Imported from MerverliPy/SelfTUI (AGENTS.md). Install upstream with npx skills add MerverliPy/SelfTUI. Copyright stays with the author.

SelfTUI — Project Instructions & Session Rule

This file governs all work in this repo. It is read at the start of every session and takes precedence over the global agent rules for anything it specifies.

Binding session rule — one fresh session per step

"Step" is defined as: one milestone from PLAN.md §10 (M0, M0a, M1a, M1b, M2, M3a, M3b, M4, M5, M6) or one clearly-scoped task the owner assigns in a message.

1. One step per session (binding). Never start a new step in the same session that completed a step. When a step is done (changes committed), the session STOPS. The next step is begun only by opening a new pi session with cwd at this repository's root.
2. Session-start ritual (read-only, in this order): AGENTS.md → PLAN.md §10 (roadmap) + §11 (risks/owner decisions) → the tail of LEDGER.md (latest handoff). Then choose exactly one step. Do not re-plan the whole build or chain steps.
3. Session-end ritual: complete the chosen step → append a handoff entry to LEDGER.md (work done · commands+exit codes · decisions · blockers · next action) → tick the milestone's exit condition in PLAN.md §10 → commit → stop. Do not open the next milestone in this session.
4. Evidence over assertion. Run the canonical checks (build/test/lint once they exist — see make targets as they are added) and report green only from real output.
5. Consult & append, never rewrite. LEDGER.md is the past/handoff source of truth; PLAN.md is the future. Append new entries; do not rewrite history.
6. Scope discipline. No unplanned scope, no speculative scaffolding. Match existing patterns. If a step is ambiguous or blocking, stop and ask rather than guessing.
7. Safety (carried from the council audit). Safety controls ship inline with the tool that exposes them; no mutation/run_command tool ships before its workspace jail, confirmation, timeout, cancellation, and tests exist.
8. Model reality (spike-verified). Default agent model is qwen3:8b (native tool_calls works). The agent tool layer must support dual dispatch — native message.tool_calls and tool-JSON parsed from content — and must handle qwen3's thinking phase. See PLAN.md §6.

Where the truth lives

• PLAN.md — future: architecture, roadmap (§10), risks/owner decisions (§11).
• LEDGER.md — past: chronological handoff log per session.
• COUNCIL-MEMO.md — the advisory verdict that re-cut the roadmap.

Agent skills

Issue tracker

Issues live in this repo's GitHub Issues (github.com/MerverliPy/SelfTUI), used via the gh CLI. See docs/agents/issue-tracker.md.

Triage labels

The five canonical triage roles use their default label strings: needs-triage, needs-info, ready-for-agent, ready-for-human, wontfix. See docs/agents/triage-labels.md.

Domain docs

Single-context layout: one CONTEXT.md at the repo root plus docs/adr/. See docs/agents/domain.md.