README peers section, docs/consolidator.md (threat model, tokens, firewall,
worked agenix example, reload/restart), the MCP tool count in both tables,
and config.example.toml commentary.
Real registry and transport against in-process peers and hand-written
listeners: plain, TLS with an rcgen CA, a wrong CA, redirects, oversize
bodies, a silent peer, the 64-peer cap and shutdown mid-follow. Two proxy
cases are ignored: the proxied path is missing its /api/v1 prefix.
The peer follow stream had no byte cap and kept a whole unterminated
frame in memory, and the MCP surface carried its own peer status table
that already disagreed with REST for peer 400.
- mcp::peers::fetch now returns proxy::PreparedResponse, so the status
comes from proxy::classify_status and the body from
proxy::prepare_response (peer byte cap, deadline, shutdown), the same
two steps REST uses. status_error, body_bytes and the local 2xx check
are gone.
- tail_journal consumes that prepared stream; a Frames cursor keeps a
partial frame from being rescanned from byte 0 on every chunk, the
entry cap truncates the result, and the end of the stream maps to the
note (entry cap, shutdown, byte cap, deadline or close).
- Prompts resolve the peer argument against the registry (unknown or
disabled peer is refused); list_peers returns the error registry()
produced; one wording for an unknown peer name.
- Tests: byte cap on an unterminated frame, entry cap, buffered body,
follow 404/500, deadline with an open stream, shutdown, frame cursor,
REST/MCP parity for peer 400/404/401/5xx, and a full golden snapshot of
the peers-off tool table. Peer scripts use the stripped path form C2
forwards.
Server instructions and the eight tool descriptions mention peers and the
untrusted-data rule only while a registry is configured. peers://status
returns the list_peers JSON and is advertised only then. diagnose_unit and
container_health_report take an optional validated peer argument.
The two status tools and the optional peer argument exist only when a
registry is configured. rmcp's disable_route hides the tools; the eight
shared schemas lose the optional peer property, and the two list tools also
lose the derived $schema key, so the peers-off list and schemas match a
build without the feature byte for byte. Both snapshot variants and the
wire-level tool list are asserted.
Peers resolve through PeerRegistry::get, the same entry point REST uses,
so the hop marker, allowlist and error mapping stay in one place. Log
tools keep the local {entries, truncated, note} shape; tail_journal
consumes the peer SSE stream. Tests cover every tool, path encoding,
error mapping and the integration path through the assembled router.
B1/B2 are merged, so `real_transport` builds the `ReqwestTransport` with the
configured limits and deny list instead of bailing. Binary tests now probe a
real loopback server: ok with latency, rejected credential, unreachable peer
and an unknown name, none of which may print the token.
The peer-less `--print-config` output is now compared against the committed
pre-peers text plus the default `[peers]` block, and a config with a peer
must leave everything before that block byte-identical. New binary tests
cover `--print-config`, `peer list`, `token create|list` with an unreadable
token file, and `serve` refusing to start while naming the peer.
Land the shared peer contracts: types, error mapping, transport trait,
fake transport and fixtures, registry/proxy/health skeletons, the
AppState.peers and Config.peers seams, the ping and peers route stubs,
and unit plus integration tests. No loops, no network, feature off by
default.
Probe the served cert with a non-verifying client and wait on the server log
instead of fixed sleeps, so a not-yet-processed SIGHUP no longer panics.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>