ci: authenticate docs preview build (#2115)

2026-07-31 21:52:05 +01:00
parent 9a3f224224
commit 9f7662d540
5 changed files with 356 additions and 6 deletions
@@ -11,6 +11,7 @@ on:
      - "docs/public/**"
      - "scripts/validate-public-docs.mjs"
      - "scripts/validate-public-docs.test.mjs"
+      - "scripts/notify-docs-workflow.test.mjs"
      - ".github/workflows/notify-docs.yml"
  workflow_dispatch:

@@ -34,13 +35,17 @@ jobs:
          persist-credentials: false

      - name: Test public docs validator
-        run: node --test scripts/validate-public-docs.test.mjs
+        run: node --test scripts/validate-public-docs.test.mjs scripts/notify-docs-workflow.test.mjs

      - name: Validate public docs
        run: node scripts/validate-public-docs.mjs

  preview:
    name: Deploy Cloudflare Pages preview
+    outputs:
+      enabled: ${{ steps.cloudflare.outputs.enabled }}
+      deployment_url: ${{ steps.deploy.outputs.deployment-url }}
+      alias_url: ${{ steps.deploy.outputs.pages-deployment-alias-url }}
    needs: validate
    if: >-
      github.event_name == 'pull_request' &&
@@ -49,8 +54,6 @@ jobs:
    timeout-minutes: 15
    permissions:
      contents: read
-      issues: write
-      pull-requests: write

    steps:
      - name: Checkout docs pull request
@@ -87,6 +90,7 @@ jobs:
      - name: Build preview from pull request docs
        working-directory: landing
        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          KANDEV_DOCS_SOURCE_PATH: ${{ github.workspace }}/docs
        run: pnpm run build:pages

@@ -126,12 +130,23 @@ jobs:
            --branch=docs-pr-${{ github.event.pull_request.number }}
            --commit-dirty=true

