Cross-tool adapter pass: Codex / Gemini / OpenCode

[sawy3r]

Goal

baton today ships slash commands for Claude Code only. Extend support to three additional CLI tools so a single ./install.sh provisions the harness into whichever directives ecosystem the adopter uses.

Target tools (in priority order):

Tool	Rules file	User-level commands dir	Format
Claude Code	~/.claude/CLAUDE.md	~/.claude/commands/*.md	markdown + frontmatter
OpenAI Codex CLI	AGENTS.md (canonical)	~/.codex/prompts/*.md	markdown
Gemini CLI	GEMINI.md or AGENTS.md	~/.gemini/commands/*.toml	TOML
OpenCode (SST)	AGENTS.md	~/.config/opencode/command/	markdown

Cursor (no user-level commands) and Aider (no slash-command surface) are explicitly out of scope.

Why

The seven rules and three role contracts are tool-agnostic prose — they describe a protocol, not a tool. Restricting the install path to Claude Code is a packaging artefact, not a design constraint. Cross-tool support multiplies the adoption surface for zero rule-content rework.

Proposed architecture — two layers

Layer 1 — universal core (tool-agnostic, single source of truth):

~/.baton/
├── role-prompts/{planner,implementer,verifier}.md
├── release-mode-template/*
├── bin/release-verify.sh
└── AGENTS-fragment.md

Role-prompt content lives here once. Every tool's slash command reads the same file.

Layer 2 — per-tool adapter shims:

adapters/
├── claude-code/   # *.md slash commands → ~/.claude/commands/
├── codex/         # *.md prompts        → ~/.codex/prompts/
├── gemini/        # *.toml commands     → ~/.gemini/commands/
└── opencode/      # *.md commands       → ~/.config/opencode/command/

Each adapter command is a thin wrapper whose only behaviour is "read ~/.baton/role-prompts/<role>.md and follow it as your governing instructions." No role-content duplication.

install.sh evolution

Multi-tool installer:

./install.sh                              # auto-detect tools, install all matching adapters
./install.sh --tools=claude-code,codex    # explicit subset
./install.sh --core-only                  # Layer 1 only, skip all adapters
./install.sh --list-detected-tools        # probe + report, no install

Detection probes:

• Claude Code: [[ -d ~/.claude ]]
• Codex: command -v codex OR [[ -d ~/.codex ]]
• Gemini: command -v gemini OR [[ -d ~/.gemini ]]
• OpenCode: command -v opencode OR [[ -d ~/.config/opencode ]]

Migration path from v0.1.x users

Anyone who installed v0.1.x onto a machine should be able to either:

• (a) Clean and re-install: rm -rf ~/.claude/baton ~/.claude/bin ~/.claude/commands/{plan,implement,verify,merge}-* then re-run the new install.sh.
• (b) Re-run install.sh directly — it should be idempotent enough to overwrite the v0.1.x Claude-Code shims in place and additionally drop Layer 1 at its new ~/.baton/ home. Adapters point at $HOME/.baton/... instead of $HOME/.claude/baton/....

Rule content does not change across this transition.

Acceptance criteria

• Layer 1 directory structure exists in the repo at core/ (or equivalent)
• adapters/claude-code/ reproduces existing v0.1.x behaviour (regression-pinned)
• adapters/codex/ lands four prompts at ~/.codex/prompts/ that successfully load ~/.baton/role-prompts/<role>.md
• adapters/gemini/ lands four .toml commands at ~/.gemini/commands/ that load the same role prompts
• adapters/opencode/ lands four .md commands at ~/.config/opencode/command/ that load the same role prompts
• install.sh auto-detects each tool and installs its adapter when present; explicit --tools= override works
• install.sh --dry-run prints all four adapters' would-be paths without copying
• At least one slice has been driven through /plan-release → /implement-slice → /verify-slice end-to-end on a non-Claude-Code tool to validate the adapter path (smoke test, not exhaustive)
• README's "Caveats" section drops the "Claude Code only" bullet
• ROADMAP.md's "Now (shipped)" section updated to reflect cross-tool support

Out of scope for this issue

• Claude Code plugin packaging (.claude-plugin/plugin.json + marketplace install). Separate issue when this lands.
• IDE-only integrations (VS Code Copilot, JetBrains AI Assistant).
• Cursor adapter (no user-level command surface; project-level only — covered if/when someone needs it).
• Aider adapter (no slash-command surface at all; documented manual paste-role-prompts workflow only).

Reference

Full design is in ROADMAP.md. This issue tracks execution.

[caioribeiroclw-pixel]

This split into universal core + thin adapters looks like the right shape.

One acceptance criterion I’d add before this grows: make install.sh --dry-run emit a small adapter fidelity report, not only paths. For each target:

• source role prompt path + content hash/version
• generated shim path
• resolution anchor ($HOME/.baton/... vs relative/project-local)
• command invocation semantics preserved or downgraded
• unsupported/lossy behavior, if any
• one smoke command that proves the adapter actually loads the same role prompt

That catches a class of bugs where all files exist but one client silently reads a stale/duplicated prompt, resolves from the wrong home dir, or loses frontmatter/command semantics.

Example table shape:

tool        role        shim                         source_hash  load_test  fidelity
claude      planner     ~/.claude/commands/plan.md   sha256:...   pass       full
codex       planner     ~/.codex/prompts/plan.md     sha256:...   pass       command-surface differs
gemini      planner     ~/.gemini/commands/plan.toml sha256:...   pass       TOML wrapper only
opencode    planner     ~/.config/opencode/...       sha256:...   pass       full

The important part is separating “installed successfully” from “same protocol preserved”. I’ve been testing this distinction in Pluribus as a portability/fidelity report for generated AI-agent context files, and the same idea maps well here: https://github.com/caioribeiroclw-pixel/pluribus/blob/main/docs/portability-fidelity-report.md

[sawy3r]

Progress (v0.4.0 / v0.4.1): Codex shipped (install-codex.sh). README/INSTALL now lead with an agent-driven install — point your agent at the repo and it places files + wires the AGENTS-fragment — which covers any tool today. Per-tool native-command targets are documented + on the ROADMAP: Claude / Codex / OpenCode use markdown commands; Gemini = TOML (needs an md→TOML transform: prompt/description, $ARGUMENTS→{{args}}); Hermes Agent = skills (~/.hermes/skills/). Instructions wire into CLAUDE.md / AGENTS.md / GEMINI.md (AGENTS.md covers most). Remaining open work here: native scripted adapters for Gemini / OpenCode / Hermes.

[sawy3r]

Related, not duplicate: swornagent/sworn#50 — Sworn's install-time delivery mechanism for multi-tool slash-command sync (brew install asks which LLM chat interface(s) you use, writes commands into each one's convention, sourced from Sworn's own binary embed instead of requiring this repo's separate install-claude.sh).

This issue (cross-tool adapters) covers the protocol-side question this repo owns: what format/convention each tool (Codex/Gemini/OpenCode/etc.) actually expects of the content. sworn#50 covers where and when Sworn writes it. sworn#50's implementation will need this issue's research as an input.