Dca Investor

DCA Investor

A boring strategist. Each Monday morning, it buys a fixed dollar amount split across a fixed basket. It does not try to time entries. It does not skip weeks because "the market looks toppy." It does not double up after a drawdown. The only signals it consults are the calendar and the buying-power floor.

This is the simplest possible strategist — and on purpose. It's the proof-of-pattern for the suite, and it's the most realistic live candidate for a $1k cash account.

Required reading before authoring/modifying: _shared/intent-schema.md, _shared/tick-decision-emitter.md, _shared/circuit-breakers.md, _shared/position-sizing.md. The numeric rules below live in references/decision-rules.md.

Strategy summary

• Universe: SPY, QQQ, BND (broad equity + tech-equity + total-bond).
• Weights: SPY 50% · QQQ 30% · BND 20%.
• Tranche: $50/week total. Per-symbol: SPY $25, QQQ $15, BND $10.
• Cadence: weekly, Monday after market open. Cron hint 0 14 * * 1 (10am ET).
• Mode: review-only until the human reviews tick logs and promotes to live.
• Exit rule: none. DCA is one-way. Sells are out of scope for this strategist.
• Director call: regime-intelligence is called only in light-gate mode — its output is logged in notes but does not gate any action. A future v2 may add a regime gate; v1 is intentionally pure.

6-phase protocol

Phase 1 — Pre-flight (BLOCKING)

Run _shared/executor's pre-flight via the executor — but capture inputs first so this strategist can short-circuit before invoking it. Specifically:

1. Determine cadence eligibility:
   • Read the last successful DCA tick timestamp from ~/.hermes/profiles/autotrader/state/dca-investor.json (if it exists).
   • If less than 6 days have elapsed since the last placement-eligible tick: skip without invoking the executor. Emit tick_decision with actions: [] and notes: "cadence_not_due; next_at <ISO>".
2. If cadence IS due, proceed to phase 2.

(The executor will still run its full pre-flight when invoked in phase 5. The cadence check is strategist-side because the executor doesn't know about per-strategist cadences.)

Phase 2 — Compose context

Invoke regime-intelligence via the Skill tool with the prompt:

"Three-line summary of the current macro regime — growth direction, inflation direction, monetary stance. No prose."

Capture the response into a regime_summary string (≤200 chars). It will be embedded in tick_decision.notes. It does NOT affect any sizing or skip decision in v1.

Phase 3 — Signal gather

Call get_equity_quotes(symbols=["SPY", "QQQ", "BND"]) to capture current prices. Used purely for intent_price (the strategist's reference price at decision time). No filtering on this — DCA proceeds at any price.

Phase 4 — Decide

Construct three intents (one per basket member):

[
  {
    "symbol": "SPY",
    "side": "buy",
    "quantity_type": "notional",
    "notional_usd": 25.00,
    "order_type": "market",
    "intent_price": <SPY quote.ask_price>,
    "max_slippage_pct": 1.0,
    "time_in_force": "gfd",
    "reason": "DCA weekly tranche — SPY 50% of $50 = $25",
    "expires_at": "<today 15:55 ET as ISO>",
    "asset_class": "equity",
    "option_leg": null
  },
  /* QQQ $15, BND $10 — same shape */
]

Notes on the intent shape:

• quantity_type: notional because the basket weights are designed in dollars, not shares. The executor routes this to Robinhood's dollar_amount parameter.
• max_slippage_pct: 1.0 is loose on purpose — DCA does not care about precise entry; we just need a sanity check that the quote hasn't fallen off a cliff.
• expires_at is end-of-RTH today. If the tick somehow runs in extended hours, the intent expires before the executor can place.

If the tick lands inside a circuit_breaker_tripped window, the executor will drop these buys with status: skipped. That's fine — DCA simply waits for the next week.

Phase 5 — Hand off to executor

Invoke _shared/executor via the Skill tool with:

{
  "strategy": "dca-investor",
  "mode": "<mode from frontmatter>",
  "context_hint": "<best guess>",
  "intents": [ ...three intents... ]
}

The executor returns:

{
  "aborted": null | "<reason>",
  "circuit_breaker_tripped": <bool>,
  "account_value": <number>,
  "buying_power": <number>,
  "actions": [ <three RealizedAction objects> ]
}

Phase 6 — Emit & persist

Write the ## tick_decision block. If any action was status: placed, update ~/.hermes/profiles/autotrader/state/dca-investor.json with the placement timestamp (so phase 1 can read it next tick).

If aborted is set: do not update state. The next tick will retry from scratch.

Interactive mode

In Claude Code or autotrader chat:

• "What does DCA say about this week?" — run phases 1-5; emit tick_decision. All actions will be reviewed (claude-code never places; chat needs EXECUTE:).
• "Skip this week" — emit tick_decision with actions: [] and notes: "user_skipped". Do NOT update the state file (so next week's tick is unaffected).
• "Increase tranche to $100" — refuse. Tranche size is a frontmatter constant. The user must edit tranche_size_usd in the SKILL.md frontmatter, not at runtime.

Promoting to live

1. Run paper ticks for at least 4 consecutive Mondays. Inspect the tick_decision log for each.
2. Confirm:
   • No aborted outcomes you don't understand.
   • Slippage on realized_review_price vs intent_price is consistently under 0.3%.
   • The basket math comes out right: SPY notional 50%, QQQ 30%, BND 20% of the tranche.
3. Edit this skill's frontmatter: mode: review → mode: live.
4. Create cron: hermes --profile autotrader cron create '0 14 * * 1' --skills strategists/dca-investor.
5. Watch the first live tick. Verify the order lands in get_equity_orders.

What this strategist will never do

• Skip a week because the market is "expensive" or "looks toppy."
• Buy extra after a drawdown.
• Sell — even to "take profits" or "rebalance."
• Trade anything outside the basket.
• Override the buying-power cap or per-name cap.
• Run on a non-Monday.
• Run from a profile other than autotrader.