Turbocharge Claude Code, Cursor, Codex, Gemini & every coding agent: faster, cheaper, with contextual understanding specific to your codebase.
TypeScript ai-agentsanthropicclaude-codeclicode-graphcodexcontext-engineeringcursordeveloper-toolsgeminiknowledge-graphllmmcpmcp-serveropen-sourceopenaitree-sitter
fix(blast): post the comment from the job that has a write token (#224)
anirudhkumar-nanonets6c980629840c
diff --git a/.github/workflows/blast-pages.yml b/.github/workflows/blast-pages.ymlindex 2a68369c..525e97f0 100644--- a/.github/workflows/blast-pages.yml+++ b/.github/workflows/blast-pages.yml@@ -1,7 +1,7 @@name: Blast viewer-# Publishes the interactive viewer for a pull request to GitHub Pages, and links it-# from the blast-radius comment.+# Publishes the interactive viewer for a pull request to GitHub Pages, and posts the+# blast-radius comment that links to it.## `pull_request_target`, not `pull_request`, because publishing needs write access to# the gh-pages branch and a `pull_request` job on a fork PR has none. That choice@@ -12,9 +12,18 @@ name: Blast viewer# package.json, no postinstall). The graph is built from the PR's source text by a# tree-sitter parser, which executes none of it.#-# A dispatch by PR number is the second entry point. It exists because a fork PR's-# comment job cannot post the link itself (a `pull_request` job on a fork gets a-# read-only token), so publishing that page has to be startable by hand.+# It posts the comment as well as the page, because a `pull_request` job on a fork+# gets a read-only token no matter what its `permissions:` block says: `blast.yml`+# can compute the radius there and cannot post it, and most of this repo's open pull+# requests come from forks. So `blast.yml` runs with `comment: false` and this is the+# single writer — two jobs editing one marked comment would race and overwrite each+# other. The cost of that choice is honest: graft's own PRs are now described by+# main's graft rather than by the code in the PR. `blast.yml` still runs the PR's+# build and writes the body to its step summary, so a renderer a PR breaks still+# fails visibly; it just no longer speaks on the PR.+#+# A dispatch by PR number is the second entry point. It backfills a pull request that+# was opened before any of this existed, without waiting for a new push.on:pull_request_target:branches: [main]@@ -22,7 +31,7 @@ on:workflow_dispatch:inputs:pr:- description: "Pull request number to publish the page for."+ description: "Pull request number to publish the page and comment for."required: truetype: string@@ -123,8 +132,13 @@ jobs:# every bubble falls back to its hub symbol, so the page reads `cli.ts` and# `exportViz` where the comment reads "CLI Entry Point" and "Visualization# Export". The point of the page is that a reviewer recognises the feature.+ #+ # `--format markdown`, not text: one run produces both the page and the+ # comment body, so the two can never describe different radii or word the+ # same area differently.node ../base/dist/cli.js blast . --base "origin/$BASE_REF" \- --name --title "PR #$PR_NUMBER" --export-viz ../site --format text > /dev/null+ --name --title "PR #$PR_NUMBER" --export-viz ../site --format markdown > ../blast.md+ cat ../blast.md >> "$GITHUB_STEP_SUMMARY"- name: Publish to gh-pagesenv:@@ -158,6 +172,43 @@ jobs:git commit -m "viz: PR #$PR_NUMBER" || { echo "nothing changed"; exit 0; }git push origin gh-pages+ # 3. The comment — posted here, and only here.+ #+ # `blast.yml` runs the PR's own code, so it can never hold a write token; on a+ # fork its comment step would 403, and most of this repo's open pull requests+ # are forks. Under `pull_request_target` the token belongs to the BASE repo,+ # so a fork PR is ordinary work. It runs after the page is pushed, so the link+ # it carries is live by the time anyone can click it.