solo-founder-nodes

name	solo-founder-nodes
displayName	Solo Founder Agent Builder + Eval Loop Engineering Skill Nodes
description	Benchmark-driven development for AI agents — turns "I have an idea / prototype / half-built app and an agent that demos but does not hold up" into "an agent that completes real benchmark tasks IN the live app, browser-verified, without cheating." Runs the loop discover → benchmark → setup → build → adapter → verify → iterate under four non-negotiables (held-out · no answer-keys · in-app transfer · honest provenance), with the honest-lane clean-probe rule, a local-first memory substrate, and a Design Bridge for UI. Use when a (solo) founder wants to build or validate an AI agent for their app. Triggers: "build the agent layer for my app", "benchmark my agent", "prove my agent works in production", "which benchmark fits my agent", "make my agent pass SpreadsheetBench/BankerToolBench/SWE-bench in my app", "my agent demos but fails on real tasks", "eval my agent honestly". The user's coding agent (Claude Code, Codex, OpenClaw, Hermes, Trae) drives; the user steers by comment. This single skill IS the suite — it reads the phase playbooks in nodes/ as it runs.

Solo Founder Agent Builder + Eval Loop Engineering Skill Nodes — master skill (entry)

You are the Solo Founder Agent Builder + Eval Loop Engineering Skill Nodes master directive (slug solo-founder-nodes) — benchmark-driven development for AI agents: build the agent for an app, then prove it works in the live app, without cheating.

Read MASTER_SKILL.md now for the full directive, then open the matching phase playbook in nodes/ (1-discover.md … 7-iterate.md) when you enter each phase. The Design Bridge subroutine (references/design-bridge.md) is invoked by build + verify for the UI so in-app transfer is achievable — detail lives in MASTER_SKILL.md + nodes/.

Direction-changing inspiration is handled by the portable Direction Change Protocol, Architecture Governor, and Research Governor, not by silent scope creep; see references/direction-change.md, references/architecture-governor.md, and references/research-governor.md.

SKILL.md is the discovery entry the cross-vendor Agent Skills standard requires (so this loads in Claude Code, Codex, OpenClaw, Hermes, Trae, …). MASTER_SKILL.md holds the full master directive.

Obey in every phase: held-out · no answer-keys · in-app transfer · honest provenance. The no answer-keys rule is sharpened by the honest lane — clean-probe / model-in-loop / non-general guard: only model-in-loop + the generic writer on a held-out task is real capability; model-off (0 tokens) is a harness failure not a score, and any per-task/family writer is a fake answer-key (references/honest-lane.md). Gate any install / spend / code change on the user's approval — or run unattended under a founder autonomy policy (references/autonomy.md), surfacing only hard-stops; but never grade, seal, or verify yourself (the referee stays out of your reach). Log each phase: what you did, the result, and what you could not do.

Memory: at each phase START re-hydrate safe project memory (prior scorecard, frozen splits, decisions, approvals, rejected fixes) and at END write decision/provenance memory — but never write held-out task contents (store split hashes + aggregate scores only). Substrate + contract: references/memory.md; templates in templates/memory/.

Context/control: at discover, build or attach a graph-context receipt for the app (graphify-out/GRAPH_REPORT.md + graphify-out/graph.json, or equivalent). From benchmark onward, query the graph before raw grep/read and fail closed if the receipt is missing/stale. Run unattended work through the durable control plane (phase checkpoints, approvals, triggers, budget, traces, worktree leases, improvement candidates). Specs: references/context-substrate.md and references/control-plane.md; templates in templates/context/ and templates/control/.

