Add Cloudflare Pages docs preview with /preview-docs slash command

[localden]

Summary

Adds per-PR documentation previews on Cloudflare Pages so reviewers can see the rendered mkdocs site before merge.

• Auto-preview: PRs from repo admins that touch docs/**, docs_src/**, mkdocs.yml or pyproject.toml build and deploy automatically on open and on every push.
• Slash command: on any other PR, an admin or maintainer can comment /preview-docs to trigger a preview. Anyone else gets a polite rejection.
• A sticky PR comment carries the preview URL (updated in place on each rebuild).
• A companion workflow deletes the Cloudflare deployments when the PR closes.

Security

mkdocs executes Python from the PR (mkdocstrings imports src/mcp, !!python/name: directives in mkdocs.yml), so the design treats the build as untrusted code:

• An authorize job checks the triggering actor's permission before any PR code is fetched. The auto-path keys on event.sender.login (the pusher), not the PR opener, so a non-admin pushing to an admin's branch does not get an auto-build.
• The build job runs PR code with no secrets and uploads the static site/ as an artifact. The deploy job downloads the artifact and runs wrangler pages deploy — PR code never shares a runner with the Cloudflare token.
• .github/zizmor.yml documents and suppresses the dangerous-triggers finding for these two workflows.

Setup required before this takes effect

1. Create a Cloudflare Pages project (Direct Upload), e.g. mcp-python-sdk-docs.
2. Repo secrets: CLOUDFLARE_API_TOKEN (scope: Account → Cloudflare Pages → Edit) and CLOUDFLARE_ACCOUNT_ID.
3. Repo variable: CLOUDFLARE_PAGES_PROJECT = the project name from step 1.

pull_request_target runs the base branch's workflow, so this only takes effect once merged to main.

Test plan

• actionlint clean on both new workflows
• zizmor .github/workflows/ exits 0
• uv sync --frozen --group docs && uv run --frozen --no-sync mkdocs build produces site/index.html + site/api/ locally
• After merge + secrets configured: open a small docs PR as an admin → preview comment appears with a working pr-<n>.<project>.pages.dev link
• Push to that PR → comment updates in place
• Comment /preview-docs from a non-admin account on a fork PR → rejection reply
• Comment /preview-docs from an admin → preview appears
• Close the PR → Cloudflare deployments for pr-<n> are deleted

[cubic-dev-ai]

4 issues found across 3 files

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[cubic-dev-ai]

1 issue found across 3 files (changes from recent commits).

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[claude]

🟡 In the cleanup loop, each curl -fsS -X DELETE runs under set -euo pipefail with no per-id error tolerance, so a single failed DELETE (transient 5xx, rate limit, or an already-deleted deployment returning 404) aborts the step and skips every remaining id — leaving the rest of the PR's preview deployments stale, which is the leak this workflow exists to prevent. A one-line || echo "failed to delete $id" per id (optionally tracking failures and exiting non-zero at the end) keeps the batch resilient while still surfacing the error.

Extended reasoning...

The bug. The delete loop in docs-preview-cleanup.yml runs:

foridin$ids;doecho"Deleting deployment $id"
curl -fsS -X DELETE -H "Authorization: Bearer $CF_API_TOKEN""$base/$id?force=true"> /dev/null
deleted=$((deleted +1))done

under set -euo pipefail. Because of -f, curl exits non-zero on any HTTP ≥ 400 (and on transient network errors), and set -e then aborts the whole step mid-loop. Every id after the failing one in $ids is never deleted.

Concrete walk-through. Suppose a long-lived PR accumulated 8 preview deployments and $ids collects all 8. (1) The PR closes and the workflow fires once. (2) DELETE succeeds for ids 1–2. (3) The DELETE for id 3 hits a transient Cloudflare 5xx or a 429 — curl -f exits non-zero. (4) set -e terminates the step immediately; ids 4–8 are never attempted. (5) The workflow only triggers on pull_request_target: closed, so there is no automatic retry — the remaining deployments stay live on Cloudflare indefinitely unless someone notices the red run and manually re-runs it from the Actions UI.

Why nothing else prevents it. The script already shows care about partial-iteration hazards (the collect-then-delete comment to avoid pagination shifts, the 200-page cap raised after review), but the delete phase has no equivalent protection: there is no || true, no retry, and no per-id error accounting.

Addressing the counterargument. It's fair that fail-fast with a visible red run is a defensible design and that a manual re-run will re-list and delete the leftovers, so the original "no retry path" framing overstates it slightly — recovery exists, it's just manual and easy to miss on a closed PR. It's also true that the trigger needs an actual API failure on a specific DELETE, which is uncommon, and the consequence is the same low-impact class already discussed on the pagination-cap thread (a stale preview; nothing breaks). That's why this is a nit rather than a blocker. Still, a workflow whose entire purpose is unattended cleanup ideally shouldn't let one bad response abandon the rest of the batch when the hardening costs one line.

Fix. Tolerate per-id failures while still reporting them, e.g.:

failed=0
foridin$ids;doecho"Deleting deployment $id"if curl -fsS -X DELETE -H "Authorization: Bearer $CF_API_TOKEN""$base/$id?force=true"> /dev/null;then
deleted=$((deleted +1))elseecho"::warning::failed to delete deployment $id"
failed=$((failed +1))fidoneecho"Deleted $deleted deployment(s) for $BRANCH ($failed failed)."
[ "$failed"-eq 0 ]

This deletes everything it can, still fails the run (visibly) if anything was left behind, and a manual re-run then only has the genuinely stuck ids to deal with.