pi-mega-compact 0.8.14 → 0.8.15
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 +44 -21
- package/dist/extensions/mega-cache-replay.test.js +4 -2
- package/dist/extensions/mega-compact-s38.test.js +317 -0
- package/dist/extensions/mega-compact.js +14 -0
- package/dist/extensions/mega-compact.test.js +37 -8
- package/dist/extensions/mega-config.js +5 -0
- package/dist/extensions/mega-events/agent-handlers.js +184 -4
- package/dist/extensions/mega-events/context-handler.js +35 -2
- package/dist/extensions/mega-events/error-classifier.js +118 -0
- package/dist/extensions/mega-events.js +1 -0
- package/dist/extensions/mega-runtime/state.js +16 -0
- package/dist/extensions/mega-teamrun.test.js +14 -1
- package/extensions/dashboard-client/src/styles/base.css +2 -1
- package/extensions/mega-cache-replay.test.ts +3 -3
- package/extensions/mega-compact-s38.test.ts +330 -0
- package/extensions/mega-compact.test.ts +41 -13
- package/extensions/mega-compact.ts +14 -0
- package/extensions/mega-config.ts +24 -0
- package/extensions/mega-dashboard.ts +7 -0
- package/extensions/mega-events/agent-handlers.ts +179 -5
- package/extensions/mega-events/context-handler.ts +31 -2
- package/extensions/mega-events/error-classifier.ts +109 -0
- package/extensions/mega-events.ts +1 -0
- package/extensions/mega-runtime/helpers.ts +5 -0
- package/extensions/mega-runtime/state.ts +17 -1
- package/extensions/mega-teamrun.test.ts +13 -1
- package/package.json +4 -1
- package/dist/extensions/dashboard-server/helpers.js +0 -37
- package/dist/extensions/dashboard-server/html/all-repos-tab.js +0 -26
- package/dist/extensions/dashboard-server/html/body-open.js +0 -23
- package/dist/extensions/dashboard-server/html/current-repo-tab.js +0 -130
- package/dist/extensions/dashboard-server/html/head-open.js +0 -16
- package/dist/extensions/dashboard-server/html/high-score-tab.js +0 -25
- package/dist/extensions/dashboard-server/html/repo-detail-modal.js +0 -26
- package/dist/extensions/dashboard-server/html/script.js +0 -259
- package/dist/extensions/dashboard-server/html/styles.js +0 -103
- package/dist/extensions/dashboard-server/html/summary-tab.js +0 -19
- package/dist/extensions/dashboard-server/html-template.js +0 -41
- package/dist/src/store/sqlite/connection.js +0 -35
- package/dist/src/store/sqlite/index-store.js +0 -167
- package/dist/src/store/sqlite/memory.js +0 -54
- package/dist/src/store/sqlite/minhash-lsh.js +0 -47
- package/dist/src/store/sqlite/sessions.js +0 -39
- package/dist/src/store/sqlite/transaction.js +0 -19
- package/dist/src/vectorStore/add.js +0 -260
- package/dist/src/vectorStore/dedup.js +0 -52
- package/dist/src/vectorStore/index.js +0 -10
- package/dist/src/vectorStore/queries.js +0 -83
- package/dist/src/vectorStore/search.js +0 -95
- package/dist/src/vectorStore/session.js +0 -19
- package/dist/src/vectorStore/store.js +0 -105
- package/dist/src/vectorStore/types.js +0 -6
- package/dist/src/vectorStore/utils.js +0 -23
|
@@ -4,6 +4,7 @@ import { recordScore } from "../../src/store/sqlite.js";
|
|
|
4
4
|
import { evaluateAndUnlockAchievements } from "../../src/store/sqlite/game-achievements.js";
|
|
5
5
|
import { isMegaCache } from "../../src/game/scoring.js";
|
|
6
6
|
import { resolveRepoRoot } from "../mega-config.js";
|
|
7
|
+
import { classifyError } from "./error-classifier.js";
|
|
7
8
|
/** Register agent/turn tracking event handlers. */
|
|
8
9
|
export function registerAgentHandlers(pi, runtime, config) {
|
|
9
10
|
// ---- Agent tracking for real-time widget + status-line updates ---------
|
|
@@ -37,10 +38,31 @@ export function registerAgentHandlers(pi, runtime, config) {
|
|
|
37
38
|
const idle = ctx.isIdle?.() ?? true;
|
|
38
39
|
const queued = ctx.hasPendingMessages?.() ?? false;
|
|
39
40
|
const now = Date.now();
|
|
41
|
+
// S38.5: read LIVE pressure from ctx.getContextUsage() instead of the
|
|
42
|
+
// stale runtime.lastCtxTokens (only updated by the `context` event).
