v2.10.0

Independent recommendations and verification records are easier to compare, and structured credential fields are now masked by name.

• Fix masking of opaque, short and non-string values under recognized sensitive keys, including nested objects and arrays. Ordinary identifiers such as primary_key remain readable; text-assignment masking keeps its existing rules.
• Report observed replacements during file/diff preparation, brief rendering and result normalization. Older observations remain unknown. Consultants now provide synthetic examples using separate variables and boolean assertions, so masking does not destroy their reproduction code.
• Add a concise recommendation separate from stance. Matching stances can recommend different actions; older history stays explicitly undeclared. Comparison and individual retrieval retain the existing size limits through nine consultants.
• Add refuted and carry bounded evidence summaries into follow-ups. Partial checks stay unverified. Legacy local notes are not forwarded until explicitly recorded again.
• Explain material gaps and changed requirements, and shorten the path from installation to the first question. Standalone plugin documentation now includes install, usage, contribution and license information.
• Make ESM test fixtures explicit for the earliest supported Node 22 release.

Install or update:

npm publication is pending the registry's one-time-password authentication. npm latest still resolves to 2.9.2. The GitHub release archive below contains v2.10.0 and can be installed directly.

npx -y --package=https://github.com/coz-a/severally/releases/download/v2.10.0/severally-mcp-2.10.0.tgz severally-mcp install

The installer updates the global runtime command and registered copies. Restart clients afterward. See the installation guide.

Validation on 2026-10-27 (JST): build, auto format, format check, ESLint, strict type checking and package-content checks passed. Linux/Node 24.20.0 ran 603 tests with 593 passes and 10 platform skips; Node 22.0.0 passed 587 with 16 platform/SQLite skips. Independent review found no Critical or Important issues. The development-candidate S3 comparison used five runs per version; all five refined received examples were reproduced using actual functions. These are narrow checks, not a competitive re-score or a latency superiority claim.

The existing Codex credential-path limitations remain documented: individual-file rename replacement, executable-containing roots and later legacy job homes are not fully covered. Live enforcement on Windows/macOS, first-login journeys, next-day continuation and the full competitive comparison remain unverified. The known generic redaction return-type annotation still does not express sensitive-property type replacement.

PR #20 and the merged main commit passed CI on Linux (Node 22/24), macOS and Windows (Node 24).

Real CLI checks on Linux/Node 24.20.0, 2026-10-27 (JST), using the host's configured models and a 180-second probe budget:

CLI	Version	Consultation	Doctor
Codex	0.164.0	Passed	Passed
Claude Code	2.1.311	Passed	Passed
Antigravity	1.4.7	Timed out after configured quota fallback	Passed
OpenCode	2.0.24	Passed	Passed

Antigravity exhausted the configured gemini-4-argon-high quota, used the operator's configured gemini-3.8-flash-high fallback, and did not finish the review in 180 seconds. This is not counted as a successful consultation. Doctor results describe only the observed probes; they do not erase the documented read gaps or establish all prohibited operations on every OS. Effective provider models/effort were not independently observed.

v2.9.2

Codex consultations run again on Windows.

• On Windows, Codex's unelevated sandbox cannot enforce the read-deny entries that protect credential paths, so it refused every consultation with windows unelevated restricted-token sandbox cannot enforce deny-read restrictions directly; refusing to run unsandboxed. The Codex adapter now asks for Codex's elevated sandbox on Windows only, keeping the credential protection. If Codex has not set that sandbox up yet, the first consultation asks for administrator approval.
• Installation, status and update instructions, installer hints and the version-mismatch fix use the npm package again (severally-mcp@<version>, @latest).

Install or update:

npx -y severally-mcp@latest install

Restart clients after installation. See the installation guide for options.

