{"id":50907000,"url":"https://github.com/randomittin/heimdall","last_synced_at":"2026-06-16T06:03:27.875Z","repository":{"id":349523606,"uuid":"1202700430","full_name":"randomittin/heimdall","owner":"randomittin","description":"Autonomous Claude Code agent — 14 specialized agents, 10-parallel execution, project-specific skill learning, mechanical checkpointing, wave-based planning with acceptance gates, caveman ultra (~75% token savings). One prompt → finished project.","archived":false,"fork":false,"pushed_at":"2026-06-12T08:42:36.000Z","size":5714,"stargazers_count":4,"open_issues_count":1,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-06-12T10:16:41.690Z","etag":null,"topics":["agent-orchestration","ai-agent","ai-coding","ai-orchestration","anthropic","automation","autonomous-agent","claude-agent","claude-ai","claude-code","claude-plugin","claude-skill","cli","coding-assistant","developer-tools","devtools","llm-agent","llm-framework","llm-tools","multi-agent"],"latest_commit_sha":null,"homepage":"https://github.com/randomittin/superx#readme","language":"Shell","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/randomittin.png","metadata":{"files":{"readme":"README.md","changelog":"CHANGELOG.md","contributing":"CONTRIBUTING.md","funding":null,"license":"LICENSE","code_of_conduct":"CODE_OF_CONDUCT.md","threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":"SECURITY.md","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":"2026-04-06T09:57:34.000Z","updated_at":"2026-06-12T08:42:40.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/randomittin/heimdall","commit_stats":null,"previous_names":["randomittin/superx","randomittin/heimdall"],"tags_count":2,"template":false,"template_full_name":null,"purl":"pkg:github/randomittin/heimdall","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/randomittin%2Fheimdall","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/randomittin%2Fheimdall/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/randomittin%2Fheimdall/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/randomittin%2Fheimdall/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/randomittin","download_url":"https://codeload.github.com/randomittin/heimdall/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/randomittin%2Fheimdall/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34393302,"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":"2026-06-16T02:00:06.860Z","response_time":126,"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":["agent-orchestration","ai-agent","ai-coding","ai-orchestration","anthropic","automation","autonomous-agent","claude-agent","claude-ai","claude-code","claude-plugin","claude-skill","cli","coding-assistant","developer-tools","devtools","llm-agent","llm-framework","llm-tools","multi-agent"],"created_at":"2026-06-16T06:03:22.278Z","updated_at":"2026-06-16T06:03:27.868Z","avatar_url":"https://github.com/randomittin.png","language":"Shell","funding_links":[],"categories":[],"sub_categories":[],"readme":"# Heimdall\n\n**Nothing ships unproven.** Verification gates for coding agents — catches what the model misses, proves it did.\n\n[![Claude Code](https://img.shields.io/badge/Claude%20Code-plugin-e056a0?style=flat-square)](https://code.claude.com)\n[![License: MIT](https://img.shields.io/badge/License-MIT-9b59b6?style=flat-square)](LICENSE)\n[![Version](https://img.shields.io/badge/version-1.1.0-00d4ff?style=flat-square)](CHANGELOG.md)\n\n## Install\n\nLead path — one command, zero to running:\n\n```bash\ncurl -fsSL https://raw.githubusercontent.com/randomittin/heimdall/v2.0.5/install.sh | bash\n```\n\n\u003e The branded one-liner `curl -fsSL https://runheimdall.dev/install | bash` is **coming** — it is a 302 to the pinned release tag on GitHub (brand on the line, verifiability intact). Until the redirect is live, the raw GitHub pinned-tag URL above is the working, byte-identical path.\n\n**Prefer to read it first?** (recommended)\n\n```bash\ncurl -fsSL https://raw.githubusercontent.com/randomittin/heimdall/v2.0.5/install.sh -o install.sh\nless install.sh        # short, function-wrapped, no sudo, no prompts, no telemetry —\n                       # resolves to a pinned tag, so what you read is what runs\nbash install.sh\n```\n\n**npm-native?** `npx runheimdall` is **coming** — a thin wrapper that fetches and runs the same pinned script. Until it is published, use the raw pinned-tag curl one-liner above.\n\n**Already in the Claude Code plugin ecosystem?**\n\n```bash\nclaude plugins marketplace add randomittin/heimdall\nclaude plugins install hmd@heimdall\n```\n\n## Then\n\n```bash\nhmd demo               # `heimdall demo` works too — hmd is the short form, same binary\n```\n\n## Inside Claude Code\n\n```\n/hmd:verify   /hmd:save   /hmd:debloat     # all commands live under the hmd: namespace\n```\n\n## What Heimdall will never do — and how to undo it\n\nNo telemetry. No network calls home. No sudo. No writes outside its own dir. MIT. Read the source.\n\n```bash\nhmd uninstall          # remove completely: deletes the plugin dir; nothing else was touched\n```\n\nReversibility is the point — `hmd uninstall` is the trust line that lowers the activation energy of a curl-pipe install.