|
|
43
|
+
// agent_end may fire without a preceding context event this turn (e.g.
|
|
44
|
+
// a sub-agent settling), so the cached value can be null/stale and the
|
|
45
|
+
// durable-trim branch would be unreachable. ctx.getContextUsage() is the
|
|
46
|
+
// authoritative live reading (mirrors context-handler.ts:86). Fall back
|
|
47
|
+
// to the cached value only if the ctx omits it.
|
|
48
|
+
const liveUsage = ctx.getContextUsage?.();
|
|
49
|
+
const liveTokens = typeof liveUsage?.tokens === "number"
|
|
50
|
+
? liveUsage.tokens
|
|
51
|
+
: runtime.lastCtxTokens ?? 0;
|
|
52
|
+
// Keep the cache fresh for snapshot()/diag regardless of which source we use.
|
|
53
|
+
if (typeof liveUsage?.tokens === "number") {
|
|
54
|
+
runtime.lastCtxTokens = liveUsage.tokens;
|
|
55
|
+
}
|
|
56
|
+
if (typeof liveUsage?.percent === "number") {
|
|
57
|
+
runtime.lastCtxPercent = liveUsage.percent;
|
|
58
|
+
}
|
|
59
|
+
if (typeof liveUsage?.contextWindow === "number") {
|
|
60
|
+
runtime.lastCtxWindow = liveUsage.contextWindow;
|
|
61
|
+
}
|
|
40
62
|
// DIAG (team-run relief): surface whether the agent is idle + over
|
|
41
63
|
// threshold at agent_end so we can see if a mid-run durable-trim trigger
|
|
42
64
|
// *should* have fired but didn't.
|
|
43
|
-
const overThreshold =
|
|
65
|
+
const overThreshold = liveTokens >= runtime.effectiveThreshold;
|
|
44
66
|
runtime.diagAgentEndIdle++;
|
|
45
67
|
runtime.logger.info("agent-end-idle", {
|
|
46
68
|
sessionId: runtime.rt.sessionId,
|
|
@@ -87,11 +109,20 @@ export function registerAgentHandlers(pi, runtime, config) {
|
|
|
87
109
|
// synchronous `piCompactWouldNoop` branch check misses a native
|
|
88
110
|
// compaction that hasn't appended its entry yet — calling
|
|
89
111
|
// ctx.compact() then races with pi and throws "Already compacted"
|
|
90
|
-
// to the user. The `
|
|
112
|
+
// to the user. The `lastNativeCompactAt` cooldown (updated by the
|
|
91
113
|
// session_compact listener for EVERY compaction, native or
|
|
92
114
|
// extension-supplied) closes that race window.
|
|
115
|
+
//
|
|
116
|
+
// S38.5: strict race guard widens the cooldown 10s -> 30s AND defers
|
|
117
|
+
// ctx.compact() via setTimeout(500) with a re-check, so pi's
|
|
118
|
+
// about-to-run native _checkCompaction can append its `compaction`
|
|
119
|
+
// branch entry first (closes the first-race-in-burst window). Gated
|
|
120
|
+
// by MEGACOMPACT_RACE_GUARD_STRICT (default true); false reverts to
|
|
121
|
+
// the v0.7.4 synchronous 10s guard. Mirrors the legacy path in
|
|
122
|
+
// context-handler.ts:258-287 so both call sites stay in sync.
|
|
123
|
+
const cooldownMs = config.raceGuardStrict ? 30_000 : 10_000;
|
|
93
124
|
const sinceCompact = now - (runtime.rt.lastNativeCompactAt ?? 0);
|
|
94
|
-
if (sinceCompact <
|
|
125
|
+
if (sinceCompact < cooldownMs) {
|
|
95
126
|
runtime.diagAgentEndDurableSkipRecent++;
|
|
96
127
|
}
|
|
97
128
|
else if (!piCompactWouldNoop(ctx)) {
|
|
@@ -103,7 +134,35 @@ export function registerAgentHandlers(pi, runtime, config) {
|
|
|
103
134
|
thresholdTokens: config.thresholdTokens,
|
|
104
135
|
queued,
|
|
105
136
|
});
|
|
106
|
-
|
|
137
|
+
if (config.raceGuardStrict) {
|
|
138
|
+
// Strict: defer ctx.compact() with a re-check so pi's
|
|
139
|
+
// about-to-run native _checkCompaction can append its
|
|
140
|
+
// `compaction` branch entry first. setTimeout(500) — pi's
|
|
141
|
+
// compaction-summary append is async I/O, so queueMicrotask
|
|
142
|
+
// would re-check before it lands.
