diff --git a/.coveragerc b/.coveragerc
index 1d3c78a..7dc1873 100644
--- a/.coveragerc
+++ b/.coveragerc
@@ -3,6 +3,16 @@ branch = True
 source = .
 
 [report]
+# Coverage policy: see docs/decisions/0004-coverage-policy.md.
+#
+# `omit` is reserved for genuinely interactive entry-point shells whose
+# bodies are `read_tty_line()` / curses prompt loops — there is no
+# behaviour to assert that a test wouldn't have to fake wholesale, so a
+# test here would inflate the number without buying confidence. This is
+# NOT a place to hide subprocess/backend orchestration: that code is
+# security-relevant and is measured via the integration suite instead
+# (run scripts/coverage.sh for the combined unit+integration number).
 omit =
     bot_bottle/cli/tui.py
+    bot_bottle/cli/init.py
     tests/*
diff --git a/.gitea/workflows/test.yml b/.gitea/workflows/test.yml
index cd563f5..0dabe0f 100644
--- a/.gitea/workflows/test.yml
+++ b/.gitea/workflows/test.yml
@@ -70,3 +70,32 @@ jobs:
 
       - name: Run integration tests
         run: python3 -m unittest discover -t . -s tests/integration -v
+
+  # Combined unit+integration coverage + the diff-coverage gate.
+  # See docs/decisions/0004-coverage-policy.md. The hard gate is diff
+  # coverage (new/changed lines >= 90%); the combined + critical reports
+  # are informational and degrade gracefully when the runner has no
+  # Docker (integration tests skip, those modules just read lower).
+  coverage:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install dev requirements
+        run: python3 -m pip install -r requirements-dev.txt
+
+      - name: Combined coverage (unit + integration)
+        run: PYTHON=python3 bash scripts/coverage.sh critical
+
+      - name: Diff-coverage gate (changed lines >= 90%)
+        run: |
+          git fetch --no-tags origin main:refs/remotes/origin/main
+          python3 scripts/diff_coverage.py --base origin/main --min 90
diff --git a/docs/decisions/0004-coverage-policy.md b/docs/decisions/0004-coverage-policy.md
new file mode 100644
index 0000000..6141e5b
--- /dev/null
+++ b/docs/decisions/0004-coverage-policy.md
@@ -0,0 +1,90 @@
+# ADR 0004: Risk-weighted coverage, not a single global target
+
+- **Status:** Accepted
+- **Date:** 2026-06-25
+- **Deciders:** didericis
+
+## Context
+
+bot-bottle is a security tool: it sandboxes agents, scans egress for
+secret exfiltration, strips credentials, and gates git pushes. A latent
+bug in that logic is expensive, so test coverage there genuinely
+matters. But the repo also contains code where coverage is a poor
+signal:
+
+- **Interactive entry-point shells** — `cli/init.py` (a `read_tty_line()`
+  prompt loop) and `cli/tui.py` (a curses picker). Their bodies are I/O;
+  a unit test has to fake the entire terminal conversation, so it
+  inflates the number without asserting behaviour that would otherwise
+  go unchecked.
+- **Subprocess / backend orchestration** — the docker / smolmachines /
+  macos-container backends shell out to `docker`, `container`, `smolvm`.
+  Mock-heavy unit tests here mostly re-assert the argv you already
+  wrote (the test passes whether or not the real teardown works), while
+  many of the missed *branches* are failure paths you cannot provoke
+  against a real daemon on cue.
+
+Chasing a single global percentage (e.g. 90%) pushes the most test
+effort onto the least safety-relevant code — exactly backwards — and
+invites performative tests written to colour a line rather than to catch
+a regression (Goodhart's law).
+
+## Decision
+
+Coverage is **risk-weighted**, measured over the **combined unit +
+integration** suites, with three rules:
+
+1. **Critical modules target ≥ 90%.** The security/logic core —
+   `egress_addon{,_core}.py`, `dlp_detectors.py`, `egress.py`,
+   `manifest*.py`, `git_gate.py`, `git_http_backend.py`, `supervise.py`,
+   `yaml_subset.py`, `bottle_state.py` — is Docker-independent and
+   unit-testable, so it carries the high bar. We ratchet toward 90% as
+   these modules are touched; new gaps in them are not acceptable.
