Pi Config

Personal global Pi configuration. Clone it to ~/.pi/agent/ for shared settings, skills, prompt templates, extensions, and model configuration.

Setup

mkdir -p ~/.pi
git clone git@github.com:HazAT/pi-config ~/.pi/agent
cd ~/.pi/agent && ./setup.sh

Add provider credentials to ~/.pi/agent/auth.json and restart Pi. setup.sh requires this repository at ~/.pi/agent/, preserves an existing settings.json, installs the configured Git-backed packages, and installs the macOS Harness runtime with uv on macOS.

Session-file workflow

The workflow is four explicit prompt commands:

1. /plan <request> — loads the planning skill, investigates interactively, records the selected approach, and stops.
2. /todos — loads the todo-writing skill, converts the selected plan into worker-ready todos, and stops.
3. /execute — reads the plan and todos, uses raw Superconductor orchestration, runs source-writing workers sequentially, verifies each focused commit, and updates todo evidence from the coordinator session.
4. /review — runs an independent local SC review team, automatically repairs blocking findings with sequential workers, and re-reviews to approval or a bounded stop.

PI_SESSION_FILE is required. Handover state is plain Markdown beside the current Pi session:

${PI_SESSION_FILE%.jsonl}.plan.md
${PI_SESSION_FILE%.jsonl}.todos.md
${PI_SESSION_FILE%.jsonl}.review.md

There is no sidecar directory, repository run store, membership system, or custom JSONL state. The current coordinator owns all writes. Delegated agents receive resolved absolute paths in their SC prompts, read the handovers directly, and return results through SC. /review requires only the session JSONL; the plan and todos files are optional context so directly implemented work can be reviewed without first running /plan or /todos.

The todos file uses stable IDs, dependency and status fields, objective acceptance criteria, verification evidence, and commit SHAs. Workers never edit it themselves.

Superconductor policy

Workflow orchestration uses the sc CLI directly. Before launching agents or teams, read skills/superconductor/SKILL.md and run sc instructions orchestration; live instructions and capability output override remembered syntax.

• Source-writing workers share the checkout and run sequentially.
• Every successful worker reads skills/worker/SKILL.md and the commit skill, creates one focused verified commit, reports its full SHA, and does not push.
• The coordinator waits for and reads each stable worker target, checks errors, verifies the commit with Git, then updates the todos file.
• Review team members are independent and read-only. They report through their SC team context; only the coordinator writes the review file.
• Synchronous teams omit --notify self; the coordinator waits for role targets and collects durable reports from sc team status without requiring the user to submit a generated follow-up.
• A NEEDS CHANGES review launches sequential repair workers automatically, verifies their commits, and runs a fresh review. Repair is bounded to two rounds.
• Dispatch or idle state never proves completion.

The workflow does not create worktrees or branches, arrange fixed panels, ship, push, open PRs, merge, or clean up sessions. Those outcomes require separate explicit requests and the relevant live SC instructions.

Repository layout

Pi discovers global prompt templates from ~/.pi/agent/prompts/*.md and skills from directories under ~/.pi/agent/skills/ that contain SKILL.md.

Verification

jq empty settings.json models.json mcp.json package.json
pi --no-extensions --no-context-files \
  --skill ./skills/plan/SKILL.md \
  --prompt-template ./prompts/plan.md \
  --list-models gpt-5.6
git diff --check

Also search the repository for names from the deleted extension and confirm no stale API or package references remain.

Update

cd ~/.pi/agent
git pull
./setup.sh