Adclaw Host AI Accounting

AdClaw Host AI Accounting

Purpose

Use this skill for AdClaw Host zero-setup onboarding work where paid Starter/Pro/Business customers receive included Host AI messages and a hosted Citedy Agent key without entering secrets manually.

The skill focuses on operational truth: quota math, D1 accounting, entitlement gates, prompt/error UX, Citedy key handoff, and redaction. It does not replace cloudflare, workers-best-practices, stripe-best-practices, architecture-patterns, or security-review; use those together when touching their domains.

Trigger Examples

• "implement Host AI quota"
• "audit Workers AI cost model"
• "review included LLM messages per plan"
• "design Citedy auto key provisioning"
• "what happens on the 501st Starter message"
• "check prompt/accounting docs for AdClaw Host AI"
• "make sure LLM/Citedy keys are not leaked"

Required Reading

Before changing code or docs, read the canonical project doc:

/root/adclaw/CF/docs/goal/host-preactivated-llm-citedy-keys.md

When implementation touches Citedy key creation, also read:

/root/saas-blog/docs/sub-project/adclaw/citedy-api-key-handoff.md

When implementation touches AdClaw provider/env storage, inspect current code instead of assuming schema:

/root/AdClaw/src/adclaw/providers/store.py
/root/AdClaw/src/adclaw/envs/store.py
/root/AdClaw/src/adclaw/app/routers/citedy.py

Workflow

1. Establish current facts from repo docs and code.
2. Identify which boundary is being changed: Host Worker, D1 accounting, AdClaw sandbox bootstrap, Citedy key service, UI copy, or docs only.
3. Apply the accounting checks in references/accounting-model.md.
4. Apply the secret and abuse checks in references/security-redaction.md.
5. Apply the UX/prompt checks in references/quota-ux.md.
6. If reviewing implementation, use references/review-checklist.md as the findings checklist.
7. Validate with the smallest safe test. Do not run saas-blog Node workloads on this VPS unless the current user explicitly overrides the host ban.

Core Decisions

• Workers AI access stays in the Host Worker through env.AI.
• The sandbox receives only a customer-scoped synthetic Host AI token.
• Never put Cloudflare account API tokens or Workers AI account credentials in a sandbox, backup, browser response, log, Sentry event, or analytics event.
• Citedy hosted auto-provisioning should create a securely random key for the deterministic adclaw-host agent record.
• Manual Get Citedy key remains fallback and rotation UX, not the normal paid hosted onboarding path.
• Host AI remains the managed default/fallback provider even when a customer adds their own provider key.
• Enforce both message limits and cost caps.
• Block before calling Workers AI when a quota or entitlement gate fails.
• R2 backups are secret-bearing if they contain /workspace/working.secret.

Default Planning Quotas

Use these only as the current planning baseline. Re-check Cloudflare pricing before changing production limits or customer-facing promises.

Starter:  300 included Host AI messages, cost cap $1.00/month
Pro:    1,500 included Host AI messages, cost cap $4.50/month
Business: 5,000 included Host AI messages, cost cap $15.00/month

Initial model:

@cf/google/gemma-4-26b-a4b-it

Pricing baseline:

$0.10 per 1M input tokens
$0.30 per 1M output tokens

Formula:

cost_usd = input_tokens / 1_000_000 * 0.10
         + output_tokens / 1_000_000 * 0.30

Output Standard

When producing an audit or implementation plan, include:

• exact files or APIs affected;
• current fact source;
• quota/accounting impact;
• secret boundary impact;
• entitlement/refund/dispute behavior;
• customer-facing limit/error behavior;
• tests or smoke evidence required.

Avoid vague conclusions like "safe" without stating which boundary was checked.