+
+2. **Subprocess/backend orchestration is covered by the integration
+   suite, not omitted.** `scripts/coverage.sh` runs unit + integration
+   under one coverage measurement so these modules are scored where they
+   are actually exercised. They stay *visible* — hiding the code that
+   tears down sandboxes and wires networks is the one place we will not
+   omit.
+
+3. **Interactive entry-point shells are omitted** (`.coveragerc`), with a
+   rationale comment. This is the only sanctioned use of `omit` besides
+   `tests/*`.
+
+The forward-looking guard is a **diff-coverage gate**
+(`scripts/diff_coverage.py`): new/changed executable lines on a branch
+must be ≥ 90% covered. This catches regressions where they are
+introduced without forcing a back-fill crusade through legacy glue. The
+gate skips lines in omitted files (there is no coverage data for them),
+so the omit list cannot launder *new* logic into the dark: anything that
+needs real testing must live outside the interactive shells to be
+scored at all.
+
+The **global percentage is informational**, not a CI gate — it would
+otherwise be hostage to the CI runner's Docker availability and to the
+omit list.
+
+## Consequences
+
+- The number we report (`scripts/coverage.sh`) means "coverage of the
+  code we consider testable, across both suites" — a dip is a real
+  regression in code we control, not noise from added CLI glue.
+- No incentive to write mock-the-mock tests for orchestration to defend
+  a global figure.
+- The omit list needs governance: an entry must be a genuinely
+  interactive shell, justified in the `.coveragerc` comment and here.
+  `cli/init.py` and `cli/tui.py` qualify; backend orchestration does
+  not.
+- CI must run the integration suite under coverage to score the
+  orchestration modules; where the runner lacks Docker those tests skip
+  and their modules read low — accepted, because the *enforced* gates
+  (critical-module standard + diff coverage) are Docker-independent.
+- "We're at N%" is now a curated figure; outsiders should read the
+  policy, not just the badge.
+
+## Links
+
+- PRs #290 (cover the egress adapter), and the coverage-policy PR that
+  introduces this record.
+- `.coveragerc`, `scripts/coverage.sh`, `scripts/diff_coverage.py`.
diff --git a/scripts/coverage.sh b/scripts/coverage.sh
new file mode 100755
index 0000000..db9556c
--- /dev/null
+++ b/scripts/coverage.sh
@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+# Combined unit + integration coverage (see docs/decisions/0004-coverage-policy.md).
+#
+# Runs the unit suite, then appends the integration suite (which skips
+# cleanly when Docker / the backend CLIs are unavailable), and prints one
+# combined report. The integration suite is what scores the subprocess /
+# backend orchestration modules, so the number here is the policy's
+# yardstick — not the unit-only badge.
+#
+# Usage:
+#   scripts/coverage.sh            # combined report
+#   scripts/coverage.sh critical   # also report just the critical modules
+set -euo pipefail
+
+cd "$(dirname "$0")/.."
+
+PY="${PYTHON:-python3}"
+
+# Critical security/logic core held to the high bar by ADR 0004.
