Agent Loop Bootstrap

Your task: set up a self-perpetuating **AGENT LOOP** in the current repo, modeled after a story-driven loop pattern.

## Outcome
Create the minimal files + GitHub Actions workflow so the loop can run one iteration at a time:
- read state from files
- choose exactly one story
- implement minimal change
- enforce guard constraints
- commit + push
- push re-triggers the workflow until no todo stories remain or PAUSED is true

## Non-negotiables
- One story per run. Small commits.
- Never commit if guard fails.
- Keep `AGENTS.md` short (~70 lines). Only include repeated-footgun info.
- Persistent memory is ONLY repo files + git commits (no hidden memory).
- Prefer docs-only scaffolding first; don't add deps unless a story requires it.

## Create these files (if missing)
- `AGENTS.md`
- `.opencode/opencode.json`
- `scripts/loop/prd.json`
- `scripts/loop/progress.txt`
- `scripts/loop/constraints.json`
- `scripts/loop/failure.json`
- `scripts/loop/guard.sh` (diff guard enforcing constraints)
- `.github/workflows/loop.yml`
- `README.md` (briefly explains the AGENT LOOP + how to run)

## AGENT LOOP rules (how to operate each iteration)
1) Load `AGENTS.md` and state files in `scripts/loop/`
2) If `- PAUSED: true` → exit immediately
3) In `prd.json`, pick the first story with `"status": "todo"` (or highest priority todo)
4) Mark it `"doing"`
5) Implement minimal changes to satisfy the story acceptance criteria
6) Run verification (if the story specifies commands; otherwise keep docs-only)
7) Run `bash scripts/loop/guard.sh scripts/loop/constraints.json`
8) If green:
   - mark story `"done"` in `prd.json`
   - append learnings to `progress.txt`
   - commit + push with a bot identity
9) If failure:
   - update `failure.json` (increment consecutiveFailures, store run URL + timestamp)
   - if failures exceed `MAX_FAILURE_RETRIES`, set `- PAUSED: true` in `AGENTS.md`
   - commit + push failure state so the next run can retry (unless paused)

## Initial PRD requirements
Initialize `scripts/loop/prd.json` with 3-6 tiny stories:
- S1: Boot the loop automation scaffold (files + workflow)
- S2: Add/verify diff guard in CI
- S3: Make README describe how to run
(keep them small enough for one run each)

## Workflow requirements
- Trigger on `workflow_dispatch` AND on pushes to `AGENTS.md` and `scripts/loop/**`
- Auto-run only on bot pushes (manual dispatch always allowed)
- Use a PAT secret for checkout/push (do not hardcode tokens)
- Install Node + jq + OpenCode
- Run exactly one agent iteration with OpenCode using:
  `AGENTS.md`, `prd.json`, `progress.txt`, `constraints.json`, `failure.json`
- Validate JSON, run guard, then commit + push

Now implement the scaffold in this repo.