teapot-coding-agent 0.29.24 → 0.29.26
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/dist/agent/agent.js +174 -10
- package/dist/agent/key-pool.js +54 -98
- package/dist/agent/llm.js +20 -20
- package/dist/master.js +27 -5
- package/package.json +1 -1
- package/public/assets/index-C5kwsT7K.css +1 -0
- package/public/assets/index-CMx58YZZ.js +30 -0
- package/public/assets/index-CMx58YZZ.js.map +1 -0
- package/public/assets/{shiki-Fj9UAiHo.js → shiki-CR1k4ds7.js} +2 -2
- package/public/assets/{shiki-Fj9UAiHo.js.map → shiki-CR1k4ds7.js.map} +1 -1
- package/public/index.html +2 -2
- package/public/assets/index-BNtTyyVy.css +0 -1
- package/public/assets/index-U117kIzB.js +0 -30
- package/public/assets/index-U117kIzB.js.map +0 -1
package/dist/agent/agent.js
CHANGED
|
@@ -191,7 +191,12 @@ Most session state is not injected into prompts — fetch it with tools instead:
|
|
|
191
191
|
- finish(goalComplete=true, summary) → goal fully achieved.
|
|
192
192
|
- ask_user(question, options?) → park the loop and wait for the operator's
|
|
193
193
|
decision (plan confirmation, ambiguity). One concrete question at a time.
|
|
194
|
-
- read_memory() /
|
|
194
|
+
- read_memory(query?) / add_memory(text) / forget_memory(query) / set_memory(content)
|
|
195
|
+
→ your durable notes (memory.md). add_memory appends and cannot lose earlier
|
|
196
|
+
notes; forget_memory removes the lines matching a substring, so a stale note can
|
|
197
|
+
be corrected instead of piling up; set_memory REPLACES the whole file. Pass a
|
|
198
|
+
query to read_memory to search rather than dumping everything. The notes are NOT
|
|
199
|
+
in your prompt — read them.
|
|
195
200
|
- get_todo() / set_todo(content) → the operator-maintained task list
|
|
196
201
|
(todo.md); check it when picking up work, keep it current as you go.
|
|
197
202
|
- When corrected, add_feedback(rule) so it sticks; review via get_feedback().
|
|
@@ -219,7 +224,11 @@ Most session state is not injected into prompts — fetch it with tools instead:
|
|
|
219
224
|
shown to the operator directly (rendered as markdown in the progress
|
|
220
225
|
panel) — do NOT also repeat the same content as a chat message; end the
|
|
221
226
|
turn right after the call (a short "reported." is fine).
|
|
222
|
-
- Be frugal: prefer small precise edits, avoid runaway loops
|
|
227
|
+
- Be frugal: prefer small precise edits, avoid runaway loops.
|
|
228
|
+
- When the operator asks you to remember something, WRITE it with add_memory. When
|
|
229
|
+
earlier context might be relevant again (a decision, a gotcha, where you left off),
|
|
230
|
+
SEARCH it with read_memory(query) and use what you find. Nothing is injected: an
|
|
231
|
+
unread note is an unwritten note.`;
|
|
223
232
|
/**
|
|
224
233
|
* A message whose ENTIRE text is parenthesised is an acknowledgement, not work.
|
|
225
234
|
*
|
|
@@ -2483,8 +2492,38 @@ export class Agent {
|
|
|
2483
2492
|
continue;
|
|
2484
2493
|
}
|
|
2485
2494
|
if (call.function.name === "read_memory") {
|
|
2495
|
+
const a = safeParse(call.function.arguments);
|
|
2496
|
+
const query = String(a.query ?? "").trim();
|
|
2486
2497
|
const mem = await fs.readFile(this.memoryFile, "utf8").catch(() => "");
|
|
2487
|
-
|
|
2498
|
+
if (!mem.trim()) {
|
|
2499
|
+
await this.answerMeta(call, "(memory.md is empty — nothing noted yet)");
|
|
2500
|
+
continue;
|
|
2501
|
+
}
|
|
2502
|
+
// #48: `query` searches instead of dumping the whole file.
|
|
2503
|
+
//
|
|
2504
|
+
// This is the one idea worth taking from the ecosystem. MCP's reference memory
|
|
2505
|
+
// server (the de-facto standard, used by Claude and offered by DSH as a
|
|
2506
|
+
// reference) exposes `search` alongside `read`, and its search is deliberately
|
|
2507
|
+
// a case-insensitive substring match — no embeddings, no model, no provider
|
|
2508
|
+
// dependency. DSH's guidance to the model is the same shape: "when historical
|
|
2509
|
+
// information may be relevant, search memory and use relevant results".