+ - name: Post or update the blast-radius comment+ env:+ GH_TOKEN: ${{ github.token }}+ PR_NUMBER: ${{ steps.pr.outputs.number }}+ REPO: ${{ github.repository }}+ # Unset, the comment simply has no link rather than a link that 404s.+ VIEWER_BASE: ${{ vars.BLAST_VIEWER_BASE }}+ run: |+ set -euo pipefail+ # One comment per PR, edited in place: a new one per push turns a busy PR+ # into a wall of stale diagrams. The marker is how it is found again and+ # stays invisible in the rendered body — the same marker the action used,+ # so a PR that already has a comment from that path is edited, not doubled.+ marker='<!-- graft-blast-radius -->'+ printf '%s\n' "$marker" > body.md+ cat blast.md >> body.md+ if [ -n "$VIEWER_BASE" ]; then+ printf '\n[**Open the interactive graph →**](%s/pr/%s/) — click an area to see its dependent symbols at file:line.\n' \+ "$VIEWER_BASE" "$PR_NUMBER" >> body.md+ fi+ # `--paginate` runs the filter once per page, so a match can print more+ # than one line; the first is the comment we own.+ existing=$(gh api "repos/$REPO/issues/$PR_NUMBER/comments" --paginate \+ --jq "map(select(.body | startswith(\"$marker\"))) | .[0].id // empty" | head -1)+ if [ -n "$existing" ]; then+ gh api -X PATCH "repos/$REPO/issues/comments/$existing" -F [email protected] >/dev/null+ else+ gh api -X POST "repos/$REPO/issues/$PR_NUMBER/comments" -F [email protected] >/dev/null+ fi+# A merged or closed PR's page is dead weight — and on a busy repo, thousands of# 1 MB files. Removed with the PR, not on a timer.cleanup:diff --git a/.github/workflows/blast.yml b/.github/workflows/blast.ymlindex e54f1b70..fc4b4688 100644--- a/.github/workflows/blast.yml+++ b/.github/workflows/blast.yml@@ -3,6 +3,13 @@ name: Blast radius# graft's own PRs run the action graft ships, so the comment format is reviewed the# same way the code is. `pull_request` (not `_target`) on purpose: the job reads the# PR's code, so it must not carry write access to secrets a fork could reach.+#+# Which is exactly why it no longer posts. A `pull_request` job on a fork holds a+# read-only token regardless of the `permissions:` block below, so this step's+# comment 403s on most of the repo's open pull requests. `blast-pages.yml` posts it+# instead, from a token that belongs to the base repo. What stays here is the part+# only this workflow can do: run the PR's OWN build, and fail if the code in the PR+# cannot produce a comment body at all — the rendered body lands in the step summary.on:pull_request:branches: [main]@@ -13,9 +20,10 @@ permissions:jobs:blast:runs-on: ubuntu-latest+ # Read-only now that it does not comment: least privilege, and it makes the+ # split with blast-pages.yml legible from the permissions alone.permissions:contents: read- pull-requests: writesteps:- uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4with:@@ -25,6 +33,9 @@ jobs:- uses: ./.github/actions/graft-blastwith:depth: "2"+ # Computed, not posted — see the header. `blast-pages.yml` owns the comment+ # so that fork PRs get one, and so two jobs never edit the same body.+ comment: "false"# graft reviews its own PRs with the code IN the PR, not with the last# release — a release cannot run a command the PR is adding.setup-cmd: npm ci && npm run build@@ -32,10 +43,8 @@ jobs:# A fork's PR gets no secrets, so naming would fail anyway; every area then# carries its hub-symbol name, which is still reviewable.name-areas: ${{ github.event.pull_request.head.repo.fork == false }}- # Linked only once Pages is actually serving: set the repo variable- # BLAST_VIEWER_BASE (e.g. https://nanonets.github.io/Graft) after pointing- # Pages at the gh-pages branch that blast-pages.yml writes. Unset, the- # comment simply has no link rather than a link that 404s.+ # Still set so the step summary shows the same body the comment will+ # carry, link included.viewer-url: ${{ vars.BLAST_VIEWER_BASE != '' && format('{0}/pr/{1}/', vars.BLAST_VIEWER_BASE, github.event.pull_request.number) || '' }}env:GRAFT_API_KEY: ${{ secrets.GRAFT_API_KEY }}