ExMCP 1.0.0-rc.7

1.0.0-rc.7 is the modern-preferred soak candidate that follows published 1.0.0-rc.6. It keeps the dual-era MCP 2026-07-28 default while packaging the post-rc.6 transport lifecycle, ACP, and security hardening that cannot promote directly from rc.6 to stable 1.0.

What changed since rc.6

• Legacy Streamable HTTP GET SSE streams persist events before delivery, retain the MCP session across ordinary connection gaps, and replay from Last-Event-ID without duplication under bounded retention.
• ACP Pi managed event handling correctly processes multi-chunk agent streams, and adapter/MCP subprocesses isolate their environments by default.
• The 2026-08-12 security harden enforces protocol lifecycle, identity, authorization, network, resource, and privacy boundaries across MCP and ACP. See SECURITY_AUDIT_2026-08-12.md.
• CI and compatibility follow-ups keep Elixir 1.17 docs/tooling green, fail closed on invalid MRTR identities, and align fixtures with server-issued sessions and strict schema/privacy behavior.

:prefer_modern remains the application default. :legacy_only preserves the legacy protocol era connection policy (rc.7 still enforces server-issued sessions and newer lifecycle/security rules). See the migration guide for rollout steps and the API diff for the public compatibility audit, including the rc.7 delta note.

Why another soak: The SSE disconnect lifecycle intentionally retains a session after an ordinary GET-stream disconnect, and the security harden changes session issuance, outbound network policy, and subprocess defaults. Those observable compatibility changes restart the stable-1.0 clock. See the 2.0 roadmap backport lane.

Installation

{:ex_mcp, "~> 1.0.0-rc.7"}

Pin the rollout policy when it must not change with a package upgrade:

config :ex_mcp, protocol_mode: :prefer_modern
# Emergency rollback / legacy protocol era (not an exact rc.5 package rollback):
# config :ex_mcp, protocol_mode: :legacy_only

Security and dependencies

The release retains the patched Mint, Mint WebSocket, Plug, Plug Cowboy, Cowboy, Cowlib, HPAX, Decimal, and related runtime dependency set from rc.6. The locked dependency set has no critical or high-severity advisories.

Cowlib 2.19.0, the newest compatible release, retains two upstream advisories:

• EEF-CVE-2026-43966 (medium) is mitigated because Plug rejects CR/LF/NUL in response-header values and Cowboy independently terminates invalid response headers before serialization.
• EEF-CVE-2026-43969 (low) affects cow_cookie:cookie/1, which ExMCP and its Plug/Cowboy response path do not call.

These acknowledgements are exact IDs in the Hex audit configuration. CI still fails on every unacknowledged advisory and builds the real Hex archive.

Qualification evidence and ownership

The release tag identifies the exact source commit. Its GitHub Actions run is the durable evidence record; the external-conformance job uploads the complete runner output as mcp-2026-07-28-conformance.

The external suite is pinned to @modelcontextprotocol/conformance@0.2.0-alpha.10 because no stable 2026-07-28-aware runner is published; its CI lane pins Node.js 22 because the harness imports fs.globSync. Official TypeScript SDK v2 interop pins @modelcontextprotocol/client, server, and node to 2.0.0. The legacy conformance baseline remains pinned independently to 0.1.16.

Load and regression budget

The qualifying workload is mix test.suite performance on OTP 27 / Elixir 1.17. It covers the committed microbenchmarks, concurrent serialization, high-volume tool request preparation, large payloads, a ten-second sustained load interval, security checks, schema compilation, subscription queue pressure, and slow-consumer behavior.

The rc.7 publish gate is current-head only:

• zero errors or timeouts in the performance and stress suites;
• no committed absolute performance threshold may regress;
• process and mailbox counts must return to their bounded post-cleanup state;
• slow subscribers must coalesce or disconnect according to policy without cross-principal delivery.

A same-runner rc.5 vs rc.7 comparison (median runtime and retained memory for shared legacy workloads, plus runner identity, commit SHAs, and raw outputs) is not an rc.7 publish gate. Record that evidence before stable 1.0.0. Any unexplained current-head breach still blocks this candidate. Runner noise must be resolved by repeating the current revision on the same machine, not by widening a threshold in the release commit.

Stable 1.0 gates still open

Publishing rc.7 restarts, but does not complete, the stable-release clock. Stable 1.0.0 requires:

1. at least seven calendar days of final modern-preferred RC use without a release-blocking regression;
2. a successful mixed-version cluster rollback drill with active subscriptions and in-flight MRTR operations; and
3. a same-runner rc.5 vs rc.7 performance comparison (median runtime and retained memory for shared legacy workloads, plus runner identity, commit SHAs, and raw outputs).

Any wire-design, public-API, or documented lifecycle change requires another RC and restarts the soak. Current master packaged as rc.7 meets that condition because of the legacy SSE persistence/disconnect-lifecycle fix and the 2026-08-12 security harden described above.

Post-merge publish checklist

1. Confirm CI green on the release commit, including the MCP conformance artifact and ACP 1.3.0 interop job.
2. Tag v1.0.0-rc.7 at that commit.
3. Create a GitHub prerelease for v1.0.0-rc.7 (prerelease checkbox on) with notes pointing at this document and CHANGELOG.md.
4. mix hex.publish
5. Start the ≥7-day modern-preferred soak.
6. Complete the mixed-version cluster rollback drill before stable 1.0.0.

rc.5 performance comparison

Deferred to stable 1.0.0. See the load-budget policy above: rc.7 publishes on current-head thresholds only.