|
|
2510
|
+
//
|
|
2511
|
+
// MEASURED why it matters here: `set_memory` was called 181 times but
|
|
2512
|
+
// `read_memory` only 97 times across every session, and only 13 of those reads
|
|
2513
|
+
// came after a write. Dumping a growing file gives the model no way to ask a
|
|
2514
|
+
// QUESTION, so a big file is all-or-nothing — either burn the whole thing into
|
|
2515
|
+
// the context or do not look at all.
|
|
2516
|
+
if (!query) {
|
|
2517
|
+
await this.answerMeta(call, mem.trim());
|
|
2518
|
+
continue;
|
|
2519
|
+
}
|
|
2520
|
+
const needle = query.toLowerCase();
|
|
2521
|
+
const hits = mem
|
|
2522
|
+
.split(/\r?\n/)
|
|
2523
|
+
.filter((line) => line.toLowerCase().includes(needle));
|
|
2524
|
+
await this.answerMeta(call, hits.length
|
|
2525
|
+
? `${hits.length} line(s) matching ${JSON.stringify(query)}:\n${hits.join("\n")}`
|
|
2526
|
+
: `(no line in memory.md matches ${JSON.stringify(query)} — ${mem.split(/\r?\n/).length} lines searched; call read_memory with no query to see all of it)`);
|
|
2488
2527
|
continue;
|
|
2489
2528
|
}
|
|
2490
2529
|
if (call.function.name === "list_skills") {
|
|
@@ -2499,7 +2538,70 @@ export class Agent {
|
|
|
2499
2538
|
const a = safeParse(call.function.arguments);
|
|
2500
2539
|
const content = String(a.content ?? "").slice(0, 32_000);
|
|
2501
2540
|
await fs.writeFile(this.memoryFile, content, "utf8");
|
|
2502
|
-
|
|
2541
|
+
// #48: this used to say "injected into future prompts", which was false —
|
|
2542
|
+
// memory.md is never injected. The model then behaved as if it already held
|
|
2543
|
+
// its notes. Say what actually happened, and what the model must do instead.
|
|
2544
|
+
await this.answerMeta(call, `memory.md REPLACED (${content.length} bytes). It is not in your prompt — call read_memory when you need it.`);
|
|
2545
|
+
continue;
|
|
2546
|
+
}
|
|
2547
|
+
if (call.function.name === "forget_memory") {
|
|
2548
|
+
const a = safeParse(call.function.arguments);
|
|
2549
|
+
const query = String(a.query ?? "").trim();
|
|
2550
|
+
if (!query) {
|
|
2551
|
+
await this.answerMeta(call, "forget_memory needs a `query` — nothing removed");
|
|
2552
|
+
continue;
|
|
2553
|
+
}
|
|
2554
|
+
// #48: notes must be REMOVABLE, not only appendable.
|
|
2555
|
+
//
|
|
2556
|
+
// MCP's reference memory makes observations "added or removed
|
|
2557
|
+
// independently" for exactly this reason: an append-only store accumulates
|
|
2558
|
+
// stale facts ("the build takes 4 minutes" stays long after it takes 40
|
|
2559
|
+
// seconds), and a memory that is confidently wrong is worse than one that is
|
|
2560
|
+
// silent. Without a delete, the only way to correct a file was `set_memory`
|
|
2561
|
+
// with the whole content re-typed — which is how 1028 bytes became 83.
|
|
2562
|
+
const prev = await fs.readFile(this.memoryFile, "utf8").catch(() => "");
|
|
2563
|
+
if (!prev.trim()) {
|
|
2564
|
+
await this.answerMeta(call, "(memory.md is empty — nothing to forget)");
|
|
2565
|
+
continue;
|
|
2566
|
+
}
|
|
2567
|
+
const needle = query.toLowerCase();
|
|
2568
|
+
const kept = [];
|
|
2569
|
+
const removed = [];
|
|
2570
|
+
for (const line of prev.split(/\r?\n/)) {
|
|
2571
|
+
if (line.toLowerCase().includes(needle))
|
|
2572
|
+
removed.push(line);
|
|
2573
|
+
else
|
|
2574
|
+
kept.push(line);
|
|
2575
|
+
}
|
|
2576
|
+
if (removed.length === 0) {
|
|
2577
|
+
await this.answerMeta(call, `(no line in memory.md matches ${JSON.stringify(query)} — nothing removed)`);
|
|
2578
|
+
continue;
|
|
2579
|
+
}
|
|
2580
|
+
// collapse the blank runs a deletion leaves behind, so the file does not
|
|
2581
|
+
// grow holes
|
|
2582
|
+
const next = kept.join("\n").replace(/\n{3,}/g, "\n\n").trimEnd() + "\n";
|
|
2583
|
+
await fs.writeFile(this.memoryFile, next, "utf8");
|
|
2584
|
+
await this.answerMeta(call, `removed ${removed.length} line(s) from memory.md:\n${removed.join("\n").slice(0, 1_500)}\n` +
|
|
2585
|
+
`(memory.md is now ${next.length} bytes — re-add with add_memory if that was wrong)`);
|
|
2586
|
+
continue;
|
|
2587
|
+
}
|
|
2588
|
+
if (call.function.name === "add_memory") {
|
|
2589
|
+
const a = safeParse(call.function.arguments);
|
|
2590
|
+
const text = String(a.text ?? "").trim().slice(0, 8_000);
|
|
2591
|
+
if (!text) {
|
|
2592
|
+
await this.answerMeta(call, "add_memory needs non-empty `text` — nothing appended");
|
|
2593
|
+
continue;
|
|
2594
|
+
}
|
|
2595
|
+
// APPEND rather than replace (#48). `set_memory` overwrites the whole file,
|
|
2596
|
+
// so a model that had not just read it silently ERASED everything it did not
|
|
2597
|
+
// re-send — MEASURED on a real session: 21 writes, with the content dropping
|
|
2598
|
+
// from 1028 to 83 bytes. Appending is the operation "remember this" actually
|
|
2599
|
+
// wants, and it cannot lose what is already there.
