iso-ai-init

name: iso-ai-init description: Initialize AI defaults. Global steps run anywhere — install/verify caveman (ultra + shrink + statusline) and shrink any allowlisted MCP server that is present and stdio-launchable. Repo-scoped steps (graphify CLI install + /graphify wiring) run only inside a git repo. A gate decides which steps apply. Use when the user runs /iso-ai-init or asks to set up AI tooling.

iso-ai-init

Set up IsaiaScope AI defaults. Some steps are global (run anywhere, even outside a git repo); others are repo-scoped (run only inside a git working tree). The manifest-driven runner decides which apply — being outside a repo is not an error, it just skips the repo-scoped steps.

Deterministic orchestration lives in scripts/init-runner.js, driven by steps.json. Each enabled Init step points at an independently owned script in templates/. Resolve paths against the skill base directory (where this SKILL.md lives), referred to below as <skill-base-dir>.

Step 0 — Deterministic Init run

Run the manifest-driven runner first:

node <skill-base-dir>/scripts/init-runner.js

The runner evaluates whether the current directory is a git repo, filters repo-scoped Init steps outside git repos, executes enabled steps in steps.json order, and prints the summary. Add or remove deterministic Init steps by editing steps.json plus the step script, not by rewriting this skill.

After the runner completes, continue to Step 3b only if the runner ran inside a git repo.

Implementation detail — Gate

The old gate remains as human-readable implementation detail and dependency documentation. It does a hard node check (node is required by caveman install, the MCP-shrink script, and the statusline merge — fail fast here, not mid-step), then prints IN_GIT_REPO=true|false on its first line plus a human-readable plan. Do not use it as the primary ordering interface; scripts/init-runner.js and steps.json own deterministic ordering and scope filtering.

bash <skill-base-dir>/templates/preflight-gate.sh

When you add future repo-scoped Init steps, add them to steps.json first. Update the gate's plan only if the human-readable plan would otherwise become misleading. (uv is checked/auto-installed inside graphify-init.sh, since it's only needed for the repo-scoped graphify step.)

Human detail — Caveman (global)

All caveman setup lives in templates/caveman-init.sh + templates/caveman-config.json.

bash <skill-base-dir>/templates/caveman-init.sh

The script handles two sub-steps:

• 1a install caveman globally if not already set up. Detection uses an installed-artifact marker (~/.claude/hooks/caveman-config.js), not command -v caveman — caveman is a Claude plugin, not a PATH binary, so command -v always fails and would re-run the installer every time. Install flags: not --all (it turns on --with-init, which writes IDE rule files into the repo); instead --non-interactive --skip-skills, run from $HOME, so the install is global-only with zero repo writes. Defaults keep hooks + caveman-shrink (the binary our wrappers need) ON. Codex gets the caveman skill via a separate global skills add -a codex (also from $HOME).

• 1b write templates/caveman-config.json → ~/.config/caveman/config.json (sets ultra globally)

• 1c — Statusline copy templates/statusline.sh → ~/.claude/statusline-command.sh, then safely merge the statusLine key into ~/.claude/settings.json (read-modify-write via node: preserves all other keys, skips if already set, backs up before writing). No hand-editing.

Wrapping MCP servers with caveman-shrink is not done here — it's owned by Step 2, which registers concrete entries from the allowlist.

Statusline shows: …/repo/dir branch ctx:75% $5.82 🪨 ULTRA 🐴 ULTRA

• ctx% red at ≥ 90% usage, magenta below
• 🪨 ULTRA → caveman mode; switches to token savings after /caveman-stats
• 🐴 ULTRA → ponytail mode (read from ~/.config/ponytail/config.json)

Human detail — MCP shrink (global)

Wrap allowlisted, token-heavy MCP servers with caveman-shrink. Runs everywhere — MCP config is global, not per-repo.

node <skill-base-dir>/templates/shrink-known-mcps.js

How it works (idempotent; backs up ~/.claude.json + ~/.claude/settings.json before any write):

• Precondition — verifies caveman-shrink is resolvable (installed by Step 1). If absent → no changes, tells you to run the caveman step first. Wrapping without the proxy would write entries that fail at MCP launch.
• Prunes bare entries — removes any upstream-less caveman-shrink MCP entry (what caveman --with-mcp-shrink registers: npx -y caveman-shrink with nothing to proxy) from top-level and per-project config. Wrapped entries (caveman-shrink + a real upstream) are kept. This is how the installer's bare registration is managed away while keeping the shrink binary.
• Carries an allowlist of names worth shrinking (ALLOWLIST array — edit to add more). It makes no assumption about any MCP's transport: whether a server is stdio or remote/HTTP is a per-machine fact, so it is checked at runtime, not hardcoded. The same name (e.g. notion) may be a local stdio server on one machine and a hosted HTTP endpoint on another.
• For each allowlisted name, based on what is actually configured on this machine:
  • present as stdio in ~/.claude.json → wrap in place (reads its own command, so paths/env are preserved).
  • present as stdio via an enabled plugin → disable the plugin's copy and add a wrapped entry built from the plugin's own launch command (no duplicate server).
  • present but remote/HTTP → skipped — caveman-shrink spawns a local child process; there is nothing to spawn for a URL. (Not an error, just not compressible here.)
  • already shrunk → skipped.
  • absent → skipped. The script never installs an MCP you don't already have.

The allowlist is the only thing to maintain. Transport and launch command are discovered, never assumed — so the skill stays agnostic to any one machine's MCP setup.

Human detail — headroom (global)

headroom (AI context compression) compresses everything the agent reads — tool output, logs, files, conversation history — before it reaches the model (60-95% fewer tokens). It is a Python CLI wired into both agents via headroom init (durable hooks + provider routing through a local proxy on 127.0.0.1:8787). It replaces rtk: same goal (fewer tokens), broader layer — whole context, not just command output. All logic lives in templates/headroom-init.sh.

bash <skill-base-dir>/templates/headroom-init.sh

The script handles three sub-steps (idempotent):

• install headroom globally with the token-savings extras only — proxy,code,mcp,ml,relevance,memory, plus pytorch-mps on darwin/arm64 (the MPS torch backend). Install order: pipx (isolated, the recommended pip-family installer) → uv-managed Python 3.13 fallback. The fallback exists because a broken system/brew Python (e.g. the macOS 26 platform.mac_ver() bug) lets pipx install "succeed" yet produces a headroom that can't run; the correctness probe is headroom --version actually executing, and uv's own Python reports its OS correctly. A post-install gate fails hard if headroom still won't run.
• Claude Code wiring — headroom init --global --memory claude writes durable proxy-keepalive hooks, sets ANTHROPIC_BASE_URL to the local proxy, and enables persistent memory. Gated on the headroom-init-claude marker in ~/.claude/settings.json. Run from $HOME → zero repo writes.
• Codex wiring — headroom init --global --memory codex does the equivalent durable hooks + provider routing for Codex. Gated on a headroom reference in ~/.codex/config.toml / ~/.codex/AGENTS.md.

Both wirings are global and re-runnable; the markers only suppress repeat output. A Claude Code / Codex restart is needed to activate proxy routing — the proxy must be up or the agent can't reach the API (headroom's keepalive hooks start it).

Human detail — ponytail (global)

ponytail (code-minimalism) steers the agent to write the minimum necessary code ("lazy senior dev": does it need to exist? is it stdlib? can it be one line?). It is a plugin for Claude Code + Codex; intensity (lite/full/ultra/off) lives in ~/.config/ponytail/config.json. We pin ultra to match caveman. All logic lives in templates/ponytail-init.sh.

bash <skill-base-dir>/templates/ponytail-init.sh

The script handles three sub-steps (idempotent):

• config — write ~/.config/ponytail/config.json with defaultMode: ultra. Gated on the file already being ultra. This is also what the statusline's 🐴 segment reads.
• Claude Code plugin — claude plugin marketplace add DietrichGebert/ponytail + claude plugin install ponytail@ponytail --scope user. Gated on a ponytail marker in ~/.claude/settings.json. Run from $HOME.
• Codex plugin — codex plugin marketplace add DietrichGebert/ponytail + codex plugin add ponytail@ponytail (official Codex CLI). Gated on ponytail appearing in codex plugin list.

A Claude Code / Codex restart is needed to activate the plugin.

Human detail — Graphify wiring (repo-scoped — skip if not in git)

Only run this step if the gate reported IN_GIT_REPO=true. Outside a git repo, skip it entirely.