+CRITICAL="bot_bottle/egress_addon.py,bot_bottle/egress_addon_core.py,\
+bot_bottle/dlp_detectors.py,bot_bottle/egress.py,bot_bottle/manifest.py,\
+bot_bottle/manifest_egress.py,bot_bottle/manifest_agent.py,\
+bot_bottle/manifest_schema.py,bot_bottle/git_gate.py,\
+bot_bottle/git_http_backend.py,bot_bottle/supervise.py,\
+bot_bottle/yaml_subset.py,bot_bottle/bottle_state.py"
+
+rm -f .coverage
+
+echo "== unit ==" >&2
+"$PY" -m coverage run -m unittest discover -t . -s tests/unit
+
+echo "== integration (skips without Docker) ==" >&2
+"$PY" -m coverage run --append -m unittest discover -t . -s tests/integration
+
+echo "== combined report ==" >&2
+"$PY" -m coverage report -m
+
+if [ "${1:-}" = "critical" ]; then
+    echo "== critical modules (ADR 0004 target: 90%) ==" >&2
+    "$PY" -m coverage report --include="$CRITICAL"
+fi
diff --git a/scripts/diff_coverage.py b/scripts/diff_coverage.py
new file mode 100755
index 0000000..546c655
--- /dev/null
+++ b/scripts/diff_coverage.py
@@ -0,0 +1,126 @@
+#!/usr/bin/env python3
+"""Diff-coverage gate (see docs/decisions/0004-coverage-policy.md).
+
+Fails if too few of the *added/changed* executable lines on this branch
+are covered. Stdlib-only by design — the project carries no runtime deps
+and we are not adding `diff-cover` to satisfy a check.
+
+Reads coverage data already produced by a `coverage run` (e.g. via
+`scripts/coverage.sh`): it shells out to `coverage json` for per-line
+data and to `git diff` for the changed lines. Lines in omitted files
+(the interactive shells) have no coverage data and are skipped, by
+policy.
+
+Usage:
+    scripts/coverage.sh                 # produce .coverage first
+    python3 scripts/diff_coverage.py    # gate against origin/main, min 90%
+    python3 scripts/diff_coverage.py --base main --min 85
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+_HUNK_RE = re.compile(r"^@@ -\d+(?:,\d+)? \+(\d+)(?:,(\d+))? @@")
+
+
+def _run(cmd: list[str]) -> str:
+    return subprocess.run(
+        cmd, check=True, capture_output=True, text=True,
+    ).stdout
+
+
+def added_lines_by_file(base: str) -> dict[str, set[int]]:
+    """Map each changed .py file to the set of line numbers added/changed
+    relative to `base`, parsed from a zero-context unified diff."""
+    diff = _run(["git", "diff", "--unified=0", f"{base}...HEAD", "--", "*.py"])
+    out: dict[str, set[int]] = {}
+    current: str | None = None
+    new_line = 0
+    for line in diff.splitlines():
+        if line.startswith("+++ b/"):
+            current = line[6:]
+            out.setdefault(current, set())
+            continue
+        hunk = _HUNK_RE.match(line)
+        if hunk:
+            new_line = int(hunk.group(1))
+            continue
+        if current is None:
+            continue
+        if line.startswith("+") and not line.startswith("+++"):
+            out[current].add(new_line)
+            new_line += 1
+        elif line.startswith("-") and not line.startswith("---"):
+            # Deletion: does not advance the new-file cursor.
+            continue
+    return out
+
+
+def coverage_json() -> dict[str, object]:
+    """Render the existing .coverage data to JSON and load it."""
+    with tempfile.NamedTemporaryFile("r", suffix=".json", delete=True) as fh:
+        _run([sys.executable, "-m", "coverage", "json", "-o", fh.name])
+        return json.load(open(fh.name, encoding="utf-8"))
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--base", default="origin/main",
+                    help="git ref to diff against (default: origin/main)")
+    ap.add_argument("--min", type=float, default=90.0,
+                    help="minimum %% of changed executable lines covered")
+    args = ap.parse_args()
+
+    if not Path(".coverage").exists():
+        print("diff-coverage: no .coverage data; run scripts/coverage.sh first",
+              file=sys.stderr)
+        return 2
+
+    added = added_lines_by_file(args.base)
+    files = coverage_json().get("files", {})
+    if not isinstance(files, dict):
+        files = {}
+
+    total = 0
+    covered = 0
+    misses: list[str] = []
+    for path, lines in sorted(added.items()):
+        info = files.get(path)
+        if not isinstance(info, dict):
+            # Omitted file or not measured (e.g. a test file) — skip by policy.