|
|
2600
|
+
const prev = await fs.readFile(this.memoryFile, "utf8").catch(() => "");
|
|
2601
|
+
const sep = prev && !prev.endsWith("\n") ? "\n" : "";
|
|
2602
|
+
const next = `${prev}${sep}${text}\n`;
|
|
2603
|
+
await fs.writeFile(this.memoryFile, next.slice(0, 32_000), "utf8");
|
|
2604
|
+
await this.answerMeta(call, `appended to memory.md (now ${next.length} bytes). Not in your prompt — call read_memory to see it.`);
|
|
2503
2605
|
continue;
|
|
2504
2606
|
}
|
|
2505
2607
|
await this.log.append("tool_call", this.currentSession, this.currentBranch, {
|
|
@@ -3532,7 +3634,8 @@ export class Agent {
|
|
|
3532
3634
|
{
|
|
3533
3635
|
role: "user",
|
|
3534
3636
|
content: `[harness] Context was compacted: ${oldCount} earlier messages were summarized. ` +
|
|
3535
|
-
"
|
|
3637
|
+
"AGENTS.md is embedded in your system prompt; memory.md, todo.md and decisions.md are NOT — "
|
|
3638
|
+
+ "call read_memory / get_todo / get_decisions if you need them.\n\n" +
|
|
3536
3639
|
`## Summary of earlier conversation\n${summary}`,
|
|
3537
3640
|
},
|
|
3538
3641
|
...this.messages.slice(cut),
|
|
@@ -3551,6 +3654,32 @@ export class Agent {
|
|
|
3551
3654
|
// Post-compact file restore (Claude Code style): re-attach the most
|
|
3552
3655
|
// recently read files so the model doesn't burn its first turns re-reading
|
|
3553
3656
|
// exactly what it had in context a moment ago. Budget: 3 files × 400 lines.
|
|
3657
|
+
if (mode === "summarize") {
|
|
3658
|
+
// #48: the durable notes go back in HERE, and only here.
|
|
3659
|
+
//
|
|
3660
|
+
// MEASURED before this: `set_memory` 181 times across every session but a
|
|
3661
|
+
// `read_memory` AFTER writing only 13 times — the notes were written and
|
|
3662
|
+
// never consulted, so the whole system was decoration. The prompt made it
|
|
3663
|
+
// worse by claiming memory.md was injected when it never was.
|
|
3664
|
+
//
|
|
3665
|
+
// Compaction is the one moment where re-attaching is free and obviously
|
|
3666
|
+
// right: the history that held the notes has just been summarised away, and
|
|
3667
|
+
// the prompt has to move anyway. Everywhere else they stay behind
|
|
3668
|
+
// `read_memory()` so the prefix cache is never disturbed.
