Build `open-dgm`: an open-source Darwin Gödel Machine framework for recursively self-improving coding agents

Build `open-dgm`: an open-source Darwin Gödel Machine framework for recursively self-improving coding agents. Goal: implement a real, inspectable, benchmarked DGM, not a vague “AI improves itself” demo. Start with a baseline coding agent, let it propose patches to its own code/config, evaluate each descendant in a sandbox, store variants in an archive, and keep evolving from both high performers and under-explored stepping stones. Use Python, MIT or Apache-2.0, clean package layout, type hints, pytest, ruff, Docker/devcontainer, CI, and CLI-first UX. Respect all licenses. Do not claim formal Gödel-machine optimality; this is empirical evolutionary search. Core properties: - Editable coding-agent codebase. - Self-improvement operator that patches the agent itself. - Immutable archive of variants, patches, traces, scores, lineage, costs, configs, prompts, tools, safety flags, and reproducibility metadata. - Parent selection balancing benchmark score and exploration. - Benchmark harness for real coding-task performance. - Sandboxing against secret access, evaluator tampering, benchmark cheating, unsafe code, and uncontrolled recursion. - Public artifacts: code, tests, docs, sample runs, lineage graphs, score charts, and reports. Architecture: 1. Agent core Implement a baseline repo-editing coding agent that can inspect files, plan, propose patches, run tests, diagnose failures, retry, and emit a final diff plus structured trace. Support providers through a thin interface: OpenAI-compatible endpoints, Anthropic if feasible, local/open-weight models via vLLM/Ollama/LM Studio-style servers, and a mock deterministic provider for CI. Capture model name, prompts, token/cost estimates, retries, tool calls, shell commands, patches, and test results. Keep provider logic out of the DGM core. 2. Self-improvement operator A parent variant receives its own source tree, architecture docs, recent traces, failure clusters, constraints, safety policy, and a target objective. It proposes a patch to its own agent code/config. Apply the patch in an isolated child workspace, run static checks/tests/safety gates, evaluate the child, and archive the result. Mutation targets should include repo navigation, patch generation, test-failure diagnosis, retry strategy, context management, tool choice, prompt templates, solution verification, model/cost routing, and benchmark robustness without touching evaluator internals. 3. Archive Use SQLite plus patch files and JSONL traces initially. Every variant is an immutable node with ID, parent ID(s), diff, config, prompts, tools, model settings, benchmark scores, per-task results, failure summaries, cost/latency, safety flags, reproducibility data, child count, and evaluator notes. CLI must list variants, show lineage, reconstruct variants, compare variants, export results, and render Mermaid/HTML lineage graphs. Historical records must not be silently mutable. 4. Parent selection Implement pluggable strategies: greedy best-parent, uniform random sampling, score-weighted sampling, DGM-style score plus inverse child-count/exploration penalty, and optional UCB/Thompson-style strategy. Default should favor high-scoring variants while preserving exploration of under-sampled branches. Leave extension points for Huxley-Gödel-style clade metaproductivity and group-evolving-agent selection. 5. Evaluation harness Create a benchmark abstraction: task source, workspace setup, agent invocation, test execution, scoring, validation, timeout/cost budget, trace capture, and reproducibility metadata. Include: - Toy local benchmark: 20–50 deterministic Python repo-editing tasks that run fast in CI without paid APIs. - Mini SWE-style adapter: local issue-to-patch fixtures plus docs for connecting real SWE-bench/SWE-bench Verified. - Polyglot-style adapter: small Exercism-like fixtures for Python, JS, Go, and Rust if feasible. Child agents may modify their own agent code, but not benchmark definitions, tests, expected answers, scoring, evaluator integrity checks, archive history, safety policy, or CI evaluator scripts. Unsafe meta-evaluator research mode, if included, must be disabled by default and clearly marked. 6. Sandboxing and safety Every child evaluation runs in an isolated workspace with no host secrets, no default network, resource limits, timeouts, filesystem allowlist, command logs, diff inspection, and reject rules. Detect and block attempts to read env vars/secrets, alter scoring/tests/evaluator, disable tests, contact external endpoints during evaluation, add hidden payloads/obfuscation, generate malware/exploit/credential-theft/persistence/evasion logic, or bypass safety controls. Document this in SECURITY.md. This is for benign coding-agent research only. 7. CLI Implement: open-dgm init open-dgm benchmark list open-dgm agent run --task TASK_ID open-dgm evolve --generations N --children-per-generation K --benchmark toy-python open-dgm archive list open-dgm archive show VARIANT_ID open-dgm archive lineage VARIANT_ID --format mermaid open-dgm compare A B open-dgm report EXPERIMENT_ID `evolve` should load archive state, select parents, create child workspaces, generate self-patches, apply them, run safety/static/test gates, evaluate benchmarks, archive accepted children, and produce a report. 8. Reports, reproducibility, docs Each experiment outputs Markdown and JSON reports, archive snapshot, lineage graph, score trajectory chart, cost/token summary, per-task table, notable patches, safety incidents, and exact reproduction command. Add structured tracing for model calls, tool calls, shell commands, patches, tests, and evaluator decisions. Ship Dockerfile/devcontainer, pyproject.toml, Makefile/task runner, CI, pytest suite, example fixtures, seeded deterministic mode, mock-provider demo, and at least one completed sample DGM run. Docs must explain DGM vs proof-based Gödel machines, archive-based open-ended search, parent selection, self-patching, anti-cheating, extension points, lineage inspection, reproducibility, security model, limitations, and non-goals. Milestones: 1. Scaffold, architecture docs, provider interface, mock provider, CLI skeleton, toy benchmark interface, CI, SECURITY.md. 2. Baseline coding-agent loop with file inspection, patching, tests, retries, traces, and deterministic toy tasks. 3. Immutable archive, reconstruction, comparison, lineage graph export, and sample archived runs. 4. Self-improvement operator: failure diagnosis, self-patch generation, child workspace, safety gates, evaluation, archive update. 5. Multi-generation evolution runner with parent-selection policies, budgets, parallel child evaluation where safe, and reports. 6. Mini SWE-style and Polyglot-style adapters plus tamper-detection tests. 7. Public demo showing baseline, evolved generations, archive, lineage graph, score chart, best variant reconstruction, and reproduction steps. 8. Optional extensions: clade metaproductivity, statistical acceptance tests, group-evolving agents, runtime self-improvement, and VeRO-style optimizer-of-agents evaluation. Acceptance criteria: - Open-source repo with passing CI. - Full local deterministic demo works without paid APIs. - Real-model mode documented but optional. - Complete sample DGM evolution run included. - Archive is inspectable and variants are reconstructable. - Evaluator tampering is detected in tests. - Safety policy is enforced by default. - Docs allow researchers to add providers, benchmarks, or parent-selection policies. - No exaggerated claims of formal optimality, AGI, or unsafe autonomy. Final deliverable: a working open-source DGM framework that makes recursive coding-agent self-improvement concrete, auditable, benchmarked, reproducible, and responsibly sandboxed.