ContextStream MCP v1.0.23

Changelog

1.0.23

• Agent Coach Kit: workflow(action="compare") marks feedback a version
  carried that its own job kept for that job only. A later job's agent could
  read such feedback in "what changed" as something to apply; it now reads
  "(this job only; not for later jobs)" in the text and job_only on the card.
  The compare reads back each carried item's scope (at most 12 reads, shared by
  both sides); one it could not read is marked "(scope not checked; ask the
  person before reusing it)", never guessed (22be1ee).
• Agent Coach Kit card (ui://widget/coach.html) meets WCAG 2.1 AA: one
  persistent aria-live region announces saves, the focus ring has at least
  3:1 contrast in light and dark, the base size is relative, the card declares
  lang="en" and honors prefers-reduced-motion, nothing scrolls sideways at
  320 px, scrolling text panels take keyboard focus as named regions, and the
  compare sides and button rows are labelled groups (22be1ee).
• Agent Coach Kit for Codex Cloud (repository only): save-output --carried REF:HOW records what each version carried, a new verdict command records
  the person's verdict on an open or closed job, metrics shows how the
  coaching is going from real jobs only (no quality score), and the skill asks
  only about what is missing or changed (5a383d4, cd77a8f).
• Agent Coach Kit package for the ChatGPT and Codex plugin directory
  (plugins/agent-coach-kit, repository only): the plugin listing, the one
  remote server https://mcp.contextstream.io/coach with OAuth (read:coach,
  write:coach), the coaching skill, and plugins/build_agent_coach_kit.py,
  which refuses a package that breaks a submission rule it can check. Nothing
  has been submitted (e0bfb25).

Who gets what: the compare label and the card reach ChatGPT and Codex when the
hosted gateway runs this version. The kit and the plugin package live in the
repository; they are not part of the npm package or the binaries. No
dependency changes.