|
|
3669
|
+
try {
|
|
3670
|
+
const notes = (await fs.readFile(this.memoryFile, "utf8")).trim();
|
|
3671
|
+
if (notes) {
|
|
3672
|
+
this.messages.splice(1, 0, {
|
|
3673
|
+
role: "user",
|
|
3674
|
+
content: `[harness] Your durable notes (memory.md), re-attached after compaction — ` +
|
|
3675
|
+
`this is what earlier you chose to remember:\n\n${notes.slice(0, 8_000)}`,
|
|
3676
|
+
});
|
|
3677
|
+
}
|
|
3678
|
+
}
|
|
3679
|
+
catch {
|
|
3680
|
+
/* no notes yet — nothing to re-attach */
|
|
3681
|
+
}
|
|
3682
|
+
}
|
|
3554
3683
|
if (this.recentReads.length && mode === "summarize") {
|
|
3555
3684
|
const restored = [];
|
|
3556
3685
|
const restoredPaths = new Set();
|
|
@@ -4018,15 +4147,37 @@ function allToolSpecs() {
|
|
|
4018
4147
|
type: "function",
|
|
4019
4148
|
function: {
|
|
4020
4149
|
name: "read_memory",
|
|
4021
|
-
description: "Read your durable notes (memory.md).",
|
|
4022
|
-
parameters: {
|
|
4150
|
+
description: "Read your durable notes (memory.md). Pass `query` to search for the lines that mention something — cheaper than reading the whole file, and the right call when you only need to check one fact.",
|
|
4151
|
+
parameters: {
|
|
4152
|
+
type: "object",
|
|
4153
|
+
properties: {
|
|
4154
|
+
query: {
|
|
4155
|
+
type: "string",
|
|
4156
|
+
description: "case-insensitive substring; omit it to get the whole file",
|
|
4157
|
+
},
|
|
4158
|
+
},
|
|
4159
|
+
},
|
|
4160
|
+
},
|
|
4161
|
+
},
|
|
4162
|
+
{
|
|
4163
|
+
type: "function",
|
|
4164
|
+
function: {
|
|
4165
|
+
name: "forget_memory",
|
|
4166
|
+
description: "Delete the lines of memory.md that match a substring. Use it to remove a note that became wrong, or to consolidate duplicates. Returns what was removed, so a mistake is visible and re-addable.",
|
|
4167
|
+
parameters: {
|
|
4168
|
+
type: "object",
|
|
4169
|
+
properties: {
|
|
4170
|
+
query: { type: "string", description: "case-insensitive substring of the line(s) to delete" },
|
|
4171
|
+
},
|
|
4172
|
+
required: ["query"],
|
|
4173
|
+
},
|
|
4023
4174
|
},
|
|
4024
4175
|
},
|
|
4025
4176
|
{
|
|
4026
4177
|
type: "function",
|
|
4027
4178
|
function: {
|
|
4028
4179
|
name: "set_memory",
|
|
4029
|
-
description: "
|
|
4180
|
+
description: "REPLACE your durable notes (memory.md) entirely — anything you do not include is lost. Prefer add_memory unless you are rewriting the whole file. Keep notes terse: decisions, gotchas, where you left off.",
|
|
4030
4181
|
parameters: {
|
|
4031
4182
|
type: "object",
|
|
4032
4183
|
properties: { content: { type: "string" } },
|
|
@@ -4034,6 +4185,18 @@ function allToolSpecs() {
|
|
|
4034
4185
|
},
|
|
4035
4186
|
},
|
|
4036
4187
|
},
|
|
4188
|
+
{
|
|
4189
|
+
type: "function",
|
|
4190
|
+
function: {
|
|
4191
|
+
name: "add_memory",
|
|
4192
|
+
description: "Append one note to memory.md without touching what is already there. Use this to remember a decision, a gotcha, or where you left off — it cannot lose earlier notes (#48).",
|
|
4193
|
+
parameters: {
|
|
4194
|
+
type: "object",
|
|
4195
|
+
properties: { text: { type: "string", description: "the note to append, one or two lines" } },
|
|
4196
|
+
required: ["text"],
|
|
4197
|
+
},
|
|
4198
|
+
},
|
|
4199
|
+
},
|
|
4037
4200
|
{
|
|
4038
4201
|
type: "function",
|
|
4039
4202
|
function: {
|
|
@@ -4084,7 +4247,7 @@ function rebuildMessagesFrom(list) {
|
|
|
4084
4247
|
const msgs = [];
|
|
4085
4248
|
const META_TOOLS = new Set([
|
|
4086
4249
|
"finish", "report_progress", "set_goal", "get_goal",
|
|
4087
|
-
"read_memory", "set_memory", "list_skills", "get_todo", "set_todo",
|
|
4250
|
+
"read_memory", "set_memory", "add_memory", "forget_memory", "list_skills", "get_todo", "set_todo",
|
|
4088
4251
|
"get_feedback", "add_feedback", "record_decision", "get_decisions",
|
|
4089
4252
|
]);
|
|
4090
4253
|
const openCalls = new Map(); // real tool_call id -> name
|
|
@@ -4240,7 +4403,8 @@ function rebuildMessagesFrom(list) {
|
|
|
4240
4403
|
msgs.splice(0, summarized, {
|
|
4241
4404
|
role: "user",
|
|
4242
4405
|
content: `[harness] Context was compacted: ${summarized} earlier messages were summarized. ` +
|
|
4243
|
-
"
|
|
4406
|
+
"AGENTS.md is embedded in your system prompt; memory.md, todo.md and decisions.md are NOT — "
|
|
4407
|
+
+ "call read_memory / get_todo / get_decisions if you need them.\n\n" +
|
|
4244
4408
|
`## Summary of earlier conversation\n${sum}`,
|
|
4245
4409
|
});
|
|
4246
4410
|
}
|
package/dist/agent/key-pool.js
CHANGED
|
@@ -1,129 +1,85 @@
|
|
|
1
1
|
/**
|
|
2
|
-
* API
|
|
3
|
-
*
|
|
2
|
+
* Multiple API keys per provider (#170) — "複数鍵登録しといて、規定のリトライ回数を超えたら
|
|
3
|
+
* 次の鍵にローテ / 全キーを回ってもまだレートリミットだったらエラーで停止".