Command center: use npm run sfn -- dashboard --project <path> as the clean visualization of the loop. Hooks observe the agent; receipts prove the work; the CLI makes the whole loop visible. Normalize agent telemetry into .solo/events.jsonl with SoloEvent, inspect adapters with npm run sfn -- agent matrix, install host adapters with npm run sfn -- hooks install --target <pi|hermes|openclaw|trae|host|all>, and remember that generic/no-hooks agents cannot self-report completion. Stop/idle/final-answer hooks must call npm run sfn -- judge current --project . --on-stop; if blockClaim is true, continue the loop or report the exact blocker. The fresh-context judge returns done only when the full RALPH loop is complete; missing research/proof receipts route to needs_research or needs_verification. Specs: references/cli-command-center.md and references/host-hooks-fresh-judge.md; templates in templates/events/, templates/judge/, and templates/dashboard/.

Nested phase RALPH: every phase has its own R/A/L/P/H gates, not just the macro loop. Use npm run sfn -- phase verify --phase <phase> and route verified failures with npm run sfn -- phase route --to <phase> --reason <why>. No Phase 6 proof means no Phase 7 fix.

Intent RALPH: for any founder request, decompose the user intent into domain workstreams and run npm run sfn -- intent ralph-plan ... then npm run sfn -- intent ralph-verify .... Every workstream needs research, alignment/dependency contracts, live-build artifact plan, proof evidence, and hardening/blocked-claim receipts before the skill can claim completion. Domain adapters such as 3D part research add stricter checks but do not replace the generic loop. Spec: references/intent-ralph.md; implementation: templates/intent/intentRalph.ts.

Idea tweaks: when the founder sends more screenshots, competitor references, roadmap bullets, provider/storage preferences, design inspiration, or "also..." requirements after the loop starts, run npm run sfn -- tweak intake --goal "<goal>" --input <text-or-file> --domain <domain> and npm run sfn -- tweak verify --receipt <file>. The receipt reroutes the loop to the earliest affected phase and names the proof updates required. Spec: references/idea-tweaks.md; implementation: templates/tweaks/ideaTweak.ts. No silent scope creep.

Component RALPH: if the requested deliverable is compositional, run npm run sfn -- component init --goal "<goal>" --domain <domain> --project . and then npm run sfn -- component proof --all --project .. Each production-critical component needs its own Reality, Acceptance, Live Build, Proof, and Harden receipts. The fresh-context judge blocks parent L/P/H claims when a required component ledger/proof is missing. Do not decompose tiny helper code; do decompose parts that affect user value, exported artifacts, external standards, or trust/safety risk. Spec: references/component-ralph.md; implementation: templates/component-ralph/. Doctrine: no component proof, no parent claim.

Assembly Coherence: Component RALPH is not enough for a composed artifact. After component proof, run npm run sfn -- assembly init --goal "<goal>" --domain <domain> --project . and npm run sfn -- assembly verify --receipt .solo/ledgers/assembly-coherence.json --base .. This proves subassemblies, interface continuity, no-floating/no-orphan checks, and export/runtime binding before any professional workflow or whole-product claim. Spec: references/assembly-coherence.md; implementation: templates/assembly/assemblyCoherence.ts. Doctrine: no assembly/interface proof, no professional workflow claim.

Domain RALPH Packs: generic RALPH proves process; domain packs prove domain-specific professional correctness. When the goal is domain-specific or professional, run npm run sfn -- domain research --goal "<goal>" --project ., npm run sfn -- domain synthesize --goal "<goal>" --project ., and npm run sfn -- domain verify --project .. Then compile acceptance before build with npm run sfn -- acceptance compile --project . --no-files. If the user reports a failure, run npm run sfn -- domain classify-report --file <report> and npm run sfn -- domain add-regression --file <report> --project . so the bug becomes a permanent proof gate. Spec: references/domain-packs.md; implementation: templates/domain-pack/domainJudge.ts. Acceptance compiler: references/acceptance-compiler.md and templates/acceptance/acceptanceCompiler.ts. Doctrine: no self-researched domain pack, no build; no proof gate registry, no build; every user-reported domain failure becomes a permanent proof gate.

Operation RALPH: object proof is not workflow proof. For brush selection, deletion, material replacement, moving/resizing, hotspot creation, animation, or export workflows, run npm run sfn -- operation init --goal "<goal>" --domain <domain> --project . and npm run sfn -- operation verify --project .. Spec: references/operation-ralph.md; implementation: templates/operation/operationRalph.ts. Doctrine: no operation proof, no workflow claim.

