bot-review

name	bot-review
description	End-to-end audit of the Cerebro bot: services, account/positions, LLM agents, ingest/news, data, postgres, redis, risk, execution, strategies, TUI. Usage: /bot-review [scope]

Bot Review Skill

Comprehensive, top-to-bottom audit of the Cerebro trading bot. Launches the senior-go-engineer agent to review each subsystem for architectural compliance (hexagonal), safety invariants (paper-first, kill-switch, risk gate), concurrency, decimal handling, error/context propagation, and test coverage — then verifies compilation, lint and tests.

Scope Map

The bot is organised as hexagonal ports-and-adapters. Each subsystem below is reviewed as an independent slice, then cross-cutting concerns are audited.

1. Composition root & lifecycle — internal/app/

• runtime.go, app.go, lifecycle.go, brokers.go, ingest.go, strategies.go, memory.go, source.go, universe.go, quote_fallback.go
• Verify: ports wired to correct adapters; errgroup supervision; graceful shutdown via context; no adapter types leaking into non-app packages.

2. Domain & ports — internal/domain/, internal/port/

• Ports: broker.go, derivatives.go, ingest.go, llm.go, notifier.go, storage.go, strategy.go, cache.go, symbol_source.go, universe.go
• Verify: domain has zero outward imports; every adapter implements a port; money/prices use shopspring/decimal (never float64).

3. Exchange adapter (account / positions / orders) — internal/adapter/binance/

• client.go, rate_limiter.go, spot/, futures/
• Verify: account + position + order endpoints use decimal; rate-limiter covers all REST/WebSocket paths; error mapping into domain errors; reconnection & heartbeat for user-data streams.

4. Execution & risk — internal/execution/, internal/risk/

• Execution: router.go, worker.go, monitor.go, idempotency.go, paper/
• Risk: gate.go, sizing.go, calendar.go
• Verify: risk gate is never bypassed; paper mode is default; kill-switch halts immediately; idempotency keys are stable; order state machine safe under retries.

5. Strategies & market data — internal/strategy/, internal/marketdata/

• Strategies: registry.go, dedup.go, mean_reversion.go, trend_following.go, volatility_breakout.go, indicators/
• Market data: hub.go, candle.go, replay.go
• Verify: signals are deterministic under replay; dedup windows sized per strategy; indicators use decimal arithmetic; backtest parity.

6. LLM agents — internal/agent/, internal/adapter/llm/

• Agents: runtime.go, copilot.go, discovery.go, screening.go, reviewer.go, risk_agent.go, pricing.go, cost.go, usage.go, tools/, prompts/
• Providers: openai.go, anthropic.go, gemini.go, fallback.go, circuit.go, errors.go
• Verify: tool allow-lists per agent; cost/usage accounting; circuit breaker
  • fallback chain; prompts versioned; no secrets in logs; Anthropic prompt cache hit-rate sane.

7. Ingest & news — internal/ingest/

• news/cryptopanic/ (RE + headless fallback + circuit — see AGENTS.md), calendar/, coinglass/, scrape/
• Verify: CryptoPanic AES key/bundle hash current; fallback cooldown wired; Redis cache keys (news:latest, news:by_asset:<CODE>) + TTLs; agent fetch_latest_news reads cache first.

8. Persistence — internal/adapter/postgres/, migrations

• pool.go, trade_store.go, agent_log_store.go, audit_store.go, log_archiver.go, scripts/migrations/
• Verify: NUMERIC for money (never FLOAT); named constraints; up/down migrations both present and idempotent; pgx pool sized & bounded; no cross-package SQL leakage.

9. Cache — internal/adapter/redis/, internal/adapter/redisx/

• Verify: key naming is namespaced; TTLs everywhere; context cancellation honoured; no unbounded key growth; cache-miss paths degrade safely.

10. ChatOps, TUI, CLI — internal/chatops/, internal/tui/, internal/cli/

• Verify: commands gated by mode (paper vs live); TUI does not call adapters directly (goes through app layer); bots handle retry + rate limits.

11. Cross-cutting — internal/observability/, internal/resilience/, internal/watchdog/, internal/config/, internal/backtest/

• Verify: structured logging via slog; no log.Printf or fmt.Println in hot paths; config load + validate runs before wiring; watchdog escalates to kill-switch on repeated faults; secrets only read from env, never YAML.

Instructions

When the user invokes /bot-review, follow these steps:

Step 1 — Resolve scope

• No arg or all: review every subsystem in the Scope Map above.
• Named subsystem (e.g. /bot-review llm, /bot-review ingest, /bot-review risk, /bot-review postgres, /bot-review execution): review only that slice plus its direct ports.
• Path (e.g. /bot-review internal/risk/gate.go): review that file/package.
• changes: review uncommitted diff via git status + git diff.

Step 2 — Snapshot the tree

Before spawning the agent, gather a fresh context snapshot:

• git status and git log -5 --oneline
• ls internal/ and the targeted subsystem directory
• Read CLAUDE.md, AGENTS.md, and relevant .claude/rules/*.md

Step 3 — Launch senior-go-engineer

Give it a prompt containing:

• The exact files/packages in scope (resolved from Step 1 + Step 2)
• The Scope Map checklist above as review criteria
• Safety invariants to enforce:
  • paper-first + triple-agreement for live
  • kill-switch halts execution
  • risk gate never bypassed
  • decimal everywhere for money
  • hexagonal boundaries intact
• Request findings grouped by severity (critical / important / minor) with file:line refs and a one-line remediation hint per finding.

Step 4 — Compile & verify

After the agent returns, run the project's verification pipeline in parallel (these come from AGENTS.md):

make build      # compile binary
make lint       # golangci-lint
make test       # unit tests
make check      # dry-run config validation

Optional, when the change set touches those areas:

make test-int                                   # integration (needs DB)
go test -tags=live ./internal/ingest/news/cryptopanic -run TestLive -v

Report any build/lint/test failures as critical findings.

Step 5 — Present report

Summarise to the user as:

1. One-line status per subsystem (✅ / ⚠ / ❌)
2. Critical findings (must-fix before live)
3. Important findings (pre-merge)
4. Minor findings / nits
5. Verification results (build, lint, test, check)
6. Suggested next actions — typically /fix critical or /test <pkg>.

Examples

• /bot-review — full bot audit across every subsystem
• /bot-review llm — agents + LLM adapters only
• /bot-review ingest — news, calendar, coinglass, scrape
• /bot-review risk — risk gate, sizing, calendar
• /bot-review execution — router, worker, monitor, paper, idempotency
• /bot-review postgres — pool, stores, migrations
• /bot-review redis — cache adapter + key usage across the codebase
• /bot-review account — binance client + account/position paths
• /bot-review changes — only uncommitted diff
• /bot-review internal/agent/screening.go — single file

Notes

• This skill is read-only. It never applies fixes. Use /fix to apply remediations to findings, and /test <pkg> to focus test runs.
• Keep the agent prompt concise: point to the Scope Map section by name rather than pasting it inline.
• If make build fails, stop and surface the compiler error first — further review is noise until the tree compiles.