strategy-review

Strategy Review Workflow

When to invoke

• User pastes or builds a StrategySpec payload and asks for a review, critique, or sanity-check.
• User asks "will this strategy work?", "is this a reasonable setup?", or similar qualitative questions about a strategy — before it has been backtested in this session.
• User references a strategy they previously saved (by name or UUID) and wants to revisit it.
• User wants to compare two or more candidate strategies against each other (invokes compare_strategies rather than backtest_strategy alone).

Do NOT invoke for:

• Pure market-data lookups (use /cryptozavr:research).
• Live-trading or order-placement questions — those don't exist in Phase 2.
• Re-runs of a backtest with different OHLCV without spec changes — just call backtest_strategy directly.

The review loop

1. Parse intent. Is the user comparing, tuning, stress-testing, or just asking "is this well-formed?"
2. Validate first (cheap). Call validate_strategy with the raw payload. If valid=False, stop — report issues with location path + message + type, suggest fixes, wait for a new payload.
3. Explain (human-readable). Call explain_strategy so the user sees their spec in plain text (entry/exit/risk sections). Use this to confirm you both understand the same strategy before spending compute on a backtest.
4. Backtest (flagship). Call backtest_strategy with the validated payload. Set initial_equity to a round number like 10000 unless the user specifies. Default limit=500; bump to limit=1500 for longer lookbacks if the user cares about multi-month behaviour.
5. Inspect metrics. Pull metrics.total_return, win_rate, max_drawdown, profit_factor, sharpe_ratio out of the response. Flag any unusual combinations (e.g. high win_rate + negative total_return → fees/slippage eating the edge).
6. Stress-test (optional). If the backtest looks good, call stress_test with the default bull/bear/chop scenarios to see whether the edge survives different regimes. If one regime catastrophically loses, call that out — it's a robustness signal.
7. Compare (optional). If the user supplied multiple candidates, call compare_strategies([...specs]) and highlight the per-metric winner using the pivoted comparison dict.
8. Save (opt-in). Ask before saving. If the user wants to keep a spec around, call save_strategy and record the returned UUID so the user can reference it later.
9. Advise — carefully. Translate numbers into observations, not orders. "This draws down 28% — that's deep enough that most humans would abandon the system mid-drawdown." Not "you should raise your stop-loss."

Tool selection matrix

Question	Tool
"Is my spec valid?"	validate_strategy
"Explain what this strategy does"	explain_strategy
"How well did it perform?"	backtest_strategy
"Does it hold up in a downturn?"	stress_test(scenarios=["bear"]) or default 3 scenarios
"Which of these three is best?"	compare_strategies
"Save this one for later"	save_strategy (returns UUID)
"Show me what I've saved"	list_strategies(limit=50)
"What's different between v1 and v2?"	diff_strategies(a, b)

All eight tools are strategy-tagged and phase-2. They are read-only (no balances moved, no orders placed) and idempotent with respect to each other — except save_strategy which persists a row (upsert-by-hash).

Rails

• No live-trading advice. Phase 2 has no execution surface. Never tell the user to enter a position.
• Don't hide poor metrics. If total_return < 0 or max_drawdown > 50%, say so first, not last.
• Don't over-interpret a single backtest. Backtests reflect the OHLCV window they ran on. Always note the window (inferred from report.period_start_ms / period_end_ms).
• Placeholder embedding is semantically meaningless. If the user asks for similarity search ("find strategies like this one"), clarify that the current embedding is a BLAKE2b placeholder — real similarity lands when the embedding provider is wired.
• Fees and slippage are opinionated defaults. BacktestEngine uses 5 bps fees + 10 bps slippage if unset. Tell the user when metrics are sensitive to these — they usually are at high turnover.

Subagent delegation

For non-trivial reviews that bounce through ≥3 tools + interpretation (e.g. "tune my EMA crossover for BTC 1h, try three period pairs, compare, explain"), dispatch the crypto-researcher subagent in a fresh context. Pass it the validated payload, the metric targets, and the rail reminders above. It keeps the main thread clean for the user's follow-up questions.

Output shape

Present findings in this order:

1. Validation — one line: ✅ Spec valid or ❌ Spec invalid: <N> issues.
2. Explanation — lift the markdown field from explain_strategy verbatim (already formatted).
3. Backtest window — venue, symbol, timeframe, range (ISO timestamps from period_start_ms).
4. Headline metrics — total_return, win_rate, max_drawdown, profit_factor, sharpe_ratio as a table.
5. Regime robustness — one bullet per scenario if stress_test was run.
6. Comparison — if compare_strategies was run, one row per strategy × metric.
7. Observations — two or three bullets, qualitative, no action items.
8. Next steps — what the user could try (tune X, add Y filter), not what they should do.