GitHub Action

GitHub Action

This repo doubles as a GitHub Action (action.yml + Dockerfile at the repo root) — a thin wrapper around the atomic-forge CLI, nothing more:

no hosted service, no bot, everything runs inside the job's own container. This is deliberately the only integration this project ships — CI is where the target audience (teams with an existing test suite) already is.

Usage

name: atomic-forge
on: [workflow_dispatch]

jobs:
  forge:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - uses: arunsoman/atomic-forge@v0.1.0
        id: forge
        with:
          tasks: tasks.json
          project-dir: forge_out
          api-key: ${{ secrets.FORGE_API_KEY }}
          base-url: https://integrate.api.nvidia.com/v1   # any OpenAI-compatible endpoint
          model: openai/gpt-oss-120b

      - name: Upload generated project + trajectory
        if: always()
        uses: actions/upload-artifact@v4
        with:
          name: forge-output
          path: |
            ${{ steps.forge.outputs.project-dir }}
            ${{ steps.forge.outputs.trajectory-path }}

      - name: Fail the job if repair didn't reach green
        if: steps.forge.outputs.success != 'true'
        run: exit 1

Inputs

Input	Default	Notes
command	run	run (generate/test/repair from tasks) or fix/fix-comment (R8/R9 — one-shot issue/comment-driven fix, fork-only PR)
tasks	tasks.json	[run] Path to the AtomicTaskBatch JSON, relative to the checkout
project-dir	forge_out	Where files get generated / cloned into
api-key	(required)	Maps to FORGE_API_KEY — pass a secret, never a literal key
base-url	(unset → OpenAI)	Maps to FORGE_BASE_URL
model	(unset)	Maps to FORGE_MODEL
test-cmd	(auto-detect)	[run] Force a specific test command
max-rounds	3	Repair rounds before giving up
samples	2	Patch candidates sampled per repair round
timeout	300	Per-command timeout, seconds
report	jsonl	[run] .forge/reports.jsonl write-back (jsonl or none)
architect	false	Opt-in planner pass before each repair round — see Environment-Bootstrap
issue-url	—	[fix] GitHub issue URL
issue-body	—	[fix] The issue body, e.g. ${{ github.event.issue.body }} — skips an in-container gh fetch
repo	—	[fix-comment] owner/repo, e.g. ${{ github.repository }}
comment-body	—	[fix-comment] The review comment text, e.g. ${{ github.event.comment.body }}
file-path	—	[fix-comment] Repo-relative file the comment was anchored to
line	—	[fix-comment] The line the comment was anchored to (optional)
source-url	—	[fix-comment] The PR/comment URL, for the PR body's "Fixes" link
dry-run	false	[fix/fix-comment] Do everything except push/open the PR

fix/fix-comment also need a GitHub token with repo scope exported as GH_TOKEN (the gh CLI's own convention — pr.py's fork-only PR flow shells out to it), e.g. env: { GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} } at the job or step level (a fine-grained PAT if forking across orgs).

Outputs

Output	Meaning
success	'true' if the repair loop reached zero failing tests, or fix/fix-comment opened (or dry-ran) a PR
project-dir	Echoes the project-dir input, for chaining into upload-artifact
trajectory-path	Path to .forge/trajectory.jsonl, if the run got far enough to write one
pr-url	[fix/fix-comment] The opened PR's URL, when success='true' and dry-run wasn't set

Review-comment-driven fix (R8/R9)

command: fix-comment answers "turn this review comment into a fix PR" — same CIE-index → testgen → repair → fork-only-PR pipeline as fix, scoped to the file (and line) the comment was left on. Wire it to a trigger yourself (this Action never auto-triggers on anything — see "What this is not" below); a mention-gated example:

name: atomic-forge fix-comment
on:
  pull_request_review_comment:
    types: [created]

jobs:
  fix-comment:
    if: contains(github.event.comment.body, '@atomic-forge fix')
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: arunsoman/atomic-forge@v0.1.0
        id: forge
        with:
          command: fix-comment
          repo: ${{ github.repository }}
          file-path: ${{ github.event.comment.path }}
          line: ${{ github.event.comment.line }}
          comment-body: ${{ github.event.comment.body }}
          source-url: ${{ github.event.comment.html_url }}
          api-key: ${{ secrets.FORGE_API_KEY }}
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}

Scope note (stated in fix.py::run_fix_from_comment's own docstring too): this targets the repo's default branch to clone and fork-PR against, NOT the exact PR branch the comment was left on — pushing onto an arbitrary external contributor's own PR branch is a separate, unvalidated GitHub-permissions problem this deliberately doesn't attempt.

What this is not

No marketplace-wide promotion, no auto-triggering on every issue/PR by default (the example above is workflow_dispatch — opt-in, on purpose; the comment-trigger example is opt-in too, gated on a mention), no separate hosted control plane. It's the CLI, containerized, with GitHub's input/output plumbing on top. If you want it to run on issues, PRs, or comments automatically, wire that trigger yourself in your own workflow file — this Action doesn't assume it for you.

Local verification

The image builds and runs the same way GitHub runs it:

docker build -t atomic-forge-action .
docker run --rm -v "$(pwd)":/workspace -w /workspace \
  -e INPUT_TASKS=tasks.json -e INPUT_PROJECT_DIR=out \
  -e INPUT_API_KEY=$FORGE_API_KEY -e INPUT_BASE_URL=$FORGE_BASE_URL -e INPUT_MODEL=$FORGE_MODEL \
  atomic-forge-action