+  publish-preview-link:
+    name: Publish docs preview link
+    needs: preview
+    if: needs.preview.outputs.enabled == 'true'
+    runs-on: ubuntu-latest
+    timeout-minutes: 4
+    permissions:
+      contents: read
+      issues: write
+      pull-requests: write
+
+    steps:
      - name: Publish docs preview link
-        if: steps.cloudflare.outputs.enabled == 'true'
        uses: actions/github-script@f28e40c7f34bde8b3046d885e986cb6290c5673b # v7
        env:
-          DEPLOYMENT_URL: ${{ steps.deploy.outputs.deployment-url }}
-          ALIAS_URL: ${{ steps.deploy.outputs.pages-deployment-alias-url }}
+          DEPLOYMENT_URL: ${{ needs.preview.outputs.deployment_url }}
+          ALIAS_URL: ${{ needs.preview.outputs.alias_url }}
        with:
          script: |
            const marker = '<!-- kandev-docs-cloudflare-preview -->';
@@ -0,0 +1,98 @@
+---
+spec: docs/specs/published-docs-preview-reliability/spec.md
+created: 2026-07-31
+status: implemented
+---
+
+# Implementation Plan: Published Docs Preview Reliability
+
+## Overview
+
+The cited preview failed because `.github/workflows/notify-docs.yml` launches
+landing without `GITHUB_TOKEN`; the renderer therefore consumes the hosted
+runner IP's 60-request unauthenticated quota and rejects any generated fallback
+community data. The repair first adds a static regression test for the workflow
+permission boundary, then authenticates the build with a read-only job token
+and moves preview-comment publication to a separate write-capable job.
+
+## Root Cause
+
+- The preview build step exports `KANDEV_DOCS_SOURCE_PATH` but not
+  `GITHUB_TOKEN`, confirmed by the failing log's
+  `[github-auth] configured=false` diagnostic.
+- Landing fetches repository metrics and contribution activity during static
+  generation. A non-success response produces fallback markup that
+  `build-pages.mjs` rejects.
+- GitHub's unauthenticated REST quota is IP-scoped and limited to 60 requests
+  per hour, so runner starting quota and concurrent consumption make the result
+  intermittent.
+- The existing preview job also owns issue and pull-request write permissions
+  for its final comment step. Passing that job token directly to the build
+  would make the preview reliable but violate least privilege.
+
+## Workflow
+
+### Static permission contract
+
+- Add `scripts/notify-docs-workflow.test.mjs` to inspect the checked-in
+  workflow and prove the preview build has an explicit token, the preview job
+  is read-only, and preview-link publication is isolated in a write-capable job
+  that does not build pull request content.
+- Add the new test to the workflow's pull-request path filter and validator
+  test command so future workflow changes keep exercising the contract.
+
+### Authenticated preview build
+
+- Restrict `jobs.preview.permissions` in
+  `.github/workflows/notify-docs.yml` to `contents: read`.
+- Export `GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}` only from the
+  `Build preview from pull request docs` step. Landing already recognizes this
+  environment variable through its existing `githubApiHeaders()` contract.
+- Export the Cloudflare-enabled flag, deployment URL, and alias URL as preview
+  job outputs.
+
+### Isolated preview comment
+
+- Move `Publish docs preview link` into a `publish-preview-link` job that needs
+  `preview`, runs only when preview deployment is enabled, and owns
+  `issues: write` plus `pull-requests: write`.
+- Pass deployment outputs through `needs.preview.outputs` and retain the
+  existing idempotent marker-based create-or-update script.
+
+## Tests
+
+- **What:** the landing build receives an explicit authenticated token while
+  its job cannot write issues or pull requests.
+  **File:** `scripts/notify-docs-workflow.test.mjs`.
+  **How:** extract the checked-in preview job and named build step from the YAML
+  text, then assert the token mapping and permission exclusions.
+- **What:** preview comment publication retains write permissions without
+  executing the landing build.
+  **File:** `scripts/notify-docs-workflow.test.mjs`.
+  **How:** extract the publication job and assert its dependency, output inputs,
+  permissions, and absence of checkout/build steps.
+- **What:** public-doc validation and the workflow contract remain green under
+  the exact command used by CI.
+  **File:** `.github/workflows/notify-docs.yml`.
+  **How:** run both Node test files together, then run the public-doc validator.
+
+## Implementation Waves And Parallel Candidates
+
+Execution is sequential because the regression test and workflow repair share
+the same permission and job-output contract.
+
+- [x] [Task 01: Authenticate and isolate docs preview publication](task-01-authenticate-and-isolate-preview.md) — done
+
+## Documentation Impact
+
+The concise repair spec records CI behavior and permissions. No public CLI,
+configuration, API, deployment, or UI documentation changes are required.
