@calltelemetry/ct-release-mcp

中文文档
0.3.24发布于 2026/11/08MITTypeScriptES Module近 30 天下载 2133

Governed Call Telemetry release planning MCP server and CLI

原始使用文档
文档目录7 节
  1. Install
  2. Package release workflow recovery
  3. Authentication
  4. Catalog source
  5. Command behavior
  6. CLI reference
  7. Development

ct-release-mcp

Governed, read-first release planning for Call Telemetry agents and operators.

@calltelemetry/ct-release-mcp exposes the same commands as stdio MCP tools, federated by ct-mcp, and as a structured CLI:

  • get_release_state
  • get_stable_nightly
  • get_release_train
  • list_versions
  • plan_release
  • show_plan
  • validate_plan
  • get_appliance_build

The package does not dispatch Unified Release and cannot create tags, releases, images, bundles, or channel markers. plan_release is the one mutating command: it dispatches only the side-effect-free Release Plan workflow. Candidate identities come only from a successful calltelemetry/ct-release Release Plan workflow artifact.

Install

Run the pinned package with no arguments as a stdio MCP server:

{
  "command": "npx",
  "args": ["-y", "@calltelemetry/[email protected]"]
}

Run any governed command directly through the same pinned package:

npx -y @calltelemetry/[email protected] get_release_state \
  --channel stable \
  --line 0.8.7

npx -y @calltelemetry/[email protected] plan_release \
  --target-branch 0.8.7-stable

Use --input-json to pass the exact MCP input object without translating field names:

npx -y @calltelemetry/[email protected] list_versions \
  --input-json '{"line":"0.8.7","limit":20,"includePrereleases":true}'

Node.js 22 or newer is required.

Package release workflow recovery

The docker.yml workflow runs only for pushes to main. After registry login but before any versioned-image push, it checks the exact package-version git tag and both GHCR versioned image aliases (vX.Y.Z and X.Y.Z). An existing release is accepted only when both aliases share one digest and both platform images identify the tagged commit as their source. If verified, it skips the build and npm dispatch. A push to a git tag does not rebuild or overwrite that image. Image builds for this repository are serialized, including workflow reruns. Pending main pushes use GitHub's expanded concurrency queue, which admits up to 100 waiting jobs; overflow is canceled and must be investigated. Since GitHub does not guarantee dispatch order, an older versioned build does not move latest when main has a different package version.

If the git tag or either image alias exists without the complete set, the digests or source commit differ, or any lookup fails, the workflow stops before pushing an image. Investigate the exact tag, image, and failed run before recovery; do not rerun a build against an existing image or move an existing tag. A package that was already published should not be published again. The hosted MCP is the Cloudflare Worker that ci.yml deploys from main.

Authentication

The server resolves GitHub authentication in this order:

  1. An explicit token supplied by an embedding application
  2. GH_TOKEN
  3. GITHUB_TOKEN
  4. The active gh auth token

The credential must be able to read the governed Call Telemetry repositories, release records, tags, workflow runs, and artifacts. plan_release also needs permission to dispatch the calltelemetry/ct-release Release Plan workflow. Tokens are never returned in tool results or error context.

User and PAT credentials are correlated to the workflow actor returned by GET /user. GitHub App installation tokens cannot call that endpoint, so the server proves installation identity through the installation-only repository listing and requires access to calltelemetry/ct-release. It then derives the App bot identity from the token-authenticated GraphQL viewer and binds both the workflow actor and triggering_actor login, database ID, and node ID to that identity. Only first-attempt runs can correlate. Optionally set CT_RELEASE_GITHUB_APP_ACTOR to a trusted App bot login (for example, ct-pr-agent[bot]) as an additional allowlist. The workflow ref defaults to main and can be configured with CT_RELEASE_GITHUB_DEFAULT_BRANCH. Installation-token dispatches remain bound to the exact App identity, repository, workflow ref, target-branch nonce, prior-run snapshot, and bounded dispatch time; a generic 403 is never accepted as installation identity.

