CRM Cleanup

Keep pipeline data from rotting after calls without letting an agent write to the CRM on vibes. The transcript and the current records are the only evidence; the reconciling agent proposes updates, and deterministic code enforces the update authority and the evidence trail. The skill never writes; the sealed proposal is what a gated CRM operator or human approver consumes.

Procedure

1. Native data.digest binds the exact transcript and record set.
2. The reconciling agent proposes field updates from the transcript, each with the target record, field, new value, and a supporting quote.
3. Deterministic enforcement checks every proposal: the record must exist in the supplied set, the field must be inside crm_schema.allowed_fields (out-of-allowlist updates are rejected with a named reason, not silently dropped), the quote must appear verbatim in the transcript, and the value must be non-empty. An unknown record or an invented quote refuses the whole run; nothing partial escapes.
4. The proposal packet carries each update's from value from the supplied records so the downstream write can detect drift before applying.

write_performed is always false and the packet names its gate (crm-operator-or-human-approver). A run with no supported updates seals no_action rather than inventing work.

Output

crm_update_proposal (runx.crm_update_proposal.v1) carries decision (proposed, no_action, refused), updates with before and after values and evidence quotes, rejected_updates, validation, and both input digests.

Inputs are transcript, crm_records, and crm_schema.

Agent task contracts

crm-cleanup-reconcile

Read transcript, crm_records, and crm_schema from step inputs. Return update_draft with updates: an array of record_id, field, to, and evidence_quote entries. Only propose updates the transcript actually supports, quote the transcript verbatim, and only target fields inside the allowlist. Return an empty updates array when the call changes nothing. Never invent records, quotes, or values.