+
+## Risks
+
+- Job outputs must preserve the exact Cloudflare action output values across
+  the new job boundary.
+- The publication job must remain skipped when Cloudflare is unconfigured or
+  preview fails.
+- Exposing the token at workflow or job scope would unnecessarily widen its
+  availability; the implementation must keep it scoped to the build step.
@@ -0,0 +1,71 @@
+---
+id: "01-authenticate-and-isolate-preview"
+title: "Authenticate and isolate docs preview publication"
+status: done
+wave: 1
+depends_on: []
+plan: "plan.md"
+spec: "../../specs/published-docs-preview-reliability/spec.md"
+---
+
+# Task 01: Authenticate and Isolate Docs Preview Publication
+
+## Acceptance
+
+- The landing build receives `${{ secrets.GITHUB_TOKEN }}` while its job has
+  only `contents: read`.
+- Preview deployment URLs flow to a separate comment-publication job with issue
+  and pull-request write permissions, and that job never checks out or builds
+  pull request content.
+- A checked-in regression test enforces both permission boundaries and runs in
+  the public-doc validation job.
+
+## Verification
+
+```bash
+node --test scripts/notify-docs-workflow.test.mjs
+node --test scripts/validate-public-docs.test.mjs scripts/notify-docs-workflow.test.mjs
+node scripts/validate-public-docs.mjs
+```
+
+## Files likely touched
+
+- `.github/workflows/notify-docs.yml`
+- `scripts/notify-docs-workflow.test.mjs`
+
+## Dependencies
+
+None.
+
+## Parallelism
+
+`sequential`; the test and workflow share one security contract.
+
+## Inputs
+
+- Repair spec scenarios and permission boundary.
+- Failing job `91236751922`, which reported
+  `[github-auth] configured=false` and rejected unavailable community data.
+- Passing job `91230452438`, which used the same unauthenticated path with more
+  starting quota.
+- Landing `scripts/build-pages.mjs`, `lib/github-community.ts`, and
+  `packages/site/src/github-api.ts` contracts on `kdlbs/landing` `main`.
+
+## Output contract
+
+Report the RED workflow-contract failure, minimal workflow repair, files
+changed, exact verification results, residual risks, and update this task plus
+`plan.md` status in the same conversation.
+
+## Result
+
+- RED coverage failed against the original workflow because the preview job
+  still had issue/pull-request write permissions, the build had no
+  `GITHUB_TOKEN`, and no isolated publication job existed.
+- The preview job now uses `contents: read`, passes `GITHUB_TOKEN` only to the
+  landing build, and exports deployment URLs to a separate write-capable
+  publication job.
+- Verified with `node --test scripts/notify-docs-workflow.test.mjs` (2 tests),
+  `node --test scripts/validate-public-docs.test.mjs scripts/notify-docs-workflow.test.mjs`
+  (60 tests), `node scripts/validate-public-docs.mjs` (41 pages),
+  `git diff --check`, and a PyYAML parse of the workflow.
@@ -0,0 +1,72 @@
+---
+status: shipped
+created: 2026-07-31
+owner: kdlbs
+---
+
+# Published Docs Preview Reliability
+
+## Why
+
+Pull request documentation previews intermittently fail when the landing build
+cannot fetch live public GitHub community data. Contributors need preview
+results to depend on the proposed documentation and publisher contract, not on
+the shared unauthenticated GitHub API quota of a hosted runner.
+
+## What
+
+- Same-repository pull request previews authenticate landing's GitHub API
+  reads with the workflow job's short-lived `GITHUB_TOKEN`.
+- The landing build receives only a token whose job permissions are restricted
+  to read-only repository contents.
+- Publishing or updating the preview link remains available through a separate
+  job with issue and pull-request write permissions.
+- The workflow retains its existing same-repository pull request gate,
+  Cloudflare deployment behavior, stable alias, and preview comment format.
+- A repository test fails if the build loses authenticated GitHub API access or
+  regains issue or pull-request write permissions.
+
+## Permissions
+
+- The preview build and Cloudflare deployment job has `contents: read` and no
+  issue or pull-request write permission. Its `GITHUB_TOKEN` is exposed only to
+  the landing build step for authenticated public GitHub API reads.
+- The preview-link publication job has `contents: read`, `issues: write`, and
+  `pull-requests: write`. It consumes only the deployment URLs exported by the
+  successful preview job and does not check out or build pull request content.
+- Fork pull requests remain excluded from the preview path.
+
+## Failure modes
+
+- A failed landing build or Cloudflare deployment fails the preview job and
+  prevents preview-link publication.
+- Missing Cloudflare configuration retains the existing warning-and-skip
+  behavior and does not start the publication job.
+- A missing deployment URL fails the publication job rather than publishing an
+  incomplete comment.
