diff --git a/CHANGELOG.md b/CHANGELOG.md
index c9734e4..bba075c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -17,6 +17,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 * **BREAKING** Bumped `MINIMUM_QSV_VERSION` from `20.0.0` to `20.1.0` (and bumped the qsv version installed by `Dockerfile.worker`, `.github/workflows/ci.yml`, and `.github/workflows/main.yml` to match). Operators must upgrade their `qsv` binary at the path configured by `ckanext.datapusher_plus.qsv_bin` before deploying this version — DP+ will refuse to start otherwise with `JobError: At least qsv version 20.1.0 required. Found 20.0.0.`. qsv 20.1.0 itself introduces no breaking changes against 20.0.0 (per the [20.1.0 release notes](https://github.com/dathere/qsv/releases/tag/20.1.0): "pipelines built on 20.0.0 will upgrade in place"), so the upgrade is binary-swap-only — no data re-ingestion is required. The reason this is still flagged as **BREAKING** is the minimum-version gate, not the qsv behavior. User-visible improvement justifying the floor bump: qsv-dateparser 0.14 → 0.15 in qsv 20.1.0 adds recognition for ISO 8601 `T`-separated datetimes without a timezone suffix (e.g. `2024-10-11T14:30:00`) — qsv 20.0.0 misclassified these as `String` during `qsv stats --infer-dates`, which in DP+ surfaced as a date-typed column being demoted to text on certain CSV shapes (one of the gaps documented in the issue #173 regression test, now closed). The regression-test matrix in `tests/test_issue_173_date_format_inference.py` is updated to the new baseline.
 * **BREAKING** Bumped `MINIMUM_QSV_VERSION` from `20.1.0` to `25.0.0` (and bumped the qsv version installed by `Dockerfile.worker`, `.github/workflows/ci.yml`, `.github/workflows/main.yml` and `.github/workflows/test.yml` to match). Operators must upgrade their `qsv` binary at the path configured by `ckanext.datapusher_plus.qsv_bin` before deploying this version — DP+ will refuse to start otherwise with `JobError: At least qsv version 25.0.0 required. Found 20.1.0.`. qsv [25.0.0](https://github.com/dathere/qsv/releases/tag/25.0.0) is the first release with what `QSVCommand` reads through `qsv-client`: structured JSON errors (`QSV_ERROR_FORMAT=json`, [qsv#4765](https://github.com/dathere/qsv/pull/4765)), so a failed qsv step reports its error kind and message instead of raw stderr, and `qsv --capabilities` for the version check. On older qsv, both fell back to parsing plain text. It also redacts API keys from `describegpt`'s attribution ([qsv#4766](https://github.com/dathere/qsv/pull/4766)). Date inference is unchanged: the issue #173 matrix in `tests/test_issue_173_date_format_inference.py` gives the same type for every column under qsv 25.0.0 as under 20.1.0, and is now pinned to 25.0.x.
 
+### Fixed
+* Pinned `fastapi<0.144`. Prefect 3.7 only requires `fastapi<1.0`, and the FastAPI releases of 2026-10-18 to 23 break it twice: from 0.155 a fresh install fails on `import prefect` with `TypeError: Body() got an unexpected keyword argument 'min_items'` (unit suite, `prefect server start` and the integration CI all die), and from 0.144 the server starts but rejects the Prefect client's `workers/filter` request with `422 Input should be a valid integer` for `limit: null`, so `ckan datapusher_plus prefect-deploy` fails. 0.143.x is what the last green CI run installed. The pin goes away once a Prefect release in range works on current FastAPI (issue #353).
+
 ### qsv 20.0.0 migration notes (operator-facing)
 The new minimum brings the following user-visible changes from qsv 4.0.0 → 20.0.0 that may affect existing CKAN deployments:
 
diff --git a/requirements.txt b/requirements.txt
index 994212c..07c1ea2 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -6,6 +6,9 @@ pandas==2.2.3
 shapely==2.1.0
 pyproj>=3.7.1
 prefect>=3.7,<3.8
+# fastapi 0.155+ breaks `import prefect` (removed min_items); 0.144-0.154 reject
+# the Prefect client's `limit: null` with 422.
+fastapi<0.144
 # importlib_metadata: declared explicitly because prefect 3.7.1's
 # ``workers/base.py`` imports it unconditionally but doesn't list it
 # as a direct dep — it was being satisfied transitively by