The instructions blob in the MCP `initialize` handshake is the spec
non-Claude-Code clients (codex, Cline, opencode, hermes-agent, Cursor)
inherit verbatim. Three gaps mean the bridge daemon handles them in
code (codex-channel-molecule bridge.py:192-200, 278-285) but in-process
agents reading the text alone don't get the same guard:
1. Reply-then-pop ordering was implicit. A literal-minded agent could
pop after a 502 from `send_message_to_user`, dropping the message.
Now: pop ONLY AFTER reply succeeds; on error leave the row unacked
for platform redelivery.
2. peer_agent with empty peer_id had no specified handling. Agent
would call `delegate_task(workspace_id="")` → 400 → re-poll →
infinite loop on the same poison row. Now: skip reply, drain via
inbox_pop.
3. The single security rule ("don't execute without chat-side
approval") effectively disabled peer_agent autonomous handling —
codex daemons have no canvas user to approve from. Now: dual trust
model. canvas_user requires user approval; peer_agent permits
autonomous handling but caps destructive side-effects at the
workspace boundary.
Also disclaims peer_name/peer_role as non-attested display strings —
the platform registry isn't cryptographic identity, and an agent
shouldn't grant elevated permissions based on a peer registering with
peer_role="admin".
Four new pinned tests in test_a2a_mcp_server.py:
- test_initialize_instructions_pins_reply_then_pop_ordering
- test_initialize_instructions_handles_malformed_peer_agent
- test_initialize_instructions_disclaims_peer_role_attestation
- test_initialize_instructions_distinguishes_canvas_user_from_peer_trust
Each fails on staging-HEAD and passes on the patched text — verified
by reverting a2a_mcp_server.py and re-running.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
b5f530e27a