Skip to content

symbaiex/symbaiex-agent-tools

v1.0.0MIT

SYMBaiEX agent enrollment, public evidence, claim verification, research jobs, signed webhooks, and authenticated MCP tools.

GitHubOpen Website

skills

Public Agent Skills — portable SKILL.md capability packages that work across Claude Code, Codex, Cursor, and other Skills-compatible agents.

This repository is also a conformant Agent Plugins 1.0 package. Root plugin.json describes the package and mcp.json points compatible clients at the public documentation and authenticated evidence Streamable HTTP MCP servers. Read AGENTS.md for the repository-wide credential and approval boundaries.

Skills

  • skills/gpt-engineer/ — the primary end-to-end GPT engineer: strict GPT-5.6 Sol/Terra/Luna research, layered engineering standards, adaptive one/three/six-lane fleets with an opt-in qualified eight-reader team mode, isolated implementation, integration, verification, private resumable run journals, content-addressed gate reuse, outcome-based fleet tuning, bounded goal persistence, and task-owned resource teardown. Native Codex custom agents are the interactive default; the official Codex SDK/app-server is the programmatic path, and the Python codex exec driver is a guarded compatibility adapter rather than the normal scheduler.
  • skills/gpt-engineer-mem/ — the memory-aware GPT engineer: bounded Claude Mem/Codex recall, live freshness checks, GPT-5.6 Sol/Terra/Luna delivery, and read-only memory-service diagnostics.
  • skills/gpt-engineer-spark/ — keep a capable lead in control while a model-pinned GPT-5.3-Codex-Spark fleet handles dependency-aware exploration, isolated candidate edits, and checks.
  • skills/claude-multi-agent/ — delegate real engineering work to Claude Code as an autonomous team or a saved native dynamic workflow with explicit research, planning, build, verification, bounded gap-closing, and child-process cleanup phases.
  • skills/gpt-orchestration/ — coordinate native agent fleets for repository-wide audits and implementation work with explicit ownership, safe concurrency, runtime-honest model handling, and independent verification.
  • skills/gpt-orchestration-build/ — take an existing audit, finding list, or implementation plan and build every confirmed item through verified waves.
  • skills/gpt-orchestration-auto/ — run a persistent /goal-style research, implementation, verification, and gap-closing loop until the outcome is complete.
  • skills/symbaiex-agent-enrollment/ — enroll and operate a user-directed SYMBaiEX agent with locally held Ed25519 credentials.
  • skills/symbaiex-evidence-search/ — search bounded public evidence and editorial records through the published REST or MCP contracts.
  • skills/symbaiex-claim-verification/ — verify stored claims against version-bound citations and inspect revision history.
  • skills/symbaiex-research-jobs/ — create and monitor bounded cited research and JSONL export jobs.
  • skills/symbaiex-webhooks/ — configure owner-scoped, HMAC-signed evidence and usage event delivery.

Install a skill

Skills in this repo follow the open Agent Skills spec: a folder with a SKILL.md (metadata + instructions) plus optional scripts/, references/, and assets/, kept under skills/<name>/ so the skills CLI (the tool behind skills.sh) auto-discovers them:

npx skills add SYMBaiEX/skills                         # interactive: pick agent + skill
npx skills add SYMBaiEX/skills --skill gpt-engineer -y
npx skills add SYMBaiEX/skills --skill gpt-engineer-mem -y
npx skills add SYMBaiEX/skills --skill gpt-engineer-spark -y
npx skills add SYMBaiEX/skills --skill claude-multi-agent -y
npx skills add SYMBaiEX/skills --skill gpt-orchestration -y
npx skills add SYMBaiEX/skills --skill gpt-orchestration-build -y
npx skills add SYMBaiEX/skills --skill gpt-orchestration-auto -y

Install the SYMBaiEX platform skills with Bun:

bunx skills add SYMBaiEX/skills --skill symbaiex-agent-enrollment -y
bunx skills add SYMBaiEX/skills --skill symbaiex-evidence-search -y
bunx skills add SYMBaiEX/skills --skill symbaiex-claim-verification -y
bunx skills add SYMBaiEX/skills --skill symbaiex-research-jobs -y
bunx skills add SYMBaiEX/skills --skill symbaiex-webhooks -y