|
|
143
|
+
const stamp = runtime.rt.lastNativeCompactAt;
|
|
144
|
+
const liveSid = runtime.rt.sessionId;
|
|
145
|
+
setTimeout(() => {
|
|
146
|
+
try {
|
|
147
|
+
if (runtime.rt.sessionId !== liveSid)
|
|
148
|
+
return; // session reset
|
|
149
|
+
const since2 = now - (runtime.rt.lastNativeCompactAt ?? 0);
|
|
150
|
+
if (runtime.rt.lastNativeCompactAt !== stamp && since2 < cooldownMs)
|
|
151
|
+
return;
|
|
152
|
+
if (piCompactWouldNoop(ctx))
|
|
153
|
+
return;
|
|
154
|
+
ctx.compact({
|
|
155
|
+
customInstructions: undefined,
|
|
156
|
+
}); // guardrails-allow PREVENT-PI-004: local ctx.compact() — no network; deferred + re-validated.
|
|
157
|
+
}
|
|
158
|
+
catch {
|
|
159
|
+
/* non-fatal */
|
|
160
|
+
}
|
|
161
|
+
}, 500);
|
|
162
|
+
}
|
|
163
|
+
else {
|
|
164
|
+
ctx.compact({ customInstructions: undefined }); // guardrails-allow PREVENT-PI-004: local ctx.compact() — no network; agent settled so no in-flight abort. Race-guarded by lastNativeCompactAt cooldown above (ctx.compact returns void → throw is surfaced by pi as compaction_end; the cooldown prevents the call entirely).
|
|
165
|
+
}
|
|
107
166
|
didDurableTrim = true;
|
|
108
167
|
}
|
|
109
168
|
}
|
|
@@ -141,6 +200,7 @@ export function registerAgentHandlers(pi, runtime, config) {
|
|
|
141
200
|
pi.on("turn_start", async (event, ctx) => {
|
|
142
201
|
runtime.currentTurn = event.turnIndex;
|
|
143
202
|
runtime.rt.lengthStopPending = false; // S28: re-arm defensively each user turn
|
|
203
|
+
runtime.rt.errorRetryCount = 0; // S38: reset error-retry counter each user turn
|
|
144
204
|
runtime.dashboard.event("turn_start", { turnIndex: event.turnIndex });
|
|
145
205
|
runtime.snapshot(ctx);
|
|
146
206
|
});
|
|
@@ -218,5 +278,125 @@ export function registerAgentHandlers(pi, runtime, config) {
|
|
|
218
278
|
runtime.rt.lengthStopPending = true;
|
|
219
279
|
runtime.dashboard.event("length_stop", { turnIndex: event.turnIndex });
|
|
220
280
|
}
|
|
281
|
+
// S38: broader error-retry safety net. S28 only catches stopReason==='length';
|
|
282
|
+
// this catches ALL other error types (provider failure, network timeout, 5xx,
|
|
283
|
+
// 429, auth, compaction-noop) that surface at turn_end. Non-fatal: wrapped in
|
|
284
|
+
// try/catch so a classifier/retry failure never breaks the agent loop.
|
|
285
|
+
// PREVENT-PI-003: retry nudge fires via pi.sendUserMessage (user-role).
|
|
286
|
+
try {
|
|
287
|
+
// (1) S28 owns length — skip the classifier entirely for it.
|
|
288
|
+
const sr = event.message?.stopReason;
|
|
289
|
+
if (sr === 'length') {
|
|
290
|
+
// S28 handles; nothing for S38 to do here.
|
|
291
|
+
}
|
|
292
|
+
else {
|
|
293
|
+
const category = classifyError(event.message);
|
|
294
|
+
if (category === null) {
|
|
295
|
+
// (3) success / normal flow / unknown-but-non-retryable — reset.
|
|
296
|
+
runtime.rt.errorRetryCount = 0;
|
|
297
|
+
runtime.rt.consecutiveErrors = 0; // S38.6: circuit-breaker reset on success
|
|
298
|
+
}
|
|
299
|
+
else if (category === 'compaction-noop') {
|
|
300
|
+
// (4) pi race / manual compact catch — NOT retryable. The compaction
|
|
301
|
+
// already succeeded via pi's native path; retrying would race again
|
|
302
|
+
// (FAIL-2026071701). Log a diagnostic, reset the counter, and surface
|
|
303
|
+
// the original error WITHOUT firing a retry nudge.
