{"id":51063486,"url":"https://github.com/ntholm86/ai-steward","last_synced_at":"2026-06-23T04:02:21.011Z","repository":{"id":366516269,"uuid":"1275201895","full_name":"ntholm86/ai-steward","owner":"ntholm86","description":"Autonomous code improvement loop - harness-captured LLM evidence, operator-held Commander's Intent, CI-enforced. Reference implementation of Principles of Earned Autonomy.","archived":false,"fork":false,"pushed_at":"2026-06-22T06:48:02.000Z","size":1612,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-06-22T08:10:39.206Z","etag":null,"topics":["ai-agents","anthropic","audit-trail","autonomous-agents","autonomous-coding","claude","code-improvement","commanders-intent","harness-protocol","improvement-loop","llm-agents","observable-autonomy","principles-of-earned-autonomy","python","self-improving"],"latest_commit_sha":null,"homepage":"https://github.com/ntholm86/principles-of-earned-autonomy","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/ntholm86.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-06-20T11:39:09.000Z","updated_at":"2026-06-22T06:48:06.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/ntholm86/ai-steward","commit_stats":null,"previous_names":["ntholm86/ai-steward"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/ntholm86/ai-steward","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ntholm86%2Fai-steward","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ntholm86%2Fai-steward/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ntholm86%2Fai-steward/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ntholm86%2Fai-steward/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/ntholm86","download_url":"https://codeload.github.com/ntholm86/ai-steward/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ntholm86%2Fai-steward/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34674702,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-23T02:00:07.161Z","response_time":65,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["ai-agents","anthropic","audit-trail","autonomous-agents","autonomous-coding","claude","code-improvement","commanders-intent","harness-protocol","improvement-loop","llm-agents","observable-autonomy","principles-of-earned-autonomy","python","self-improving"],"created_at":"2026-06-23T04:02:18.863Z","updated_at":"2026-06-23T04:02:21.003Z","avatar_url":"https://github.com/ntholm86.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# ai-steward\n\nAn autonomous code improvement loop with structural accountability.\n\n**Purpose 1 — Proof:** A reference implementation of [Principles of Earned Autonomy](https://github.com/ntholm86/manifesto). Demonstrates that autonomous delegation can be structurally trustworthy — not through promises, but through Observable Autonomy (harness-captured LLM evidence), Commander's Intent (operator-written destination), and Convergence Is Silence (stop when done, not when tired).\n\n**Purpose 2 — Tool:** Genuinely useful. Write a destination → run the loop → review the staged diff → commit or discard. Works on any codebase. Cost is measured, not claimed (~$0.018/cycle on claude-haiku-4-5).\n\n## How it works\n\n```\n┌──────────────────────────────────┐\n│  OPERATOR                        │\n│  .acm/destination.md           │  ← Commander's Intent: what + why\n│  Reviews staged diffs            │\n│  Commits or discards             │\n└──────────────┬───────────────────┘\n               │\n               ▼\n┌──────────────────────────────────┐\n│  PIPELINE                        │\n│  PRE-FLIGHT → SCAN → IMPLEMENT   │\n│           → VERIFY → RECORD      │\n│                                  │\n│  • PRE-FLIGHT: git clean, tests  │  ← zero LLM tokens\n│  • SCAN: read files + dest → LLM │  ← one cheap model call\n│  • IMPLEMENT: apply change → LLM │  ← one cheap model call\n│  • VERIFY: syntax, size, tests   │  ← zero LLM tokens\n│  • RECORD: append trail entry    │  ← zero LLM tokens\n└──────────────┬───────────────────┘\n               │\n               ▼\n┌──────────────────────────────────┐\n│  HARNESS (localhost:8474)        │\n│  Intercepts all LLM API calls    │\n│  Writes .acm/sessions/*.jsonl  │  ← BEFORE response is processed\n│  Hash-chained, tamper-evident    │  ← agent cannot fabricate evidence\n└──────────────────────────────────┘\n```\n\nEvery cycle produces two records in the target repo's `.acm/` directory:\n- `audit-trail.md` — the agent's reasoning (what it proposed and why)\n- `sessions/\u003culid\u003e.jsonl` — the harness-captured LLM evidence (independent, unmodifiable)\n\nThe agent cannot claim it did something it did not do.\n\n## Quickstart\n\n**Prerequisites:** Python 3.12+, [llm-harness-proxy](https://github.com/ntholm86/llm-harness-proxy) running on `localhost:8474`, `ANTHROPIC_API_KEY` set.\n\n```bash\npip install -e \".[dev]\"\n```\n\n**In your target repo**, create `.ai-steward.yaml`:\n\n```yaml\nmodels:\n  analyze: claude-haiku-4-5\n  propose: claude-haiku-4-5\n  implement: claude-haiku-4-5\n  verify: claude-haiku-4-5\n  judge: claude-haiku-4-5\n```\n\nAnd `.acm/destination.md` — write what you want the codebase to become and why:\n\n```markdown\n# Destination — my-project\n\nWhat this is for and what a good improvement looks like.\nThe agent reads this before every SCAN. Be honest about trade-offs.\n```\n\n**Run:**\n\n```bash\nai-steward run /path/to/your/repo\n```\n\nThe loop proposes one change, applies it, verifies tests pass, and stages the diff. You inspect and commit or discard.\n\n## V1 status\n\n- Self-targeting proven: ai-steward runs against its own repository\n- Observable Autonomy structural: harness sessions co-located with trail in `.acm/`\n- Commander's Intent structural: SCAN reads `.acm/destination.md` before every proposal\n- 66 tests, mypy-clean, CI on GitHub Actions\n- Cost baseline: ~$0.018/cycle (SCAN + IMPLEMENT, claude-haiku-4-5)\n\n**V2 (not yet implemented):** model-family independence (proposer and verifier from different families), configurable verify commands (for non-Python repos), multi-cycle convergence tracking.\n\n## Observable Autonomy\n\nThe harness proxy writes the session ledger **before** the LLM response is processed. The agent cannot selectively omit calls, retroactively modify evidence, or claim a reasoning path it did not take. The trail entry (agent-authored) and the JSONL session (harness-captured) are co-located but distinct — two trust levels, one directory.\n\n```\n.acm/\n  audit-trail.md          ← agent's claim about what happened\n  sessions/\n    01KV....jsonl         ← independent proof of what the LLM was asked and answered\n```\n\n## Development\n\n```bash\npip install -e \".[dev]\"\npytest          # 66 tests\nmypy src/       # type-check\n```\n\nCI runs both on push via `.github/workflows/ci.yml`.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fntholm86%2Fai-steward","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fntholm86%2Fai-steward","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fntholm86%2Fai-steward/lists"}