|
|
4
4
|
*
|
|
5
|
-
* ##
|
|
5
|
+
* ## The algorithm, deliberately small
|
|
6
6
|
*
|
|
7
|
-
* A
|
|
8
|
-
*
|
|
9
|
-
*
|
|
10
|
-
*
|
|
11
|
-
*
|
|
7
|
+
* A request is made with the current key. The HTTP client already retries a rate limit
|
|
8
|
+
* a fixed number of times on that key (its own backoff ladder). When those retries are
|
|
9
|
+
* spent and the failure is STILL a rate limit, the pool moves to the next key and the
|
|
10
|
+
* whole ladder runs again on it. When the last key is exhausted too, the error is
|
|
11
|
+
* raised and the turn stops.
|
|
12
12
|
*
|
|
13
|
-
*
|
|
13
|
+
* That is the entire policy:
|
|
14
14
|
*
|
|
15
|
-
*
|
|
16
|
-
* in a row while another sits idle; with a small pool that is the difference
|
|
17
|
-
* between spreading load and not. A cursor per pool makes consecutive picks
|
|
18
|
-
* distinct.
|
|
15
|
+
* key1 [retry × N] -> key2 [retry × N] -> … -> raise
|
|
19
16
|
*
|
|
20
|
-
*
|
|
21
|
-
* temporary statement about one key, so discarding it would be wrong and blocking
|
|
22
|
-
* the whole provider would be worse. It returns to the pool when the cooldown
|
|
23
|
-
* lapses — no timer, just a timestamp compared on read, because a timer per key
|
|
24
|
-
* would keep resetting and complicate shutdown.
|
|
17
|
+
* ## What this intentionally does NOT do
|
|
25
18
|
*
|
|
26
|
-
*
|
|
27
|
-
*
|
|
28
|
-
*
|
|
29
|
-
*
|
|
30
|
-
*
|
|
19
|
+
* An earlier version of this file tracked per-key "recovered at" timestamps, parsed
|
|
20
|
+
* provider reset hints, classified a 429 as transient-vs-exhausted, and persisted spent
|
|
21
|
+
* accounts to disk. None of that was asked for, none of it was generic (the hint
|
|
22
|
+
* formats are provider-specific), and the operator's actual ask is the loop above. A
|
|
23
|
+
* simpler rule is also easier to reason about when it goes wrong: with this policy the
|
|
24
|
+
* only state is "which key am I on", and the only outcome is success or a raised error.
|
|
31
25
|
*
|
|
32
|
-
*
|
|
33
|
-
* behaviour is predictable and the pool does not reshuffle behind their back.
|
|
26
|
+
* ## Provider-agnostic
|
|
34
27
|
*
|
|
35
|
-
*
|
|
36
|
-
*
|
|
28
|
+
* Nothing here knows about any particular endpoint. "Rate limited" is the HTTP 429 that
|
|
29
|
+
* every OpenAI-compatible provider uses; the caller decides which statuses qualify.
|
|
37
30
|
*/
|
|
38
|
-
/**
|
|
39
|
-
export const KEY_COOLDOWN_MS = 60_000;
|
|
40
|
-
/** Build a pool, dropping duplicates and blanks while keeping the config's order. */
|
|
31
|
+
/** Build a pool from config order, dropping blanks and duplicates. */
|
|
41
32
|
export function makeKeyPool(keys) {
|
|
42
33
|
const seen = new Set();
|
|
43
34
|
const unique = [];
|
|
44
35
|
for (const k of keys) {
|
|
45
|
-
const t = k.trim();
|
|
36
|
+
const t = typeof k === "string" ? k.trim() : "";
|
|
46
37
|
if (!t || seen.has(t))
|
|
47
38
|
continue;
|
|
48
39
|
seen.add(t);
|
|
49
40
|
unique.push(t);
|
|
50
41
|
}
|
|
51
|
-
return { keys: unique,
|
|
42
|
+
return { keys: unique, current: 0 };
|
|
52
43
|
}
|
|
53
|
-
/**
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
*/
|
|
59
|
-
export function nextKey(pool, now) {
|
|
60
|
-
if (pool.keys.length === 0)
|
|
44
|
+
/** The key in use, or null for an empty pool (or no pool at all). */
|
|
45
|
+
export function currentKey(pool) {
|
|
46
|
+
// tolerate a missing pool: most callers simply have no `keyPool`, and making every
|
|
47
|
+
// one of them guard first is how a `undefined.keys` crash reaches a 400 (measured)
|
|
48
|
+
if (!pool)
|
|
61
49
|
return null;
|
|
62
|
-
|
|
63
|
-
// try each key at most once, starting from the cursor: the first not cooling down
|
|
64
|
-
for (let i = 0; i < n; i++) {
|
|
65
|
-
const idx = (pool.cursor + i) % n;
|
|
66
|
-
const key = pool.keys[idx];
|
|
67
|
-
const until = pool.cooldownUntil.get(key) ?? 0;
|
|
68
|
-
if (until <= now) {
|
|
69
|
-
pool.cursor = (idx + 1) % n;
|
|
70
|
-
return key;
|
|
71
|
-
}
|
|
72
|
-
}
|
|
73
|
-
// EVERY key is cooling down. Refusing to send cannot succeed, so hand out the one
|
|
74
|
-
// that recovers soonest — see rule 3.