Catalog source

0.3.15 reads scripts/release-storage.json from the release repository through the existing authenticated GitHub client. Each operation resolves its catalog URL once from the configured Cloudflare R2 release origin (https://get.calltelemetry.com), and reports that same source. Configuration and fetch failures stop planning before branch snapshot or dispatch. Configuration is read again for the next operation; a failed read does not become a cached source.

The configuration read is bounded to 16 KiB. Provider, schema, bucket, region, endpoint and public origin must agree. Catalog fetch bounds, train validation, plan identity and the shared MCP/CLI command registry retain their existing contracts.

Command behavior

MCP tool names and CLI subcommands are identical. CLI field flags use the MCP camelCase name as canonical spelling and accept a kebab-case alias, such as --includePrereleases or --include-prereleases. Boolean flags require an explicit true or false; integers must be safe integers. --input-json cannot be combined with field flags. The adapter-only, valueless --pretty flag may appear anywhere after a command and indents JSON without changing the MCP-equivalent input or result.

  • get_release_state reads the public appliance catalog, the lifecycle and composition ledgers, GitHub releases, and component tags. Catalog and release records are primary authority; component tags are labeled secondary evidence. It exposes the promoted channel tip (tip_marker ?? tip), downloadable catalogDownloadTip, the bundle-only fallback flags (catalogTipFallback/catalogTipMarkerDeployable), materialized candidate head (catalogVersionTip), and variant. A newer candidate is not promoted authority and is not drift. A sanctioned bundle-only fallback (tip_fallback: true, tip_marker_deployable: false) is also not drift: the marker stays the release pin while catalogDownloadTip names the older appliance. A final channel that its lifecycle ledger explicitly marks unpublished returns channelStatus: "not_published" instead of false catalog/owner drift. Secondary component tag lag is returned in advisories; drift is reserved for authoritative disagreements. Secondary evidence reads use GitHub's release-line matching-ref endpoint instead of crawling every component tag page. When trains/<line>.yaml owns stable-nightly, get_release_state also surfaces stableNightly with explicit non-qualification (qualifiedForStable: false) while keeping catalogTip strictly bound to promoted mainline authority (null when channel is stable-nightly).
  • get_stable_nightly (CLI alias: stable-nightly) reads the opt-in channels/stable-nightly.json pointer from the R2 release origin with a 256 KiB bound, rejects duplicate JSON keys and non-integer numbers, verifies the canonical JSON SHA-256 admission_id, enforces qualified_for_stable: false and package_origin provenance, verifies that the pointer line owns stable-nightly in trains/<line>.yaml, and evaluates readyForStablePromotion against promotion_receipts.stable (security and uat verification lanes). HTTP 404 returns status: "uninitialized" rather than treating an unpublished nightly channel as an operational crash.
  • get_release_train returns the separately versioned lifecycle and component composition documents with their exact Git SHAs.
  • list_versions lists catalog and ct-release versions for one release line, including clearly labeled draft identity reservations, and includes the active stable-nightly prerelease record (qualified_for_stable: false) when includePrereleases is true on a line that owns stable-nightly.
  • plan_release dispatches only the side-effect-free Release Plan workflow, then promptly returns its exact run ID and pre-dispatch authority snapshot. It also validates the requested stable branch against the lifecycle and component manifests, snapshots the exact calltelemetry/cisco-cdr branch SHA, and dispatches the resulting four-field release target without adding caller-facing inputs. This exact-target contract became mandatory in calltelemetry/ct-release commit 579ff34720cf1a273201f12174b242ec926ebc39 (PR #1570); v0.3.4 was tagged but not published after its trusted-publisher workflow identity was rejected. @calltelemetry/[email protected] is the first published coordinated caller version, and 0.3.6 adds strict validation of the target persisted in current plan artifacts and dispatch inputs. 