@magpie-community/opencode-kiro-auth 0.1.22 → 0.1.24

package/README.md

@@ -145,10 +145,17 @@ request:
   IDE sends it: `additionalModelRequestFields` with `thinking.type:
   "adaptive"` and `output_config.effort` (or `reasoning.effort`), and no
   budget. So xhigh and max stay apart there (#538).
+  Thinking is asked for only when the request's `thinking` is `enabled` or
+  `adaptive`. An effort with no `thinking`, or with it `disabled` (Claude
+  Code's title requests, or a turn with thinking off), is sent with no
+  thinking at any level.
 - Anthropic's hosted web search is answered with Kiro's own web search
   (see [Web search](#web-search)).
 - Kiro's failures keep their meaning:
-  - input too long → 400
+  - input too long → 400 "input is too long for the model's context", whether
+    Kiro says it as an HTTP answer, as an exception in the stream, or by
+    stopping with `MODEL_CONTEXT_WINDOW_EXCEEDED` before a word (after some
+    of the reply, it ends with stop reason `model_context_window_exceeded`)
   - no capacity → 503
   - usage limit → 429
   - throttling → "rate limited"

package/index.mjs

@@ -842,10 +842,15 @@ function documentText(p) {
 }
 
 // thinking is whether the model is asked to think aloud, and for how
-// long: when the caller asked for it, of a Claude model (or Auto).
+// long: when the caller asked for it, of a Claude model (or Auto). Only a
+// thinking of type enabled or adaptive asks for it, as Anthropic's API
+// reads a request: an effort with no thinking (Claude Code's title and
+// other side requests, or its main turn with thinking turned off), or with
+// thinking disabled, is sent no thinking at any level (#538), as magpie's
+// gateway reads it too.
 function thinking(req, model) {
   const effort = req.output_config?.effort ?? ""
-  const on = req.thinking?.type === "enabled" || req.thinking?.type === "adaptive" || (effort && effort !== "low")
+  const on = req.thinking?.type === "enabled" || req.thinking?.type === "adaptive"
   if (!on || !thinks(String(model).toLowerCase())) return 0
   const budget = { low: 10000, high: 30000, xhigh: 50000, max: 50000 }[effort]
   if (budget) return budget
@@ -1150,6 +1155,16 @@ function sendKiro(a, body, signal) {
   })
 }
 
+// tooLongSaid is whether Kiro's words say the conversation is past the
+// model's window: its ValidationException "Input is too long." with reason
+// CONTENT_LENGTH_EXCEEDS_THRESHOLD, as an HTTP answer or as an exception
+// in the event stream, or a reply it stopped with MODEL_CONTEXT_WINDOW_EXCEEDED.
+const tooLongSaid = (msg) => /CONTENT_LENGTH_EXCEEDS_THRESHOLD|MODEL_CONTEXT_WINDOW_EXCEEDED|input is too long/i.test(msg)
+// TOO_LONG heads a too-long failure in the words magpie's gateway takes
+// for one (tooLong): another member with a larger window is asked, no
+// account rests, and with none left the agent is told to compact.
+const TOO_LONG = "input is too long for the model's context: "
+
 // failure is the status and message for an error Kiro answered with:
 // {"message": "...", "reason": "..."}.
 function failure(status, body) {
@@ -1159,8 +1174,7 @@ function failure(status, body) {
     if (e?.message) msg = e.message + (e.reason ? ` (${e.reason})` : "")
   } catch {}
   msg ||= STATUS_TEXT[status] ?? ""
-  if (msg.includes("CONTENT_LENGTH_EXCEEDS_THRESHOLD") || msg.toLowerCase().includes("input is too long"))
-    return { status: 400, message: "input is too long for the model's context: " + msg }
+  if (tooLongSaid(msg)) return { status: 400, message: TOO_LONG + msg }
   if (msg.includes("INSUFFICIENT_MODEL_CAPACITY")) return { status: 503, message: msg }
   if (msg.includes("MONTHLY_REQUEST_COUNT") || msg.includes("USAGE_LIMIT")) return { status: 429, message: "usage limit reached: " + msg }
   return { status, message: msg }
@@ -1408,8 +1422,15 @@ async function* events(body, model, budget) {
   // Kiro says how full the context is rather than how many tokens
   if (!usage.input && pct > 0 && window > 0) usage.input = Math.floor((pct / 100) * window)
   if (!usage.output) usage.output = Math.floor((said + 3) / 4)
+  const full = /^model_context_window_exceeded$/i.test(stop)
+  if (full && !said && !tools) {
+    // stopped for the window before a word: the conversation doesn't fit
+    yield { kind: "error", text: "the reply stopped as the model's context window was full (MODEL_CONTEXT_WINDOW_EXCEEDED)" }
+    return
+  }
   let reason = "end_turn"
   if (tools > 0) reason = "tool_use"
+  else if (full) reason = "model_context_window_exceeded" // cut short, as Anthropic's API says it
   else if (/^max_tokens$/i.test(stop)) reason = "max_tokens"
   else if (/^content_filtered$/i.test(stop)) reason = "refusal"
   yield { kind: "stop", reason, usage }
@@ -1449,7 +1470,14 @@ const QUOTA_WORDS = /quota|insufficient|balance|credit|billing|exceeded|rate.?li
 // magpie's built-in relay gives it: a status, not a stream, so another
 // account can take over — a 429 when it reads as a quota or rate limit,
 // else a 502.
-const failedBefore = (text) => errorResponse(QUOTA_WORDS.test(text) ? 429 : 502, text)
+//
+// A conversation past the model's window is a 400 that says so, whichever
+// way Kiro said it, before the quota words ("exceeded"): as a 502, magpie
+// rested the account, asked it again, and told a lone agent of a vendor
+// error, so it never compacted (ergou0954_54426 on Discord: Kiro in a
+// group routed by intent, "超出上下文", the turn failed).
+const failedBefore = (text) =>
+  tooLongSaid(text) ? errorResponse(400, TOO_LONG + text) : errorResponse(QUOTA_WORDS.test(text) ? 429 : 502, text)
 
 // reply answers the Messages request from Kiro's events: a stream of
 // Messages' events, or one message. names gives back the caller's name of
@@ -1882,8 +1910,10 @@ async function* searching(creds, auth, req, model, budget, signal, hosted, first
           else said.push({ type: "text", text: e.text })
         }
         if (e.kind === "error") {
-          // in a turn after a search, a status, not a finished answer (reply)
-          yield turn > 1 ? { ...e, status: 502 } : e
+          // in a turn after a search, a status, not a finished answer
+          // (reply): one past the window says so, as failedBefore does
+          if (turn > 1 && tooLongSaid(e.text)) yield { ...e, status: 400, text: TOO_LONG + e.text }
+          else yield turn > 1 ? { ...e, status: 502 } : e
           return
         }
         yield e

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@magpie-community/opencode-kiro-auth",
-  "version": "0.1.22",
+  "version": "0.1.24",
   "description": "Kiro subscriptions (Free, Pro, Pro+, Power) in OpenCode and magpie",
   "type": "module",
   "main": "./index.mjs",