Repository navigation
Conversation
The interface diff ran on any PR that touched the root lockfile, which is almost every TypeScript PR, and covered only `everything`. - scripts/lib/interface-diff-servers.mjs: the server table (how each of the seven boots for a probe) and the selection: a server's own files, the root tsconfig (TypeScript servers), a package-lock.json edit that changes a package the server resolves (its tree walked in both lockfiles), and every server when the diff tooling or the pinned mcp-server-diff changes. - scripts/interface-diff.mjs: --server, --all and --list; builds each selected base (one npm ci for the TypeScript ones, uv sync for the Python ones) and probes Python servers through uv run --no-sync; one report per server under <report-dir>/<server>/. - everything-mcp-diff.yml -> mcp-interface-diff.yml: a `changes` job lists the servers, `diff` and `comment` fan out per server with a per-server sticky comment and artifact. The security split, the release and dispatch triggers and the fork fallback are kept. Both version-PR workflows dispatch it, since the CalVer stamps now touch a diffed server. - The gate's stage diffs only the servers changed since the merge-base, through the same selection. - Docs: quality-gate.md, AGENTS.md, RELEASING.md, the pre-push-gate and release skills. Closes #5428 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: cliffhall <cliff@futurescale.com>
|
|
| Tools | Prompts | Resources | Resource templates | |
|---|---|---|---|---|
| Base | 22 | 4 | 8 | 5 |
| Head | 22 | 4 | 8 | 5 |
Updated by this run for b3f714938d0bfb0bd5046ac4916e8393d20c4731.
|
| Tools | Prompts | Resources | Resource templates | |
|---|---|---|---|---|
| Base | 16 | 0 | 0 | 1 |
| Head | 16 | 0 | 0 | 1 |
Updated by this run for b3f714938d0bfb0bd5046ac4916e8393d20c4731.
|
| Tools | Prompts | Resources | Resource templates | |
|---|---|---|---|---|
| Base | 1 | 1 | 0 | 0 |
| Head | 1 | 1 | 0 | 0 |
Updated by this run for b3f714938d0bfb0bd5046ac4916e8393d20c4731.
|
| Tools | Prompts | Resources | Resource templates | |
|---|---|---|---|---|
| Base | 14 | 2 | 0 | 0 |
| Head | 14 | 2 | 0 | 0 |
tools
--- base/tools.json
+++ target/tools.json
+ tools[git_log].inputSchema.properties.max_count.description: "Maximum number of commits to show. Must be at least 1."
+ tools[git_log].inputSchema.properties.max_count.minimum: 1Updated by this run for b3f714938d0bfb0bd5046ac4916e8393d20c4731.
|
| Tools | Prompts | Resources | Resource templates | |
|---|---|---|---|---|
| Base | 3 | 0 | 0 | 0 |
| Head | 3 | 0 | 0 | 0 |
Updated by this run for b3f714938d0bfb0bd5046ac4916e8393d20c4731.
|
| Tools | Prompts | Resources | Resource templates | |
|---|---|---|---|---|
| Base | 9 | 0 | 1 | 0 |
| Head | 9 | 0 | 1 | 0 |
Updated by this run for b3f714938d0bfb0bd5046ac4916e8393d20c4731.
|
| Tools | Prompts | Resources | Resource templates | |
|---|---|---|---|---|
| Base | 1 | 0 | 0 | 0 |
| Head | 1 | 0 | 0 | 0 |
Updated by this run for b3f714938d0bfb0bd5046ac4916e8393d20c4731.
There was a problem hiding this comment.
🟡 Changes recommended
The renamed workflow is not yet dispatchable on the default branch, blocking required interface-diff checks on version PRs, and C-quoted Git paths can cause changed servers to be skipped and falsely pass the diff.
2 open issues
What changed in this PR
Extends MCP interface diffs from everything to all seven reference servers, while selecting only servers affected by a change, as requested in #5428.