|
|
75
|
-
let bestIdx = pool.cursor;
|
|
76
|
-
let bestUntil = Infinity;
|
|
77
|
-
for (let i = 0; i < n; i++) {
|
|
78
|
-
const idx = (pool.cursor + i) % n;
|
|
79
|
-
const until = pool.cooldownUntil.get(pool.keys[idx]) ?? 0;
|
|
80
|
-
if (until < bestUntil) {
|
|
81
|
-
bestUntil = until;
|
|
82
|
-
bestIdx = idx;
|
|
83
|
-
}
|
|
84
|
-
}
|
|
85
|
-
pool.cursor = (bestIdx + 1) % n;
|
|
86
|
-
return pool.keys[bestIdx];
|
|
87
|
-
}
|
|
88
|
-
/** Mark a key rate-limited: it is skipped for `ms`, then returns to the pool. */
|
|
89
|
-
export function coolDown(pool, key, now, ms = KEY_COOLDOWN_MS) {
|
|
90
|
-
if (!pool.keys.includes(key))
|
|
91
|
-
return;
|
|
92
|
-
const until = now + ms;
|
|
93
|
-
// never SHORTEN an existing cooldown: a second 429 while one is pending must not
|
|
94
|
-
// bring the key back early
|
|
95
|
-
const cur = pool.cooldownUntil.get(key) ?? 0;
|
|
96
|
-
if (until > cur)
|
|
97
|
-
pool.cooldownUntil.set(key, until);
|
|
98
|
-
}
|
|
99
|
-
/** Keys usable right now (diagnostics / the UI). */
|
|
100
|
-
export function availableKeys(pool, now) {
|
|
101
|
-
return pool.keys.filter((k) => (pool.cooldownUntil.get(k) ?? 0) <= now);
|
|
50
|
+
return pool.keys[pool.current] ?? null;
|
|
102
51
|
}
|
|
103
52
|
/**
|
|
104
|
-
*
|
|
53
|
+
* Move to the next key, or null when every key has already had its turn.
|
|
105
54
|
*
|
|
106
|
-
*
|
|
107
|
-
*
|
|
108
|
-
*
|
|
109
|
-
*
|
|
55
|
+
* Null is the "全キーを回ってもまだレートリミットだった" signal: the caller stops and raises
|
|
56
|
+
* the last error instead of cycling forever. No wraparound on purpose — wrapping would
|
|
57
|
+
* re-try keys whose retry ladder was just spent, turning a hard limit into an infinite
|
|
58
|
+
* loop.
|
|
110
59
|
*/
|
|
111
|
-
export function
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
if (next === null || next === tried)
|
|
60
|
+
export function nextKey(pool) {
|
|
61
|
+
const next = pool.current + 1;
|
|
62
|
+
if (next >= pool.keys.length)
|
|
115
63
|
return null;
|
|
116
|
-
|
|
64
|
+
pool.current = next;
|
|
65
|
+
return pool.keys[next];
|
|
66
|
+
}
|
|
67
|
+
/** Start over from the first key (a fresh turn gets the full cycle again). */
|
|
68
|
+
export function resetPool(pool) {
|
|
69
|
+
pool.current = 0;
|
|
70
|
+
}
|
|
71
|
+
/** How many keys this pool holds. */
|
|
72
|
+
export function poolSize(pool) {
|
|
73
|
+
return pool.keys.length;
|
|
117
74
|
}
|
|
118
75
|
/**
|
|
119
|
-
*
|
|
76
|
+
* Is this failure a rate limit — the only kind rotating can fix?
|
|
120
77
|
*
|
|
121
|
-
* 429 is the
|
|
122
|
-
* exhausted key
|
|
123
|
-
*
|
|
124
|
-
*
|
|
125
|
-
* outage affects every key equally, so rotating just burns the pool.
|
|
78
|
+
* 429 is the universal spelling. 402 ("payment required") and 403 are included because
|
|
79
|
+
* an exhausted subscription and a quota-blocked key answer that way, and a second
|
|
80
|
+
* account is exactly what resolves them. Everything else (5xx, 400, a network error) is
|
|
81
|
+
* NOT rotated: it would fail identically on every key, so rotating just burns the cycle.
