BUILDCONNECTGOVERNEVOLVESHIPA guided explainer. Practical workflows first. Experimental capabilities clearly labeled.02 / 20 · THE FACTORYYour repo. Your agent.Build a harness you own. Keep the model replaceable.01REPOSITORYStart with a GitHub URLor a blank slate.02METAHARNESSGenerate the skills, toolsand rules for your project.03YOUR PRODUCTShip your own brandedCLI and npm package.npx metaharness --wizardREADME.md · What this isYour own package: CLI + agents + tools + policy + memory.03 / 20 · DISCOVERUnderstand before executing.Static analysis turns project structure into a starting plan.01ANALYZEInspect manifests, file layoutand detected language mix.02ASSESSReview agent fit, tool risk,build readiness and run cost.03RECOMMENDGet an agent topology andsuggested skills to review.metaharness score <repo>README.md · Score any repo; USAGE.md · genomeAnalysis proposes a plan. Inferred commands stay disabled until reviewed.04 / 20 · STUDIODesign it in your browser.A local browser workflow, with public repo data fetched from GitHub.01REPO → HARNESSAnalyze a repository andedit its recommendations.02CREATE + AUTHORCreate a harness or authora skill, agent or command.03VERIFY + EXPORTInspect a supplied archive.Download your own .zip.ruvnet.github.io/metaharness/USERGUIDE.md · What can I do with it?Review recommendations locally, then download a branded package.05 / 20 · SCAFFOLDEverything around the model.Generate a focused starting point, then make it yours.01AGENT CONTENTPrompts, skills, commandsand project instructions.02TOOL SURFACEYour CLI, MCP adapter,host configuration and hooks.03OPERATING RULESScoped memory, governanceand release provenance.npx metaharness my-bot --template vertical:codingREADME.md · What it gives youTrim unnecessary tools and edit prompts before using the harness.06 / 20 · HOSTSMeet developers where they work.One harness, host-specific integrations.01CODING HOSTSClaude Code · OpenAI CodexGitHub Copilot · OpenCode02AGENT RUNTIMESpi.dev · Hermes · OpenClawPrime Agent03ISOLATION + CIRVM · GitHub ActionsInteractive work and CI/CDnpx metaharness my-bot --host codex --host claude-codeREADME.md · Hosts; host capabilities differAdapters translate the same harness into each host’s integration format.07 / 20 · TEMPLATES / 1Start with the right domain.Choose a focused workflow instead of shipping every tool.01ENGINEERINGminimal · coding · aidevops · repo-maintainer02KNOWLEDGEresearch · ruvieweducation03PROFESSIONALtrading · legal · healthReview domain safeguards.npx metaharness --listREADME.md · VerticalsTemplate prompts and policies are starting points for your domain.08 / 20 · TEMPLATES / 2Build for your business.Domain agents, skills and commands remain yours to customize.01CUSTOMERSsupport · crmsales02GROWTHmarketingadvertising03NEW FRONTIERSbusiness · agenticsgaming · exoticnpx metaharness my-bot --template vertical:supportREADME.md · Verticals; examples-packages/Keep the useful agents; remove the ones your team does not need.09 / 20 · GOVERNANCEGive tools clear boundaries.MCP is an adapter surface, with access denied by default.01MODESOff, local stdio,or authenticated remote MCP.02PERMISSIONSNetwork, shell and writesrequire explicit policy.03LIMITS + AUDITTime limits, call budgets,approval gates and logging.harness mcp-scan <path>README.md · MCP; Prime Agent deny-lists require a sandboxA denied tool stays denied. Host enforcement capabilities still matter.10 / 20 · MEMORY + FEDERATIONKeep context scoped.Project memory and peer coordination are explicit capabilities.01PROJECT MEMORYA separate namespace foryour harness and its work.02FEDERATIONManage peers and use thekernel federation primitives.03SECRETSWire credentials deliberately.Keep secrets out of artifacts.harness federate · harness secretsUSAGE.md · commands; ARCHITECTURE.md · kernelProject state is scoped; credentials and peer trust require configuration.11 / 20 · MODEL ROUTINGSpend where quality needs it.Learn routing decisions from your own evaluation outcomes.01ROUTEChoose a model predictedto meet your quality bar.02LEARNFeed back success, latencyand cost from actual runs.03TRAINUse your own logs; optionalnative tiny-dancer training.npm i @metaharness/routerREADME.md · router; gains depend on workloadA quality target guides model selection; outcomes improve the routing fit.12 / 20 · DARWIN MODEEvolve the harness.Keep model weights frozen while testing harness changes.01PROPOSEMutate configuration;RefineMutator cites failures.02EVALUATERun sandboxed comparisonsagainst a measured baseline.03RETAIN / REVERTKeep improvements only afterthe configured safety gates.npm run evolveREADME.md · Darwin Mode; --no-darwin skips integrationCandidate → sandbox test → compare → retain or revert. No guaranteed lift.13 / 20 · GOVERNED VARIATIONExplore difficult work safely.AVO adds a governed inspect, edit, execute and repair loop.01CONTROLImmutable capabilities andbounded execution budgets.02RECOVERBranch, repair or revert.Quarantine unsafe candidates.03PROVERVF memory, lineage andsigned replay receipts.