poly-copy-trading

name

poly-copy-trading

description

Cogni poly copy-trade mirror pipeline specialist. Load when working on the mirror loop itself — `mirror-coordinator`, `wallet-watch`, `poly_copy_trade_{targets,config,fills,decisions}` tables, poll cadence, v0 live-money caps, RLS on copy-trade tables, shared poller refactor (task.0332), status-sync / sync-truth cache (task.0328), v1 hardening bucket (task.0323), or Phase 4 streaming prep (task.0322). Also triggers for: 'mirror this wallet', 'why didn't the mirror fire', 'flip copy_trade_config', 'tracked wallet add', 'mirror skip reason=already_placed noise', 'fills ledger status drift', 'cap exceeded daily', 'WebSocket streaming poly', 'target ranker'. For provisioning trading wallets / CLOB creds / CTF approvals see `poly-auth-wallets`; for CLOB order semantics / Data-API / wallet screening see `poly-market-data`.

Poly Copy-Trading Pipeline

You are the expert for the mirror loop — everything that turns "target wallet traded" into "our wallet placed a mirror order." Trading-wallet provisioning, signing, and CLOB semantics live in sibling skills.

Architecture in one pass

wallet-watch source                   (WS wake-up + Data-API drain, or legacy source in old branches)
    │
    ▼
features/wallet-watch                 (normalize → Fill; no decisions, no writes)
    │
    ▼
mirror-pipeline.ts::runMirrorTick     (dedup, target policy, position context)
    │
    ▼ INSERT_BEFORE_PLACE              (correctness gate — at-most-once)
poly_copy_trade_fills row lands       (status=pending, client_order_id set)
    │
    ▼
PolyTradeExecutor.placeIntent         (authorizeIntent → CLOB adapter)
    │
    ▼
ledger updated + decision row written (placed | skipped | error)

Invariant order matters. Flipping INSERT and PLACE means a successful CLOB submit whose ledger row never committed → next poll double-mirrors. Never reorder.

Durable flowchart: Poly Order And Position Lifecycle § Mirror Decision To Limit Order.

Key code landmarks

• nodes/poly/app/src/features/copy-trade/mirror-pipeline.ts — runMirrorTick glue
• nodes/poly/app/src/features/copy-trade/plan-mirror.ts — pure mirror decision policy
• nodes/poly/app/src/features/wallet-watch/ — wallet observation source, no decisions
• nodes/poly/app/src/bootstrap/jobs/copy-trade-mirror.job.ts — poll shim + target-config hydration
• nodes/poly/app/src/features/copy-trade/target-source.ts — dbTargetSource + envTargetSource; target policy fields
• nodes/poly/packages/db-schema/src/copy-trade.ts — schema for poly_copy_trade_* tables
• nodes/poly/app/src/features/trading/order-ledger.ts — ledger writes + sync-truth reads

Tables (RLS-scoped, per-tenant)

Table	Purpose
poly_copy_trade_targets	Per-tenant tracked target wallets plus target policy fields. RLS via created_by_user_id.
poly_copy_trade_fills	The ledger. Row per decision. idempotency_key = keccak256(target_id + ':' + fill_id).
poly_copy_trade_decisions	Append-only audit log of every decide() outcome. Never updated or deleted.

dbTargetSource.listForActor(actorId) — RLS-clamped via appDb, used by per-user routes. dbTargetSource.listAllActive() — under serviceDb, the ONE sanctioned BYPASSRLS read across tenants. Only called by the mirror-poll enumerator in container.ts. If you find a second caller, that is a bug.

Runtime config — candidate-a

Enable switch: there is no per-tenant kill-switch table. An active target row plus active wallet connection plus active grant is the opt-in gate. Remove or disable the target row to stop mirroring that wallet.

Poll cadence: 30s. Warmup backlog: 60s. Hardcoded in copy-trade-mirror.job.ts. Bounds our latency floor at ~30-60s (task.0322 P4 swaps to CLOB WebSocket).

Target policy: poly_copy_trade_targets carries mirror_filter_percentile and mirror_max_usdc_per_trade. The enumerator threads these through buildMirrorTargetConfig into MirrorTargetConfig.sizing.

Grant caps: poly_wallet_grants carries per_order_usdc_cap, daily_usdc_cap, and hourly_fills_cap. These are enforced by authorizeIntent, downstream of planMirrorFromFill.

Tracked wallets — add via dashboard + or POST /api/v1/poly/copy-trade/targets; remove via − or DELETE /api/v1/poly/copy-trade/targets/[id]. Never seed wallets via env vars (the Phase-A envTargetSource exists for local-dev only).