|
|
304
|
+
runtime.rt.errorRetryCount = 0;
|
|
305
|
+
runtime.rt.consecutiveErrors = 0; // S38.6: circuit-breaker reset
|
|
306
|
+
runtime.dashboard.event('compaction_noop_diagnostic', {
|
|
307
|
+
turnIndex: event.turnIndex,
|
|
308
|
+
sessionId: runtime.rt.sessionId,
|
|
309
|
+
});
|
|
310
|
+
runtime.logger.info('compaction-noop-diagnostic', {
|
|
311
|
+
sessionId: runtime.rt.sessionId,
|
|
312
|
+
turnIndex: event.turnIndex,
|
|
313
|
+
});
|
|
314
|
+
}
|
|
315
|
+
else {
|
|
316
|
+
// (5) transient or permanent — retry with exponential backoff.
|
|
317
|
+
// S38.7: hard-stop switch — bypass ALL retry logic when set.
|
|
318
|
+
if (config.errorRetryHardStop) {
|
|
319
|
+
runtime.rt.errorRetryCount = 0;
|
|
320
|
+
runtime.dashboard.event('error_retry_disabled', {
|
|
321
|
+
category,
|
|
322
|
+
turnIndex: event.turnIndex,
|
|
323
|
+
reason: 'hard-stop',
|
|
324
|
+
});
|
|
325
|
+
return; // early exit — no retry
|
|
326
|
+
}
|
|
327
|
+
// S38.6: circuit-breaker — stop retrying after too many consecutive errors.
|
|
328
|
+
runtime.rt.consecutiveErrors++;
|
|
329
|
+
if (runtime.rt.consecutiveErrors > config.maxConsecutiveErrors) {
|
|
330
|
+
runtime.dashboard.event('error_retry_circuit_open', {
|
|
331
|
+
consecutive: runtime.rt.consecutiveErrors,
|
|
332
|
+
max: config.maxConsecutiveErrors,
|
|
333
|
+
turnIndex: event.turnIndex,
|
|
334
|
+
});
|
|
335
|
+
runtime.logger.warn('error-retry-circuit-open', {
|
|
336
|
+
sessionId: runtime.rt.sessionId,
|
|
337
|
+
consecutive: runtime.rt.consecutiveErrors,
|
|
338
|
+
max: config.maxConsecutiveErrors,
|
|
339
|
+
});
|
|
340
|
+
return; // early exit — circuit breaker tripped
|
|
341
|
+
}
|
|
342
|
+
const max = category === 'transient'
|
|
343
|
+
? config.autoRetryTransientMax
|
|
344
|
+
: config.autoRetryPermanentMax;
|
|
345
|
+
// max === 0 disables the category entirely (revert to S28-only).
|
|
346
|
+
if (max <= 0) {
|
|
347
|
+
runtime.rt.errorRetryCount = 0;
|
|
348
|
+
}
|
|
349
|
+
else {
|
|
350
|
+
runtime.rt.errorRetryCount++;
|
|
351
|
+
if (runtime.rt.errorRetryCount > max) {
|
|
352
|
+
// Exhausted — surface the error, reset for the next burst.
|
|
353
|
+
runtime.dashboard.event('error_retry_exhausted', {
|
|
354
|
+
category,
|
|
355
|
+
count: runtime.rt.errorRetryCount,
|
|
356
|
+
max,
|
|
357
|
+
turnIndex: event.turnIndex,
|
|
358
|
+
});
|
|
359
|
+
runtime.logger.info('error-retry-exhausted', {
|
|
360
|
+
sessionId: runtime.rt.sessionId,
|
|
361
|
+
category,
|
|
362
|
+
count: runtime.rt.errorRetryCount,
|
|
363
|
+
max,
|
|
364
|
+
});
|
|
365
|
+
runtime.rt.errorRetryCount = 0;
|
|
366
|
+
}
|
|
367
|
+
else {
|
|
368
|
+
// S38: fire the retry nudge. Each error turn fires its own
|
|
369
|
+
// nudge up to `max` — we intentionally do NOT debounce on the
|
|
370
|
+
// backoff window (errorRetryUntil) here, because that would
|
|
371
|
+
// suppress turns 2..max on a fast-erroring provider (the exact
|
|
372
|
+
// scenario this feature targets) and the user could see zero
|
|
373
|
+
// retries followed by error_retry_exhausted. The per-turn cap
|
|
374
|
+
// (errorRetryCount <= max) plus the session circuit breaker
|
|
375
|
+
// (consecutiveErrors > maxConsecutiveErrors) already bound the
|
|
376
|
+
// loop, so a tight turn_end storm cannot busy-loop unbounded.
|
|
377
|
+
// errorRetryUntil + errorRetryBackoffMs() are retained on the
|
|
378
|
+
// runtime for future/optional pacing but are not gating.
