feat(agent-loop): activate retryLoopReminder via LoopState.recentCalls

Extends LoopState with a sliding window of the last N ExecutedCalls
(oldest-first), capped at LOOP_STATE_RECENT_CALLS_WINDOW = 5. The loop
maintains the window automatically; reminderChannel producers read it
without touching internal state.

This activates retryLoopReminder which was shape-only in faa472be9.
The guard now fires end-to-end: when round >= 3 and the tail-2 calls
both returned success:false, the LLM sees a "stop retrying, write a
summary instead" <reminder> on the next turn. The tail-2 check rather
than window-wide is deliberate — a flaky run with intermittent success
(F, F, F, OK, F) is not a retry loop, just flaky tools.

Why window=5: retry loops usually manifest within 2-3 consecutive
rounds; a 5-deep window gives room for burst-detection and
stale-tool heuristics without bloating the reminder channel. Cap
keeps the reminder producers O(5) regardless of loop length.

Tests: 3 new (sliding-window cap + slide + order in shared-ai, retry
composition + budget+retry chain + tail-only heuristic in mana-ai).
Total agent-loop tests now 74 across both packages.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Till JS 2026-04-23 15:02:40 +02:00
parent fc635f9830
commit 8f283726b1
6 changed files with 171 additions and 24 deletions

View file

@ -101,8 +101,8 @@ Details zum Deep-Research-Flow: [`docs/reports/gemini-deep-research.md`](../../d
Claude-Code-inspirierte Primitive in `runPlannerLoop` (live in `@mana/shared-ai`, siehe [`docs/plans/agent-loop-improvements-m1.md`](../../docs/plans/agent-loop-improvements-m1.md)) und deren Konsumierung hier:
- [x] `reminderChannel` wired via `buildReminderChannel()` in `src/planner/reminders.ts`. Erster Live-Producer: `tokenBudgetReminder` — warnt ab 75% Tagesbudget, eskaliert ab 100% mit "JETZT abschliessen"-Prompt. Round-usage wird on-the-fly drauf addiert, so dass der Warn-Level mitwandert.
- [x] `retryLoopReminder` — Shape fertig, aber dormant: LoopState exponiert heute nur `lastCall`, nicht ein Failure-Window. Aktiviert automatisch sobald shared-ai LoopState um `recentResults` erweitert.
- [x] `reminderChannel` wired via `buildReminderChannel()` in `src/planner/reminders.ts`. Live-Producer 1: `tokenBudgetReminder` — warnt ab 75% Tagesbudget, eskaliert ab 100% mit "JETZT abschliessen"-Prompt. Round-usage wird on-the-fly drauf addiert, so dass der Warn-Level mitwandert.
- [x] `retryLoopReminder` live — feuert ab Round 3 wenn die letzten 2 Tool-Calls beide fehlschlugen. Liest das `recentCalls`-Sliding-Window (5 Einträge, oldest-first) aus `LoopState`.
- [x] `POLICY_MODE` env (off/log-only/enforce, default log-only) für die mana-ai-seitige Freitext-Inspection (`detectInjectionMarker`). Rate-Limit und destructive-opt-in sind hier NICHT aktiv — tools werden nur als PlanSteps aufgezeichnet, die echte Enforcement passiert im Webapp-Client.
- [ ] Parallel-Reads im Server-Tick haben keinen Effekt, weil `SERVER_TOOLS` per Konstruktion propose-only ist. Könnte relevant werden sobald mana-ai die vollständige tool-registry absorbiert (M4 des Personas-Plans).

