Issue to PR

atomic-forge fix <github_issue_url> is the fully autonomous one-shot: hand it an issue, and (with CIE + an LLM endpoint in your env) it fetches the issue, clones the repo, generates a failing regression test from the issue text, repairs against that test, and on green opens a PR — from your fork, never pushing to origin.

export FORGE_MODEL=qwen3.5:cloud FORGE_BASE_URL=http://localhost:11434/v1 FORGE_API_KEY=ollama
pip install git+https://github.com/arunsoman/atomic-forge.git \
            git+https://github.com/arunsoman/cie.git pytest

atomic-forge fix https://github.com/mahmoud/boltons/issues/123 --dry-run   # try without pushing
atomic-forge fix https://github.com/mahmoud/boltons/issues/123            # fix + open PR from your fork

The pipeline, in order

1. preflight — bails immediately if the LLM env or CIE isn't available (CIE is required for fix, not optional).
2. fetch + clone — reads the issue via gh, shallow-clones the repo.
3. bootstrap — the Bootstrap-Gate proves "at least one test in this repo is discoverable and executable" before any repair logic runs: deterministic stack probe (6 ecosystems), and — opt-in, Docker-only — a Repo2Run-style agentic fallback for repos no registered stack detects.
4. CIE populates the graph — cie index the checkout; CIE serves graph tools over MCP ([Statement-Level-Graph]]-aware localization).
5. regression test generated from the issue text, grounded in the real signatures, and the harness validates it reproduces the bug (fails on the buggy code on an assertion — a collection/import error or a test that passes on buggy code is rejected and no PR is raised).
6. forge repairs — the real Repair-Loop re-runs that test each round until green or --max-rounds (default 5).
7. PR from your fork — on green, forks the repo, pushes the fix branch to the fork only (never to origin), and opens fork → upstream via gh.

--dry-run does everything except the push/PR. Local file as issue body: --issue-body-file bug.txt (or - for stdin).

fix-comment: the same pipeline from a review comment

PR-review-comment-driven fixes (Environment-Bootstrap): scoped to the commented file/line instead of a full issue.

atomic-forge fix-comment --repo owner/repo \
  --comment-body "This throws TypeError when the list is empty" \
  --file src/queue.py --line 88

Safety properties

• Fork-only pushes. The fix branch goes to your fork; origin is never pushed to. The gate to raising anything is a validated failing regression test, not model confidence.
• Reproducible validation — the generated test must fail on the buggy code (assertion failure, not collection error) before repair starts; a fix is only PR-able if it turns that test green and the repo's own suite stays green.
• Every action is on the record — append-only JSONL trajectory (Checkpointing-and-Resumability for the run record).