growmax-skills

Growmax team Claude Code skills and workflows. Auto-discovers skills/, commands/, agents/, and hooks/. Ships: github-repo-analyzer (skill); terse-mode (skill — low-token terse reply style, in-house alternative to caveman-style compressors; trigger with 'terse mode'/'be terse'/'caveman mode', levels lite|full|off); senior-agentic-enginner (skill — Senior Agentic Engineer persona for designing, reviewing, and evolving a project's own agentic system: routes each new requirement to the smallest correct layer — agent, skill, workflow, context, or platform adapter — keeps the canonical architecture platform-neutral across Claude/Codex/Hermes/OpenCode, and guards against responsibility duplication, monolithic agents, and context bloat; use for 'add a new agent/skill/workflow', 'restructure the agentic system', or migrating CLAUDE.md/AGENTS.md into agents+skills+workflows); learn-app + product-manager (per-repo product notebook — /learn-app builds and maintains docs/product/ from a code census + optional read-only walk, with an async question ledger the human answers in-file; the product-manager agent answers questions and writes specs from the notebook with per-claim provenance, never inventing product truth); feature-review (end-of-development multi-agent review — TDD compliance, contextual-creation UX, UI-standards compliance (compose-never-fork + tokens, per .claude/UI-STANDARDS.md), 10k-row scalability + security sweep, advisory architecture ideas — one verified scorecard with a per-finding HIGH/MEDIUM/LOW confidence report, a run ledger, and a periodic-calibration process for improving precision over time); ux-audit + ux-migrate (census UI drift against the repo's UI standard into an approvable backlog, then pay it down autonomously by composing shared components — mirrors e2e-map + e2e-batch; a write-time hook nudges the mechanical rules for free); ui-builder (pre-build shared-component resolver — /ui-builder maps a UI request to the exact shared component per .claude/UI-STANDARDS.md Table 1 with a reuse/extend/new verdict BEFORE code is written; ships the ui-builder-resolver agent and a prompt-time reminder hook, gated to repos that ship the rulebook); app-cartograph (map a live app into durable context — navigation + business flows + knowledge graph — via a code-seeded frontier and a write-safe human-in-the-loop walk; runs upstream of E2E); e2e-flow (gated multi-agent E2E test-generation for a single flow, web + api); e2e-map + e2e-batch (whole-app flow census → one approval → autonomous suite generation); and the bug-fix pipeline: confirm-bug (turn a raw bug report into a human-confirmed RED reproduction — read-only diagnosis with a cross-surface divergence table and ranked root causes, a human ruling on correct behavior, then a deterministic env-gated failing spec + repro/BUG-id/ contract that is frozen as the grader; reports UNREPRODUCIBLE instead of faking a red), fix-bug (fix the root cause against that frozen grader in a fresh session, then have an independent read-only validator grade it — 'fixed' requires the runner green AND the named assertion actually passing AND the repro byte-identical to its confirmed commit; on PASS the repro is promoted out of its env gate into a permanent regression test), bugfix (the whole lap as ONE command — a thin chain of confirm-bug into fix-bug, same three interrupts; safe because context isolation lives in the subagents and grader integrity in the pushed tag, not in session boundaries), and validate-fix (the same grader standalone, for a fix you didn't watch or as a CI check — default FAIL, and an edited repro fails regardless of how green the run is). The pipeline holds a 3-interrupt budget per bug (ruling+strategy · confirm-red · ship): a risk tier tunes rigor instead of adding gates, repros are matrices (boundary + cross-tenant rows, not one case), and a mechanical mutation check proves the repro would fail without the fix — assurance that costs no human attention. The graders are themselves adversarially evaluated: regression/bugfix-pipeline/ ships a zero-dependency fixture app and four eval cases — reproducible, unreproducible, a counterfeit fix branching on the repro's env gate, and a repro tampered after confirmation — plus a protect-repro PreToolUse hook (self-tested) and a pushed confirmation tag (repro-BUG-id) as the validator's tamper-proof baseline. gmax (the agentic-development operating layer): /gmax-setup installs 6 personas + 7 skills + standards + workflow.config.md into any project non-destructively — one gated spine (understand → design → plan → build waves → senior review) with three human gates; detects the project's EXISTING knowledge base and points at it, never creates one; /gmax-update refreshes only the canonical files via .gmax-version diffing, never touching project-filled config, standards, or KB; harness shims for Claude Code + OpenCode, zero-shim for Hermes; regression-covered at regression/gmax-setup/. gsecure (the agentic unit-test-cause build system): /gsecure-setup installs 7 test-* role agents + the unit-testing skill + fill-in references/standards/void-plan templates into any project non-destructively — gathers the project's stack, existing test tooling/structure, knowledge base, issue tracker, and static gates by recon, adopts whatever already exists, and asks the developer for exactly what is missing without ever suggesting test tools itself; produces test-causes (a spec of what a unit promises and what would falsify it, derived from intent, implementation read only as evidence) where a red test is the intended product, never touches production code, and files a git issue instead when it turns up a real defect; harness shims for Claude Code + OpenCode, zero-shim for Hermes; work starts per-project with /test-build <batch>.