Requirements and Roadmap

Competitive requirements survey — autonomous issue→PR / generate-test-repair agents

Source: competitive scan of SWE-agent, OpenHands, Aider, Cognition Devin, GitHub Copilot coding agent, Sweep.dev, Codegen.com, Factory.ai (Droid), and Google Jules (2026). Each competitor's headline feature is restated as a requirement (R1–R15) below. Full detail — status against the actual atomic-forge codebase, and the state-of-the-art research literature backing each one — now lives in its own document under Agent-Computer-Interface per requirement below.

#	Requirement	Status	Doc
R1	Purpose-built agent-computer interface (structured navigate/view/edit/execute) instead of raw shell	✅ Done	Agent-Computer-Interface
R2	Independent critic/gate that can withhold a low-confidence patch	✅ Already met	Critic-Verification-Gate
R3	Planner (strong model) / executor (cheap model) role split	✅ Done (opt-in, default off)	Planner-Executor-Split
R4	Whole-codebase symbol map for structural awareness	✅ Already met	Repo-Scale-Context
R5	Auto-commit each accepted edit with a descriptive message	✅ Already met	Auto-Commit-Messages
R6	Persistent sandbox with terminal, editor, and browser	Not a goal (by design)	Persistent-Sandbox
R7	Multi-channel task intake (chat, issue tracker, UI)	✅ Done (stdin path)	Multi-Channel-Intake
R8	Fix PR generated directly from a review comment	✅ Done	Review-Comment-Driven-Fix
R9	Zero new developer tooling — operate inside the host platform	✅ Done	Zero-Friction-Integration
R10	Self-review that the patch resolves the stated issue	✅ Already met (fix pipeline)	Self-Review-Issue-Resolution
R11	Scale to large/enterprise monorepos without prohibitive indexing cost	Statement-level graph shipped 2026-08-29	Enterprise-Scale-Indexing
R12	Terminal-native CLI fitting into existing CI/CD	✅ Already met	CLI-CI-Native
R13	Fully async, per-task isolated, parallel execution	✅ Done (within-task); cross-task still open	Parallel-Execution
R14	Execute the test suite to select a patch, not model judgment	✅ Already met (+ correctness fix)	Execution-Guided-Repair
R15	No training on private code by default	✅ Done (--local-only)	Data-Privacy-No-Training
R16	Bootstrap any GitHub repo (language/build-agnostic) to a runnable state before repair begins	Deterministic gate + checkpoint + C/C++ shipped; agentic fallback open	Environment-Bootstrap

atomic-forge's own claimed differentiators (for reference, not sourced from competitors)

Not matched by name in the competitor set above — worth keeping front-and-center rather than diluting with the R1–R15 backlog:

• Machine-checked task contract — AtomicTask with a required test_triad (positive/negative/recovery), enforced by pydantic at construction time.
• Crash-safe, resumable runs — every phase transition checkpointed to SQLite before work starts; resume re-hashes files on disk and regenerates only what changed.
• Blast-radius gate — statically rejects a winning patch that changes or removes a function/method signature while an external caller still depends on the old one. Directly supported by research in Critic-Verification-Gate.
• 7-way verdict taxonomy (passed/failed/partial/timeout/ lint_error/crashed/skipped) instead of a pass/fail boolean, with full phase-by-phase run history. See Execution-Guided-Repair for a research-backed upgrade path (execution-trace-level signal).
• Adaptive concurrency control — ramps LLM-call parallelism up by 1 per success, down by 2 on a 429, with a monotonic counter to avoid a race between an in-flight success and a rate-limit step-down.

Implementation pass (2026-08-29) — what actually happened

Before implementing, R1–R16 were re-audited against the real codebase (the original docs were written from the README alone). Result: several "Partial"/"Not implemented" rows were already substantively met in code — tools.py was already an ACI (R1), the blast-radius gate already fed rejections back into the next round's prompt (R2), codegraph.py was already a precomputed, incrementally-hashed SQLite graph with multi-hop depth support (R4, most of R11), fix.py already re-verified the generated test's result independently rather than trusting the repair loop's self-report (R10), and the repair loop's execution-based selection was already more sophisticated than scoped, including flake-tolerance (R14). Those got their status corrected rather than re-implemented.

Genuine gaps were then implemented, tested (191→201 tests passing across the pass, all green), and in several cases validated by actually building and running the Docker image, not just reading the YAML:

• R1: view_window(center_line, radius) convenience, both backends + the ripgrep reference.
• R3: opt-in architect_mode (one extra planning call before K-sampling), default OFF — SAFEdit's counter-signal means this shouldn't ship default-on without a live-LLM benchmark this environment can't run.
• R7: --issue-body-file - reads the bug description from stdin.
• R8: run_fix_from_comment — review-comment-driven fix, same pipeline as fix, scoped to the commented file (CLI: fix-comment).
• R9: action.yml gained a command input (run/fix/fix-comment), entrypoint.sh dispatches accordingly, Dockerfile now installs gh+CIE+mcp (previously entirely absent, so fix/fix-comment could never have worked via the Action at all).
• R13: K-sampled repair attempts now run in parallel (ThreadPoolExecutor, default on) — required making CodeGraph's SQLite connection thread-safe (check_same_thread=False + an RLock).
• R15: --local-only refuses to run against a non-loopback/private LLM endpoint — makes the "nothing has to leave your machine" claim enforced, not just possible.
• R16: registered Java (Maven/Gradle), Go, and Rust stacks alongside the existing Python/Node — fix <url> now bootstraps 5 ecosystems instead of 2.

A real, pre-existing correctness bug was found and fixed along the way (not something introduced by this pass, but exposed by it): a write→retest→write→retest sequence — exactly what K-sampling and multi-round repair both do — could intermittently evaluate a STALE cached .pyc instead of the just-written fix, silently corrupting the execution-based candidate selection R14 depends on being trustworthy. Root-caused and fixed in sandbox.py::_purge_pycache; see Parallel-Execution and Execution-Guided-Repair for the full writeup. Verified 15/15 clean (was ~50% flaky) after the fix.

What's left, honestly, not silently deferred:

• R6 — deliberately not built; a doc-only decision (see its own file).
• R11 — statement-level def-use graph granularity (per ARISE). The "iterative query" half of the original plan turned out to already be met (callers/callees already accept depth and BFS-traverse). The remaining piece is a real schema/parser change to a module several other requirements now depend on being correct (R1, R4, R13) — not attempted partially/untested in this pass.
• R16's harder half — C/C++ (no single dominant build-marker) and the Repo2Run-style agentic bootstrap fallback for repos matching none of the 5 registered stacks (~1-2 week estimate in the original plan; genuinely out of reach for one pass).
• Any change gated on a live LLM benchmark comparison (R3's SAFEdit-style validation) — this environment has no configured LLM endpoint/credentials, so those flags are shipped correctly-wired-but-conservatively-off rather than validated.