|
|
379
|
+
runtime.dashboard.event('error_retry', {
|
|
380
|
+
category,
|
|
381
|
+
count: runtime.rt.errorRetryCount,
|
|
382
|
+
max,
|
|
383
|
+
turnIndex: event.turnIndex,
|
|
384
|
+
});
|
|
385
|
+
runtime.logger.info('error-retry', {
|
|
386
|
+
sessionId: runtime.rt.sessionId,
|
|
387
|
+
category,
|
|
388
|
+
count: runtime.rt.errorRetryCount,
|
|
389
|
+
max,
|
|
390
|
+
});
|
|
391
|
+
// PREVENT-PI-003: user-role sendUserMessage only.
|
|
392
|
+
pi.sendUserMessage('[mega-compact] the last turn ended with an error; please retry.');
|
|
393
|
+
}
|
|
394
|
+
}
|
|
395
|
+
}
|
|
396
|
+
}
|
|
397
|
+
}
|
|
398
|
+
catch {
|
|
399
|
+
/* non-fatal: a classifier/retry failure never breaks the agent loop */
|
|
400
|
+
}
|
|
221
401
|
});
|
|
222
402
|
}
|
|
@@ -208,10 +208,43 @@ export function registerContextHandler(pi, runtime, config) {
|
|
|
208
208
|
// NATIVE compaction just fired (avoids racing pi and surfacing a spurious
|
|
209
209
|
// "Already compacted" / "Nothing to compact" toast). Uses lastNativeCompactAt
|
|
210
210
|
// (NOT lastCompactAt, which runCompact also stamps for our own checkpoint).
|
|
211
|
+
// S38.5: strict race guard widens the cooldown 10s -> 30s (gated by
|
|
212
|
+
// MEGACOMPACT_RACE_GUARD_STRICT; false reverts to v0.7.4 10s).
|
|
213
|
+
const cooldownMs = config.raceGuardStrict ? 30_000 : 10_000;
|
|
211
214
|
const sinceCompact = Date.now() - (runtime.rt.lastNativeCompactAt ?? 0);
|
|
212
|
-
if (sinceCompact <
|
|
215
|
+
if (sinceCompact < cooldownMs || piCompactWouldNoop(ctx))
|
|
213
216
|
return;
|
|
214
|
-
|
|
217
|
+
// S38.5: defer ctx.compact() with a re-check so pi's about-to-run native
|
|
218
|
+
// _checkCompaction can append its `compaction` branch entry first (closes
|
|
219
|
+
// the first-race-in-burst window). setTimeout(500) — pi's compaction-summary
|
|
220
|
+
// append is async I/O, so queueMicrotask would re-check before it lands.
|
|
221
|
+
// Non-strict (v0.7.4) keeps the synchronous call.
|
|
222
|
+
if (config.raceGuardStrict) {
|
|
223
|
+
const stamp = runtime.rt.lastNativeCompactAt;
|
|
224
|
+
const liveSid = runtime.rt.sessionId;
|
|
225
|
+
setTimeout(() => {
|
|
226
|
+
try {
|
|
227
|
+
if (runtime.rt.sessionId !== liveSid)
|
|
228
|
+
return; // session reset
|
|
229
|
+
const since2 = Date.now() - (runtime.rt.lastNativeCompactAt ?? 0);
|
|
230
|
+
if (runtime.rt.lastNativeCompactAt !== stamp && since2 < cooldownMs)
|
|
231
|
+
return;
|
|
232
|
+
if (piCompactWouldNoop(ctx))
|
|
233
|
+
return;
|
|
234
|
+
ctx.compact({
|
|
235
|
+
customInstructions: undefined,
|
|
236
|
+
}); // guardrails-allow PREVENT-PI-004: local ctx.compact() — no network; deferred + re-validated.
|
|
237
|
+
}
|
|
238
|
+
catch {
|
|
239
|
+
/* non-fatal */
|
|
240
|
+
}
|
|
241
|
+
}, 500);
|
|
242
|
+
}
|
|
243
|
+
else {
|
|
244
|
+
ctx.compact({
|
|
245
|
+
customInstructions: undefined,
|
|
246
|
+
}); // race-guarded by lastNativeCompactAt cooldown (ctx.compact returns void → not catchable; the cooldown prevents the call)
|
|
247
|
+
}
|
|
215
248
|
return;
|
|
216
249
|
}
|
|
217
250
|
// S16 LIVE trim: collapse the compacted region to a summary + recent anchor.
|
|
@@ -0,0 +1,118 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* error-classifier.ts — S38.2 error classification for retry logic.
|
|
3
|
+
*
|
|
4
|
+
* Classifies turn-end error/stop signals into retry categories.
