Cron-driven autonomous loop for the live test dashboard

scripts/dashboard-loop.prompt.txt

@@ -0,0 +1,21 @@
+GraphDone autonomous continuous-improvement iteration. The live test dashboard (http://localhost:3199, `npm run dashboard`) is your guiding light. Repos: Core = /home/scubasonar/Code/graphdone-repos/GraphDone-Core (work on branch dev), Cloud = /home/scubasonar/Code/graphdone-repos/GraphDone-Cloud (branch main). Read the memory index at /home/scubasonar/.claude/projects/-home-scubasonar-Code-graphdone-repos/memory/MEMORY.md first. Do ONE short, safe, well-scoped iteration, then stop and exit.
+
+CONTEXT: The mechanical upkeep (keeping the dashboard server alive + running test suites to feed it fresh data) has ALREADY been done by the shell driver before you were invoked. Your job is the judgment part: look at the data and make at most one concrete improvement.
+
+STEPS:
+1. Read the current dashboard state: `curl -s http://localhost:3199/api/state`. Identify the newest runs, any failing/warn checks, and any regression vs prior runs in the metric trends (FPS, pass-rate, durations, drift).
+2. If there is a CONCRETE, low-risk, well-scoped fix or improvement (a failing/warn check to fix, a flaky test to stabilize, a clear perf regression, a small dashboard enhancement): implement it. Write or adjust a test first where feasible. Run the relevant gate — when you touch data, the graph, or deletion paths run `TEST_URL=http://localhost:3127 npm run test:smoke` (only if the dev stack at :3127 is up; if it is not up, restrict yourself to changes provable by `node --test` / unit tests and do NOT claim the smoke gate passed).
+3. If (and only if) your change is complete and its gate is green, commit on a feature branch and merge to dev via the branch-protection bypass:
+   gh api -X DELETE repos/GraphDone/GraphDone-Core/branches/dev/protection/required_pull_request_reviews
+   gh pr merge <n> --merge --delete-branch
+   gh api -X PATCH repos/GraphDone/GraphDone-Core/branches/dev/protection/required_pull_request_reviews -F required_approving_review_count=1
+   Use scoped `git add <paths>` (NEVER `git add -A`). End the commit message with:
+   Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
+4. Record anything durable in memory (one file per fact, update MEMORY.md index).
+
+HARD CONSTRAINTS (never violate):
+- NEVER deploy to Cloudflare (do not run scripts/deploy.sh) — deploys are user-triggered only.
+- NEVER read, log, expose, or commit secrets (ADMIN_PASSWORD, GUEST_JWT_SECRET, RESEND key); admin login stays secret-backed only; no end-user passwords.
+- NEVER weaken auth, tenant isolation, or any security control. Tear down any test fixtures you create.
+- NEVER merge anything that fails THE GATE or whose gate you could not actually run.
+- If there is nothing safe and concrete to do this iteration, do nothing further and exit cleanly. Do not invent risky or speculative work. Brevity over volume.

