Repository navigation
fix(protocols): a cut standalone IPv6 extension header rebuilds with values its parse did not read #2761
Description
Describe the bug
A standalone IPv6 extension header cut short parses with fewer octets in some fields than its rebuild holds. The rebuilt octets are exact, but from_data(info).info != info. The parse reads only the captured octets; the rebuild reads at the declared width (_uncaptured, #2703). Framed in IPv6, nothing differs.
Reproduction
from pcapkit.protocols.internet.ipv6_route.protocol import IPv6_Route
H = bytes([59,2,2,1, 1,2,3,4]) + bytes(15) + b'\x01'
p = IPv6_Route(H[:6]); r = IPv6_Route.from_data(p.info)
p.info.reserved, r.info.reserved, bytes(r) == H[:6] # (b'\x01\x02', b'\x01\x02\x00\x00', True)Also: IPv6-Frag reserved_octet at cut 1 (b'' vs b'\x00'), HOPOPT/IPv6-Opts options gain a PadN the parse lacks (cuts 1, 2, 4–6), IPv6_Ext length at cut 1 (1 vs 2).
Expected behavior
The rebuild's info equals the parse's. test-coverage.rst (Round-Trip Invariants) requires a cycle to reproduce its input exactly; here the octets match, but the values do not.
Additional context
Likely site: Protocol.__post_init__ / replay_short_read: make a standalone short-read parse and its rebuild read the header the same way.
Activity
JarryShaw commentedon Nov 4, 2026
Decision needed: does a round trip whose octets match but whose parsed values differ count as a defect? The Round-Trip Invariants section speaks of octets, so it doesn't settle this case.
from_data(info).info == info, and write the rule ontotest-coverage.rst.Recommendation: option 1. A rebuild that invents a PadN option, or zero-fills
reserved, reports fields the capture never held. That is the same class of defect #2756 removed from lists.This touches
pcapkit/protocols/protocol.py, and #2490's member reorder is running across the protocol modules now. So a worker starts on option 1 once that reorder's PR lands, unless you rule option 2 first.JarryShaw commentedon Nov 4, 2026
take 1
JarryShaw commentedon Nov 4, 2026
Option 1, reconciled: a round trip has to reproduce the parsed values as well as the octets.
from_data(info).info == infofor IPv6-Routereserved, IPv6-Fragreserved_octet, HOPOPT/IPv6-Optsoptions(no invented PadN) andIPv6_Extlength.test-coverage.rst(Round-Trip Invariants) gets that rule written into it.Protocol.__post_init__/replay_short_read). That is allowed alongside Reorder protocol class members to the member-order convention #2490's reorder, so a worker starts now rather than after Reorder protocol class members to the member-order convention #2490 lands.JarryShaw commentedon Nov 4, 2026
#2767 fixes every field this issue lists except one; that one needs a decision.
The remaining gap:
IPv6_Extlengthat cut 1. The parse reads the header at 1 octet, while the rebuild reads it at the 2 octets thatmake()packs, and the data model does not record which width the parse was given. No shared rule fits all the cases. A cut Ethernet header differs only in its next layer.IPv6_Extdiffers only inlength. An IGMPv3 Report cut inside IPv4 needs IPv4's declared width.__short_read__and rebuild at that width. This adds one newDatametadata key, and that key could replace_uncapturedand_short_of_header.lengthinsideIPv6_Ext.read. That is a one-line, per-protocol change, close to the length clamp ruled out under fix(transport): the SCTP schema on its own repacks a final chunk cut inside its padding at full width #2680.Recommendation: (a). It is the only shared fix. #2767 does not close this issue; the gap stays open here.
fix(protocols): rebuild a cut standalone header with the values its p…
JarryShaw commentedon Nov 4, 2026
take (a)
JarryShaw commentedon Nov 4, 2026
Option (a), reconciled: a cut header's data model records the width its parse read the header at, alongside
__short_read__, andfrom_datarebuilds at that width. This closes theIPv6_Extlengthgap at cut 1 without a per-protocol clamp.to_dict()output in the same way as__short_read__. It must not change any other cut, framed or standalone._uncaptured(fix(protocols): a layer rebuilt by from_data from a cut capture verifies before its short read is replayed #2703) and_short_of_header(fix(protocols): a cut header parses with a Raw next layer but rebuilds with NoPayload #2717) is checked in the same change. The swap is made only if every test of those issues stays green. Otherwise both stay, and the PR says why.test-coverage.rstthen holds without exception.A worker is starting now; the PR will close this issue.
docs(changelog): correct the #2770 and #2767 entries (#2449)
fix(protocols): rebuild a cut header at the width its parse read it (#…