+            continue
+        executed = set(info.get("executed_lines", []))
+        missing = set(info.get("missing_lines", []))
+        executable = lines & (executed | missing)
+        for ln in sorted(executable):
+            total += 1
+            if ln in executed:
+                covered += 1
+            else:
+                misses.append(f"{path}:{ln}")
+
+    if total == 0:
+        print("diff-coverage: no measured changed lines to check — pass")
+        return 0
+
+    pct = 100.0 * covered / total
+    print(f"diff-coverage: {covered}/{total} changed lines covered ({pct:.1f}%)")
+    if misses:
+        print("uncovered changed lines:", file=sys.stderr)
+        for m in misses:
+            print(f"  {m}", file=sys.stderr)
+    if pct + 1e-9 < args.min:
+        print(f"diff-coverage: below {args.min:.0f}% threshold", file=sys.stderr)
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/tests/unit/test_cli_dispatch.py b/tests/unit/test_cli_dispatch.py
new file mode 100644
index 0000000..4acbc37
--- /dev/null
+++ b/tests/unit/test_cli_dispatch.py
@@ -0,0 +1,82 @@
+"""Unit: top-level CLI dispatch in bot_bottle.cli.main (ADR 0004).
+
+`cli/__init__.py` is dispatch + exit-code mapping, not interactive I/O,
+so it carries real unit tests rather than being omitted like the
+`cli/init` / `cli/tui` shells."""
+
+from __future__ import annotations
+
+import io
+import unittest
+from unittest.mock import patch
+
+import bot_bottle.cli as climod
+from bot_bottle.cli import main
+from bot_bottle.log import Die
+from bot_bottle.manifest import ManifestError
+
+
+class TestMainDispatch(unittest.TestCase):
+    def test_no_args_prints_usage_returns_2(self) -> None:
+        with patch("sys.stderr", io.StringIO()):
+            self.assertEqual(2, main([]))
+
+    def test_help_flags_return_0(self) -> None:
+        with patch("sys.stderr", io.StringIO()):
+            self.assertEqual(0, main(["-h"]))
+            self.assertEqual(0, main(["--help"]))
+
+    def test_unknown_command_dies(self) -> None:
+        with patch("sys.stderr", io.StringIO()):
+            with self.assertRaises(Die):
+                main(["definitely-not-a-command"])
+
+    def test_handler_return_code_passthrough(self) -> None:
+        def handler(_rest: list[str]) -> int:
+            return 7
+
+        with patch.dict(climod.COMMANDS, {"x": handler}):
+            self.assertEqual(7, main(["x"]))
+
+    def test_handler_none_return_becomes_0(self) -> None:
+        def handler(_rest: list[str]) -> int | None:
+            return None
+
+        with patch.dict(climod.COMMANDS, {"x": handler}):
+            self.assertEqual(0, main(["x"]))
+
+    def test_args_forwarded_to_handler(self) -> None:
+        seen: list[list[str]] = []
+
+        def handler(rest: list[str]) -> int:
+            seen.append(rest)
+            return 0
+
+        with patch.dict(climod.COMMANDS, {"x": handler}):
+            main(["x", "a", "b"])
+        self.assertEqual([["a", "b"]], seen)
+
+    def test_manifest_error_maps_to_1(self) -> None:
+        def boom(_rest: list[str]) -> int:
+            raise ManifestError("bad manifest")
+
+        with patch.dict(climod.COMMANDS, {"x": boom}), patch("sys.stderr", io.StringIO()):
+            self.assertEqual(1, main(["x"]))
+
+    def test_die_maps_to_its_code(self) -> None:
+        def boom(_rest: list[str]) -> int:
+            raise Die(3)
+
+        with patch.dict(climod.COMMANDS, {"x": boom}):
+            self.assertEqual(3, main(["x"]))
+
+    def test_keyboard_interrupt_maps_to_130(self) -> None:
+        def boom(_rest: list[str]) -> int:
+            raise KeyboardInterrupt()
+
+        with patch.dict(climod.COMMANDS, {"x": boom}):
+            self.assertEqual(130, main(["x"]))
+
+
+if __name__ == "__main__":
+    unittest.main()