Skip to content

v1.5.0

Choose a tag to compare

@azmaveth azmaveth released this 21 Sep 21:41
· 599 commits to master since this release
5fc6d9f

Added

  • The scheduled ACP ecosystem drift check now pins and monitors the newly
    published zai-org/ZCode repository alongside the Claude, Codex, and Pi
    adapter references. The pin and the check are repo-only; no library code or
    behavior changes.
  • Pi adapter golden-transcript characterization suite:
    ExMCP.Test.PiGolden (test/support/acp/pi_golden.ex) and seven golden
    areas under test/ex_mcp/acp/adapters/pi/characterization/ with fixtures
    under test/fixtures/acp/pi/, pinning the adapter's RPC envelopes,
    control-group ordering, stream-event conversion, prompt scheduling,
    configuration updates, slash commands and session-store safety rules
    before the Pi.* boundary extractions. Repo-only; nothing ships to Hex.
  • ExMCP.Client.subscribe_notifications/3 and
    ExMCP.Client.unsubscribe_notifications/2, a public delivery path for
    legacy-era (MCP 2024-11-05 through 2025-11-25) list-change and
    resource-update notifications. A listener registers a filter using the same
    keys as listen/3, minus taskIds, and receives matching notifications as
    {:ex_mcp_notification, listener, method, params}. Listened resource URIs
    are subscribed on the server once and shared across listeners, released
    when the last listener naming them goes away or its subscriber exits, and
    re-subscribed after an automatic reconnect with a
    {:ex_mcp_notification_reconnected, listener, result} report. On a modern
    peer the call returns {:error, :use_listen}. Previously the client parsed
    these notifications and dropped them with a warning about an unknown
    subscription; that warning is now a debug-level message when no listener
    matches. Modern subscription routing and subscribe_resource/3 are
    unchanged (#44).

Changed

  • Internal: .dialyzer_ignore.exs drops fifteen filters that no longer match
    a warning, and documents that entries must be verified in both MIX_ENV=dev
    and MIX_ENV=test because the files under test/ are only analyzed in the
    test environment. No analysis result changes.

  • ACP reference-adapter parity, from the 2026-09-20 drift review of
    claude-agent-acp and codex-acp (both still on ACP SDK 1.4.0):

    • Claude SDK adapter: a host may opt a session out of bypassPermissions
      with _meta.claudeCode.options.allowDangerouslySkipPermissions: false
      on session/new, session/load, or session/resume. The mode leaves
      the catalog, session/set_mode refuses it, and an active bypass mode is
      clamped to default through a set_permission_mode control. true
      cannot override the adapter's own option. (claude-agent-acp#1129)
    • Claude SDK adapter: an AskUserQuestion custom answer no longer replaces
      a selection. It joins a multi-select in the CLI's own quoted form, and
      for a single-select it rides beside the pick as the tool's per-question
      annotations[question].notes, so a client that presents the box as a
      notes field cannot make the selection disappear. (#1031, #1131)
    • Codex adapter: item/tool/requestUserInput forms follow codex-acp#299.
      The elicitation message is always "Codex needs your input to continue.",
      the question text is the field title and the header its description,
      every question is required, an isOther question with options gains a
      "None of the above" choice and a separate <id>_note field tagged
      _meta.codex.role: "user_note", and note text returns to Codex as
      user_note: <text> beside the selection instead of replacing it. A
      whitespace-only answer is dropped. Wire-visible for clients that read
      the previous <id>__other field or _meta.codex.isOtherAnswer.
    • Codex adapter: session/load pages a thread's older turns through
      thread/turns/list when the resume result carries a backwards cursor,
      replaying the full history in order instead of only the newest hundred
      turns; the initial page is now requested newest-first and reversed. A
      server without the method, or a failed page, still loads with the
      initial page. (codex-acp#481)
    • ExMCP.ACP.Adapter translations may return
      {:messages_and_reply_and_write, messages, result, iodata, state}, and
      session/load and session/resume accept {:reply_and_write, ...}.
    • Deferred with recorded decisions (pre-standard _meta extensions, not
      in ACP SDK 1.4.0): authStatus, recommendedValue, asyncTasks,
      tool-call name, and both compaction mechanisms. Not applicable:
      codex-acp#471, since ExMCP forwards MCP elicitations without a
      synthetic tool call. See docs/POST_1_0_MAINTENANCE_PLAN.md.
  • Internal: the three Mint-based pinned HTTP clients
    (ExMCP.Internal.PinnedHTTPClient, ExMCP.Authorization.PinnedHTTPClient,
    and ExMCP.Transport.HTTP.BoundedClient) now share one pure response
    reducer, ExMCP.Internal.HTTPResponseReducer, for Mint event accumulation
    and bounded-body decisions. Each client keeps its own DNS pinning, TLS,
    compression and framing policy, error atoms, and return shapes. No public
    API or behavior change.