Validation on 2026-10-26: formatting, lint, strict type checks, build and package-content checks passed on Linux/Node.js 24.20.0; the suite passed 568 tests with 10 platform-specific skips and no failures. PR CI passed on Linux (Node.js 22 and 24), macOS and Windows (Node.js 24). The Windows change follows Codex's source; a live Codex consultation on Windows has not been verified yet.

v2.9.1

Credential reads are better protected for Codex and OpenCode consultations.

• Codex now denies the default .codex credential directory as a whole when its executable lives elsewhere. Existing, replaced, and newly created children stay denied in the Linux checks. Canonical-root deduplication avoids sandbox startup failures caused by mounting the same directory through aliases.
• OpenCode receives a local pre-read guard in its synthetic HOME. Each read checks its requested path, real path, and the current targets of protected roots. Allowed reads use canonical paths; unresolvable or protected paths are refused. No user plugins are inherited.
• Installer status and version-mismatch hints point to GitHub Release archives.
• Regression coverage includes both hook interfaces, alias swaps, protected roots created or retargeted after launch, and the guard emitted by the bundled server. English and Japanese installation and protection guides are synchronized and reviewed; Japanese prose is checked with yomiyasu.

Install or update:

npx -y severally-mcp@2.9.1 install

Restart clients after installation. See the installation guide for options.

Validation on 2026-10-26: local formatting, lint, strict type checks, build, and package-content tests passed on Linux/Node.js 24.20.0; the suite passed 567 tests with 10 platform-specific skips and no failures. PR CI passed on Linux (Node.js 22 and 24), macOS and Windows (Node.js 24). This is offline test coverage, not a claim of live CLI enforcement on those operating systems. Synthetic probes used Codex 0.164.0 and OpenCode 2.0.24 on Linux. An authenticated Codex consultation completed with the configured and launched gpt-6.1-sol model and low effort; the provider's effective model and effort were not independently observed. OpenCode read probes used a localhost mock model, not a real model.

Release-time Codex diagnostics: the direct sandbox probe refused the protected-root read, and the normal live-check consultation completed. The model-driven doctor checks were inconclusive because no tool calls or control-file quotation were observed; they are not counted as passed.

Known limits remain: Codex cannot deny arbitrary .env filenames, and individual-file rename replacements, executable-containing root snapshots and future legacy job homes retain their documented gaps. OpenCode's hook and file open are separate operations, so replacing canonical paths between them remains outside the guard's guarantee. The guard requires a CLI that loads its plugin. Actual OpenCode V1 loading and macOS/Windows CLI enforcement have not been verified.

v2.9.0

Failed consultations can now be retried individually, and saved consultations are easier to find, continue and review.

• consult_start({ retry_of: job_id }) replays a failed consultation in a new chain, using its saved inputs and current files. Successful peers are left intact, and the original failure remains linked. Resolve authentication or quota failures before retrying.
• consult_list filters by consultant, question or chain and returns the latest completed job and remaining rounds. Old consultant names remain searchable after configuration changes.
• Markdown exports begin with verified points, evidence, decision effects and unchecked counts. Retry exports retain links and available earlier context. Verification remains separate from adopting advice.
• Consultation instructions are shorter, with code-context selection and recovery details in the reference guides.
• Read-deny rules include recognized relocated credential paths. Codex resolves path aliases before retaining access to its executable entry. Known Codex file-replacement/new-entry gaps, OpenCode symlink aliases and Windows descendant-cleanup limits remain documented.
• status diagnoses optional SQLite support for OpenCode credential recovery without rejecting ordinary login. Doctor/live-check can record structured compatibility evidence and explicitly test follow-ups and cancellation.
• English/Japanese installation and usage guides are updated; README logos are included in the package.

Install or update from the attached package:

npx -y -p https://github.com/coz-a/severally/releases/download/v2.9.0/severally-mcp-2.9.0.tgz severally-install

Restart clients after installation. See the installation guide for options.