0.3.7 introduced direct catalog reads for package-side planning; its GCS default is retired in 0.3.15. The hosted Release Plan authority must independently use the same canonical source; this package change alone does not prove a plan succeeds. Current artifacts require a candidate-bound R2 release_bundle on https://get.calltelemetry.com; retired gcs_bundle artifacts cannot qualify current release planning. 0.3.12 stages the plan-artifact transition: contract version 1 remains readable only in its legacy no-train shape, while contract version 2 requires the governed top-level train. Exact release targets may identify either the cisco-cdr/Web or ct-quasar/Vue branch-latest snapshot; the selected repository's source SHA must match that immutable snapshot. 0.3.13 supports rotating GitHub App installation tokens without weakening workflow-run correlation or accepting unrelated PAT authorization failures. The hosted workflow rejects a plan if that bound authority changes before artifact creation. Use show_plan after completion to read the immutable artifact. An explicitly active, unpublished train without a public channel can plan its first release with a null catalog tip. This requires no contradictory catalog artifact or published final and preserves the exact catalog, lifecycle, and composition snapshot; it never borrows another line's channel or creates a public marker. 0.3.23 adds contract version 4 for authenticated Quasar image identities; versioned Quasar tags require the exact source, producer workflow, recipe, build-input, platform, candidate, and attempt projection. Legacy short-SHA snapshots remain supported, while contract version 3 stays reserved for the separate 0.8.6 fixed-snapshot shape.
  • show_plan retrieves and verifies a successful plan artifact by workflow run ID.
  • get_appliance_build reads one calltelemetry/ai-workspace Build OVA Appliance run by id, or the newest run whose title contains the version, image format, and correlation id. It returns status, the current step, and the failed step. It does not dispatch or publish.
  • validate_plan re-reads current authority and fails closed on changed train SHAs, invalid candidate identity, dropped or changed dispatch pins, or known authority drift. For a still-unpublished train with no channel, current state reports channel: null, channelOwner: null, and channelStatus: "not_published". An explicitly hidden channel (customer_visible: false) with a null promoted marker may validate a newer, immutable draft-only mainline RC under a v2 or v4 plan contract. The active lifecycle train must be explicitly unpublished, and its latest_stable must equal the same-line published previous final. Only CATALOG_LEDGER_MISMATCH and CATALOG_RELEASE_MISMATCH entries with matching owner, requested line, null marker and previous-final contexts qualify; unknown visibility or incomplete/mismatched drift context fails closed. Beta promotion must be false in both the plan and dispatch inputs, and OVA dispatch must be false with no enabling override. An accepted plan returns the nonblocking DRAFT_RC_WITHOUT_STABLE_MARKER entry in its optional advisories array; rejected plans omit it, and current state retains the missing-marker drift. This exception never authorizes a final, alias, OVA dispatch, automatic promotion, or any other authority disagreement.

Named variant RCs such as 0.8.7-i18n-rc17 remain separate from the normal channel versions. They are sorted by their numeric RC suffix and never satisfy normal channel ownership checks. Normal <line>-stable planning always binds to a mainline channel even when a variant preview is the only customer-visible channel on that line.

plan_release returns after exact run correlation rather than waiting for the hosted workflow, keeping it inside ordinary MCP request deadlines. While the run is queued or active, show_plan returns retryable PLAN_RUN_PENDING; a failed workflow returns PLAN_RUN_FAILED with its run URL.

The plan schema retains the producer's readiness candidate and OCI index/amd64/arm64 digests in the authenticated receipt. Frozen plans require a complete digest set bound to a readiness candidate; legacy 0.8.6 fixed-pin plans may carry null digests and an empty candidate. Historical artifacts that omit these fields remain readable. Malformed or partial bindings and unrecognized fields remain errors.