Changes:
- Shares server selection between CI and the local gate, including resolved dependency-tree comparisons.
- Adds server-specific probes, reports, artifacts and sticky PR comments.
- Updates version-workflow dispatches, tests and maintainer documentation.
| File | Description |
|---|---|
scripts/version-pr-ci.test.mjs |
Updates dispatch targets and expected status names. |
scripts/lib/interface-diff-servers.test.mjs |
Tests dependency resolution and server selection. |
scripts/lib/interface-diff-servers.mjs |
Defines server launch settings and selection rules. |
scripts/interface-diff.test.mjs |
Tests new flags, launch commands and reports. |
scripts/interface-diff.mjs |
Builds and probes selected servers with individual reports. |
RELEASING.md |
Documents interface-diff dispatches for version PRs. |
docs/quality-gate.md |
Explains selective diffs in CI and the local gate. |
AGENTS.md |
Updates workflow references and gate descriptions. |
.github/workflows/version-packages.yml |
Dispatches the renamed interface-diff workflow. |
.github/workflows/prepare-python-release.yml |
Adds interface-diff dispatches for Python version PRs. |
.github/workflows/mcp-interface-diff.yml |
Adds selection and per-server diff and comment matrices. |
.claude/skills/release/SKILL.md |
Updates release status and dispatch instructions. |
.claude/skills/pre-push-gate/SKILL.md |
Documents per-server reruns and troubleshooting. |
Review effort: Balanced
| for workflow in typescript.yml python.yml dco.yml mcp-interface-diff.yml; do | ||
| gh workflow run "$workflow" --ref "$BRANCH" |
There was a problem hiding this comment.
Fixed in b3f7149: the file keeps its name, everything-mcp-diff.yml, so both version workflows still dispatch a file that exists on main. Its display name ("MCP Interface Diff") and status context (Interface diff (dispatched)) still change, because those come from the file content at the dispatched ref. The workflow header now explains why the name stays. The rename itself is filed as #5463, staged across a milestone merge.
| const lines = (text) => text.split("\n").filter(Boolean); | ||
| let baseLock = null; | ||
| try { | ||
| baseLock = JSON.parse(git("show", `${sha}:package-lock.json`)); | ||
| } catch { | ||
| // No lockfile at the base: every locked package reads as changed. | ||
| } | ||
| const headFile = path.join(repoRoot, "package-lock.json"); | ||
| return { | ||
| changedFiles: [ | ||
| ...lines(git("diff", "--name-only", "--no-renames", sha)), | ||
| ...lines(git("ls-files", "--others", "--exclude-standard")), |
There was a problem hiding this comment.
Fixed in b3f7149: both git diff --name-only and git ls-files --others now use -z and split on NUL. They are also called through run rather than git(), which trims its output and could clip a leading space. Checked with an untracked src/time/tests/test_é.py: git prints it C-quoted ("src/time/tests/test_\303\251.py"), and --list now selects time.
…elimited - GitHub dispatches only a workflow whose file is on the default branch, and both version workflows dispatch this one by file name from v2/main, so the rename to mcp-interface-diff.yml would break their dispatch until the next milestone merge. The file stays everything-mcp-diff.yml; its display name and status context still change. - git diff --name-only and ls-files C-quote a path holding a non-ASCII character, a tab or a newline, which then matched no server directory and passed with nothing to diff. Both now use -z. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: cliffhall <cliff@futurescale.com>
|
Copilot round 1: both findings were defects in what this PR added, and both are fixed in b3f7149 (gate green).
