Proactive and schedule-driven; each run is stateless — there is no local ledger file. The one thing that spans runs is a pending bulk-update request: its state lives in {{alert_channel}} itself (the posted draft plus any approval reply), not in the sandbox.

Step 0 — Confirm access

No local state to resume — each run starts with a clean sandbox. The one exception lives in Slack, not the sandbox: check {{alert_channel}} for an unresolved "Pending bulk update" post from a previous run before assuming this run's batches are new (see Step 5). Confirm the HubSpot and Clearbit connectors are live before touching any records:

text

GET /crm/v3/objects/contacts?limit=1
GET /crm/v3/objects/deals?limit=1

If either connector fails, stop and report the failure in the summary instead of running a partial sweep.

Step 1 — Pull the working set

Pull contacts and open deals touched or created in the last 24h, plus any contact record with a null in a required field:

text

GET /crm/v3/objects/contacts/search
  filterGroups: [{ propertyName: "lastmodifieddate", operator: "GTE", value: <24h ago> }]
GET /crm/v3/objects/contacts/search
  filterGroups: [{ propertyName: "jobtitle", operator: "NOT_HAS_PROPERTY" }, ...]
GET /crm/v3/objects/deals/search
  filterGroups: [{ propertyName: "dealstage", operator: "NEQ", value: "closedlost|closedwon" }]

Step 2 — Find and merge duplicate contacts

Two contacts are the same record when they match on any of:

Same primary email address.
Same first + last name AND same company domain.

A duplicate is often an older, untouched record that falls outside the Step 1 working set (not modified in the last 24h, no missing field). For each contact in the working set, run a targeted HubSpot search against the full contact base — by primary email, and separately by (firstname + lastname + company domain) — to surface any existing match outside the 24h window, not just duplicates among contacts already pulled in Step 1.

For each confirmed pair:

Pick the primary: the record with the older createdate, or the one with more associated deals/tickets if createdate ties.
Merge via HubSpot's contact merge endpoint so all associations (deals, tickets, notes, timeline) roll onto the primary.
If the two records disagree on a field HubSpot can't auto-resolve (e.g. conflicting job titles), keep the primary's value and note the conflict in the summary rather than guessing.

Step 3 — Fill missing fields from enrichment

Required fields: jobtitle, industry, numemployees (company size). For each contact missing one or more:

Query Clearbit by email or company domain.
If Clearbit returns a value, write it to the missing field.
Batch the writes for this pass. If the batch for a single field would touch more than {{bulk_update_threshold}} records, do not write it — go to Step 5 instead. Fills below the threshold write immediately, per-record.

Step 4 — Flag stale deals

A deal is stale when it hasn't moved stage and has had no logged activity (email, call, note, meeting) in 21+ days. For each match:

Set a hygiene_stale custom property (or tag) to true with the last activity date.
Never change dealstage, never close the deal — flagging only.

Step 5 — Human approval gate for bulk updates

First, resolve any pending request from a previous run: search {{alert_channel}} for the agent's own most recent "Pending bulk update: " post.

No pending post found, or the last one already carries this agent's own "Applied"/"Resolved" reply → this field has nothing outstanding; if Step 3 produced a new batch over {{bulk_update_threshold}} records for it, go to "New batch" below.
A pending post exists with no resolution reply yet → check whether an authorized approver (not the agent) has replied "approve"/"approved" in that thread:
- Approved — re-verify each record in the posted sample still needs the fill (skip any that changed since the post), write the batch for that field, reply in-thread confirming exactly what was applied, then continue to Step 6. Do not also start a "New batch" for the same field this run.
- No approval reply yet — do not write, and do not post a duplicate request for the same field. Note it as still pending in this run's summary (Step 6) and move on.

New batch — when Step 3 produces a batch over {{bulk_update_threshold}} records for a field with no unresolved pending post:

Do not write any record in that batch.
Draft an approval request: field name, record count, and a sample of 5–10 affected records with old → new values.
Post it to {{alert_channel}} as "Pending bulk update: " and stop — do not write, do not wait inside this run. The batch is applied on a later run, once that run's Step 5 finds an explicit approval reply in this thread; never retry-write without one.

Step 6 — Post the summary

Post one summary to {{alert_channel}} covering:

Duplicate contacts merged (count + any unresolved conflicts).
Fields filled from enrichment (count per field).
Deals flagged stale (count, with links).
Any bulk update pending approval, what it is, and why it's waiting.

HubSpot hygiene