\n\n---\n\n## What it is\n\nHeimdall is a verification layer for coding agents. It brings and wires the canonical *external* oracle for a domain, makes every gate **falsifiable** — proven able to go red before it is trusted green — and catches the bug class that emits no local signal: ordering races, whole-sequence invariants, and missing subsystems that sail through a naive green test suite. It keeps a growing corpus of real failure cases and replays them on every change, so a regression that once shipped can never ship twice.\n\nThe delta Heimdall sells is verification, not generation. With the model held constant, raw agents and Heimdall produce roughly equivalent code; what Heimdall adds is the proof that the code is correct — and the receipt that proves the proof can fail.\n\n---\n\n## Five-minute entry doors\n\nEach of these runs on a fresh machine with only the documented prerequisites. Doors marked *coming* are specified but not yet shipped.\n\n```bash\nhmd demo                          # scaffold a real full-stack task; --run builds it and ends in a summary card + reel + a follow-up prompt\nheimdall-debloat --report-only    # zero-risk: point it at any repo, get a bloat scorecard, change nothing\nheimdall-bench                    # reproduce the public benchmark table on your own machine (dry by default, zero API spend)\n```\n\nA fourth door — `heimdall spec`, which turns a spec into a wave-planned, gate-verified build — is **coming**; it is specified but not yet a shipped subcommand, so it is not listed as a runnable command above.\n\n`hmd demo` (or `heimdall demo`) and `heimdall-bench` are **safe to run sight-unseen**: the demo\ndefaults to dry (it scaffolds the task + prints the paste-ready command and\nexecutes nothing), and `heimdall-bench` defaults to a dry pass that validates the\nsuite and prints the capture plan **without spending a single API token**. Opt in\nto the real thing with `hmd demo --run` and `heimdall-bench --live`.\n\nThe deeper proofs already in the tree are runnable today too:\n\n```bash\nbin/falsify exchange-lob          # inject every mutant; assert the gate goes red on each (score must be 1.0)\nbin/corpus run                    # replay the failure corpus; print the real catch-rate\n```\n\n---\n\n## The story that built the bar\n\nHeimdall's own golden reference once carried a wrong byte — the H/C flag bits inverted in the Game Boy CPU trace — and it survived **every same-author layer**: self-verify, mutation proofs, and a 100% corpus catch-rate all stayed green on the wrong value. It was caught only by adversarial cross-family review anchored to an external source (the Pan Docs), then confirmed by a 3-model blind consensus that would have failed the old value 3-0. The corpus dipped from 9/9 to 7/9 the instant the reference was corrected, then recovered to 9/9 — and that dip is on the public record, because a verification system that can't show you its own failures can't be trusted with yours.\n\nReceipts: [corpus dip log](evals/corpus/CORPUS-STATUS.md) · [golden provenance + blind consensus](evals/oracles/emulator-gb/fixtures/golden/VERIFICATION.md).\n\n---\n\n## Status — failures visible on purpose\n\nThe live flagship status table is at [`evals/flagship/STATUS.md`](evals/flagship/STATUS.md), with the ❌ rows kept in view: the descoped timer subsystem the coverage matrix predicted on day zero, and the calibration row where Heimdall shows no delta. We publish where it still fails — watch us close it.\n\n---\n\n## Running it on your own work\n\nOnce installed, run Heimdall over your own task in auto mode (the default — a background safety classifier blocks prompt injection and risky escalation while still letting Heimdall work autonomously):\n\n```bash\ncd /path/to/your/project\nheimdall --auto \"build a real-time dashboard with auth and charts\"\n```\n\nThe installer is idempotent — re-run it any time to update.\n\n\u003e **`--dangerously-skip-permissions`** exists as a flag but is **not** the default. It hands an agent full autonomy with no safety classifier in the loop; prefer `--auto`. Only reach for skip-permissions in a throwaway sandbox you are willing to lose.\n\n### Prerequisites\n\n- **Claude Code** 1.0+ with valid auth ([install guide](https://docs.claude.com/en/docs/claude-code/setup))\n- **Git**\n- **`jq`** (for the gate + state helpers): `brew install jq`\n\n---\n\n## Contributing\n\nThe two contribution doors are **stack packs** and **oracle packs**:\n\n- **Stack packs** ([`skills/stacks/`](skills/stacks/)) teach Heimdall a framework's conventions, build/test commands, and gotchas. Adding one is the fastest way to make Heimdall fluent in a stack you know.\n- **Oracle packs** ([`evals/oracles/`](evals/oracles/)) add a falsifiable external gate for a new domain — a reference implementation, a mutant set, and the proof the gate goes red on every mutant.\n\nSee [CHANGELOG.md](CHANGELOG.md) for release history and [`docs/superpowers/specs/`](docs/superpowers/specs/) for design docs.\n\n---\n\n## License\n\n[MIT](LICENSE)\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Frandomittin%2Fheimdall","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Frandomittin%2Fheimdall","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Frandomittin%2Fheimdall/lists"}