For the complete GPT Engineer workflow in Codex and Claude Code, install it globally, register the bundled provider-native model profiles, then restart both clients:

npx skills add https://github.com/SYMBaiEX/skills \
  --skill gpt-engineer --agent codex claude-code --global --yes
python3 ~/.agents/skills/gpt-engineer/scripts/bootstrap.py --provider codex --upgrade --global
python3 ~/.agents/skills/gpt-engineer/scripts/bootstrap.py --provider codex --check --global
python3 ~/.agents/skills/gpt-engineer/scripts/audit_routing.py --cwd /path/to/repo --runtime --json

Install the memory-aware variant alongside the base engineer. It reuses the base profile setup when available, remains usable as a standalone orchestration contract, and never configures or restarts Claude Mem automatically:

npx skills add https://github.com/SYMBaiEX/skills \
  --skill gpt-engineer gpt-engineer-mem \
  --agent codex claude-code --global --yes
python3 ~/.agents/skills/gpt-engineer-mem/scripts/memory_preflight.py --json

The profile bootstrap is deliberately separate from skills.sh. It never edits provider configuration; --upgrade replaces only bundled agent-profile destinations. Use --provider all --upgrade only when Claude profiles are explicitly wanted. For project-local profiles and conservative Codex hooks, replace --global with /path/to/repository.

Do not activate a copied Luna model catalog as routine setup. Custom catalogs freeze upstream model metadata and the runtime audit rejects stale or unattested overrides. Prefer stock native routing; use the guarded CLI adapter when native Luna is temporarily unavailable.

Long GPT Engineer runs use private state outside the checkout by default. Inspect or resume that state with run_journal.py; opt a repository into a local .engineer control directory only with an explicit init. Command evidence and fleet outcome analysis are separate, fail-closed tools:

python3 ~/.agents/skills/gpt-engineer/scripts/run_journal.py --help
python3 ~/.agents/skills/gpt-engineer/scripts/cache_gates.py --help
python3 ~/.agents/skills/gpt-engineer/scripts/join_fleet_outcomes.py --help

These tools store normalized lifecycle, requirement, gate, and outcome evidence—not raw prompts, responses, reasoning, child transcripts, stderr, or telemetry logs.

Install and register the Codex-only Spark fleet separately:

npx skills add https://github.com/SYMBaiEX/skills \
  --skill gpt-engineer-spark --agent codex --global --yes
python3 ~/.agents/skills/gpt-engineer-spark/scripts/bootstrap.py --global
python3 ~/.agents/skills/gpt-engineer-spark/scripts/bootstrap.py --check --global

Spark fallback writers use isolated candidate copies and return reviewable change bundles. The capable main agent integrates those bundles, owns the final repository checks, and reclaims task-owned agents, subprocesses, listeners, and temporary worktrees. Shared MCP services and other active tasks are never cleanup targets.

Install the Claude team/workflow adapter and bootstrap it into a target repository:

npx skills add https://github.com/SYMBaiEX/skills \
  --skill claude-multi-agent --agent claude-code --global --yes
bash ~/.agents/skills/claude-multi-agent/scripts/bootstrap.sh --global
cd /path/to/repository
bash ~/.agents/skills/claude-multi-agent/scripts/run-workflow.sh \
  "Research, implement, verify, and gap-close this engineering goal"

The saved .claude/workflows/gpt-engineer-dynamic.js uses Claude's native workflow runtime. The default runner starts from the exact clean HEAD and returns an isolated candidate patch; exit 3 means the outer engineer must integrate and verify it. The outer GPT Engineer remains responsible for any transition to Codex Sol, Terra, Luna, or Spark.

Project hooks/settings are optional: run bootstrap.sh /path/to/repository, then commit those files before an isolated workflow run (or explicitly use IN_PLACE=1). Workflow evidence defaults to a unique directory under ${CLAUDE_CONFIG_DIR:-~/.claude}/workflow-runs/, outside the repository.

Or just copy the skill folder into your own agent's skill directory (e.g. .claude/skills/, .codex/skills/, or wherever your agent looks for skills — see the skills CLI's supported-agent table for the exact path per agent).