|
|
5
|
+
* Ordering matters: compaction-noop is matched BEFORE generic transient
|
|
6
|
+
* so a pi race / manual compact catch is never misclassified.
|
|
7
|
+
*/
|
|
8
|
+
/** S38.2: classify a turn-end error/stop signal into a retry category.
|
|
9
|
+
*
|
|
10
|
+
* `length` is returned as null — S28 owns the max-output-token length stopReason
|
|
11
|
+
* exclusively (its agent_end nudge path is separate and must not be doubled).
|
|
12
|
+
*
|
|
13
|
+
* @param message the event.message (a pi AgentMessage) or an error string
|
|
14
|
+
* @returns 'transient' | 'permanent' | 'compaction-noop' | null (success/unknown)
|
|
15
|
+
*/
|
|
16
|
+
export function classifyError(message) {
|
|
17
|
+
// Resolve a searchable text blob from a pi AgentMessage or raw string.
|
|
18
|
+
let text = '';
|
|
19
|
+
if (typeof message === 'string') {
|
|
20
|
+
text = message;
|
|
21
|
+
}
|
|
22
|
+
else if (message && typeof message === 'object') {
|
|
23
|
+
const m = message;
|
|
24
|
+
const sr = typeof m.stopReason === 'string' ? m.stopReason : '';
|
|
25
|
+
// S28 guard: length stopReason is handled exclusively by the S28 path.
|
|
26
|
+
if (sr === 'length')
|
|
27
|
+
return null;
|
|
28
|
+
// Success / normal tool flow — not an error, nothing to retry.
|
|
29
|
+
if (sr === 'stop' || sr === 'toolUse' || sr === 'tool_use')
|
|
30
|
+
return null;
|
|
31
|
+
const parts = [];
|
|
32
|
+
if (sr)
|
|
33
|
+
parts.push(sr);
|
|
34
|
+
const c = m.content;
|
|
35
|
+
if (typeof c === 'string')
|
|
36
|
+
parts.push(c);
|
|
37
|
+
else if (Array.isArray(c)) {
|
|
38
|
+
for (const b of c) {
|
|
39
|
+
if (b && typeof b === 'object' && 'text' in b) {
|
|
40
|
+
parts.push(String(b.text ?? ''));
|
|
41
|
+
}
|
|
42
|
+
}
|
|
43
|
+
}
|
|
44
|
+
if (m.error) {
|
|
45
|
+
// Extract message from error objects for pattern matching.
|
|
46
|
+
const err = m.error;
|
|
47
|
+
if (typeof err === 'string') {
|
|
48
|
+
parts.push(err);
|
|
49
|
+
}
|
|
50
|
+
else if (err && typeof err === 'object') {
|
|
51
|
+
const errObj = err;
|
|
52
|
+
if (typeof errObj.message === 'string') {
|
|
53
|
+
parts.push(errObj.message);
|
|
54
|
+
}
|
|
55
|
+
else {
|
|
56
|
+
parts.push(JSON.stringify(err));
|
|
57
|
+
}
|
|
58
|
+
}
|
|
59
|
+
}
|
|
60
|
+
// S38: detect mid-response errors where the stream died without a
|
|
61
|
+
// proper stopReason (empty/undefined) — this catches provider failures
|
|
62
|
+
// that cut off the response mid-stream, INCLUDING the case where the
|
|
63
|
+
// provider emitted partial content before dying (a truncated response
|
|
64
|
+
// with no stop reason is a mid-stream death, not a success).
|
|
65
|
+
// A genuine success ALWAYS carries stop/tool_use/toolUse (which
|
|
66
|
+
// short-circuit to null at the top), so any message reaching here with a
|
|
67
|
+
// falsy stopReason is a stream failure → retryable.
|
|
68
|
+
if (!sr) {
|
|
69
|
+
return 'transient';
|
|
70
|
+
}
|
|
71
|
+
text = parts.join(' ');
|
|
72
|
+
}
|
|
73
|
+
if (!text)
|
|
74
|
+
return null;
|
|
75
|
+
const s = text.toLowerCase();
|
|
76
|
+
// --- compaction-noop (ORDER FIRST: pi race / manual compact catch) ---
|
|
77
|
+
// FAIL-2026071701: these are NOT retryable — the compaction already
|
|
78
|
+
// succeeded via pi's native path; retrying would race again.
