diff --git a/swh/loader/cvs/rlog.py b/swh/loader/cvs/rlog.py
--- a/swh/loader/cvs/rlog.py
+++ b/swh/loader/cvs/rlog.py
@@ -46,6 +46,7 @@
 
 import calendar
 import re
+import string
 import time
 from typing import BinaryIO, Dict, List, NamedTuple, Optional, Tuple
 
@@ -74,7 +75,7 @@
     state: str
     branches: None
     revnumstr: None
-    commitid: None
+    commitid: Optional[str]
 
 
 class RlogConv:
@@ -413,6 +414,21 @@
         return time.strptime(timestr, "%Y-%m-%d %H:%M:%S %z")[:-1] + (0,)
 
 
+def _parse_commitid(commitid: bytes) -> Optional[str]:
+    s = commitid.decode("ascii").strip()
+    # Strip "commitid: " tag and the trailing semicolon.
+    s = s[len("commitid: ") : -len(";")]
+    # The commitid itself contains digit and ASCII letters only:
+    for c in s:
+        if (
+            c not in string.digits
+            and c not in string.ascii_lowercase
+            and c not in string.ascii_uppercase
+        ):
+            raise ValueError("invalid commitid")
+    return s
+
+
 def _parse_log_entry(fp) -> Tuple[Optional[revtuple], Optional[bytes], Optional[bytes]]:
     """Parse a single log entry.
 
@@ -480,6 +496,12 @@
             raise ValueError("invalid year")
     date = calendar.timegm(tm)
 
+    commitid = match.group(6) or None
+    if commitid:
+        parsed_commitid = _parse_commitid(commitid)
+    else:
+        parsed_commitid = None
+
     # return a revision tuple compatible with 'rcsparse', the log message,
     # and the EOF marker
     return (
@@ -492,7 +514,7 @@
             ),  # state, usually "Exp" or "dead"; non-ASCII data here would be weird
             None,  # TODO: branches of this rev
             None,  # TODO: revnumstr of previous rev
-            None,  # TODO: commitid
+            parsed_commitid,
         ),
         log,
         eof,
diff --git a/swh/loader/cvs/tests/data/dino-commitid.tgz b/swh/loader/cvs/tests/data/dino-commitid.tgz
new file mode 100644
index 0000000000000000000000000000000000000000..0000000000000000000000000000000000000000
GIT binary patch
literal 0
Hc$@<O00001

literal 0
Hc$@<O00001

diff --git a/swh/loader/cvs/tests/test_loader.py b/swh/loader/cvs/tests/test_loader.py
--- a/swh/loader/cvs/tests/test_loader.py
+++ b/swh/loader/cvs/tests/test_loader.py
@@ -625,3 +625,87 @@
     }
 
     check_snapshot(DINO_SNAPSHOT, loader.storage)
+
+
+DINO_SNAPSHOT2 = Snapshot(
+    id=hash_to_bytes("a9d6ce0b4f22dc4fd752ad4c25ec9ea71ed568d7"),
+    branches={
+        b"HEAD": SnapshotBranch(
+            target=hash_to_bytes("150616a2a3206f00a73f2d6a017dde22c52e4a83"),
+            target_type=TargetType.REVISION,
+        )
+    },
+)
+
+
+def test_loader_cvs_split_commits_by_commitid(swh_storage, datadir, tmp_path):
+    """Conversion of RCS history which needs to be split by commit ID"""
+    # This repository has some file revisions which use the same log message
+    # and can only be told apart by commit IDs. Without commit IDs, these commits
+    # would get merged into a single commit in our conversion result.
+    archive_name = "dino-commitid"
+    archive_path = os.path.join(datadir, f"{archive_name}.tgz")
+    repo_url = prepare_repository_from_archive(archive_path, archive_name, tmp_path)
+    repo_url += "/dino"  # CVS module name
+
+    loader = CvsLoader(
+        swh_storage, repo_url, cvsroot_path=os.path.join(tmp_path, archive_name)
+    )
+
+    assert loader.load() == {"status": "eventful"}
+
+    assert_last_visit_matches(
+        loader.storage, repo_url, status="full", type="cvs", snapshot=DINO_SNAPSHOT2.id,
+    )
+
+    check_snapshot(DINO_SNAPSHOT2, loader.storage)
+
+    stats = get_stats(loader.storage)
+    assert stats == {
+        "content": 18,
+        "directory": 36,
+        "origin": 1,
+        "origin_visit": 1,
+        "release": 0,
+        "revision": 18,
+        "skipped_content": 0,
+        "snapshot": 18,
+    }
+
+
+def test_loader_cvs_pserver_split_commits_by_commitid(swh_storage, datadir, tmp_path):
+    """Conversion via pserver which needs to be split by commit ID"""
+    # This repository has some file revisions which use the same log message
+    # and can only be told apart by commit IDs. Without commit IDs, these commits
+    # would get merged into a single commit in our conversion result.
+    archive_name = "dino-commitid"
+    archive_path = os.path.join(datadir, f"{archive_name}.tgz")
+    repo_url = prepare_repository_from_archive(archive_path, archive_name, tmp_path)
+    repo_url += "/dino"  # CVS module name
+
+    # Ask our cvsclient to connect via the 'cvs server' command
+    repo_url = f"fake://{repo_url[7:]}"
+
+    loader = CvsLoader(
+        swh_storage, repo_url, cvsroot_path=os.path.join(tmp_path, archive_name)
+    )
+
+    assert loader.load() == {"status": "eventful"}
+
+    assert_last_visit_matches(
+        loader.storage, repo_url, status="full", type="cvs", snapshot=DINO_SNAPSHOT2.id,
+    )
+
+    check_snapshot(DINO_SNAPSHOT2, loader.storage)
+
+    stats = get_stats(loader.storage)
+    assert stats == {
+        "content": 18,
+        "directory": 36,
+        "origin": 1,
+        "origin_visit": 1,
+        "release": 0,
+        "revision": 18,
+        "skipped_content": 0,
+        "snapshot": 18,
+    }