@metaharness/avoREADME.md · AVO; stronger performance claim remains gatedTool execution stays within capability, budget and promotion boundaries.14 / 20 · EXPERIMENTAL MEMORYReuse outcomes, not raw episodes.Field memory is opt-in and requires deployment configuration.01PACKED ATTRACTORSOutcome caps, supportquarantine and decay.02ADMISSIONAuthenticated principals,compatible storage and adapter.03BOUNDARIESSingle process by default.Distributed use needs support.npx metaharness my-bot --field-memoryREADME.md · field-memory; experimental, off by defaultDeployment must supply storage, adapter, identity verification and secret key.15 / 20 · WEIGHT-EFTTrain the cheaper tier.An optional weight-tuning path complements harness evolution.01EXPORTGold-resolved archives becomestandard SFT / DPO datasets.02TUNELoRA-tune an open modelin the cheaper routing tier.03CHECKSeparate training and evals.Reject overfit candidates.metaharness weight-eftREADME.md · Weight-EFT; GPU-gated, no universal gain promisedTune on training data; evaluate on separate tasks before accepting a model.16 / 20 · EXPERIMENTAL ARCReason through ChatGPT.ARC-AGI-3 controller plus remote MCP and an Apps canvas.01OBSERVE + REMEMBERExact observations, beliefs,evidence-backed memory.02ACT + SUPERVISEGuarded actions, checkpointsand supervisor interventions.03TRACK + COMPAREHash-chained receipts;optional ARC-specific AVO.@metaharness/arc-agi-3-chatgptREADME.md · private / experimental; no ARC performance claimChatGPT is the reasoning host. These ARC packages make no score claim.17 / 20 · VERIFYMake readiness inspectable.Use focused checks before trusting or publishing a harness.01HEALTHdoctor · diag · validatescore · genome02SECURITYaudit · mcp-scanthreat-model · sbom03REVIEWcompare · export-configoia-manifest · diag --bundleharness validate · harness audit · harness sbomREADME.md + USAGE.md · command referenceReadiness checks expose problems; they do not replace a security review.18 / 20 · SHIP + MAINTAINYour brand. Verifiable releases.Keep ownership while preserving evidence and update paths.01SIGN + PUBLISHEd25519 witness sign / verify.npm provenance; IPFS optional.02UPGRADE + EJECTReview template drift or liftcustom content out of Ruflo.03PORTABLE KERNELRust core with WASM andNAPI-RS adapter surfaces.harness sign · harness verify · harness upgradeUSAGE.md · publishing / eject; ARCHITECTURE.md · kernelPublish your package; consumers install your brand with npx.19 / 20 · SESSIONS + DAILY TOOLSResume with an audit trail.Opt into recoverable sessions and inspect the tools you expose.01SESSION LOGSCrash-recoverable JSONLsession recording, opt-in.02REPLAYRust and TS/WASM replaysproduce matching state hashes.03TOOL OPERATIONSMCP list / invoke commands;Bash, Zsh and Fish completion.--sessions · harness mcp ls · harness completionsREADME.md · sessions; USAGE.md · daily commandsResume from logs and compare replay state hashes across runtimes.20 / 20 · MEASURE + BEGINThe harness is the product.Evaluate quality and cost on your own work. Then ship.01BENCHMARKDRACO research evaluationand coding benchmark tooling.02COMPARE VALUECost-Pareto leaderboard;inspect scope and caveats.03START SMALLChoose a repo you know.Generate, review, validate.npx metaharness --wizardREADME.md · Quality gates / leaderboard; open Studio to startChoose one known repo. Review the output. Run its checks. Then iterate.Your repo. Your agent.A factory for custom agent harnesses you own.Analyze repos · Scaffold agents · Browser StudioHost adapters · Domain templates · MCP policyMemory · Routing · Darwin evolution · AVOExperimental field memory · Weight-EFT · ARCVerify · Sign · Publish · Upgrade · Benchmarknpx metaharness --wizardEXPLORE METAHARNESS · cognitum.one/metaharness20 CHAPTERS + INTRO · 3 MIN 28 SEC

Sunbelt Computer Software

PL/B Language Development and Support

FilesExpand file tree

metaharness-walkthrough.svg

Latest commit

History

metaharness-walkthrough.svg

File metadata and controls