The PR body is updated to match. |
Closes #5428
Description
The MCP interface diff now runs only for the servers a change touches, and it covers all seven servers, not just
everything.scripts/lib/interface-diff-servers.mjs, used by both CI and the gate). Changed paths are read NUL-delimited, so Git's C-quoting cannot hide a change. A server is diffed when one of these changes:src/<server>/(for a Python server this includes itsuv.lock)tsconfig.json, which every TS server extendspackage-lock.jsonedit that changes a package that server resolves. The workspace's dependency tree is walked in the base and head lockfiles the way Node resolves it, so a lockfile rewrite that reaches no server selects none.win-shell-args, the workflow) or the pinnedmcp-server-diffversion changes. A transitive bump undermcp-server-diffdoes not count, because it moves no probe.scripts/interface-diff.mjs--server <name>(repeatable),--all, and--list, which prints the selection as JSON and diffs nothingnpm ciand are then built one by one; each Python base gets its ownuv sync --frozen --no-devuv run --frozen --no-sync --project src/<server>, using the same "read an environment, don't build one" rule as the smoke stagefilesystem, a throwaway repo forgit, a pinned zone fortime, and a throwaway graph file formemory<report-dir>/<server>/everything-mcp-diff.ymlkeeps its file name (renaming it needs to be staged across a milestone merge; see CI: rename everything-mcp-diff.yml to mcp-interface-diff.yml, staged across a milestone merge #5463) but is now "MCP Interface Diff"changesjob resolves the base and runsnpm run interface-diff -- --list(no install needed)diffandcommentfan out as a matrix, one leg per server. Each server gets its own sticky comment (marker<!-- mcp-interface-diff:<server> -->) and its ownmcp-diff-report-<server>artifact.releaseandworkflow_dispatchtriggers, and the fork-PR fallback to the summary and artifactcommentchecks each matrix server name against the known list before using it, because that list came from a job that ran PR codeInterface diff (dispatched)interface-diffstage uses the same selection against the merge-base withorigin/v2/main.docs/quality-gate.md,AGENTS.md(the tree entry and the gate description),RELEASING.md, thepre-push-gateandreleaseskills, andversion-pr-ci.test.mjs.Server Details
Motivation and Context
The issue gives two examples. #5426 (a
filesystemminimatch bump) and #5427 (a release audit touching only the rootpackage.jsonandpackage-lock.json) both ran theeverythingdiff, because the workflow'spathsfilter matched the root lockfile. The other six servers had no interface diff at all.How Has This Been Tested?
This change has no client-facing surface, so the evidence is targeted probes.
Selection replayed on real merges (each merge commit against its first parent, with the real lockfiles):
.changeset/…,package-lock.json,src/filesystem/package.jsoneverythingfilesystempackage.json,package-lock.json(aqsoverride bump)everythingeverything.qsis in its resolved tree via express; the other servers don't load it.src/everything/…everythingeverythingsrc/sequentialthinking/…sequentialthinkingEvery server probes cleanly:
npm run interface-diff -- --allagainstorigin/v2/maintakes about 13 seconds and exits 0. The head counts (tools / prompts / resources / templates):A Python change is detected: I edited the description of the
get_current_timetool in the working tree and rannpm run interface-diff -- --server time. Exit 0, and it reported:This PR's own CI run changes the tooling, so the
changesjob should select all seven servers and post seven sticky comments. That is the live check of the matrix and comment jobs.Tests:
scripts/lib/interface-diff-servers.test.mjs(new) covers lockfile resolution and closure, selection by own files,tsconfigand tooling, and lockfile edits that must not select a server. It also pins that the server table matches thesrc/directories.scripts/interface-diff.test.mjscovers the new flags,chooseServersandstartCommand.npm run local:gateexits 0.Breaking Changes
None for users. For maintainers:
Interface diff (dispatched)(the workflow file name is unchanged)everythingsticky comment uses a new marker, so a PR that is open at merge gets a new comment beside the old one--report-dirnow writes<dir>/<server>/report.mdTypes of changes
Checklist
npm run changeset) if this changes what a TypeScript server publishes (not applicable: nothing published changes)docs/quality-gate.mdand thepre-push-gateskill)Additional context
--no-renames) and includes uncommitted and untracked files locally, because the gate probes the working tree's build.🤖 Generated with Claude Code