This step has two parts: 3a deterministic wiring (graphify-init.sh — installs/updates the CLI, wires the skill, hooks, gitignore) and 3b the actual deep graph build (an LLM step the skill drives, not the script). Like caveman, the wiring logic lives in a deterministic template script (no inline assembly):

bash <skill-base-dir>/templates/graphify-init.sh

graphify-init.sh (idempotent) does graphify's officially recommended repo setup:

• installs/auto-updates the graphify CLI;
• runs graphify claude install --project + graphify codex install --project — writes a ## graphify section into repo-local CLAUDE.md / AGENTS.md telling the agent to prefer graphify query "<q>" over grepping. On Claude Code this also adds a PreToolUse query-nudge hook (.claude/settings.json) that fires only before grep/find-style Bash calls and just suggests querying the graph — read-only, no rebuild, no git: categorically unlike a commit/husky hook;
• installs auto-update git hooks via graphify hook install — native post-commit + post-checkout scripts that rebuild the graph via AST on each commit/checkout. No LLM, no husky — plain .git/hooks. Doc/concept (LLM) changes still need a manual /graphify --update; the hook refreshes the code graph only;
• gitignores graphify-out/ (the graph artifacts) plus the regenerated/machine-specific wiring: the per-repo skill copies (.claude/skills/graphify/, .agents/skills/graphify/ — drift between graphify versions, codex copy ships buggy) and .codex/hooks.json (bakes in a machine-specific absolute graphify path). The portable guidance — the ## graphify section in CLAUDE.md/AGENTS.md and the .claude/settings.json query-nudge hook — stays committed.

It only drives the graphify CLI binary (its own interpreter — no python guessing). It does not build the graph itself — there is no CLI build verb; the full deep/semantic build is orchestrated by the /graphify skill (LLM subagents), which Step 3b runs next.

Step 3b — Build / refresh the deep graph (repo-scoped)

After wiring, build the graph at its best quality. The build is not a CLI call — invoke the /graphify skill on the repo root with --mode deep (richest semantic + INFERRED edges; the standing default per the ## graphify rule). This is an LLM step and may take a while on a large repo — expected, the user wants completeness over cost.

• graphify-out/graph.json absent → initial deep build: /graphify . --mode deep.
• graphify-out/graph.json present → refresh it: re-run /graphify . --mode deep (full deep re-extract — graphify update . is AST-only and would not refresh semantic edges, so it is not a substitute here).

Either way the command is the same (/graphify . --mode deep); only the framing differs. Skip only if the gate reported IN_GIT_REPO=false.

After the build, sweep leftover root scratch: rm -f ./.graphify_*.json. graphify drops pipeline intermediates as .graphify_{detect,ast,analysis,extract,labels,semantic}.json in the repo root and does not reliably clean them — confirmed on 0.8.18 (latest) after a normal completed deep build, so this is the common case, not a rare interruption. graphify-init.sh (3a) sweeps these before the build (clears the previous run's); run the same sweep after the build to clear the one this run just produced. Root-only glob — never graphify-out/ or the committed .graphify_version skill files.

Step 4 — Summary

Report only the steps that actually ran (omit graphify if it was gated out):

✓ [global] Caveman ultra + shrink + statusline (--all)
✓ [global] MCP shrink — allowlisted servers wrapped if present + stdio (remote/HTTP skipped)
✓ [global] headroom installed (savings extras) + Claude Code + Codex wired (durable hooks + proxy routing)
✓ [global] ponytail installed (ultra) + Claude Code + Codex wired
✓ [repo  ] Graphify CLI installed/updated + native always-on wiring (CLAUDE.md/AGENTS.md + query-nudge hook)
  · auto-update git hook installed (post-commit/post-checkout, AST rebuild)
  · graphify-out/ gitignored
  · deep graph built/refreshed via /graphify --mode deep

Then surface the natural follow-up (a pointer only — do not run it; it is an interactive skill outside this skill's deterministic scope):

• Only if IN_GIT_REPO=true AND docs/agents/ does not already exist: the engineering workflow skills (to-issues, to-prd, triage, diagnose, tdd, improve-codebase-architecture) need per-repo config (issue tracker, triage labels, domain docs). Suggest: run /setup-matt-pocock-skills. Omit this line if docs/agents/ is already present (already configured) or outside a git repo.

Remind user: restart Claude Code to activate the statusline, shrink wrappers, headroom proxy routing, ponytail plugin, and skill wiring.