Semantics-preserving routing for agents, MCP servers, and models.
GaugeMesh invokes resilireplay@0.7.0 as an optional external CLI. It is not a
runtime library and its source is not vendored.
The release-gated path uses gaugemesh mcp-stdio, tool docs-a__search, safety
class read-only, one retry, a 3000 ms request bound, --no-regression, and
separate evidence for one genuine clean control plus every fault exposed by the
published 0.7.0 mcp test command. Each fault invocation first runs its own
unmodified control. GaugeMesh derives the single clean-control row from the
first actual clean trace, then runs all twelve faults explicitly; omitting
--fault would select mcp-tool-error, not a clean-only run. For every fault
GaugeMesh requests a dry-run JSON plan, validates its 64-hex plan SHA-256, then
supplies that exact digest with --approve. Each invocation is an argument
array, not a shell command.
ResiliReplay 0.7.0 applies these faults to its recorded trace after receiving a real MCP response; it does not mutate the MCP wire or GaugeMesh server. In particular, the timeout row substitutes a synthetic timeout trace event and then makes one real retry. Its PASS therefore proves that bounded harness behavior, cleanup, and duplicate-effect observation—not termination or recovery from a real request timeout.
Observed local result on 2026-09-06 from the corrected PR source:
| Scenario | Matrix result |
|---|---|
| clean control | PASS |
| synthetic timeout trace mutation plus real retry | PASS |
| synthetic deterministic-error trace mutation plus real retry | PASS |
| malformed tools list | FAIL |
| renamed tool | FAIL |
| missing tool | FAIL |
| incompatible argument schema | FAIL |
| oversized content | FAIL |
| protocol-version mismatch | FAIL |
| invalid JSON-RPC ID | FAIL |
| malicious canary instruction | FAIL |
| permission/capability mismatch | FAIL |
| canary leakage attempt | FAIL |
The clean control passed. All twelve fault runs observed the requested trace
mutation, included a passing clean prerequisite, completed cleanup, and
reported zero duplicate effects. The required clean-control and two synthetic
retry rows passed. The combined matrix result is therefore PARTIAL, not a
claim that every mutation recovered or that a wire-level timeout was tested.
Its combined evidence SHA-256 is
sha256:7d6f7a0c7ce1373b93d60374e9e75dc4d90a7ec306fb31b8e5bc703fbef34817.
The individual plan and evidence digests are emitted by
gaugemesh verify --resilireplay.
An equal-environment comparison of base 57b8240 and candidate 1ff927e found
no status regression. The
discovery, schema, size, protocol, JSON-RPC-ID, and capability mutations are
post-response trace changes, so their failures leave the corresponding target
behavior unproven rather than demonstrating a GaugeMesh rejection. The two
canary rows are intentional negative controls outside 0.7.0’s two-fault retry
contract; their detected synthetic injections remain FAIL and are not waived.
The public 0.7.0 command output did not identify an MCP-RES v0.2 profile or
evidence class. GaugeMesh records mcpRes: null and makes no MCP-RES compliance,
official MCP certification, security certification, or exactly-once claim.