All MCP and CLI command errors use a structured { error: { code, message, retryable, context } } shape. CLI command results and errors are one-line JSON on stdout by default, or indented JSON with --pretty. Help and version are plain text. CLI exit codes are 0 for a successful command, 2 for CLI usage or input-schema errors, and 1 for operational failures. A successful validate_plan call can still return valid: false; that is the governed result when authoritative sources disagree.

Use release-control dispatch --plan-run <run-id> only after show_plan and validate_plan succeed. That separate dispatcher owns Unified Release. This package deliberately stops at authoritative planning and validation.

CLI reference

ct-release-mcp get_release_state --channel <channel> --line <line>
ct-release-mcp get_stable_nightly [--line <line>]
ct-release-mcp stable-nightly [--line <line>]
ct-release-mcp get_release_train --line <line>
ct-release-mcp list_versions --line <line> [--limit <n>] [--includePrereleases true|false]
ct-release-mcp plan_release --targetBranch <branch> [--version <candidate>]
ct-release-mcp show_plan --planRunId <run-id>
ct-release-mcp validate_plan --planRunId <run-id>
ct-release-mcp get_appliance_build [--runId <run-id>] [--version <version>] [--correlationId <id>] [--imageFormat qcow2|ova]

Run ct-release-mcp --help or ct-release-mcp <command> --help for the live reference generated from the shared command registry.

Development

npm ci
npm run validate

版本与发布记录

@calltelemetry/ct-release-mcp 共发布 34 个版本,首次发布于 2026/08/25;过去 12 个月发布了 34 个版本。

版本发布日期距今
0.3.24latest2026/11/081 天前
0.3.232026/11/062 天前
0.3.222026/11/027 天前
0.3.212026/10/2515 天前
0.3.202026/10/2317 天前
0.3.192026/10/2218 天前

关于 @calltelemetry/ct-release-mcp 的常见问题

@calltelemetry/ct-release-mcp 是什么?

Governed Call Telemetry release planning MCP server and CLI

如何安装 @calltelemetry/ct-release-mcp?

在项目目录运行 npm install @calltelemetry/ct-release-mcp,或使用 pnpm add / yarn add / bun add。它还提供命令行工具 ct-release-mcp,可以用 npx ct-release-mcp 直接运行。

@calltelemetry/ct-release-mcp 支持 TypeScript 吗?

支持。包内自带 TypeScript 类型声明(package.json 声明了 types 入口),安装后即可获得类型提示,无需再安装 @types 包。

@calltelemetry/ct-release-mcp 支持 ESM 还是 CommonJS?

以 ES Module 形式发布,推荐使用 import 引入。在 CommonJS 项目中,Node.js 20.19+ / 22.12+ 可以直接 require() 不含顶层 await 的 ES Module,更早的版本需要改用动态 import()。

@calltelemetry/ct-release-mcp 需要什么版本的 Node.js?

package.json 的 engines 字段要求 Node.js >=22.0.0。版本不符时 npm 默认只给出警告,开启 engine-strict 后会拒绝安装。

@calltelemetry/ct-release-mcp 可以商用吗?

@calltelemetry/ct-release-mcp 使用 MIT 许可。宽松许可:允许商业使用、修改和再分发,通常只需在分发时保留原版权与许可声明。以上为概括,不构成法律意见,请以许可原文为准。

@calltelemetry/ct-release-mcp 的下载量有多少?

根据 npm 官方下载统计,@calltelemetry/ct-release-mcp 在 2026/10/09 – 2026/11/07 这 30 天内被下载 2,133 次,日均约 71 次。下载量会把 CI 构建、镜像同步和作为其他包依赖的安装一并计入,适合横向比较热度,不能直接等同于使用人数。

@calltelemetry/ct-release-mcp 还在维护吗?

最新版本 0.3.24 发布于 2026/11/08(1 天前);至今共发布 34 个版本,其中过去 12 个月发布了 34 个;首次发布于 2026/08/25。发布频率只是参考信号:功能稳定的工具可能很少需要更新,建议结合源码仓库的 issue 和提交记录判断。