{
    "name": "nudge-triggered",
    "description": "Nudge-triggered compression: the model emits growth text across turns until billion-context-pi's context-usage nudge is injected, then the fake LLM detects the nudge and compresses. A low modelContextLimit makes the nudge fire quickly. Validates the nudge detection + baseline-recording path.",
    "acpConfig": { "modelContextLimit": 1500 },
    "turns": [
        {
            "respond": "nudge-compress",
            "userText": "Let's work on a long design discussion. Go ahead and start explaining the data ingestion pipeline in detail.",
            "growthText": "Working on the data ingestion pipeline design. The ingestion service accepts events from upstream producers via a Kafka topic (events-raw, 12 partitions, replication factor 3). The consumer group ingest-workers (8 instances) reads in batches of up to 500 records with a 5-second max poll interval. Each batch is deserialized from Avro at src/ingest/decode.ts:24 against the schema registry (Confluent, cached at src/ingest/schema-cache.ts:16, TTL 10 minutes). Validation runs a JSON-Schema check at src/ingest/validate.ts:30 and a domain-rules check at src/ingest/domain-rules.ts:42; invalid records are diverted to the poison-topic for manual review (src/ingest/poison.ts:18). Valid records are enriched (geo-IP lookup at src/ingest/enrich.ts:28 via MaxMind GeoIP2, user-agent parsing at src/ingest/enrich.ts:60) then written to the events table in a single batched INSERT at src/ingest/persist.ts:34. Idempotency is guaranteed by a deterministic event_id (SHA-256 of producer+seq, src/ingest/id.ts:12) with an ON CONFLICT DO NOTHING clause. Backpressure: if the DB write exceeds 800ms the consumer pauses polling for 2 seconds (src/ingest/backpressure.ts:20). Metrics: ingest_lag, ingest_validation_errors, ingest_persist_latency exported at src/ingest/metrics.ts:24. This is part one of the pipeline design discussion.",
            "topic": "Data Ingestion Pipeline",
            "summary": "## Data Ingestion Pipeline\nKafka topic events-raw (12p, rf3) → ingest-workers (8 instances, batch 500, 5s poll). Avro decode src/ingest/decode.ts:24 + schema registry cache src/ingest/schema-cache.ts:16 (10m). Validate JSON-Schema src/ingest/validate.ts:30 + domain src/ingest/domain-rules.ts:42; poison topic src/ingest/poison.ts:18. Enrich geo-IP src/ingest/enrich.ts:28, UA src/ingest/enrich.ts:60. Batched INSERT src/ingest/persist.ts:34, event_id=SHA256(prod+seq) ON CONFLICT DO NOTHING src/ingest/id.ts:12. Backpressure src/ingest/backpressure.ts:20 (DB>800ms→pause 2s). Metrics src/ingest/metrics.ts:24.",
            "range": "all"
        },
        {
            "respond": "nudge-compress",
            "userText": "Continue with the stream processing layer.",
            "growthText": "Continuing the pipeline design with the stream-processing layer. Downstream of ingestion, the stream-processing service consumes the events table's logical replication slot (src/stream/wal.ts:30, pgoutput protocol) and maintains real-time aggregates. Windowed aggregations (1-minute tumbling windows for request counts, 5-minute for revenue) are computed in a rocksdb-backed state store at src/stream/state.ts:28; the state is checkpointed to S3 every 60 seconds (src/stream/checkpoint.ts:42) for exactly-once recovery. Late events (arriving after the window closed) are routed to a side-output for up to 10 minutes then dropped (src/stream/late.ts:18). The join between the events stream and the reference dim_customers table is a lookup join cached for 5 minutes (src/stream/join.ts:52). Watermarks are derived from the event timestamp with a 30-second bounded out-of-orderness (src/stream/watermark.ts:24). The processed aggregates are written to the realtime_dashboard table and also pushed to a WebSocket fan-out at src/stream/ws-push.ts:30 so the UI updates live. Scaling is per-partition: each of the 12 partitions is owned by exactly one worker (cooperative rebalancing src/stream/rebalance.ts:40), and state migration on rebalance restores from the S3 checkpoint. This concludes the stream-processing discussion and continues our pipeline design.",
            "topic": "Stream Processing Layer",
            "summary": "## Stream Processing Layer\nLogical-repl WAL slot src/stream/wal.ts:30 → aggregates. 1min/5min tumbling windows in rocksdb state src/stream/state.ts:28; checkpoint S3/60s src/stream/checkpoint.ts:42 (exactly-once). Late events side-output 10min src/stream/late.ts:18. Lookup join dim_customers cached 5min src/stream/join.ts:52. Watermarks 30s out-of-order src/stream/watermark.ts:24. Output realtime_dashboard + WS fan-out src/stream/ws-push.ts:30. 12 partitions, cooperative rebalance src/stream/rebalance.ts:40, state restore from S3.",
            "range": "all"
        },
        {
            "respond": "nudge-compress",
            "userText": "Keep going, the context is getting full — wrap up and consolidate.",
            "growthText": "Continuing to generate pipeline design content to fill the context window. The batch layer complements the stream layer: a nightly Spark job at src/batch/nightly.ts:40 recomputes the last 24 hours from the raw events table to correct any drift in the real-time aggregates, writing corrections to the batch_corrections table. The reconciliation merge at src/batch/reconcile.ts:28 applies batch values where they disagree with stream values, logging every override. Schemas evolve via the schema registry with backward-compatible Avro reads (src/schema/evolve.ts:22); a breaking change requires a new topic. Replay support lets us re-process the last 7 days from the raw topic into a repaired consumer (src/ops/replay.ts:35) while normal ingestion continues. Data quality SLAs (completeness >99.9%, freshness <60s) are monitored by the great_expectations suite at tests/data-quality/ and the results feed the deploy gate (src/deploy/dq-gate.ts:30). This is more pipeline design content being generated to exercise the context window and trigger the compression nudge so we can validate the nudge pipeline end to end.",
            "topic": "Batch & Replay Layer",
            "summary": "## Batch & Replay\nNightly Spark src/batch/nightly.ts:40 recompute 24h → batch_corrections. Reconcile merge src/batch/reconcile.ts:28 (log overrides). Schema evolve backward-compat src/schema/evolve.ts:22 (breaking=new topic). Replay last 7d src/ops/replay.ts:35. DQ SLAs completeness>99.9%/freshness<60s via great_expectations tests/data-quality/, gate src/deploy/dq-gate.ts:30.",
            "range": "all"
        },
        {
            "respond": "text",
            "text": "Compression complete. The pipeline design discussion has been consolidated.",
            "auto": true
        }
    ],
    "verify": {
        "minBlockCount": 1,
        "minNudgeCount": 1,
        "nudgeBaselineSet": true
    }
}