ChangeFlow — Apify CrewAI engine (private staging)

Status: Staged source, NOT deployed, NOT a paid product, NOT connected to the live app.

This is a compatible reconstruction of the hosted contractor change-order workflow, not an export or verified line-for-line copy of its private CrewAI Studio configuration. The reference endpoint remains https://contractor-change-order-flow-9a19a395-48c2--062940c2.crewai.com.

Why it exists

The existing CrewAI Basic hosting allotment is shared across users and capped at 50 executions/month. This private Python Actor shifts workflow orchestration to Apify, where consumption can scale with usage subject to account credits, billing and quotas. Apify execution, model calls and storage may cost money; this code does not make the workflow free or guarantee profit. The model is accessed through Apify's documented OpenRouter proxy and its injected APIFY_TOKEN; there are no embedded credentials.

Inputs

The six current ChangeFlow fields are customer, job, contract_total, notes, exclusions, tax_rate_percent. The optional quoted_change_amount is a contractor-supplied quote (positive for additions, negative for credits). Prices are strictly validated and calculated with decimal arithmetic, not written by the model. Each model completion is also constrained to 1,400 output tokens and agents to at most two iterations each; an Apify runtime/spend limit must still be set before a paid run.

Output dataset item: { "result": { ...draft... }, "execution": { "engine": "crewai_on_apify", "model": "...", "usage_metrics": ... } }.

The Actor only prepares a draft. It never marks a customer-approved change, issues an invoice, or records a billable change. That stays in the existing application after verifiable customer acceptance. A human must check job-specific scope and legal requirements.

Verification, offline

Install test dependency pydantic>=2.11,<3.0 then run:

python -m unittest discover -s tests -v
python -m compileall -q src

Before enabling production

1. Compare 3–5 identical representative requests against the original hosted workflow, review omissions and hallucinations, and reconcile with the original exported CrewAI configuration when available.
2. Deploy as a private Apify Actor and run one explicitly authorized, capped-cost test, confirming balance, token/model charges, Actor compute and storage costs, response schema, and latency.
3. Calculate actual cost per successful order, including retries, billing fees, support burden and monthly baseline. Set both a safe overall spend ceiling and per-customer limits before enabling Stripe checkout.
4. Add a server-side Apify dispatcher/poller and a controlled feature flag to ChangeFlow. Preserve the currently connected CrewAI API as a fallback; do not expose Apify tokens to browsers.
5. Verify job-selection, voice notes, consent/approval evidence, accepted-change ledger, exports and subscription enforcement end to end; verify user separation and idempotency.

Do not publish an Apify Store listing or set up Stripe subscriptions merely because this source exists. No purchases, Apify plan upgrades or charges have been authorized by this stage.