feat: datetime/date comparisons for row_filters (#87)

.dumplingconf.example

@@ -126,9 +126,19 @@ delete = [
 #   not_like / not_ilike  — negation of like / ilike (prefer over regex look-around)
 #   regex / iregex        — Rust regex crate (not PCRE); look-around / backrefs rejected at load
 #   not_regex / not_iregex — negation of regex / iregex
-#   lt / lte / gt / gte   — numeric compare (values parsed as f64)
+#   lt / lte / gt / gte   — numeric compare (values parsed as f64);
+#                          set format = "datetime" for ISO/Postgres timestamps,
+#                          or format = "date" for calendar dates (YYYY-MM-DD).
+#                          Also valid on eq / neq. Unparseable cells fail closed.
 #   is_null / not_null    — NULL check (no value needed)
 
+# Example: keep ~18 months of orders by timestamp (plus NULL created):
+# [row_filters."public.listing_order"]
+# retain = [
+#   { column = "created", op = "gte", value = "2025-02-14", format = "datetime" },
+#   { column = "created", op = "is_null" },
+# ]
+
 # ---------------------------------------------------------------------------
 # [[column_cases]] — conditional per-column strategy overrides
 #

AGENTS.md

@@ -192,6 +192,8 @@ Two code paths:
 
 JSON path traversal is supported: `payload.profile.tier` (dot) and `payload__profile__tier` (Django double-underscore). Array elements are traversed by evaluating each item's fields, so list-of-dicts structures work naturally.
 
+Comparison operators (`lt` / `lte` / `gt` / `gte`, and optionally `eq` / `neq`) accept `format = "datetime"` or `format = "date"` to compare ISO-8601 / Postgres timestamp text (or calendar dates) instead of numbers. Thresholds are validated at config load; unparseable cell values fail closed (do not match).
+
 ### Residual PII Scanning (`scan.rs`)
 
 `ScanningWriter` wraps the output `Write` stream, intercepting bytes and passing them to `OutputScanner`. The scanner applies regex detectors line-by-line for:

CHANGELOG.md

@@ -7,6 +7,10 @@ and this project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.ht
 
 ## [Unreleased]
 
+### Added
+
+- **Datetime / date predicate comparisons**: `lt` / `lte` / `gt` / `gte` / `eq` / `neq` accept optional `format = "datetime"` (ISO-8601 / Postgres timestamp text as instants) or `format = "date"` (calendar dates). Thresholds are validated at config load; unparseable cell values fail closed ([#87](https://github.com/ababic/dumpling/issues/87)).
+
 ## [0.8.0] - 2026-08-14
 
 ### Added

README.md

@@ -328,9 +328,11 @@ Supported predicate operators:
 | `not_like` / `not_ilike` | Negation of `like` / `ilike` (prefer these over negative lookahead) |
 | `regex` / `iregex` | [Rust `regex`](https://docs.rs/regex/) crate (`iregex` is case-insensitive). **Not PCRE** — look-around (`(?!…)`, `(?=…)`, etc.), backreferences, and possessive quantifiers are unsupported and **rejected at config load** (fail closed). |
 | `not_regex` / `not_iregex` | Negation of `regex` / `iregex` |
-| `lt` / `lte` / `gt` / `gte` | Numeric compare (values parsed as numbers) |
+| `lt` / `lte` / `gt` / `gte` | Numeric compare by default (values parsed as numbers). Set `format = "datetime"` to compare ISO-8601 / Postgres timestamp text as instants, or `format = "date"` for calendar dates (`YYYY-MM-DD`). Unparseable cell values fail closed (do not match). |
 | `is_null` / `not_null` | No value needed |
 
+`format = "datetime"` / `"date"` is also valid on `eq` / `neq` (instant or date equality). Thresholds are validated at config load; cell values that do not parse do not match.
+
 Invalid or unsupported `regex` / `iregex` / `not_regex` / `not_iregex` patterns fail config load (and `dumpling lint-policy`) instead of silently matching nothing.
 
 For “scrub unless staff domain” policies, prefer a default scrub plus a `keep` case (see Conditional cases cookbook above), or scrub via `not_ilike` / `not_iregex` cases without a default rule:
@@ -376,6 +378,18 @@ delete = [
 ]
 ```
 
+Retain a recent window on timestamp columns without year-prefix `like` workarounds:
+
+```toml
+[row_filters."public.listing_order"]
+retain = [
+  { column = "created", op = "gte", value = "2025-02-14", format = "datetime" },
+  { column = "created", op = "is_null" },
+]
+```
+
+`format = "datetime"` accepts common `pg_dump` shapes (`2025-03-14 12:34:56.789+00`, `T` separators, `Z`, short offsets). Date-only thresholds are treated as midnight UTC. Use `format = "date"` to compare calendar dates only.
+
 Row filtering works for both `INSERT ... VALUES (...)` and `COPY ... FROM stdin` rows.
 
 ---

docs/src/configuration.md

@@ -379,9 +379,11 @@ salt = "${file:/run/secrets/dumpling_hmac_key}"
 | `not_like` / `not_ilike` | Negation of `like` / `ilike` (prefer these over negative lookahead) |
 | `regex` / `iregex` | [Rust `regex`](https://docs.rs/regex/) crate (`iregex` is case-insensitive). **Not PCRE** — look-around, backreferences, and possessive quantifiers are unsupported and **rejected at config load** (fail closed). |
 | `not_regex` / `not_iregex` | Negation of `regex` / `iregex` |
-| `lt` / `lte` / `gt` / `gte` | Numeric compare (values parsed as numbers) |
+| `lt` / `lte` / `gt` / `gte` | Numeric compare by default. With `format = "datetime"`, compare ISO-8601 / Postgres timestamp text as instants; with `format = "date"`, compare calendar dates. Unparseable cells fail closed. |
 | `is_null` / `not_null` | No value needed |
 
+`format = "datetime"` / `"date"` also applies to `eq` / `neq`. Threshold strings are validated at config load.
+
 Invalid or unsupported regex patterns fail config load and surface as `invalid-regex-predicate` in `dumpling lint-policy`. For “scrub unless allowlisted domain” policies, prefer a default scrub plus a `keep` case, or positive `not_ilike` / `not_iregex` cases — see the README *Conditional per-column cases* and *Row filtering* sections.
 
 Nested JSON targeting is supported in predicate `column` values via either: