Skip to content

Source: docs/COMPARISON.md

Comparison: External Baseline vs Codex Workflows

Date: 2026-02-13

Source baseline

The external baseline README describes a broad toolkit with:

  • 20 specialized agents
  • 36 skills
  • 11 workflows

Codex Workflows is adapted to Codex skill mechanics and repository-local .agent integration.

What the external baseline does strongly

  • Rich specialist catalog (agents, skills, workflows)
  • Strong orchestration concepts
  • Command-driven workflow style
  • Large knowledge surface for many domains

Previous gap in this repository

Before this refactor, codex-workflows had:

  • 1 skill
  • 1 compact playbook file
  • no automation scripts
  • no local template scaffolding

Refactor outcomes in this repository

Now includes:

  • structured per-workflow references
  • routing matrix and orchestration gates
  • deterministic and fast router variants
  • local .agent bootstrap script
  • reusable .agent template bundle
  • compatibility manifest and drift checks
  • workflow parity checks
  • CI validation and unit tests
  • codex-native domain packs (backend/frontend/security/qa)
  • explicit architecture docs

Practical difference now

Benchmark-level robustness is approximated by combining:

  • a strong router skill
  • codified workflow knowledge base
  • automation scripts
  • project-level template scaffolding
  • full compatibility pack mirroring .agent structure (packs/antigravity-compat/.agent)
  • scriptable sync path for future updates (scripts/sync_compat_pack.py)

This keeps compatibility with Codex while preserving the workflow mindset.

Codex-native default track

Bootstrap now defaults to:

bootstrap_project_agent.py --profile codex-native

This profile boots from an independent codex-native template root:

  • skills/codex-workflows/templates/codex-native/.agent
  • with all 13 rewritten workflow definitions in workflows/*.md
  • with native specialist agents in agents/*.md
  • with native capability skills in skills/*/SKILL.md
  • antigravity-compat remains opt-in interoperability mode.

Quality is enforced by:

  • skills/codex-workflows/scripts/check_codex_native_quality.py
  • skills/codex-workflows/scripts/check_codex_native_assets.py
  • CI execution in .github/workflows/ci.yml
  • similarity threshold guard (--max-similarity 0.35) versus compatibility workflows.

References are now codex-native aligned:

  • skills/codex-workflows/references/workflows/*.md
  • byte-parity with templates/codex-native/.agent/workflows/*.md
  • explicitly separated from compatibility parity checks.

Compatibility Mode

For near-equivalent structure to external benchmark projects, bootstrap with:

bootstrap_project_agent.py --profile antigravity-compat

This installs a complete .agent tree including agents, skills, workflows, rules, scripts, and shared assets.

Compatibility parity is maintained between:

  • skills/codex-workflows/templates/.agent/workflows/*.md
  • skills/codex-workflows/packs/antigravity-compat/.agent/workflows/*.md

Remaining expansion path

  1. Expand codex-native workflow depth with stack-specific execution branches.
  2. Add benchmark thresholds and trend tracking for routing/bootstrap latency.
  3. Add broader end-to-end sample implementations with automated golden checks.

Licencia MIT