Skip to content

Releases: editor-code-assistant/eca

0.115.5

19 Mar 21:55
3b1c1e8

Choose a tag to compare

  • Fix chat getting stuck when LLM streaming connection hangs or is cancelled.

0.115.4

19 Mar 19:42
4c27f54

Choose a tag to compare

  • Fix auto-compact race condition where stale messages were sent to model after compaction, causing context overflow in subagents. #357

0.115.3

19 Mar 18:40
edd2346

Choose a tag to compare

  • Fix MCP server auth being invalidated when only URL query parameters change.
  • Fix MCP HTTP transport using stale auth token instead of reading latest from state, and retry initialization with token refresh on 403/4xx errors.
  • Fix MCP initialize request sending server config name as client name instead of "ECA".

0.115.2

18 Mar 19:25
cae9884

Choose a tag to compare

  • Fix STDIO MCP server deadlock caused by list_changed notification handlers blocking the reader thread.
  • Consider Anthropic internal server error as a retriable error.
  • Improve MCP error logging to show error code, message, and data instead of null.
  • Improve error handling when MCP promtps fail on server side.
  • Use rewrite-json to edit jsons without losing formatting.
  • Allow shell output redirections to /tmp/ in plan/explorer agents instead of denying them.
  • Fix crash when hook returns additionalContext and message has string content (e.g. /compact command).

0.115.1

17 Mar 17:18
a12b3b4

Choose a tag to compare

  • Improve OAuth callback error page to show error code, description, and error URI from the authorization server response.
  • Bump plumcp to 0.2.0-beta4, simplify MCP tool call error handling now that HTTP 400/404/500 are returned as JSON-RPC errors.
  • Fix MCP server incorrectly marked as running when HTTP transport returns 4xx (e.g. 401/403) during initialization.

0.115.0

17 Mar 03:27
d1f2dcf

Choose a tag to compare

  • Improve native image size for smaller eca binaries
    • Drop UPX compression for all native image builds.
    • Optimize native image for size with -Os.
  • Add optional model and variant parameters to spawn_agent tool, allowing users to run subagents on a different model or variant than the current conversation.
  • Improve MCP OAuth spec compliance: add Protected Resource Metadata discovery (RFC 9728), OIDC Discovery 1.0 fallback for authorization server metadata, fix scope parameter formatting, include resource parameter (RFC 8707) in authorization and token requests, and support clientId in MCP server config for pre-registered OAuth applications.
  • Add trust mode: clients can send trust: true in chat/prompt to auto-accept tool calls that would require manual approval.
  • Improve MCP tool call error handling

0.114.2

15 Mar 13:19
1d9a862

Choose a tag to compare

  • Fix completion requests failing with OpenAI subscription models by always streaming on the Responses API. eca-emacs#183
  • Disable upx for Windows binaries. #362

0.114.1

14 Mar 14:29
740bb44

Choose a tag to compare

  • Fix native image on Windows not running on older CPUs by enabling -march=compatibility. #362
  • Fix simple summaries in tool calls for ollama.
  • Add MCP server instructions to LLM context. #361

0.114.0

14 Mar 00:57
5828c7e

Choose a tag to compare

  • Delete chats older than 7 days on server startup.
  • Use human-readable workspace cache directory names (e.g. my-project_a1b2c3d4), with automatic migration from old hash-only format.
  • Show MCP server URL in the details page and allow editing command/args/url inline with mcp/updateServer endpoint.
  • Add mcp/updateServer to support change mcp commands/urls from client UI.

0.113.1

13 Mar 22:52
927021e

Choose a tag to compare

  • Fix MCP server threads blocking ECA shutdown when stuck during initialization; startup now uses daemon threads with interrupt-based cancellation for clean exit.
  • Re-initialize MCP session and retry tool call on HTTP 404 (session expired) or 5xx (e.g. pod swap dropping the SSE stream), per MCP Streamable HTTP spec.
  • Fix preRequest hook all_messages containing only the current turn instead of the full conversation history.