ci(sync-about): read-only PR counter check, no bot fix-up push (#155)

## Problem
On every indicator PR the `sync-about` workflow found `docs/README.md` lagging `lib.rs` (the wiring only bumped `README.md`) and pushed a `wickra-bot` *"sync indicator count"* commit onto the PR head. That push uses `GITHUB_TOKEN`, which **triggers no workflows**, so it moved the PR head onto a commit with no CI run — and the **Codecov patch status** (keyed to the PR head sha) stopped surfacing on the PR.

## Fix
- The indicator wiring (`ScriptHelpers/_common.py` `wire_readme_counter`) now bumps **both** `README.md` and `docs/README.md` in the author's code commit, so the counter is already correct when CI runs.
- This workflow's PR flow is reduced to a **read-only check** that fails loud (fork and same-repo PRs alike) if either counter is stale, and **never pushes**.
- The `GITHUB_TOKEN` job permission drops from `contents: write` back to `read` (OpenSSF Scorecard: Token-Permissions). The removed `ctx` step + push steps are gone.
- The `main`/tag outward syncs (About description, docs/webpage/wiki/org) are **unchanged** — they use the `ABOUT_SYNC_TOKEN` PAT, not `GITHUB_TOKEN`.

## Effect
Indicator PRs keep their head on the code commit → the Codecov patch status surfaces again. No functional change to merged-main state (the counts still land, now inside the squash-merged code commit).
This commit is contained in:
kingchenc
2026-06-03 15:40:24 +02:00
committed by GitHub
parent a3a1ae4dba
commit 46dc8f5a00
+40 -105
View File
@@ -41,17 +41,17 @@ name: Sync indicator count
# `RollingVwap`, so the mod-count under-reports by one. lib.rs is the
# single source of truth for what the bindings reach.
#
# Design: keep README in sync *before* a PR is merged, by pushing a
# fix-up commit to the PR head branch. After squash-merge into main
# the bot commit is folded into the single signed merge commit, so
# main's history never shows an unsigned "sync indicator count" entry.
# Design: on PRs this workflow is a READ-ONLY check. The indicator wiring
# (ScriptHelpers/_common.py wire_readme_counter) bumps both README.md and
# docs/README.md inside the author's code commit, so the counter is already
# correct by the time CI runs. If it is not, the check below fails loud and
# asks the author to re-run the wiring — it never pushes a fix-up commit.
#
# The push to PR head uses the default `GITHUB_TOKEN`, whose pushes
# explicitly do NOT trigger downstream workflows (anti-recursion
# policy). So a counter fix-up does not re-trigger ci.yml on the PR
# — it does, however, re-trigger sync-about.yml on the next PR
# `synchronize` event, which is what we want (a no-op if the counter
# is now correct).
# (An earlier version pushed a GITHUB_TOKEN "sync indicator count" commit to
# the PR head. Because GITHUB_TOKEN pushes trigger no workflows, that commit
# moved the PR head onto a commit with no CI run, which hid the Codecov patch
# status — keyed to the PR head sha — from the PR. Keeping the counter in the
# code commit avoids that entirely.)
on:
push:
branches: [main]
@@ -73,53 +73,24 @@ permissions:
jobs:
sync:
runs-on: ubuntu-latest
# The only GITHUB_TOKEN write in this workflow: pushing the counter fix-up
# commit onto a same-repo PR head branch (git push origin HEAD:<ref>).
# This workflow never writes to wickra-lib/wickra with GITHUB_TOKEN: the PR
# flow is a read-only check, and the main/tag flow writes only to other
# repos (About metadata, docs, webpage, wiki, org) through the fine-grained
# ABOUT_SYNC_TOKEN PAT. So GITHUB_TOKEN stays read-only (OpenSSF Scorecard:
# Token-Permissions).
permissions:
contents: write
contents: read
pull-requests: read
steps:
# On PRs from forks the head ref lives in another repo; pushing
# back to it from this workflow is blocked by GitHub. We still
# want the PR to surface the missing counter, so the check below
# falls back to a hard failure when push isn't possible.
- name: Determine if push to PR head is possible
id: ctx
# Untrusted PR contexts (head.ref / head.repo.full_name are attacker
# controlled on fork PRs) are passed through the environment, never
# interpolated straight into the shell, so a crafted branch name cannot
# inject commands (OpenSSF Scorecard: Dangerous-Workflow).
env:
EVENT_NAME: ${{ github.event_name }}
HEAD_REPO: ${{ github.event.pull_request.head.repo.full_name }}
BASE_REPO: ${{ github.repository }}
HEAD_REF: ${{ github.event.pull_request.head.ref }}
run: |
if [ "$EVENT_NAME" = "pull_request" ]; then
if [ "$HEAD_REPO" = "$BASE_REPO" ]; then
echo "can_push=true" >> "$GITHUB_OUTPUT"
echo "head_ref=$HEAD_REF" >> "$GITHUB_OUTPUT"
else
echo "can_push=false" >> "$GITHUB_OUTPUT"
echo "head_ref=" >> "$GITHUB_OUTPUT"
fi
else
echo "can_push=false" >> "$GITHUB_OUTPUT"
echo "head_ref=" >> "$GITHUB_OUTPUT"
fi
