Expansion signals

The outcome

The renewal stops being the first time anyone looks at a customer. Every Monday:

1. A play picks up each won deal entering its renewal window from a native gtm_opportunities model: the deal is won, has an account, and closed ten to twelve months ago. A deal enters that window once, so each renewal is judged once.
2. An agent judges the account. In SQL over gtm_opportunities it reads the account's purchase history: whether a newer win already renewed it, the cadence across its wins, the last price paid. It reads your expansion plays from the workspace context and runs up to three web searches for dated public events. It answers renewal, expansion, repeat_purchase, at_risk or none, with a reason of at most three sentences, the play to run, and every URL it relied on.
3. The play writes the judgment onto the account record by its id: expansion_signal, expansion_reason and a expansion_signal_at stamp. It writes nothing else: no deal, no contact, no owner.
4. One digest lands in Slack that afternoon: at-risk accounts first, then expansion, repeat purchase and plain renewal, each with the reason as written on the account and its owner. A ledger keeps it to one post a week.

The worked example runs on Cargo's native accounts and deals models, so it deploys with no CRM connector. Customers who live in HubSpot, Salesforce or Attio take the crm-backed variation: the same play, with connector-backed models and the write going back to the CRM record.

Example

Every Monday, flag which customers near renewal are ready to expand and post the digest to #cs-weekly.

Illustrative output, fictional records:

:seedling: *Expansion signals, week of 2026-10-05*
_6 accounts in their renewal window: 1 at risk, 2 ready to expand._

*Tailspin Logistics* (tailspin.example) · at_risk · New CDO joined Sep 22 from a
  competitor's customer; renewed every 12 months, last at 24,000. Play: executive
  check-in. Sources: tailspin.example/news/new-cdo · owner 4411
*Fabrikam* (fabrikam.example) · expansion · Raised a Series B on Sep 9 and posted
  four data-engineer roles; renewed every 12 months, last at 38,000. Play: annual
  prepay offer. Sources: fabrikam.example/press/series-b · owner 4402
*Northwind Traders* (northwind.example) · expansion · Hired a Head of Compliance on
  Sep 30. Play: governance add-on demo. Sources: northwind.example/team · owner 4402
*Contoso* (contoso.example) · repeat_purchase · Ordered every 6 months, last at
  9,500 on Apr 2. Play: renewal call at the last price. · owner 4419
*Litware* (litware.example) · renewal · No outside event. · owner 4411

Six won deals entered their renewal window this week, so six accounts carry a signal and a reason; a seventh account had already renewed early and was written none.

Put it in your project

This folder is a worked example: real project resources written for some other company. The job is to end up with the code your company would have written, in your project, and an agent does the adapting. If the cargo-project skill is in your session it carries the long form of this; if not, this is enough.

1. Install it — the CLI does the copy. From inside the project (the folder holding cargo.json), cargo-ai project add cookbook/expansion-signals writes this example to infra/expansion-signals/ and this procedure to .agents/skills/expansion-signals/. No project yet? cargo-ai project init <dir> --cookbook expansion-signals && cd <dir> && npm install does both. If you are reading this from the project's .claude/skills/, the install already happened — start at step 2.
2. Reconcile it with what is already declared. If the project already has a Slack or Anthropic connector, or a gtm_accounts or gtm_opportunities model, rewire the imports to the existing one and drop the copy; two resources with one slug is a collision at deploy. If the customer book lives in a CRM, take the crm-backed variation now, before the first deploy.
3. Fill the models. A native model starts empty. Load accounts, then won deals with their account_id, with cargo-ai storage record create-bulk --model-uuid <uuid> --records '[{"data":{…}}]', or point a sync at them. A won deal with no account is never judged.
4. Write your expansion plays into the workspace context. references/expansion-plays.md is the example; copy it to context/expansion-plays.md and rewrite it for what you sell.
5. Adapt. Work the sections below in order: What should not change is what you argue back about; What you can change is what you offer unprompted; What you will be asked is the floor, and you derive before you ask. Record what you changed and why under ## Decisions in your copy of this file.
6. Check, then plan. node --import tsx evals/contract.mjs from this skill's folder, then npm run check && cargo-ai project plan from the project root. Show the diff, and deploy only on an explicit yes. The play ships disabled.
7. Pilot. Run the play by hand on five deals from the window, read each written reason against its sources, then enable the play, execute it once (an added play does not backfill), and let the digest post the next Monday.

What you will be asked

Derive before you ask. An input with a lookup is looked up, not asked.

What you can change

What should not change

• Customer status is computed from won deals, not kept as a flag. (infra/agents/analyst.ts) A hand-kept "customer" column is a second answer to the question the deals already answer, and it drifts: an account that renewed early would be judged as if it had not.
• The play writes the account by its id, and only the three expansion columns. (infra/plays/flag-expansion.ts) A write to a deal or owner turns a judgment into a forecast change nobody approved. In the crm-backed shape, a unify step between the play and the write reports success while nothing lands. The contract fails on either.
• Bare column slugs on the write. The read side calls them custom__expansion_signal; the write takes expansion_signal. A prefixed slug is silently dropped while the node still reports success.
• The analyst reads, never writes. (infra/agents/analyst.ts) An agent that can write decides its own routing, and a missing signal could then be a failed run, a skip or a choice.
• changeKinds: ["added"] stays. It is what judges each renewal once. Without it every won deal in the window is re-judged every Monday for two months, and each run overwrites the last reason.
• The digest re-judges nothing. (infra/agents/digest.ts) It reads the reason the account carries. A second judgment in Slack is how the record and the channel disagree about one account.
• The ledger, not Slack history, is the digest's dedupe. A deleted post or another bot's message would otherwise read as "already sent" and drop a week in silence.

Done when

• node --import tsx evals/contract.mjs passes
• cargo-ai project plan reports the gtm_accounts, gtm_opportunities and expansion_digests models, the two agents, the play, two connectors and three folders, with the play disabled and no CRM connector
• seeded accounts and won deals, with two deals inside the window, produced two analyst runs and two accounts carrying a signal, a reason and a stamp; every event named in a reason has a dated source in it
• an account whose newer win already renewed it (closed more than two months after the trigger) was written none with that reason, and an account with a second win within two months of the trigger was judged on its whole purchase history instead
• a second play execution the same week judged neither deal again
• the digest posted once in the locked channel, listed exactly the accounts stamped that week with a signal other than none, and a re-run that week posted nothing
• no deal or contact record changed

What it costs

The model writes are native and the research is web search inside the analyst, at most three per account. Immediately before the plan, read the live price of the one connector action:

• cargo-ai orchestration action list postMessage --kind connector --integration-slug slack

Say it out loud. The recurring cost is one analyst run per won deal entering its window each week, plus one digest run, billed as LLM tokens through the Anthropic connector. It scales with how many customers are near renewal, not with the size of the book. Count the window before enabling: a SELECT count(*) over gtm_opportunities with is_won true and close_date ten to twelve months back.

Composes into

crm-enrichment (firmographics on the same company records), win-loss-review (what closed-won says about who expands), and track-job-changes (a champion who left is the first at-risk signal).