Validation: formatting, lint, strict type checks, build and package-content checks passed. The Linux suite passed 554 tests with 10 skips and no failures. PR CI passed on Linux (Node.js 22 and 24), macOS and Windows (Node.js 24). This release does not claim new live coverage for every CLI, model or operating system.

v2.8.1

An OpenCode answer now records what the consultant's tool calls were given, like Codex and Antigravity, and failure details are masked.

• tool_inputs now covers OpenCode: the paths and patterns its file tools were given, with the same limits and masking as the other consultants. Its tool calls are counted once per call id.
• The detail of a failed consultation is masked like its message, so a credential in a CLI's error output does not reach the view or the history.
• Time spent preparing a CLI counts against the first attempt's budget, not on top of it.
• doctor prints each CLI's version.
• README: Install is split into Requirements and Installing, and the contents list is shorter. Plugin descriptions use the same sentence as package.json.

Install or update from the attached package:

npx -y -p https://github.com/coz-a/severally/releases/download/v2.8.1/severally-mcp-2.8.1.tgz severally-install

Restart the clients after installation. See the installation guide for options.

Validation: format, lint, strict type checks, build and generated artifacts passed. The full Linux suite passed 526 tests with 10 skips and no failures. PR CI passed on Linux (Node.js 22 and 24), macOS (Node.js 24) and Windows (Node.js 24).

Live checks on 2026-10-24, Linux 7.0 (npm run doctor): Codex 0.163.0, Claude Code 2.1.310, Antigravity 1.4.7 and OpenCode 2.0.24 each wrote no file, returned no protected file's contents and called no sub-agent tool (Codex: the protected-file check is not enforced for it). live-check --target opencode completed with openrouter/deepseek/deepseek-v4.1-flash-1017. One earlier OpenCode doctor run with that model answered outside the required JSON shape (invalid_output); the next two runs passed.

v2.8.0

A Codex or Antigravity answer now records what the consultant's tool calls were given, so you can see which commands and files its answer rests on.

• tool_inputs keeps the shell commands Codex ran and the paths Antigravity's file tools opened: at most 30, secrets masked before each is cut to 200 characters, command output never kept. It also counts how many of the handed-over files a call names; a name in a call is not proof the file was read. The group view shows only that count (exposed_files_in_tool_inputs).
• A follow-up that asked for text only (expose_from_parent: false) is no longer told to cancel and resend files, and a follow-up to a round that exposed nothing is no longer offered expose_from_parent: true.
• usage_totals counts consultations that reported no usage (cancelled, timed out) under unreported, so the totals do not read as everything that ran.
• The Skill's first example hands over a directory of code with the question; the caveat on what a consultant can see and a few sentences of the consult_start description are shorter.
• README: the status example shows real model names, code fences carry a language (checked by a test), and follow-ups have their own section.

Install or update from the attached package:

npx -y -p https://github.com/coz-a/severally/releases/download/v2.8.0/severally-mcp-2.8.0.tgz severally-install

Restart the clients after installation. See the installation guide for options.

Validation: format, lint, strict type checks, build and generated artifacts passed. The full Linux suite, including real packed-package installation, passed 522 tests with 10 Windows-only skips and no failures. PR CI passed on Linux (Node.js 22 and 24), macOS (Node.js 24) and Windows (Node.js 24).

v2.7.0

A follow-up hands over the files of the earlier round again, as they are now, so rechecking a finding against the current code no longer needs a flag.

• Follow-ups copy again what the earlier round exposed, with any expose_paths added; a directory is listed afresh. context.expose_from_parent: false sends none of them. When the consultant's name now points at another CLI or model, nothing is copied unless the lead passes true.
• A Codex answer given without running any command on the exposed files says so, and consultants are asked to read the files the question turns on.
• A number counts as an HTTP status only where a status is written, so a token count or line number no longer classifies a failure.
• Descendants left in the process group are stopped once the consultant's CLI exits.
• The same-lineage note names the model that answered after a fallback.
• consult_get({ group_id, detail: "full" }) adds the brief and the question whole within the usual budget; each member's whole answer is consult_get({ job_id }).
• The README's first question moves to the Asking section, in English and Japanese.