Observability — mirror signals

Signal	Where	Good state
poly.mirror.poll.singleton_claim	Loki	Fires exactly once per pod start
poly.wallet_watch.fetch	Loki	raw=N, fills=N, phase=ok when a drain runs
poly.wallet_watch.ws.wakeup_total	Metrics	Increments when watched assets trade
poly.mirror.decision outcome=placed	Loki	Emitted when mirror fires
poly.mirror.decision outcome=skipped reason=already_placed	Loki	Dedup. Noisy — reducing noise is in task.0323 §1.
poly.mirror.decision outcome=skipped reason=position_cap_reached	Loki	Hitting target-policy cap. Expected under load; investigate if spiking.
poly.mirror.poll.tick_error	Loki	ZERO. Any hit = bug.
poly_copy_trade_fills	poly DB	Row per mirror decision

Status-sync gotcha (task.0323 §2 / task.0328): poly_copy_trade_fills.status=open is written at INSERT time and never re-read from CLOB. Actual CLOB state may be filled, canceled, or partial. Don't trust the ledger's status column alone — cross-check Data-API /positions?user=<addr> or the synced_at staleness window exposed via /api/v1/poly/sync-health.

MCP-down fallback: scripts/loki-query.sh '{env="candidate-a",service="app",pod=~"poly-node-app-.*"} | json | event=~"poly.mirror.*"' 30 — see top-level poly-dev-manager skill for details.

Idempotency + fill_id — frozen

• fill_id = data-api:<tx>:<asset>:<side>:<ts> — assembled in polymarket-source.ts. Shape is frozen. Phase 4 adds clob-ws:<…> as a sibling scheme; never mix schemes within one fill.
• idempotency_key = keccak256(target_id + ':' + fill_id) — written to ledger and submitted as client_order_id to CLOB. Invariant: two poll ticks seeing the same on-chain fill produce the same key → CLOB dedups → at-most-once.
• The target_id in the key is the UUIDv5 derived from target_wallet, not the DB row PK (UUIDv4). This was the source of revision-1 bug #2 on task.0318 Phase A — HTTP routes had a drift between target_id (UUIDv5 for ledger correlation) and params.id (UUIDv4 DB row PK for DELETE). Same term, different space.

v1 + Phase 4 roadmap pointers

• task.0323 v1 hardening — cursor persistence, CTF SELL (wallet-owned, overlaps with poly-auth-wallets), status-sync, metrics, alerting. In Review at time of writing.
• task.0328 sync-truth cache — DB-as-CLOB-cache with synced_at + /sync-health. Done.
• task.0332 shared poller — replace per-wallet setInterval with one poll loop + TargetSubscriptionRouter. Blocks Phase 3 multi-tenant scale. Needs Design.
• task.0322 Phase 4 — CLOB WebSocket dual-path ingestion + hot signer + target ranker + counterfactual PnL. Needs Design.

Active bugs that affect this pipeline

• bug.0335 — shared operator BUY empty reject on candidate-a. Every autonomous mirror attempt rejected with empty CLOB response. Likely operator-wallet balance / allowance / keys — not a code bug in this pipeline. Ground-truth in poly-auth-wallets. Until resolved, candidate-a mirror attempts always fail at the CLOB submit step.
• bug.0329 — SELL on neg_risk market empty reject. Any position opened on neg_risk becomes roach-motel until resolution. Blocks close-position. Root cause: missing CTF approval (see poly-auth-wallets).

Anti-patterns specific to the mirror

• "I placed a trade from my own per-tenant wallet and the mirror should see it." No — the mirror watches the TARGET wallet, not your trading wallet. Placing from yourself validates CLOB + signing, not the mirror.
• Adding a second BYPASSRLS read somewhere other than the container.ts enumerator. Cross-tenant reads are a cross-tenant isolation hole; the enumerator is the audited seam.
• Generalizing scripts/experiments/place-polymarket-order.ts (scope-narrow $1 post-only dress-rehearsal) instead of using privy-polymarket-order.ts. The dress-rehearsal has deliberate guardrails.
• Smuggling P4 scope (WebSocket, ranking, streaming) into v0 or v1. Keep the fill_id shape + idempotency formula fixed.
• Silencing outcome=skipped reason=already_placed at the coordinator layer. It's noise at the Loki level but correctness-preserving at the code level. Fix per task.0323 §1 (batch / sample), not by removing the skip branch.
• Changing v0 caps via kubectl set env. Argo reverts on next sync. Caps are in code. task.0347 is the work item to lift them to tenant-config.

Enforcement rules

• INSERT_BEFORE_PLACE is the correctness gate — never reorder.
• fill_id and idempotency formulas are frozen.
• Per-tenant RLS on poly_copy_trade_* tables must stay intact. Any new table joins through billing_accounts.owner_user_id EXISTS-pattern (same shape as llm_charge_details), not direct created_by_user_id coupling.
• Caps are defense-in-depth. Even when task.0347 lifts them to per-tenant config, the code must enforce whichever is smaller: configured cap or hardcoded safety cap.