|
|
79
|
+
if (/already compacted/.test(s))
|
|
80
|
+
return 'compaction-noop';
|
|
81
|
+
if (/compaction failed/.test(s))
|
|
82
|
+
return 'compaction-noop';
|
|
83
|
+
if (/nothing to compact/.test(s))
|
|
84
|
+
return 'compaction-noop';
|
|
85
|
+
if (/auto[\s-]?compaction failed/.test(s))
|
|
86
|
+
return 'compaction-noop';
|
|
87
|
+
// --- transient (retryable) ---
|
|
88
|
+
if (s.includes('error') && !/\b(permanent|invalid request|malformed|bad request|auth|unauthorized|invalid (api )?key|permission)\b/.test(s)) {
|
|
89
|
+
return 'transient'; // generic pi stopReason 'error' / 'aborted'
|
|
90
|
+
}
|
|
91
|
+
if (s.includes('aborted'))
|
|
92
|
+
return 'transient';
|
|
93
|
+
if (/max(imum)? output token/.test(s))
|
|
94
|
+
return 'transient';
|
|
95
|
+
if (/rate[\s.-]?limit|429|too many requests/.test(s))
|
|
96
|
+
return 'transient';
|
|
97
|
+
if (/5\d\d|internal server|bad gateway|service unavailable/.test(s))
|
|
98
|
+
return 'transient';
|
|
99
|
+
if (/network|timeout|connection (lost|refused|reset)|stream (interrupted|closed|ended|failed)|disconnected/.test(s))
|
|
100
|
+
return 'transient';
|
|
101
|
+
// --- permanent (NOT retryable beyond 1) ---
|
|
102
|
+
if (/auth|unauthorized|invalid (api )?key|permission/.test(s))
|
|
103
|
+
return 'permanent';
|
|
104
|
+
if (/invalid request|malformed|bad request/.test(s))
|
|
105
|
+
return 'permanent';
|
|
106
|
+
// Unknown — do not retry (avoid busy-looping on an unclassified signal).
|
|
107
|
+
return null;
|
|
108
|
+
}
|
|
109
|
+
/** S38.2: exponential backoff for error-retry nudges (5s,10s,20s,30s,30s cap).
|
|
110
|
+
* count is 1-based (the retry about to fire). */
|
|
111
|
+
export function errorRetryBackoffMs(count) {
|
|
112
|
+
switch (count) {
|
|
113
|
+
case 1: return 5_000;
|
|
114
|
+
case 2: return 10_000;
|
|
115
|
+
case 3: return 20_000;
|
|
116
|
+
default: return 30_000; // cap from the 4th retry onward
|
|
117
|
+
}
|
|
118
|
+
}
|
|
@@ -49,6 +49,9 @@ export class MegaRuntime {
|
|
|
49
49
|
recallInjections: 0,
|
|
50
50
|
cacheHitTokens: 0,
|
|
51
51
|
lengthStopPending: false,
|
|
52
|
+
errorRetryCount: 0,
|
|
53
|
+
errorRetryUntil: 0,
|
|
54
|
+
consecutiveErrors: 0,
|
|
52
55
|
};
|
|
53
56
|
// v0.8.6 cache-stability: the cached live-trim view for the current
|
|
54
57
|
// compaction epoch. Set after a fresh runCompact + computeLiveTrimCut, and
|
|
@@ -486,6 +489,16 @@ export class MegaRuntime {
|
|
|
486
489
|
cacheHit: { sessionSec: sec(this.rt.cacheHitTokens), totalSec: sec(cacheHitsTotalTokens) },
|
|
487
490
|
},
|
|
488
491
|
model,
|
|
492
|
+
// S38.8: error-retry state for the dashboard "retries" tile. The field is
|
|
493
|
+
// declared on DashboardSnapshot (mega-dashboard.ts) and surfaced here so the
|
|
494
|
+
// dashboard can render live retry/circuit-breaker status alongside the event
|
|
495
|
+
// stream (which already carries per-retry events).
