Doordash Spend Guard
npx claude-code-templates@latest --skill doordash/doordash-spend-guard Content
DoorDash Spend Guard
A budget stated in a prompt is advisory — the model can forget it mid-session
and prompt injection can override it. This skill makes the budget
deterministic: every checkout and cart mutation goes through one audited
wrapper (dd-guard.sh) that prices the cart, checks the policy, and refuses
out-of-policy actions with a machine exit code. The companion PreToolUse hook
(doordash-spend-guard hook component) denies raw dd-cli order checkout-url /
dd-cli cart add-items calls that try to bypass the wrapper.
Unofficial community skill built on DoorDash's
doordash-oss/doordash-cli. Requiresdd-cliinstalled and logged in. Wrapper needspython3.
Honest scope
- Caps apply to the pre-fee subtotal that
cart showexposes; fees, tip and tax land on the DoorDash payment page. Documented headroom, not loopholes. - The ledger records intent when a checkout URL is issued; the human may
abandon the page.
/doordash-budgetreconciles intents againstdd-cli order history, marking entriespaidorabandoned. - Enforcement teeth come from the hook + wrapper combo. Install both; the skill alone is guidance.
Components in this bundle
| Piece | Role |
|---|---|
| this skill | protocol: always route through dd-guard, report headroom |
scripts/dd-guard.sh → installs to .claude/skills/doordash-spend-guard/scripts/dd-guard.sh |
the deterministic gate |
hook doordash/doordash-spend-guard |
denies bypasses of the wrapper |
command /doordash-budget |
view/edit policy, reconcile ledger |
State (created on first run)
~/.claude/dd-guard/limits.json— policy. Human-edited only. Never edit this file yourself; when the user asks for a change, direct them to/doordash-budgetwhich confirms interactively.
{
"per_order_max": 40,
"daily_max": 60,
"weekly_max": 150,
"monthly_max": 400,
"cooldown_minutes": 45,
"allowed_hours": { "start": 7, "end": 23 }
}~/.claude/dd-guard/ledger.jsonl— one line per priced action:{"ts": "...", "cart_uuid": "...", "subtotal": 24.5, "status": "intent"}(statusbecomespaid/abandonedafter reconciliation).
Protocol (follow ALWAYS while this skill is installed)
- Session start / before building a cart: report headroom —
bash .claude/skills/doordash-spend-guard/scripts/dd-guard.sh statusprints spend-to-date vs each cap. Tell the user their remaining budget before adding items. - Adding items: route through the wrapper —
bash .claude/skills/doordash-spend-guard/scripts/dd-guard.sh add-items <the exact dd-cli cart add-items args>The wrapper execs the realdd-cli cart add-items ...and re-prices the cart afterwards. - Checkout: NEVER call
dd-cli order checkout-urldirectly. Usebash .claude/skills/doordash-spend-guard/scripts/dd-guard.sh checkout <cart-uuid>- exit 0 → it printed the checkout URL and appended a ledger intent line.
- exit 2 → blocked; it printed a human-readable reason (which cap, by how
much). Relay the reason verbatim, then help within policy — e.g. suggest
a cheaper reorder from
dd-cli order historyor trimming the cart withdd-cli cart remove-item.
- When blocked, do not negotiate with the wrapper. No editing
limits.json, no retrying with a fresh cart to reset the cooldown, no
splitting one order into several to duck the per-order cap. If the user
wants a different policy, point them to
/doordash-budget. - If the wrapper reports the cart subtotal as unparseable, it blocks by
design (conservative default). Show the user the raw
cart showoutput and ask them to confirm the total explicitly before any manual override.
Reading the ledger
Answer "how much have I spent this week?" from the ledger, not memory:
bash .claude/skills/doordash-spend-guard/scripts/dd-guard.sh status
# or raw:
cat ~/.claude/dd-guard/ledger.jsonl | jq -s '[.[] | select(.status != "abandoned")] | map(.subtotal) | add'Remind the user that subtotals exclude fees/tips and intents may not have
been paid — /doordash-budget reconciles.