scripts/dashboard-loop.sh

@@ -0,0 +1,63 @@
+#!/usr/bin/env bash
+# Autonomous continuous-improvement loop for GraphDone, driven by cron so it
+# survives session limits AND reboots. Each invocation (cron passes "once"):
+#   1. ensures the live test dashboard server is up (the guiding light)
+#   2. self-checks the dashboard (its own unit tests)
+#   3. runs ONE bounded headless `claude -p` improvement iteration — the agent
+#      decides whether to run a real test suite (which is what feeds the
+#      dashboard fresh run data); the driver never fabricates synthetic runs.
+# Single-instance via flock. Set NO_AGENT=1 to skip step 3 (mechanical only).
+#
+#   bash scripts/dashboard-loop.sh once
+set -uo pipefail
+
+CORE="/home/scubasonar/Code/graphdone-repos/GraphDone-Core"
+STATE="$CORE/test-artifacts/dashboard"
+LOG="$STATE/loop.log"
+LOCK="$STATE/loop.lock"
+PORT="${DASHBOARD_PORT:-3199}"
+AGENT_TIMEOUT="${AGENT_TIMEOUT:-1500}"
+
+# cron runs with a minimal env — put nvm node + ~/.local/bin (claude) on PATH
+export PATH="/home/scubasonar/.nvm/versions/node/v20.20.2/bin:/home/scubasonar/.local/bin:$PATH"
+export NODE_NO_WARNINGS=1
+
+mkdir -p "$STATE"
+log() { echo "[$(date -Is)] $*" >> "$LOG"; }
+
+# single instance: bail if a previous iteration is still running
+exec 9>"$LOCK"
+if ! flock -n 9; then log "skip: previous iteration still running"; exit 0; fi
+
+cd "$CORE" || { log "fatal: cannot cd $CORE"; exit 1; }
+log "── iteration start (pid $$) ──"
+
+health() { curl -s -o /dev/null -w '%{http_code}' "http://localhost:$PORT/api/state" 2>/dev/null; }
+
+# 1) keep the dashboard alive
+if [ "$(health)" != "200" ]; then
+  log "dashboard down → starting"
+  nohup npm run dashboard >/tmp/gd-dash.log 2>&1 & disown
+  for _ in 1 2 3 4 5 6; do sleep 1; [ "$(health)" = "200" ] && break; done
+fi
+log "dashboard health=$(health)"
+
+# 2) self-check the dashboard tool (cheap; does NOT fabricate runs)
+if node --test tests/lib/dashboard/ >>"$LOG" 2>&1; then log "dashboard lib tests ok"; else log "dashboard lib tests FAILED"; fi
+
+# 3) one bounded headless agent iteration (skippable)
+if [ "${NO_AGENT:-0}" = "1" ]; then
+  log "NO_AGENT=1 → skipping agent iteration"
+elif command -v claude >/dev/null 2>&1 && [ -f "$CORE/scripts/dashboard-loop.prompt.txt" ]; then
+  log "agent iteration start (timeout ${AGENT_TIMEOUT}s)"
+  if timeout "$AGENT_TIMEOUT" claude -p "$(cat "$CORE/scripts/dashboard-loop.prompt.txt")" \
+      --permission-mode bypassPermissions >>"$LOG" 2>&1; then
+    log "agent iteration done"
+  else
+    log "agent iteration ended (rc=$?)"
+  fi
+else
+  log "no claude CLI or prompt file → mechanical upkeep only"
+fi
+
+log "── iteration complete ──"

tests/lib/dashboard/README.md

@@ -76,3 +76,15 @@ Read-only and **bound to `127.0.0.1`** (no LAN exposure, no auth surface). No
 command-execution endpoints. Media is served only from within a run's own report
 directory (real-path checked, traversal/symlink-escape rejected) and only for an
 allowlist of media extensions; `/static/` serves a fixed three-file allowlist.
+
+## Autonomous loop (`scripts/dashboard-loop.sh`)
+
+So the dashboard keeps living and improving across sessions/reboots, a cron-driven
+loop runs `scripts/dashboard-loop.sh once` every ~30 min (system crontab, tagged
+`# graphdone-dashboard-loop`). Each iteration: (1) ensures the dashboard server is
+up, (2) self-checks it (`node --test tests/lib/dashboard/`), (3) runs ONE bounded
+headless `claude -p` improvement iteration against `scripts/dashboard-loop.prompt.txt`.
+A `flock` lock prevents overlap; the agent prompt has hard constraints (never
+deploy, never touch secrets, never merge a failing gate). Run mechanical-only with
+`NO_AGENT=1 bash scripts/dashboard-loop.sh once`. Disable the loop by removing the
+`graphdone-dashboard-loop` lines from `crontab -e`. Logs: `test-artifacts/dashboard/loop.log`.