|
|
496
|
+
retries: {
|
|
497
|
+
errorRetryCount: this.rt.errorRetryCount,
|
|
498
|
+
consecutiveErrors: this.rt.consecutiveErrors,
|
|
499
|
+
maxConsecutiveErrors: this.config.maxConsecutiveErrors,
|
|
500
|
+
errorRetryHardStop: this.config.errorRetryHardStop,
|
|
501
|
+
},
|
|
489
502
|
diag: {
|
|
490
503
|
ctxFastGate: this.diagCtxFastGate,
|
|
491
504
|
liveTrimFires: this.diagLiveTrimFires,
|
|
@@ -742,6 +755,9 @@ export class MegaRuntime {
|
|
|
742
755
|
recallInjections: 0,
|
|
743
756
|
cacheHitTokens: 0,
|
|
744
757
|
lengthStopPending: false,
|
|
758
|
+
errorRetryCount: 0,
|
|
759
|
+
errorRetryUntil: 0,
|
|
760
|
+
consecutiveErrors: 0,
|
|
745
761
|
};
|
|
746
762
|
this.trimCache = null; // v0.8.6: never replay a stale trim into a new session
|
|
747
763
|
this.statusKey = undefined;
|
|
@@ -38,6 +38,12 @@ function harness() {
|
|
|
38
38
|
process.env.MEGACOMPACT_DEBUG = "true";
|
|
39
39
|
process.env.MEGACOMPACT_THRESHOLD_TOKENS = "50";
|
|
40
40
|
process.env.MEGACOMPACT_FAST_GATE_PCT = "1";
|
|
41
|
+
// Strict race-guard mode double-counts diagAgentEndDurable (sync++ at branch
|
|
42
|
+
// entry + deferred++ in the setTimeout(500) callback = 6, not 3), lands
|
|
43
|
+
// compactCalls after the synchronous assertions, and leaks timers that hang
|
|
44
|
+
// `node --test`. The strict deferred path is covered by the two S38.5 tests
|
|
45
|
+
// in mega-compact.test.ts. Use the synchronous v0.7.4 path here.
|
|
46
|
+
process.env.MEGACOMPACT_RACE_GUARD_STRICT = "false";
|
|
41
47
|
process.env.MEGACOMPACT_ANCHOR_USER_MESSAGES = "1";
|
|
42
48
|
process.env.MEGACOMPACT_DURABLE_TRIM_FLOOR = "0"; // piCompactWouldNoop must not skip
|
|
43
49
|
process.env.MEGACOMPACT_MEMORY_AUTO_REVIEW = "false";
|
|
@@ -138,6 +144,13 @@ test("control: session_before_compact supplies a durable compaction (parent sett
|
|
|
138
144
|
assert.equal(res.compaction.firstKeptEntryId, "e2", "reuses pi's boundary (PREVENT-PI-002)");
|
|
139
145
|
});
|
|
140
146
|
test("cleanup", async () => {
|
|
141
|
-
|
|
147
|
+
// Race closeVectorIndex with a timeout to prevent 40-min hangs.
|
|
148
|
+
try {
|
|
149
|
+
await Promise.race([
|
|
150
|
+
closeVectorIndex(),
|
|
151
|
+
new Promise((r) => setTimeout(r, 3000)),
|
|
152
|
+
]);
|
|
153
|
+
}
|
|
154
|
+
catch { /* ignore */ }
|
|
142
155
|
rmSync(baseTmp, { recursive: true, force: true });
|
|
143
156
|
});
|
|
@@ -59,7 +59,7 @@ function harness() {
|
|
|
59
59
|
// the percent-basis grewEnough (>=10) never trips (no re-compact → pure replay).
|
|
60
60
|
const usage = { tokens: 200000, contextWindow: 200000, percent: 100 as number | null };
|
|
61
61
|
|
|
62
|
-
const handlers: Record<string, Function> = {};
|
|
62
|
+
const handlers: Record<string, Function[]> = {};
|
|
63
63
|
const compactCalls: any[] = [];
|
|
64
64
|
|
|
65
65
|
function msg(role: string, text: string, toolName?: string): AgentMessage {
|
|
@@ -106,7 +106,7 @@ function harness() {
|
|
|
106
106
|
}
|
|
107
107
|
|
|
108
108
|
const pi = {
|
|
109
|
-
on: (ev: string, h: Function) => { handlers[ev] = h; },
|
|
109
|
+
on: (ev: string, h: Function) => { if (!handlers[ev]) handlers[ev] = []; handlers[ev].push(h); },
|
|
110
110
|
registerCommand: () => {}, registerTool: () => {}, registerShortcut: () => {},
|
|
111
111
|
registerFlag: () => {}, getFlag: () => undefined, registerMessageRenderer: () => {},
|
|
112
112
|
registerEntryRenderer: () => {}, sendMessage: () => {}, sendUserMessage: () => {},
|
|
@@ -121,7 +121,7 @@ function harness() {
|
|
|
121
121
|
mod.default(pi);
|
|
122
122
|
const { lastRuntime } = require("./mega-events.js") as { lastRuntime: any };
|
|
123
123
|
|
|
124
|
-
const fire = (ev: string, event: any, ctx: any) => handlers[ev](event, ctx);
|
|
124
|
+
const fire = async (ev: string, event: any, ctx: any) => { let r: any; for (const h of handlers[ev] || []) r = await h(event, ctx); return r; };
|
|
125
125
|
return {
|
|
126
126
|
stateDir, handlers, compactCalls, fire, ctx: makeCtx, usage, buildSession,
|
|
127
127
|
runtime: lastRuntime, // MegaRuntime with diag* counters + rt + trimCache
|