@magpie-community/opencode-kiro-auth 0.1.23 → 0.1.24
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +4 -1
- package/index.mjs +30 -5
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -152,7 +152,10 @@ request:
|
|
|
152
152
|
- Anthropic's hosted web search is answered with Kiro's own web search
|
|
153
153
|
(see [Web search](#web-search)).
|
|
154
154
|
- Kiro's failures keep their meaning:
|
|
155
|
-
- input too long → 400
|
|
155
|
+
- input too long → 400 "input is too long for the model's context", whether
|
|
156
|
+
Kiro says it as an HTTP answer, as an exception in the stream, or by
|
|
157
|
+
stopping with `MODEL_CONTEXT_WINDOW_EXCEEDED` before a word (after some
|
|
158
|
+
of the reply, it ends with stop reason `model_context_window_exceeded`)
|
|
156
159
|
- no capacity → 503
|
|
157
160
|
- usage limit → 429
|
|
158
161
|
- throttling → "rate limited"
|
package/index.mjs
CHANGED
|
@@ -1155,6 +1155,16 @@ function sendKiro(a, body, signal) {
|
|
|
1155
1155
|
})
|
|
1156
1156
|
}
|
|
1157
1157
|
|
|
1158
|
+
// tooLongSaid is whether Kiro's words say the conversation is past the
|
|
1159
|
+
// model's window: its ValidationException "Input is too long." with reason
|
|
1160
|
+
// CONTENT_LENGTH_EXCEEDS_THRESHOLD, as an HTTP answer or as an exception
|
|
1161
|
+
// in the event stream, or a reply it stopped with MODEL_CONTEXT_WINDOW_EXCEEDED.
|
|
1162
|
+
const tooLongSaid = (msg) => /CONTENT_LENGTH_EXCEEDS_THRESHOLD|MODEL_CONTEXT_WINDOW_EXCEEDED|input is too long/i.test(msg)
|
|
1163
|
+
// TOO_LONG heads a too-long failure in the words magpie's gateway takes
|
|
1164
|
+
// for one (tooLong): another member with a larger window is asked, no
|
|
1165
|
+
// account rests, and with none left the agent is told to compact.
|
|
1166
|
+
const TOO_LONG = "input is too long for the model's context: "
|
|
1167
|
+
|
|
1158
1168
|
// failure is the status and message for an error Kiro answered with:
|
|
1159
1169
|
// {"message": "...", "reason": "..."}.
|
|
1160
1170
|
function failure(status, body) {
|
|
@@ -1164,8 +1174,7 @@ function failure(status, body) {
|
|
|
1164
1174
|
if (e?.message) msg = e.message + (e.reason ? ` (${e.reason})` : "")
|
|
1165
1175
|
} catch {}
|
|
1166
1176
|
msg ||= STATUS_TEXT[status] ?? ""
|
|
1167
|
-
if (msg
|
|
1168
|
-
return { status: 400, message: "input is too long for the model's context: " + msg }
|
|
1177
|
+
if (tooLongSaid(msg)) return { status: 400, message: TOO_LONG + msg }
|
|
1169
1178
|
if (msg.includes("INSUFFICIENT_MODEL_CAPACITY")) return { status: 503, message: msg }
|
|
1170
1179
|
if (msg.includes("MONTHLY_REQUEST_COUNT") || msg.includes("USAGE_LIMIT")) return { status: 429, message: "usage limit reached: " + msg }
|
|
1171
1180
|
return { status, message: msg }
|
|
@@ -1413,8 +1422,15 @@ async function* events(body, model, budget) {
|
|
|
1413
1422
|
// Kiro says how full the context is rather than how many tokens
|
|
1414
1423
|
if (!usage.input && pct > 0 && window > 0) usage.input = Math.floor((pct / 100) * window)
|
|
1415
1424
|
if (!usage.output) usage.output = Math.floor((said + 3) / 4)
|
|
1425
|
+
const full = /^model_context_window_exceeded$/i.test(stop)
|
|
1426
|
+
if (full && !said && !tools) {
|
|
1427
|
+
// stopped for the window before a word: the conversation doesn't fit
|
|
1428
|
+
yield { kind: "error", text: "the reply stopped as the model's context window was full (MODEL_CONTEXT_WINDOW_EXCEEDED)" }
|
|
1429
|
+
return
|
|
1430
|
+
}
|
|
1416
1431
|
let reason = "end_turn"
|
|
1417
1432
|
if (tools > 0) reason = "tool_use"
|
|
1433
|
+
else if (full) reason = "model_context_window_exceeded" // cut short, as Anthropic's API says it
|
|
1418
1434
|
else if (/^max_tokens$/i.test(stop)) reason = "max_tokens"
|
|
1419
1435
|
else if (/^content_filtered$/i.test(stop)) reason = "refusal"
|
|
1420
1436
|
yield { kind: "stop", reason, usage }
|
|
@@ -1454,7 +1470,14 @@ const QUOTA_WORDS = /quota|insufficient|balance|credit|billing|exceeded|rate.?li
|
|
|
1454
1470
|
// magpie's built-in relay gives it: a status, not a stream, so another
|
|
1455
1471
|
// account can take over — a 429 when it reads as a quota or rate limit,
|
|
1456
1472
|
// else a 502.
|
|
1457
|
-
|
|
1473
|
+
//
|
|
1474
|
+
// A conversation past the model's window is a 400 that says so, whichever
|
|
1475
|
+
// way Kiro said it, before the quota words ("exceeded"): as a 502, magpie
|
|
1476
|
+
// rested the account, asked it again, and told a lone agent of a vendor
|
|
1477
|
+
// error, so it never compacted (ergou0954_54426 on Discord: Kiro in a
|
|
1478
|
+
// group routed by intent, "超出上下文", the turn failed).
|
|
1479
|
+
const failedBefore = (text) =>
|
|
1480
|
+
tooLongSaid(text) ? errorResponse(400, TOO_LONG + text) : errorResponse(QUOTA_WORDS.test(text) ? 429 : 502, text)
|
|
1458
1481
|
|
|
1459
1482
|
// reply answers the Messages request from Kiro's events: a stream of
|
|
1460
1483
|
// Messages' events, or one message. names gives back the caller's name of
|
|
@@ -1887,8 +1910,10 @@ async function* searching(creds, auth, req, model, budget, signal, hosted, first
|
|
|
1887
1910
|
else said.push({ type: "text", text: e.text })
|
|
1888
1911
|
}
|
|
1889
1912
|
if (e.kind === "error") {
|
|
1890
|
-
// in a turn after a search, a status, not a finished answer
|
|
1891
|
-
|
|
1913
|
+
// in a turn after a search, a status, not a finished answer
|
|
1914
|
+
// (reply): one past the window says so, as failedBefore does
|
|
1915
|
+
if (turn > 1 && tooLongSaid(e.text)) yield { ...e, status: 400, text: TOO_LONG + e.text }
|
|
1916
|
+
else yield turn > 1 ? { ...e, status: 502 } : e
|
|
1892
1917
|
return
|
|
1893
1918
|
}
|
|
1894
1919
|
yield e
|