Setup

Every step changes a file outside this repo, so each one needs the user's explicit yes. If they say no, skip that step and go on.

Run setup with Bash as PY "${CLAUDE_SKILL_DIR}/../../scripts/setup.py" <flags>. PY is python on Windows and python3 elsewhere; use py -3 only if both fail.

1. Tier

Ask which plan the user is on. It sets when the statusline turns amber and red and when yah says to wrap:

• pro: amber 100k, wrap 120k, wrap now 160k
• max5 (default): 120k / 150k / 200k
• max20: 150k / 200k / 260k
• api (API key, pay per token): 80k / 100k / 150k

2. Preview

Run setup.py --tier <tier> --dry-run and show the output verbatim. It writes nothing. Point out any existing statusLine it would replace (it is backed up and --uninstall restores it).

3. Apply

On yes, run setup.py --tier <tier> --yes and show the output. The statusline shows from the next render.

3b. Quality profile (max20 only, optional)

Only when the tier is max20, offer: "Turn on the quality profile? Every session starts with ultracode on, which runs xhigh effort with workflows and costs more per turn. yah run sessions build at high effort, and each phase gets one critique of its plan before the build and one merge-blocker judge on its green PR, both on Fable (on Opus once Fable's weekly bar is at 50%). yah keeps it lean: workflows capped at medium size (<10 agents), and a once-per-session rule to use workflows only for genuinely parallel work, with the model and effort for each job taken from roles in config.json." On yes, run setup.py --ultracode --yes and show the output. It changes nothing else. --uninstall restores the previous values.

3c. Auto-update (optional)

Offer: "Turn on auto-update for yah? Claude Code leaves it off for plugins from third-party marketplaces, so yah stays at the version you installed until you run claude plugin update yah@you-are-here. With it on, Claude Code fetches new versions in the background once a session starts, and the next session loads them." On yes, run setup.py --auto-update and show the output. It changes only that setting. --uninstall restores the previous value.

3d. Auto-merge (optional)

Offer: "Let yah run merge its own PRs? It is off by default, and you merge every PR. With it on, only the yah run driver merges, never the model, and only a green PR you authored for a plan phase: at least one check ran and all passed, GitHub says it is mergeable with a clean merge state, it is not a draft, and no change request is newer than its last commit. It makes a merge commit pinned to the commit it checked (never a squash), then retargets any PR stacked on it, then deletes the merged branch, never a protected one. A run that merges ends with exit 8; with --plan it goes on to the next phase." On yes, run setup.py --auto-merge and show the output. It sets only auto_merge in config.json. --uninstall restores the previous value.

3e. GPT (optional)

Only when codex is on PATH (command -v codex; on Windows where codex), offer: "Send multi-source web research and read-only reviews to GPT? /yah:gpt research <question> hands a self-contained brief to OpenAI's codex CLI on your ChatGPT plan, in a read-only sandbox, and returns a short answer with cited URLs and the path of the full report. It is off by default. Turning it on checks that codex is signed in and makes one small live call." On yes, run setup.py --codex and show the output; if it says codex is not signed in, tell the user to run codex login and offer again. It sets only codex in config.json. --uninstall restores the previous value.

Once codex is on, tell the user that OpenAI's Codex plugin gives interactive reviews (/codex:review, /codex:adversarial-review) and that they install it themselves with /plugin marketplace add openai/codex-plugin-cc, then /plugin install codex@openai-codex. Warn them never to run /codex:setup --enable-review-gate: that gate is a Stop hook that can loop Claude and Codex and drain both plans' limits.

4. Launcher (optional)

Explain: yah <project> opens Claude Code in that project from any folder, so project hooks and memory load; yah alone lists projects. On yes, find the shell's rc file (~/.zshrc, ~/.bashrc, ~/.config/fish/config.fish, or on Windows the path printed by powershell -NoProfile -Command '$PROFILE') and run setup.py --install-launcher <rcfile>. On Windows also run it with the yah.cmd path that setup printed, so yah works in cmd.exe too. Tell the user to open a new terminal.

5. Working rules (optional)

Show the user the contents of ${CLAUDE_SKILL_DIR}/../../RULES.md. On yes, run setup.py --install-rules to append them to their user CLAUDE.md inside a marked block. Re-running does not duplicate it.

6. Done

Say: to undo all of this, run /yah:setup again and ask to uninstall, which runs setup.py --uninstall (restores the previous statusLine, removes the launcher and rules blocks, keeps the data folder).