  • mint moves to 1.10.1, which fixes EEF-CVE-2026-82672 (unvalidated
    chunk-size line tail in the HTTP/1 client). mix hex.audit fails on 1.10.0
    since the advisory was published on 2026-09-19.

  • Internal dependency-cycle reductions with no public API change. Seven of
    the nine mix xref graph --format cycles components reported on Elixir
    1.19.5 / OTP 28.3.1 are gone (MessageProcessor/MethodHandlers,
    Content.Validation/Rules, SchemaPolicy/SchemaRemoteResolver,
    ExMCP/ClientConfig, Server.Subscriptions/Tasks, the
    Internal.SessionStore behaviour and its ETS/DETS backends, and the
    VersionRegistry/Protocol.Methods/ErrorCodes/Types version cycle),
    each through a narrow dependency inversion or a small @moduledoc false
    lower-level module (ExMCP.Tasks.StoreCall,
    ExMCP.Internal.SessionStore.Factory, ExMCP.Internal.RevisionCatalog).
    The two large client and transport cycles remain; see
    docs/POST_1_0_MAINTENANCE_PLAN.md.

  • The Codex ACP adapter is split into ExMCP.ACP.Adapters.Codex.Permissions,
    Codex.Content, and Codex.MCP (internal modules) with no wire-visible
    change: the golden transcript fixtures are byte-identical, and the public
    ExMCP.ACP.Adapters.Codex API and state struct are unchanged. See the
    Codex status in docs/POST_1_0_MAINTENANCE_PLAN.md.

Fixed

  • Pi adapter: three agent-controlled payloads no longer crash the adapter.
    A streamed tool call carried under partial.content[contentIndex] is now
    selected with Enum.at/2 (Access cannot index a list, so the clause
    written to support that shape always raised), and a get_available_models
    payload whose models is not a list, or whose entries are not maps, drops
    the unusable data instead of raising. Found by the new Pi characterization
    gate and pinned by three golden scenarios.

  • Codex adapter: session/load history replay now handles app-server v2's camelCase agentMessage items. Previously a replayed agentMessage reached the streaming completion path with no session state and crashed the load; only the legacy agent_message spelling replayed.

  • ExMCP.SessionManager no longer logs at info when it starts or when it
    sweeps expired sessions; both are debug. The :ex_mcp application boots
    before a stdio server can suppress logging, and the default Elixir logger
    writes to stdout, so in a release that did not configure :stdio_mode
    ahead of boot the startup line landed in the protocol stream before the
    first frame. A subprocess test now boots the application's supervision tree
    under the default logger and asserts stdout stays empty. The configuration
    guide documents config :ex_mcp, stdio_mode: true together with
    config :logger, :default_handler, config: [type: :standard_error] as the
    stdio deployment setting that closes the window for every application in
    the VM.

  • The stdio transports no longer let the process locale translate protocol
    frames. ExMCP.Server.StdioServer and ExMCP.ACP.Agent.Transport.Stdio
    now detect, on every read and write, whether each device is a character
    or a byte device and use the matching calls, through one shared framing
    owner, which is byte-exact for UTF-8 on every supported OTP and survives
    the VM flipping a device to latin1 after undecodable input (on OTP 27 that
    flip still ends the session under a UTF-8 locale; see the transport
    guide). Previously,
    under a C locale or no locale (launchd, systemd, minimal MCP host
    environments) the MCP stdio server double-encoded non-ASCII input and
    emitted \x{...} escapes for non-ASCII output, producing invalid JSON,
    while under a UTF-8 locale the ACP agent transport corrupted non-ASCII
    input. An echo tool masked the MCP case because the write reversed the
    read's damage. A byte-order mark before the first frame is now stripped
    instead of hanging the session at initialize. Diagnosed by deepfates in
    #41. A shared non-ASCII payload corpus now runs byte-exact through every
    transport, with a locale matrix on the stdio subprocess test (#52).