diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index c9ed769..de8161d 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -45,7 +45,21 @@ jobs:
           python-version: ${{ matrix.python-version }}
       - run: sudo apt-get update && sudo apt-get install -y --no-install-recommends build-essential swig libpcsclite-dev
       - run: pip install -e ".[dev]"
-      - run: pytest -q -m "not real_card"
+      # Bare --cov picks up [tool.coverage.run] source; the project-wide floor comes
+      # from [tool.coverage.report] fail_under, so there is one place to edit when
+      # ratcheting. Both matrix entries are measured: src carries no version-specific
+      # branches today, so 3.10 and 3.14 are expected to report the same number, and a
+      # divergence is itself worth seeing.
+      - run: pytest -q -m "not real_card" --cov --cov-report=term-missing --cov-report=xml
+      # Second, tighter floor on the layer that actually needs to move: the command
+      # modules hold 1699 of the 2158 missed statements. coverage.py has no per-file
+      # fail_under, so this is a separate report over the same data.
+      - run: coverage report --include='src/cryptnox_id_cli/cli/commands/*' --fail-under=36
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02  # v4.6.2
+        if: ${{ !env.ACT && matrix.python-version == '3.14' }}
+        with:
+          name: coverage
+          path: coverage.xml
 
   build:
     runs-on: ubuntu-latest
diff --git a/.gitignore b/.gitignore
index f6912a5..8605695 100644
--- a/.gitignore
+++ b/.gitignore
@@ -14,6 +14,7 @@ env/
 .pytest_cache/
 .ruff_cache/
 .coverage
+coverage.xml
 htmlcov/
 
 # Docs build
diff --git a/CHANGELOG.md b/CHANGELOG.md
index d35f653..5120e1c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -14,6 +14,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   mechanisms.
 - `perso generate-key --create-key-object` - the same dev/eval fallback
   `import-key` already had.
+- Coverage measurement in CI: `pytest-cov` in the `dev` extra, a project-wide
+  floor from `fail_under`, and a tighter floor over the command modules. With
+  branch coverage enabled both floors compare the combined statement-and-branch
+  percentage, which sits below line coverage alone.
 
 ## [1.0.3] - 2026-08-31
 
diff --git a/README.md b/README.md
index 9c4db9b..5032cfe 100644
--- a/README.md
+++ b/README.md
@@ -163,6 +163,17 @@ pytest -q -m "not real_card"
 Unit tests use a mock transport — no reader or card required. Real-card tests
 are opt-in (`-m real_card`) and never run in CI.
 
+To measure coverage locally, the same way CI does:
+
+```bash
+pytest -q -m "not real_card" --cov --cov-report=term-missing
+```
+
+CI enforces two floors: a project-wide one from `fail_under` in `pyproject.toml`,
+and a tighter one over `src/cryptnox_id_cli/cli/commands/`. Because branch
+coverage is enabled, both compare the **combined** statement-and-branch
+percentage that `coverage report` prints, which sits below line coverage alone.
+
 ---
 
 ## License
diff --git a/pyproject.toml b/pyproject.toml
index a208b70..eb71024 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -53,6 +53,7 @@ Changelog = "https://github.com/cryptnox/cryptnox-id-cli/blob/main/CHANGELOG.md"
 [project.optional-dependencies]
 dev = [
     "pytest>=7.4",
+    "pytest-cov>=5.0",
     "ruff>=0.4",
     "mypy>=1.8",
     "pyinstaller>=6.0",
@@ -103,4 +104,20 @@ markers = [
     "real_card: test requires a physical card on the ACS reader (opt-in)",
     "destructive: test mutates the card; requires CRYPTNOX_ALLOW_DESTRUCTIVE_TESTS=1",
 ]
+# Deliberately no --cov here: it would make a bare `pytest` fail for anyone without
+# pytest-cov installed, and slow every local run. CI passes the flags explicitly.
 addopts = "-ra"
+
+[tool.coverage.run]
+source = ["cryptnox_id_cli"]
+branch = true
+
+[tool.coverage.report]
+show_missing = true
+skip_covered = true
+# NOTE: with branch = true, this gate compares the COMBINED statement-and-branch
+# percentage that `coverage report` prints - not line coverage alone. At the time
+# this gate was introduced the three numbers were 60.25% combined, 64.11% line,
+# 43.62% branch, so reading it as a line-coverage floor overstates the suite by
+# four points. Raise it in the same commit as the tests that earn the increase.
+fail_under = 60