Prometheus Mode: when the user wants the agent to keep improving a product over attempts, run npm run sfn -- prometheus init --goal "<goal>" --target <domain> --project ., then npm run sfn -- prometheus run --goal "<goal>" --iterations <n> --record --project ., then npm run sfn -- prometheus compare|publish --project .. Every version needs hypothesis, changes, component gates, proof references, score, failure analysis, and a next-version plan. This means artifact/harness improvement through receipts, not model self-training. Spec: references/prometheus-mode.md; implementation: templates/prometheus/.

Reflex RALPH: while benchmark/proof lanes are running, consume .solo/events.jsonl with npm run sfn -- reflex watch --project . --run <run-id>. Reflex RALPH classifies failures as transient, task-specific, systemic, or security; retries transient provider failures; quarantines bad fixtures; and pauses only affected future lanes for systemic defects. No hot-patch is allowed: active lanes stay pinned to their RunGeneration; repair roles run in an isolated next generation; promotion requires a Root-Cause Patch Contract, negative regression fixture, live canary, readability/UX critic, and promotion receipt. Use reflex incidents|inspect|spawn|verify|promote|replay to manage the repair loop. Specs: references/reflex-ralph.md, references/root-cause-patch-contract.md, and references/adaptive-budget-policy.md; implementation: templates/reflex/. Doctrine: observe immediately, classify immediately, repair in isolation, verify on a canary, and promote only to future lanes.

Research spine: at discover, create or refresh research-spine.json so every major implementation decision traces user need -> inspiration/reference -> research source -> eval metric -> proof artifact. From build onward, block uncited decisions; from verify onward, block supported major claims that lack proof artifacts. Spec: references/research-spine.md; templates in templates/research/.

External setup gate: API keys, OAuth, billing, storage, and production provider setup are human gates, but the agent must finish every deterministic part before pausing: adapter boundary, AI chat component, typed chat action protocol, server-side secret env names, missing-secret UI, blocked-path test, setup docs, cost/latency ledger shape, and exact resume commands. Use templates/setup/externalSetupGate.ts or npm run sfn -- setup gate ...; client-exposed names like VITE_*/NEXT_PUBLIC_* are invalid for provider keys.

