You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Resolve the carried-over blocker and remaining suggestions from Hai Huang's
re-review:
Must-fix (blocker, carried over):
- applyMCPResponseBodyMod (cmf_body.go): the MCP response path had the
identical multi-part leak the A2A path had — it rewrote only the first
result.content[] text block and break'd, forwarding blocks[1..] verbatim
with mutated=true. Now collects all text blocks and fails closed on >1
(the read side only inspects the first; one redacted value can't safely
rewrite many), mirroring the A2A/inference guards. Added a multi-block
fail-closed test asserting the planted secret is not forwarded.
Suggestions:
- MCPRequestBodyMod: replace the len(NewArguments)==0 no-op inference with
explicit ArgsSet/URISet flags. A "strip all arguments" redaction returns
an empty-but-set map and must apply, not silently forward the original
args. Added strip-all-applies + unset-no-op tests.
- Docs: fix the misleading "Authorization stripped" comments in
manager_cpex.go and headers.go — Authorization is deliberately KEPT for
CPEX identity plugins. Made the audit-log redaction + session-API binding
guidance a non-optional MUST.
- Supply chain: assert the FFI tarball VERSION matches CPEX_FFI_VERSION and
FFI_ABI matches the new pinned CPEX_FFI_ABI file at build time (dropped
the stale TODO). Wired CPEX_FFI_ABI through the demo Makefile and CI
build matrix. Runtime ABI assertion already exists in the Go binding's
abi.go init().
Verified: go test ./plugins/cpex/ passes; go vet clean; authbridge-cpex
image rebuilds with sha256 + cosign + VERSION/ABI all verified; hr-cpex
demo 9/9 scenarios pass on kind (Bob SSN visible, Eve redacted, denies
enforce, taint + cross-principal isolation intact).
Also retitled the PR to "Feat:" to satisfy the org verify-pr-title gate.
Signed-off-by: Frederico Araujo <frederico.araujo@ibm.com>
0 commit comments