|
|
126
82
|
*/
|
|
127
|
-
export function
|
|
128
|
-
return status === 429 || status ===
|
|
83
|
+
export function isRateLimitStatus(status) {
|
|
84
|
+
return status === 429 || status === 402 || status === 403;
|
|
129
85
|
}
|
package/dist/agent/llm.js
CHANGED
|
@@ -6,7 +6,7 @@
|
|
|
6
6
|
* than hand-rolled.
|
|
7
7
|
*/
|
|
8
8
|
import OpenAI from "openai";
|
|
9
|
-
import {
|
|
9
|
+
import { isRateLimitStatus, nextKey } from "./key-pool.js";
|
|
10
10
|
const clients = new WeakMap();
|
|
11
11
|
/**
|
|
12
12
|
* Rotate to another key after a KEY-level failure (#170), returning a config that
|
|
@@ -19,15 +19,15 @@ const clients = new WeakMap();
|
|
|
19
19
|
* candidates are the failed one: retrying with the SAME credential cannot help, and
|
|
20
20
|
* the caller's own staircase is the right response then.
|
|
21
21
|
*/
|
|
22
|
-
export function rotateKey(cfg
|
|
22
|
+
export function rotateKey(cfg) {
|
|
23
23
|
const pool = cfg.keyPool;
|
|
24
|
-
if (!pool
|
|
24
|
+
if (!pool)
|
|
25
25
|
return null;
|
|
26
|
-
const next =
|
|
26
|
+
const next = nextKey(pool);
|
|
27
27
|
if (!next)
|
|
28
|
-
return null;
|
|
29
|
-
// a new apiKey means a new OpenAI client below, which is correct: the credential
|
|
30
|
-
//
|
|
28
|
+
return null; // every key has had its turn
|
|
29
|
+
// a new apiKey means a new OpenAI client below, which is correct: the credential is
|
|
30
|
+
// baked into the client at construction
|
|
31
31
|
return { ...cfg, apiKey: next };
|
|
32
32
|
}
|
|
33
33
|
/**
|
|
@@ -228,18 +228,15 @@ export function providerHeaders(baseUrl, sessionId) {
|
|
|
228
228
|
function client(cfg) {
|
|
229
229
|
let c = clients.get(cfg);
|
|
230
230
|
if (!c) {
|
|
231
|
-
// #170:
|
|
232
|
-
//
|
|
233
|
-
//
|
|
234
|
-
//
|
|
235
|
-
// and `withKeyRotation` picks the next key; the agent's own ladder (5s/5s/30s) is
|
|
236
|
-
// still there as the outer backstop for a 5xx, which no key can fix.
|
|
237
|
-
const pooled = (cfg.keyPool?.keys.length ?? 0) > 1;
|
|
231
|
+
// #170: the SDK's retry ladder is deliberately KEPT for a pooled provider too.
|
|
232
|
+
// The operator's rule is "規定のリトライ回数を超えたら次の鍵にローテ", i.e. each key gets
|
|
233
|
+
// its own retries before the pool moves on — so `maxRetries` is the per-key budget
|
|
234
|
+
// and must not be zeroed.
|
|
238
235
|
c = new OpenAI({
|
|
239
236
|
baseURL: cfg.baseUrl,
|
|
240
237
|
apiKey: cfg.apiKey,
|
|
241
238
|
timeout: cfg.timeoutMs ?? 120_000,
|
|
242
|
-
maxRetries:
|
|
239
|
+
maxRetries: cfg.maxRetries ?? 4, // per-key: rotation happens after these are spent
|
|
243
240
|
defaultHeaders: providerHeaders(cfg.baseUrl, cfg.sessionId),
|
|
244
241
|
});
|
|
245
242
|
clients.set(cfg, c);
|
|
@@ -259,21 +256,24 @@ function client(cfg) {
|
|
|
259
256
|
*/
|
|
260
257
|
async function withKeyRotation(cfg, run) {
|
|
261
258
|
let cur = cfg;
|
|
262
|
-
const attempts = Math.max(1, cfg.keyPool?.keys.length ?? 1);
|
|
263
259
|
let lastErr;
|
|
264
|
-
for (let i = 0; i <
|
|
260
|
+
for (let i = 0; i < Math.max(1, cfg.keyPool?.keys.length ?? 1); i++) {
|
|
265
261
|
try {
|
|
262
|
+
// The client's own retry ladder runs inside `run`, so by the time this catch
|
|
263
|
+
// fires the FIRST key has already been retried its configured number of times.
|
|
264
|
+
// Only then do we move on — that is the operator's "規定のリトライ回数を超えたら".
|
|
266
265
|
return { value: await run(cur), cfg: cur };
|
|
267
266
|
}
|
|
268
267
|
catch (err) {
|
|
269
268
|
lastErr = err;
|
|
270
269
|
const status = err?.status;
|
|
271
|
-
if (typeof status !== "number" || !
