Skip to content

0002-target-2026-07-28-only.md

Latest commit

 

History

History
66 lines (52 loc) · 3.29 KB

0002-target-2026-07-28-only.md

File metadata and controls

66 lines (52 loc) · 3.29 KB
Raw

ADR-0002 — Target MCP 2026-07-28 exclusively

  • Status: Accepted
  • Date: 2026-08-04

Context

The SDK currently speaks 2024-11-05: an initialize handshake, a notifications/initialized sink, tools/list, tools/call. The current spec revision is 2026-07-28, the largest break since MCP launched. Relevant to a stdio server:

  • The initialize / notifications/initialized handshake is removed. Protocol version and client capabilities ride on every request in params._meta["io.modelcontextprotocol/protocolVersion"] and …/clientCapabilities.
  • server/discover is a MUST: it advertises supported versions, capabilities and identity, and doubles as the stdio backward-compatibility probe.
  • Every result carries resultType ("complete" or "input_required").
  • List results are CacheableResults: ttlMs and cacheScope are required.
  • Servers never initiate JSON-RPC requests; sampling/elicitation/roots become MRTR InputRequiredResults.
  • Error codes were repartitioned: -32000..-32019 implementation-defined, -32020..-32099 spec-reserved (HeaderMismatch -32020, MissingRequiredClientCapability -32021, UnsupportedProtocolVersion -32022).

Two shapes were on the table: dual-era (support both handshakes) or latest-only.

Decision

Implement 2026-07-28 and nothing else. No initialize, no notifications/initialized, no legacy fallback path in the core. A request whose _meta protocol version is not 2026-07-28 is rejected with UnsupportedProtocolVersionError (-32022) listing what we support.

Consequences

Good

  • The core stays small enough to read in one sitting. Dual-era support would mean two result builders (with and without resultType/ttlMs), two dispatch tables, and a state flag recording which era the peer is in — state that a stateless protocol was specifically redesigned to remove.
  • Statelessness maps exactly onto a bash read loop. There is no session to keep, so there is no session bug to have.
  • Being early on a spec whose whole point is "no handshake, no session" is the position worth holding for an SDK whose selling point is that it is 300 lines of shell.

Bad / accepted

  • Clients that predate 2026-07-28 cannot talk to this server, including editors still on 2025-06-18/2025-11-25. This is the real cost and it is deliberate: the version lives in assets/*_config.json, so a user who needs a legacy client today edits data, not code — but the core will still reject legacy requests, so a genuine legacy client needs a shim, not a config edit.
  • The spec is young; details may move. Mitigated by ADR-0003: conformance is asserted against the official published schema.json, so drift shows up as a failing test rather than as a silent incompatibility.

Alternatives considered

  • Dual-era support behind MCP_LEGACY_COMPAT=1 — rejected as the default shape, but it is the natural extension point if real users show up on old clients: it is additive (a second dispatch branch), and nothing in this decision forecloses it.
  • Stay on 2025-11-25 until clients catch up — rejected: it keeps the session and handshake machinery that is the worst fit for a shell, in exchange for compatibility the demo server does not currently need.