+- GitHub API failure after authenticated access remains a real build failure;
+  the workflow does not publish stale or fallback community data as a valid
+  preview.
+
+## Scenarios
+
+- **GIVEN** a same-repository pull request changes public documentation,
+  **WHEN** the landing preview build requests public Kandev repository data,
+  **THEN** the request uses the job-scoped `GITHUB_TOKEN` under read-only
+  contents permissions.
+- **GIVEN** a successful configured Cloudflare preview deployment, **WHEN** the
+  deployment exports its URL and stable alias, **THEN** a separate write-capable
+  job creates or updates the pull request preview comment.
+- **GIVEN** the preview build executes pull request documentation, **WHEN** that
+  process inspects its job token, **THEN** the token has no issue or pull-request
+  write permission.
+- **GIVEN** Cloudflare preview deployment is not configured or the preview job
+  fails, **WHEN** downstream jobs are evaluated, **THEN** no preview-link
+  publication job runs.
+
+## Out of scope
+
+- Changing landing's community-data fallback or artifact validation logic.
+- Supporting preview deployments for fork pull requests.
+- Changing Cloudflare credentials, project configuration, deployment aliases,
+  or the production rebuild hook.
@@ -0,0 +1,94 @@
+import assert from "node:assert/strict";
+import fs from "node:fs/promises";
+import path from "node:path";
+import test from "node:test";
+import { fileURLToPath } from "node:url";
+
+const repoRoot = path.resolve(
+  path.dirname(fileURLToPath(import.meta.url)),
+  "..",
+);
+const workflowPath = path.join(
+  repoRoot,
+  ".github/workflows/notify-docs.yml",
+);
+
+function extractJob(workflow, name) {
+  const marker = `  ${name}:\n`;
+  const start = workflow.indexOf(marker);
+  assert.notEqual(start, -1, `workflow job not found: ${name}`);
+
+  const bodyStart = start + marker.length;
+  const remainder = workflow.slice(bodyStart);
+  const nextJob = remainder.search(/\n  [a-z0-9-]+:\n/);
+  return remainder.slice(0, nextJob === -1 ? remainder.length : nextJob);
+}
+
+function extractStep(job, name) {
+  const marker = `      - name: ${name}\n`;
+  const start = job.indexOf(marker);
+  assert.notEqual(start, -1, `workflow step not found: ${name}`);
+
+  const bodyStart = start + marker.length;
+  const remainder = job.slice(bodyStart);
+  const nextStep = remainder.search(/\n      - name: /);
+  return remainder.slice(0, nextStep === -1 ? remainder.length : nextStep);
+}
+
+const tokenExpression = String.raw`\$\{\{ secrets\.GITHUB_TOKEN \}\}`;
+
+test("docs preview build uses an authenticated read-only token", async () => {
+  const workflow = await fs.readFile(workflowPath, "utf8");
+  const preview = extractJob(workflow, "preview");
+  const build = extractStep(preview, "Build preview from pull request docs");
+  const validate = extractJob(workflow, "validate");
+  const testStep = extractStep(validate, "Test public docs validator");
+
+  assert.match(
+    preview,
+    /permissions:\n      contents: read\n\n    steps:/,
+    "preview job permissions must be limited to contents: read",
+  );
+  assert.match(build, new RegExp(`GITHUB_TOKEN: ${tokenExpression}`));
+  assert.doesNotMatch(
+    preview.replace(build, ""),
+    /GITHUB_TOKEN/,
+    "GITHUB_TOKEN must remain scoped to the landing build step",
+  );
+  assert.match(testStep, /scripts\/validate-public-docs\.test\.mjs/);
+  assert.match(testStep, /scripts\/notify-docs-workflow\.test\.mjs/);
+  assert.match(workflow, /- "scripts\/notify-docs-workflow\.test\.mjs"/);
+});
+
+test("preview comment publication is isolated from pull request content", async () => {
+  const workflow = await fs.readFile(workflowPath, "utf8");
+  const preview = extractJob(workflow, "preview");
+  const publication = extractJob(workflow, "publish-preview-link");
+  const publishStep = extractStep(publication, "Publish docs preview link");
+
+  assert.match(
+    preview,
+    /outputs:\n      enabled: \$\{\{ steps\.cloudflare\.outputs\.enabled \}\}\n      deployment_url: \$\{\{ steps\.deploy\.outputs\.deployment-url \}\}\n      alias_url: \$\{\{ steps\.deploy\.outputs\.pages-deployment-alias-url \}\}/,
+  );
+  assert.match(publication, /needs: preview/);
+  assert.match(
+    publication,
+    /if: needs\.preview\.outputs\.enabled == 'true'/,
+  );
+  assert.match(publication, /permissions:\n      contents: read\n      issues: write\n      pull-requests: write\n/);
+  assert.match(
+    publishStep,
+    new RegExp(
+      `DEPLOYMENT_URL: ${String.raw`\$\{\{ needs\.preview\.outputs\.deployment_url \}\}`}`,
+    ),
+  );
+  assert.match(
+    publishStep,
+    new RegExp(
+      `ALIAS_URL: ${String.raw`\$\{\{ needs\.preview\.outputs\.alias_url \}\}`}`,
+    ),
+  );
+  assert.match(publishStep, /uses: actions\/github-script@/);
+  assert.doesNotMatch(publication, /actions\/checkout@/);
+  assert.doesNotMatch(publication, /pnpm run build:pages/);
+});