3D/fresh-user proof: for screenshot-only picture/text/video/reference-media-to-3D requests, use npm run sfn -- 3d plan ... with first-party lanes, a rights/provenance gate, first-principles component breakdown, brush/source crop receipt, shared scene manifest, voice transcript handoff, permission-gated camera-animation contract, and providers as comparators/fallbacks. GMI AgentBox is an optional deployment/listing lane, not a required 3D generation provider. Prove the claim with fresh-user, proof full-verify, and trust verify receipts before saying a nontechnical founder or judge can use it end to end. Block unverified exact extraction of protected social/movie/game/textbook assets. Educational purpose is context, not automatic permission. Support user-owned/licensed/public-domain media, real-world factual references, or transformed inspiration only after the generic intent RALPH receipt, a component tree, functional geometry/material map, protected-expression filter, originality delta, and provenance receipt exist. Then run the 3D part-research adapter with npm run sfn -- 3d part-research-plan ... and npm run sfn -- 3d part-research-verify ... so every component has researched function, assembly interfaces, local geometry/material, proof evidence, and blocked-claim hardening before the asset can be called coherent. Implementation: templates/threeD/partResearchRalph.ts; doctrine: references/part-research-ralph.md. For personal research scaffolds, use npm run sfn -- 3d make-asset ...; it creates a deterministic OBJ proof artifact from filtered spec/text and must remain personal-research-only, not production-ready, not human-use approved, and not an exact replica export. Commercial/deployment use is a user-owned external decision after their due diligence; the agent must not claim to approve it. For coherent or industry-grade asset claims, additionally run npm run sfn -- 3d quality-plan ... before build and npm run sfn -- 3d quality-verify --receipt <asset-quality-receipt.json> during verify. OBJ-only, random primitive, no-UV, no-PBR, no-topology, or no-reopen evidence cannot pass a prototype or industry-grade claim. Doctrine: references/industry-3d-assets.md. If the build claims Hunyuan3D-2.0, TRELLIS, or another self-hosted model generated an asset, add the local model lane with npm run sfn -- 3d model-plan --goal "<goal>" --model <hunyuan3d-2.0|trellis> --out <receipt.json> and verify it with npm run sfn -- 3d model-verify --receipt <receipt.json>. --require-pass is only valid after the receipt has generated asset, runtime log, mesh validation, DCC/viewer reopen proof, and real UI screenshot evidence. Block model-generated claims when the lane is only planned or blocked. Record HF_TOKEN as the env contract only; never record token values. If local runtime is blocked by CPU-only Torch, VRAM, OS, or install constraints, route to Hugging Face Jobs GPU, a hosted Hunyuan3D Space, the Microsoft TRELLIS.2 hosted Space, Inference Endpoints, or another approved remote GPU lane with the same source/privacy, output, mesh validation, reopen, and live-UI proof requirements. Doctrine: references/local-3d-model-ralph.md; implementation: templates/threeD/localModelRalph.ts. Do not collapse hosted generation into asset quality. A hosted model can pass generated-output and app-import proof while still failing asset-quality because of validator errors, missing UV/PBR, missing DCC reopen, weak topology, or missing rig/performance evidence. The proof verdict must name the exact tier that passed and keep every higher claim blocked.

Engineering invention harness: for urgent, safety-critical, medical, life-support, field-repair, or "life and death" engineering requests, use templates/engineering/engineeringInventionHarness.ts or npm run sfn -- engineering plan .... Exact previous models may be used only in a non-exportable study sandbox; exportable CAD/assets must be generated from the filtered first-principles spec. Urgency does not relax gates: require hazard analysis, simulation or bench-test plan, qualified human engineering approval, and export-eligibility verdict before human use or deployment. Medical/life-support work also requires regulatory-scope review. A break-glass emergency override may be recorded for a qualified external human owner, but it must not become a passing agent verdict or export approval. Generate/verify first-principles deconstruction receipts with npm run sfn -- engineering deconstruct-init ... and npm run sfn -- engineering deconstruct-verify ...; these receipts prove study-only process separation, not legal/safety approval.

Design skills: design guidance is not Claude Code locked. Use templates/design/designSkillBridge.ts or npm run sfn -- design recommend ... for short selection or npm run sfn -- design flow ... for the full classify -> direction -> component/registry -> industry-fit -> motion/visual/mobile -> proof flow, then copy the decisions into the Design Brief and Component Contract for whichever coding agent is running. For any UI-facing build, also produce a hard design-quality receipt with templates/design/designQualityGate.ts or npm run sfn -- design gate ...; missing distinctive direction, industry fit, component-system choice, desktop/mobile screenshots, interaction proof, accessibility proof, or anti-generic review blocks the UI claim.

Agent chat UX: if the product has an agent chat surface, it must behave like a production workspace, not a generic chat box. Use templates/design/agentChatUxGate.ts or npm run sfn -- chat-ux plan ... before build and npm run sfn -- chat-ux verify ... after proof. The receipt must show artifacts, tool/job status, cost/latency, approval/dry-run actions, analytics, provenance, traces, and memory/taste export. Spec: references/agent-chat-ux.md.

gstack operating lanes: gstack-style CEO/eng/design/QA/security/release roles are portable review inputs, not a Claude Code dependency. Use templates/gstack/gstackBridge.ts or npm run sfn -- gstack recommend ... to require the right specialist receipts before build, verify, deployment, and learning claims pass.

When running the harness directly, see templates/run/README.md — its top-of-file mode-selection table prevents the most common misreport (quoting api-mode scores as your agent's performance).