AI host adapter matrix
The four PACE agents (Planner, Generator, Reviewer, Stabilizer) and the two end-to-end pipeline workflows (e2e-from-scratch, e2e-from-recording) can run on any modern LLM host. Each host has a thin adapter layer that maps the canonical prompt in tests/agents/ to that host's native invocation mechanism.
Single-agent adapters
| Host | Adapter location | How to invoke |
|---|---|---|
| Claude Code | .claude/agents/*.md + .claude/commands/*.md | Use the Agent tool by name, or type /<command> as a slash command in the chat. |
| Codex CLI | .codex/agents/*.toml | Mention the agent name in chat — for example, "Have playwright-planner write a spec for…". |
| Gemini CLI | .gemini/commands/*.toml | Type /<command> as a slash command. Gemini CLI inlines the canonical prompt via @{file} at invocation time. |
| Antigravity | .agent/skills/playwright-*/SKILL.md | Agents are auto-discovered by semantic match against the SKILL.md description; invoke by name or let the orchestrator match automatically. |
| Any other LLM | n/a — no adapter file | Paste the relevant canonical prompt from tests/agents/playwright-{planner,generator,reviewer,stabilizer}.md directly into chat, then supply your input. |
Pipeline workflow adapters
The two pipeline workflows orchestrate multiple agents in sequence. Orchestration logic (steps, halt rules, loop limits, final report format) is defined once in tests/agents/workflows/ and is host-neutral.
| Host | Adapter location | e2e-from-scratch | e2e-from-recording |
|---|---|---|---|
| Claude Code | .claude/commands/e2e-*.md | /e2e-from-scratch "<scenario>" | /e2e-from-recording <recording-path> |
| Codex CLI | .codex/agents/e2e-*.toml | Run e2e-from-scratch with this scenario: … | Run e2e-from-recording for <recording-path> |
| Gemini CLI | .gemini/commands/e2e-*.toml | /e2e-from-scratch "<scenario>" | /e2e-from-recording <path> |
| Antigravity | .agent/workflows/e2e-*.md | /e2e-from-scratch (then provide the scenario) | /e2e-from-recording (then provide the recording path) |
Notes on per-host behaviour
- Claude Code dispatches subagents via the
Agenttool — each phase (Planner, Generator, Reviewer) runs as a separate subagent. - Codex CLI spawns playwright-* subagents at depth 1 (siblings), fitting Codex's default
agents.max_depthconstraint. - Antigravity calls the playwright-* skills via Antigravity's backtick skill-invocation syntax.
- Gemini CLI has no native subagent chaining — all phases run continuously in one conversation context.
Halt rules are identical across all hosts: vague scenario, smoke-run failure after three retries, or CRITICAL/IMPORTANT Reviewer findings. Workflows never commit on their own — that step is always left to the user.
Project guidance files
Each host also has a project-guidance file (the host's equivalent of CLAUDE.md) that is kept in sync with the others:
| File | Host |
|---|---|
CLAUDE.md | Claude Code |
AGENTS.md | Codex CLI |
GEMINI.md | Gemini CLI |