, keyAccount, remoteConfig, cycleOf, typeUnions, systemAsUser, systemKey, DEFAULT_SYSTEM, EDGE_READS, fitShape, replyAfterTools, AFTER_TOOLS_REFUSED, enterpriseUsageOf, accountName, notDesktop, textOnly, runsOutOf, packagesOf, endOf, enterpriseHeaders, enterpriseOf, liveModels, browserSignIn, withSystem, asReasoning, normalizeToolChoice, fitTools, fitted, renamed, unsaid, FILTER_SAYS, unflagged, usageOf, desktopHeld, explained, REFUSED_HINT, RESTRICTED_HINT, EXHAUSTED_HINT, REFUSED_DOC, commaPaused, commaPausedOn, opened, freeCredits, creditsOf, priceOf, fresh, SITES, renewing, renewed, SESSION, attend, scopeOf, turnKey, conversationKey, signatureOf, isMessages, withoutUnsigned, withoutSignedThinking, prepared, answeredBefore, nativeOf, modelOf, configModels, freeWindowOf, tokensShort, freeWindowRows, freeWindowLedger, readLedgerTail, pinned, pinning, rowTime, rowAccount, ofAccount, windowOffset, windowReset, FREE_WINDOW_MODELS, FREE_WINDOW_CN, FREE_WINDOW_AI, FREE_WINDOW_SPAN, FREE_WINDOW_GAP, FREE_WINDOW_LEAD, checkinOf, API_MESSAGES: NPM_MESSAGES, API_CHAT: NPM }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@magpie-community/opencode-workbuddy-auth",
-  "version": "0.2.13",
+  "version": "0.2.15",
   "description": "OpenCode provider plugins for WorkBuddy's plans: WorkBuddy (CodeBuddy, copilot.tencent.com) and WorkBuddy AI (www.workbuddy.ai)",
   "type": "module",
   "main": "./index.mjs",