diff --git a/CHANGELOG.md b/CHANGELOG.md
index d136cbc..00bbc39 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,6 +9,13 @@ surface is governed by [`COMPATIBILITY.md`](COMPATIBILITY.md).
 
 ## [Unreleased]
 
+### Added
+- **SDK: resume a run after a crash.** `Runtime.resume_run(run_id)` attaches to an
+  existing governed run (it neither creates a new run nor cancels on error), so a
+  Python agent can pick its work back up from the last checkpoint after a `SIGKILL`.
+  The run keeps its server-side budget and already-spent usage, so it can't
+  overspend by restarting. See the [SDK README](sdks/python/README.md#resume-after-a-crash).
+
 ## [0.2.0] - 2026-06-04
 
 A frictionless-adoption release: a one-line CLI install, three runnable key-free
diff --git a/sdks/python/README.md b/sdks/python/README.md
index eea4ede..8bc57a7 100644
--- a/sdks/python/README.md
+++ b/sdks/python/README.md
@@ -36,6 +36,25 @@ with rt.governed_run(name="research",
 When the governor halts the run (token / dollar / loop / time budget), the next
 `run.step()` — or a proxied model call — raises `rk.BudgetExceeded`.
 
+## Resume after a crash
+
+The daemon reloads non-terminal runs on restart with the budget and usage they had
+already spent, so a `SIGKILL`'d run keeps enforcing without re-spending. Reattach to
+it by id with `resume_run` and pick your work back up from the last checkpoint:
+
+```python
+with rt.resume_run(run_id) as run:          # attaches; never creates or cancels
+    cp = run.latest_checkpoint()            # the state you saved before the crash
+    start = cp["payload"]["cursor"] if cp else 0
+    for i in range(start, total):           # skip the steps you already paid for
+        run.step()                          # counts against the SAME budget
+        # ... your work ...
+        run.checkpoint("step", {"cursor": i + 1})
+```
+
+The run resumes against whatever budget it had left, so it can't overspend by
+restarting — `run.step()` still raises `rk.BudgetExceeded` at the original ceiling.
+
 ## Human-in-the-loop tools
 
 Gate side-effecting tools on human approval (the daemon's policy decides what needs
diff --git a/sdks/python/riskkernel/runtime.py b/sdks/python/riskkernel/runtime.py
index d8dd856..b32d303 100644
--- a/sdks/python/riskkernel/runtime.py
+++ b/sdks/python/riskkernel/runtime.py
@@ -157,6 +157,33 @@ def governed_run(self, name: Optional[str] = None,
         finally:
             _current_run.reset(token)
 
+    @contextmanager
+    def resume_run(self, run_id: str):
+        """Attach to an existing governed run by id — the resume path after a crash.
+
+        Unlike governed_run, this neither creates nor cancels the run: the daemon
+        reloads non-terminal runs on restart with the budget and usage they had
+        already spent, so enforcement continues without re-spending. Fetch
+        ``run.latest_checkpoint()`` to pick your work back up where it left off::
+
+            with rt.resume_run(run_id) as run:
+                cp = run.latest_checkpoint()
+                start = cp["payload"]["cursor"] if cp else 0
+                for i in range(start, total):
+                    run.step()                       # counts against the SAME budget
+                    ...
+                    run.checkpoint("step", {"cursor": i + 1})
+
+        Raises APIError(404) if the run id is unknown.
+        """
+        data = self.client.get_run(run_id)
+        run = Run(self.client, data, self._poll, self._timeout)
+        token = _current_run.set(run)
+        try:
+            yield run
+        finally:
+            _current_run.reset(token)
+
 
 # Module-level default runtime, configured from the environment, for the
 # decorator/convenience API.
diff --git a/sdks/python/tests/test_sdk.py b/sdks/python/tests/test_sdk.py
index f3d6a9d..2948f6f 100644
--- a/sdks/python/tests/test_sdk.py
+++ b/sdks/python/tests/test_sdk.py
@@ -58,6 +58,9 @@ def _read(self):
 
     def do_GET(self):
         p = self.path
+        if p == "/v1/runs/run-1":
+            return self._send(200, {"id": "run-1", "name": "t", "status": "running",
+                                    "usage": {"tokens": 0, "loops": STATE.steps}})
         if p.startswith("/v1/memory/facts"):
             return self._send(200, [{"namespace": "dev", "key": "db", "value": "sqlite"}])
         if p.startswith("/v1/memory/entry"):
@@ -133,6 +136,22 @@ def test_governed_run_and_step_budget(self):
                 run.step()
             self.assertEqual(cm.exception.reason, "loop_budget_exceeded")
 
+    def test_resume_run_attaches_without_creating(self):
+        # The post-crash path: resume an existing run by id. resume_run attaches a
+        # handle (no new run, no cancel) that keeps stepping/checkpointing against
+        # the SAME run, and reads the checkpoint it left off at.
+        with self.rt.governed_run(name="t", budget=self.rt.budget(loops=5)) as run:
+            run.step()                                    # one step before the "crash"
+            run.checkpoint("before-crash", {"cursor": 1})
+            rid = run.id
+
+        with self.rt.resume_run(rid) as resumed:
+            self.assertEqual(resumed.id, rid)             # same run, not a new one
+            self.assertEqual(rk.current_run().id, rid)    # set as the current run
+            cp = resumed.latest_checkpoint()
+            self.assertEqual(cp["payload"]["cursor"], 1)  # where we left off
+            self.assertEqual(resumed.step(), 2)           # continues the same step count
+
     def test_checkpoint_roundtrip(self):
         with self.rt.governed_run(name="t") as run:
             run.checkpoint("after", {"cursor": 7})