sdlc-loop

# sdlc-loop Drive the backlog autonomously. The Python helpers own state/budget + the backlog source; you run each goal. The source is config-selected (`.sdlc/config.json` → `discovery.source`): **local goal files** (default) or **GitHub issues** (`source: github`, needs an authenticated `gh`). You run the loop the same way either way — the helper handles where goals come from and how status is recorded. First, reset the per-run budget: `python3 "${CLAUDE_SKILL_DIR}/scripts/loop.py" start .sdlc` Then repeat until the helper says stop: 1. `goal=$(python3 "${CLAUDE_SKILL_DIR}/scripts/loop.py" next .sdlc)` 2. If output is `DONE` (backlog empty) or `BUDGET` (a per-run budget hit: iterations always; wall-clock minutes / reported tokens when `config.json` sets them) → STOP. If the host surfaces token usage to you, report it between goals — `loop.py spend .sdlc <tokens>` — so `budget.max_tokens` can actually enforce; never guess a number (no report = no token cap). 3. Otherwise: first **recall prior art** — if the knowledge graph is enabled, run the `sdlc-context` pre-flight to pull a cited brief from the graph + past issues + conventions, so the goal starts informed by history instead of a flushed window (no-op when the KG is off). **Match the model to the goal** — run `python3 "${CLAUDE_SKILL_DIR}/../sdlc-model/scripts/predict.py" resolve "$goal" .sdlc`. If it prints a tier (`haiku`/`sonnet`/`opus`/`fable`), that tier is the GOAL's ceiling — run its pha...