|
|
270
|
+
if (typeof status !== "number" || !isRateLimitStatus(status))
|
|
272
271
|
throw err;
|
|
273
272
|
const rotated = rotateKey(cur);
|
|
273
|
+
// no key left: "全キーを回ってもまだレートリミットだったらエラーで停止"
|
|
274
274
|
if (!rotated)
|
|
275
|
-
throw err;
|
|
276
|
-
console.warn(`[teapot]
|
|
275
|
+
throw err;
|
|
276
|
+
console.warn(`[teapot] rate limited (HTTP ${status}) on this key — moving to the next of ${cur.keyPool?.keys.length ?? 1} for this provider`);
|
|
277
277
|
cur = rotated;
|
|
278
278
|
}
|
|
279
279
|
}
|
package/dist/master.js
CHANGED
|
@@ -5,7 +5,7 @@
|
|
|
5
5
|
*/
|
|
6
6
|
import { readFileSync, existsSync, mkdirSync, writeFileSync, readdirSync, statSync, lstatSync, renameSync, unlinkSync, openSync, closeSync, readSync, fsyncSync, } from "node:fs";
|
|
7
7
|
import { randomUUID } from "node:crypto";
|
|
8
|
-
import {
|
|
8
|
+
import { currentKey, makeKeyPool, resetPool } from "./agent/key-pool.js";
|
|
9
9
|
import { sessionTitleFrom } from "./session-title.js";
|
|
10
10
|
import path from "node:path";
|
|
11
11
|
import os from "node:os";
|
|
@@ -524,6 +524,8 @@ export class Master {
|
|
|
524
524
|
if (keys.length === 0)
|
|
525
525
|
return undefined;
|
|
526
526
|
const existing = this.keyPools.get(name);
|
|
527
|
+
// reuse while the KEY LIST is unchanged, so the pool keeps its position in the
|
|
528
|
+
// cycle across spawns; a changed list is a new pool
|
|
527
529
|
if (existing && existing.keys.length === keys.length && existing.keys.every((k, i) => k === keys[i])) {
|
|
528
530
|
return existing;
|
|
529
531
|
}
|
|
@@ -531,14 +533,29 @@ export class Master {
|
|
|
531
533
|
this.keyPools.set(name, pool);
|
|
532
534
|
return pool;
|
|
533
535
|
}
|
|
534
|
-
/**
|
|
535
|
-
|
|
536
|
+
/**
|
|
537
|
+
* Diagnostics: each provider's keys and which one is in use.
|
|
538
|
+
*
|
|
539
|
+
* Deliberately shows only position, not "spent until" timestamps: this pool has no
|
|
540
|
+
* such state — a key is rotated past once its retries are spent, and a new turn
|
|
541
|
+
* starts the cycle again.
|
|
542
|
+
*/
|
|
543
|
+
keyPoolReport() {
|
|
536
544
|
return [...this.keyPools.entries()].map(([provider, pool]) => ({
|
|
537
545
|
provider,
|
|
538
546
|
keys: pool.keys.length,
|
|
539
|
-
|
|
547
|
+
current: pool.current,
|
|
540
548
|
}));
|
|
541
549
|
}
|
|
550
|
+
/** start every provider's cycle over from its first key */
|
|
551
|
+
resetKeyPool(provider) {
|
|
552
|
+
const names = provider ? [provider] : [...this.keyPools.keys()];
|
|
553
|
+
for (const n of names) {
|
|
554
|
+
const pool = this.keyPools.get(n);
|
|
555
|
+
if (pool)
|
|
556
|
+
resetPool(pool);
|
|
557
|
+
}
|
|
558
|
+
}
|
|
542
559
|
writeConfigAtomic(body) {
|
|
543
560
|
// #144: `/dev/null` is a real, intentional value here — many callers (and
|
|
544
561
|
// every test that does not care about persistence) pass it as "no config
|
|
@@ -839,7 +856,12 @@ export class Master {
|
|
|
839
856
|
const pool = inlineKey ? undefined : this.keyPoolFor(provName, prov);
|
|
840
857
|
const llm = {
|
|
841
858
|
baseUrl: ac.baseUrl ?? prov?.baseUrl ?? this.config.llm.baseUrl,
|
|
842
|
-
|
|
859
|
+
// the pool's CURRENT key is the sticky one: a restart must resume on the
|
|
860
|
+
// account the operator's cycle had reached, not start over on a spent first one
|
|
861
|
+
apiKey: inlineKey ??
|
|
862
|
+
currentKey(pool) ??
|
|
863
|
+
prov?.apiKey ??
|
|
864
|
+
this.config.llm.apiKey,
|
|
843
865
|
model: ac.model ?? prov?.model ?? this.config.llm.model,
|
|
844
866
|
timeoutMs: 120_000,
|
|
845
867
|
...(pool && pool.keys.length > 1 ? { keyPool: pool } : {}),
|