ContextStream MCP v1.0.25

Changelog

1.0.25

• Scope alerts you cannot miss. init and context now put a scope_alerts
  array in the structured result as well as [SCOPE_MISMATCH] in the text,
  because clients such as Claude Code hand the model only the structured part.
  An alert covers a folder or repository that belongs to another project,
  several projects claiming one repository, and a scope that is unresolved or
  needs setup. It is rebuilt on every call and never deduplicated, so it stays
  until the scope is fixed, and in the wire budget it is the last structured
  field to be dropped (e1c49e6).
• Working set: a session has a primary project (the one bound to the working
  folder) and related projects the same work touches. The server returns
  working_set (primary, related, suggested) with init and context, shown
  as a [WORKING_SET] line. A project joins as related when your message names
  its checkout, when an edit lands in its checkout, or with
  project(action="working_set", add=[...]); remove=[...] and clear=true
  take projects out. A folder that belongs to a related project is cross-project
  work, not a mismatch. The project tool's input schema gains add, remove
  and clear; every earlier field and action is still accepted (e1c49e6).
• Search covers related projects: the primary project's results come first,
  then a labelled [RELATED_PROJECTS] section (related_project_results in the
  structured result) with up to 5 results from each of up to 3 related
  projects, read within 2.5 seconds. The related list comes from a per-caller
  cache that init and context fill (5 minutes) or one bounded API read, so it
  works on the stateless hosted transport. To search one related project fully,
  pass its project_id (e1c49e6).
• Hooks judge scope. PostToolUse notices an edit in another linked checkout,
  records that project as related on the server and in a local per-session
  copy, and tells the agent once per project. PreToolUse checks every
  scope-sensitive ContextStream call: a project outside the working set, or
  another workspace, gets [SCOPE_MISMATCH] with the exact fix, reads included;
  the primary and related projects are quiet, and tickets stay quiet inside the
  workspace. This replaces the generic "Scope alignment warning", which also
  fired on deliberate cross-project writes. The managed rules and the prompt
  reminders explain primary and related projects and the two notices (e1c49e6).
• Search keeps duplicate_count and duplicate_locations in compact output.
  The API now folds identical chunks found at different paths into one result;
  when the structured payload went over its budget the compact form kept only
  an allowlist of metadata and dropped both fields. Compact output now keeps the
  count and shows up to 5 path:start-end locations (then "+N more"), and the
  text adds a line such as "identical chunk also in 5 other files: ...".
  Uncompacted output is unchanged (b0deb40).
• Agent Coach Kit: the /coach server instructions and the workflow and
  workflow_save descriptions now say when the kit applies. A host that adds
  the server by address (the pilot route) receives no skill file, so these
  texts are all it reads to decide when to use the kit, and they said what the
  kit is but not when it applies. They now say it applies when the person
  starts or continues a job for work they repeat, reacts to a draft, compares
  versions or asks what earlier jobs taught, and that a one-off request or an
  ordinary question uses none of the tools. The wording describes and does not
  command, the form ChatGPT flagged in 1.0.23, and a test keeps "do not use",
  "never use" and "recurring workflow" (which made ChatGPT create a scheduled
  task of its own) out of all four texts.
• workflow_save lists the parameters of record_verdict (the job, a verdict
  of better, same or worse, and the person's words), so a verdict on a whole
  job or version is not taken for feedback on its content.
• The text grows, unlike 1.0.24: the instructions 737 to 985 characters,
  workflow 1331 to 1528, workflow_save 2054 to 2237; workflow_delete is
  unchanged at 518. The tools, their parameters and their behaviour are
  unchanged.
• The plugin skill, the Codex Cloud skill and the AGENTS snippet carry the same
  boundary (repository only), and the review run sheet has the reviewer type
  the first case without choosing the plugin first, so the recording shows
  whether ChatGPT reaches for it unprompted (repository only).
• Test fixtures in the compact-search tests use neutral paths and documentation
  UUIDs (67a2f62); the public-boundary check had failed on main since those
  tests were added (b0deb40).

Who gets what: the scope alerts, the working set, the related-project search
and the project tool's new parameters reach hosted clients when the hosted
gateway runs this version, and need the ContextStream API that returns
working_set (live). The hook behaviour and the new rules text reach a machine
after contextstream-mcp update (the rules also need a rules refresh). The
coach text reaches ChatGPT and Codex with the hosted gateway; the kit and the
plugin package live in the repository and are not part of the npm package or
the binaries.

Dependencies: rustls, the TLS library the client and the hosted gateway use,
moves from 0.23.45 to 0.23.46 in Cargo.lock for RUSTSEC-2026-0346 (Debug
output of client session values exposed TLS secrets), RUSTSEC-2026-0347
(buffered application data sent after an ECH rejection), RUSTSEC-2026-0348
(Brotli certificate decompression could allocate up to 1 GiB before
authentication) and RUSTSEC-2026-0349 (an unauthenticated ServerHello could
downgrade later connections to classical key exchange), all published on
2026-10-14. No other dependency changes.