# On PRs we check out the *head* commit (not the merge ref) so
# any fix-up commit we make goes onto the PR branch itself. On
# push events we check out the default ref. fetch-depth: 0 lets
# us push back without "shallow update not allowed".
# On PRs we check out the PR *head* commit (the author's code, not the
# merge ref) so the counter check validates exactly what will land. On
# push events we check out the default ref. No push is made, so a shallow
# checkout is enough.
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
with:
fetch-depth: 0
fetch-depth: 1
ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.ref || github.ref }}
repository: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name || github.repository }}
# Default GITHUB_TOKEN is fine for the same-repo PR-branch
# push; the About / Wiki steps re-authenticate with the PAT
# below where needed.
- name: Count indicators
id: count
@@ -139,69 +110,33 @@ jobs:
# ----- PR flow ---------------------------------------------------
- name: Check README counter (PR)
- name: Check README counter (PR, read-only)
if: github.event_name == 'pull_request'
id: pr_check
run: |
n="${{ steps.count.outputs.count }}"
if grep -qE "^${n} streaming-first indicators" README.md \
&& grep -qE "\*\*${n} indicators\*\*" docs/README.md; then
echo "matches=true" >> "$GITHUB_OUTPUT"
echo "README + docs/README counter already at ${n}; nothing to do."
else
echo "matches=false" >> "$GITHUB_OUTPUT"
echo "README/docs counter does not match ${n}; will fix up."
ok=true
if ! grep -qE "^${n} streaming-first indicators" README.md; then
echo "::error::README.md does not say '${n} streaming-first indicators' — lib.rs exports ${n}. Re-run the indicator wiring (it bumps README.md), then push again."
ok=false
fi
- name: Fix counter on fork PR head (read-only, fail loud)
if: github.event_name == 'pull_request' && steps.pr_check.outputs.matches == 'false' && steps.ctx.outputs.can_push == 'false'
run: |
n="${{ steps.count.outputs.count }}"
echo "::error::README.md / docs/README.md say a different indicator count than lib.rs (${n}). This PR is from a fork, so the workflow cannot push the fix; please set README.md to '${n} streaming-first indicators' and docs/README.md to '**${n} indicators**', then push again."
exit 1
- name: Patch README on PR head
if: github.event_name == 'pull_request' && steps.pr_check.outputs.matches == 'false' && steps.ctx.outputs.can_push == 'true'
id: pr_patch
run: |
n="${{ steps.count.outputs.count }}"
# docs/README.md carries the count in its docs.wickra.org pointer prose
# ("**N indicators**"); keep it in sync with README's prose count.
sed -i -E "s/[0-9]+ (streaming-first )?indicators/${n} \1indicators/g" README.md docs/README.md
# Bump the banner cache-buster so GitHub's Camo proxy refetches the org
# profile image (regenerated with the new count by .github/banner.yml)
# instead of serving a stale cached copy. (README only — docs has no banner.)
sed -i -E "s|(wickra-banner\.webp\?v=)[0-9]+|\1${n}|" README.md
if git diff --quiet; then
echo "No README changes after sed (counter regex did not match anything); skipping push."
echo "changed=false" >> "$GITHUB_OUTPUT"
else
echo "changed=true" >> "$GITHUB_OUTPUT"
if ! grep -qE "\*\*${n} indicators\*\*" docs/README.md; then
echo "::error::docs/README.md does not say '**${n} indicators**' — lib.rs exports ${n}. Re-run the indicator wiring (it bumps docs/README.md), then push again."
ok=false
fi
if [ "$ok" = "true" ]; then
echo "README.md + docs/README.md counter already at ${n}; nothing to do."
else
exit 1
fi
- name: Commit & push counter fix to PR head
if: github.event_name == 'pull_request' && steps.pr_patch.outputs.changed == 'true'
# head_ref still carries the (untrusted) PR branch name forwarded by the
# ctx step; pass it through the environment so the push refspec cannot be
# used to inject shell commands (OpenSSF Scorecard: Dangerous-Workflow).
env:
COUNT: ${{ steps.count.outputs.count }}
HEAD_REF: ${{ steps.ctx.outputs.head_ref }}
run: |
git config user.name "wickra-bot"
git config user.email "wickra-bot@users.noreply.github.com"
git add README.md docs/README.md
git commit -m "chore: sync indicator count to ${COUNT}"
git push origin "HEAD:${HEAD_REF}"
# ----- main / tag flow ------------------------------------------
#
# After a PR squash-merges, this workflow runs again on the push
# to main. README is already correct (it was fixed on the PR
# branch before the merge); the only outward syncs left are the
# GitHub About description (repo metadata, not a commit) and the
# wiki repo (separate repo, no main history pollution). README is
# not touched on main any more.
# After a PR squash-merges, this workflow runs again on the push to main.
# README.md / docs/README.md are already correct (the indicator wiring
# bumped them in the merged code commit); the only outward syncs left are
# the GitHub About description (repo metadata, not a commit) and the docs /
# webpage / wiki / org repos (separate repos, no main history pollution).
# The wickra repo's own README is not touched on main any more.
- name: Update GitHub About (description + homepage)
if: github.event_name != 'pull_request'