Install or update from the attached package:

npx -y -p https://github.com/coz-a/severally/releases/download/v2.7.0/severally-mcp-2.7.0.tgz severally-install

Restart the clients after installation. See the installation guide for options.

Validation: format, lint, strict type checks, build and generated artifacts passed. The full Linux suite, including real packed-package installation, passed 515 tests with 10 Windows-only skips and no failures. PR CI passed on Linux (Node.js 22 and 24), macOS (Node.js 24) and Windows (Node.js 24).

v2.6.2

Group consultation previews include the answering model and findings with their grounds, making the evidence available in the first standard response for short answers.

• Keep the existing 19,500-character response budget, accounting for indentation, separators and model names. Shortened text and omitted entries remain explicit; full answers and selected finding IDs remain available.
• Update skills and MCP descriptions to avoid fetching grounds already shown whole.
• Add official CLI installation and login links. Clarify parent-history inheritance, external communication, server network defaults and existing filesystem/cancellation limits in the English and Japanese guides.

Install or update from the attached package:

npx -y -p https://github.com/coz-a/severally/releases/download/v2.6.2/severally-mcp-2.6.2.tgz severally-install

Restart the clients after installation. See the installation guide for options.

Validation: auto format, lint, strict type checks, build, generated artifacts, documentation checks and independent code review passed. The full Linux suite, including real packed-package installation, passed 509 tests with 9 Windows-only skips and no failures. PR CI passed on Linux (Node.js 22 and 24), macOS (Node.js 24) and Windows (Node.js 24).

v2.6.1

Follow-up consultations keep more of the evidence needed to revisit a decision, within the existing history size limit.

• Preserve stance, decision-changing conditions, unresolved disagreements, and check methods and expected signals across rounds. Checked evidence gets priority, and omissions are explicit.
• Classify mixed authentication/rate-limit errors as authentication failures, avoiding retries and retaining other consultants’ successful answers.
• Tighten OpenCode protected-path rules for worktree-relative reads and disable glob and grep, which can bypass per-file read rules. Symlink aliases remain outside these lexical path restrictions; the guides describe this limitation.
• Make the check → record → report workflow explicit in skills and MCP tool descriptions. Refresh the English and Japanese documentation.

Install or update from the attached package:

npx -y -p https://github.com/coz-a/severally/releases/download/v2.6.1/severally-mcp-2.6.1.tgz severally-install

Restart the clients after installation. See the installation guide for options.

Validation: the full Linux suite passed 506 tests with the real packed-package installation enabled (9 Windows-only skips, no failures). Build, formatting, lint, strict type checks, and an independent review passed. PR CI passed on Linux (Node.js 22 and 24), macOS (Node.js 24), and Windows (Node.js 24).

v2.6.0

A self-contained question can now start a consultation without separate context.facts. Review and debate keep their proposal requirements.

• Finished consultations show verdicts and reflections saved by other sessions, including in follow-up briefs and exports.
• Concurrent record updates preserve unrelated edits. Locks recover after a writer exits, and failed saves leave no unsaved changes in memory.
• JSONC rejects unfinished block comments and comma-only containers while preserving CR, LF and CRLF line endings.
• The installer shows next steps for the clients it set up. English and Japanese guides and skills include question-only consultations, targeted follow-ups and current file exposure.

Install from the attached package:

npx -y -p https://github.com/coz-a/severally/releases/download/v2.6.0/severally-mcp-2.6.0.tgz severally-install

Restart the clients after installation. See the installation guide for options.

Validation: formatting, lint and strict type checks passed. On Linux, the full suite with the packed-package installation check enabled passed 498 tests, with 9 Windows-only skips and no failures. The final MCP server checks passed 25/25.

Required CI passed on Linux (Node.js 22 and 24), macOS (Node.js 24), and Windows (Node.js 24).