View file

@ -49,10 +49,26 @@ function makeState(overrides: Partial<LoopState> = {}): LoopState {
round: 1,
toolCallCount: 0,
usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
recentCalls: [],
...overrides,
};
}
function mkExecutedCall(
success: boolean,
toolName = 'create_thing'
): {
round: number;
call: { id: string; name: string; arguments: Record<string, unknown> };
result: { success: boolean; message: string };
} {
return {
round: 1,
call: { id: crypto.randomUUID(), name: toolName, arguments: {} },
result: { success, message: success ? 'ok' : 'boom' },
};
}
// ─── tokenBudgetReminder ──────────────────────────────────────────
describe('tokenBudgetReminder', () => {
@ -121,21 +137,50 @@ describe('tokenBudgetReminder', () => {
describe('retryLoopReminder', () => {
it('is silent before round 3', () => {
expect(retryLoopReminder({ round: 2, lastFailures: [true, true] })).toBeNull();
expect(
retryLoopReminder({
round: 2,
recentCalls: [mkExecutedCall(false), mkExecutedCall(false)],
})
).toBeNull();
});
it('warns when the last 2 calls failed at round >= 3', () => {
const msg = retryLoopReminder({ round: 3, lastFailures: [true, true] });
const msg = retryLoopReminder({
round: 3,
recentCalls: [mkExecutedCall(false), mkExecutedCall(false)],
});
expect(msg).not.toBeNull();
expect(msg).toContain('fehlgeschlagen');
});
it('stays silent when only one of the last 2 failed', () => {
expect(retryLoopReminder({ round: 4, lastFailures: [false, true] })).toBeNull();
expect(
retryLoopReminder({
round: 4,
recentCalls: [mkExecutedCall(true), mkExecutedCall(false)],
})
).toBeNull();
});
it('stays silent with fewer than 2 failures recorded', () => {
expect(retryLoopReminder({ round: 5, lastFailures: [true] })).toBeNull();
it('stays silent with fewer than 2 calls recorded', () => {
expect(retryLoopReminder({ round: 5, recentCalls: [mkExecutedCall(false)] })).toBeNull();
});
it('looks only at the TAIL 2 — a flaky run with intermittent success is not a retry loop', () => {
// 5 calls: F, F, F, OK, F → tail-2 is [OK, F] → silent
expect(
retryLoopReminder({
round: 5,
recentCalls: [
mkExecutedCall(false),
mkExecutedCall(false),
mkExecutedCall(false),
mkExecutedCall(true),
mkExecutedCall(false),
],
})
).toBeNull();
});
});
@ -164,6 +209,40 @@ describe('buildReminderChannel', () => {
expect(out[0]).toContain('90%');
});
it('fires retryLoopReminder end-to-end through the channel', () => {
const channel = buildReminderChannel({
agent: makeAgent({ maxTokensPerDay: 1_000_000 }), // budget silent
mission: makeMission(),
pretickUsage24h: 0,
});
const out = channel(
makeState({
round: 4,
recentCalls: [mkExecutedCall(false), mkExecutedCall(false)],
})
);
expect(out).toHaveLength(1);
expect(out[0]).toContain('fehlgeschlagen');
});
it('can fire budget + retry together (composition)', () => {
const channel = buildReminderChannel({
agent: makeAgent({ maxTokensPerDay: 10_000 }),
mission: makeMission(),
pretickUsage24h: 9_000,
});
const out = channel(
makeState({
round: 3,
usage: { promptTokens: 500, completionTokens: 500, totalTokens: 1_000 },
recentCalls: [mkExecutedCall(false), mkExecutedCall(false)],
})
);
expect(out).toHaveLength(2);
expect(out[0]).toContain('ausgeschoepft'); // budget first
expect(out[1]).toContain('fehlgeschlagen'); // retry second
});
it('uses the updated totalTokens each round (re-evaluated)', () => {
const channel = buildReminderChannel({
agent: makeAgent({ maxTokensPerDay: 10_000 }),

View file

@ -69,17 +69,22 @@ export function tokenBudgetReminder(ctx: ReminderContext, roundUsage: number): s
/**
* Nudge the planner to end when it is clearly iterating without new
* information: 3+ rounds in and the last 2 tool-calls returned
* `success: false`. This is a heuristic guard against infinite re-try
* loops where the LLM keeps calling the same failing tool with slightly
* different arguments.
* `success: false`. Heuristic guard against infinite retry loops where
* the LLM keeps calling the same failing tool with slightly different
* arguments.
*
* Reads the `recentCalls` sliding window from LoopState the last 5
* executed calls in oldest-first order. We only look at the tail 2
* because a run that mixes failures and successes is not a true retry
* loop, it's just flaky tools.
*/
export function retryLoopReminder(state: {
readonly round: number;
readonly lastFailures: readonly boolean[];
readonly recentCalls: readonly { readonly result: { readonly success: boolean } }[];
}): string | null {
if (state.round < 3) return null;
const recent = state.lastFailures.slice(-2);
if (recent.length === 2 && recent.every((f) => f)) {
const tail = state.recentCalls.slice(-2);
if (tail.length === 2 && tail.every((ec) => !ec.result.success)) {
return (
`Die letzten 2 Tool-Calls sind fehlgeschlagen. Brich die ` +
`Wiederholung ab — formuliere stattdessen einen Summary-Text, ` +
@ -100,19 +105,10 @@ export function retryLoopReminder(state: {
*/
export function buildReminderChannel(ctx: ReminderContext): ReminderChannel {
return (state) => {
const failures: boolean[] = [];
// We don't get the full executedCalls in LoopState (intentional —
// the channel is meant to be cheap), but `lastCall` is exposed.
// For retry-loop detection we'd ideally track the last N; for now
// the single lastCall is enough to skip 2-round miss signals, so
// this producer is effectively dormant until we extend LoopState.
// Left in place so the shape is right for M2 follow-ups.
if (state.lastCall) failures.push(!state.lastCall.result.success);
const out: string[] = [];
const budget = tokenBudgetReminder(ctx, state.usage.totalTokens);
if (budget) out.push(budget);
const retry = retryLoopReminder({ round: state.round, lastFailures: failures });
const retry = retryLoopReminder({ round: state.round, recentCalls: state.recentCalls });
if (retry) out.push(retry);
return out;
};