Runs before the agent uses a tool for Bash, Edit, Write and Agent tool calls, executing preflight-bash.sh, dispatcher-guard.sh, worktree-write-guard.sh and agent-worktree-guard.sh via bash (4 commands). From judgemind/judgemind.
Weekly LLM carry-forward audit across CA counties — runs scripts/auditllmcarryforward.py, applies per-axis thresholds, files agent/ready follow-ups for trips, and posts a heartbeat comment on the long-lived audit-log issue. Triggered by the daemon scheduler weekly.
Periodic codebase health audit — reviews recent PRs, checks for dead code, test gaps, performance issues, security concerns, dependency health, and CI pipeline health. Files issues for findings. Triggered by the dispatcher every 20 merged PRs.
Diagnose a dispatcher-v3 task-runner failure. Reads the agent's session transcript from S3 (compact), falling back to raw jsonl or CloudWatch Logs. Takes side effects directly with full peer-agent authority — close issue, file follow-up, comment, edit labels, or re-add agent/ready to retry. Writes…
Periodic dispatcher operational-health probes — runs convergence-regression, bad-outcome-streak, and site-health checks and files GitHub issues for any findings. Triggered by the daemon scheduler every 6 hours.
Daily dispatcher failure-summary report — queries last-24h aggregations from dispatcher., renders phone-readable markdown, and auto-PRs to docs/dispatcher-daily/YYYY-MM-DD.md. Triggered by the daemon scheduler at 12:00 UTC.
Dispatcher startup subagent — runs the heavy one-shot startup queries (stale-worktree sweep, stale-assignment cleanup, agent/ready queue scan, trust-check, orphan-PR triage) in an isolated subagent context and returns a compact markdown summary. Keeps the dispatcher's main context clear of 100k tokens of raw GitHub…
File a GitHub issue in judgemind/judgemind with the repo's house conventions locked in — conventional-commits title, body written to a temp file, Verify lines on every acceptance criterion, proper labels, and correct blocking dependencies. Thin guardrail — the full authoring reference lives in…
Ralph loop — iterative work-review cycle with fresh context each iteration. Spawns a worker subagent to implement, then runs up to three reviewers sequentially (Gemini standard, Gemini adversarial, Claude). Loops until all required reviewers agree to SHIP or max iterations reached. Called by /task for implementation…
Take a screenshot of a page on dev.judgemind.org for visual iteration. Use when working on frontend tasks to see what the page looks like, verify layout changes, or debug UI issues.
Periodic data quality spot-check — a senior engineer doing a high-level review for problems a human reviewer would notice. Discovers, root-causes, and (when scoped) fixes inline.
Fix-conflict phase for the per-phase /task-v2 pipeline. Reads a rebase-conflict bundle, resolves conflicts semantically against updated origin/main content, and produces either a resolved-file set OR an unresolvable signal.
Operational handler for the per-phase /task-v2 pipeline. Executes non-coding tasks (script runs, DB queries, gh actions, data rebuilds) without creating a PR. Emits a structured verdict for the dispatcher to advance the agent pipeline.
Plan phase for the per-phase /task-v2 pipeline. Reads an issue + comments, produces a plan document, scope-check findings, and a go/no-go signal. Called once per task by the dispatcher daemon before /task-v2-ralph.
Ralph phase for the per-phase /task-v2 pipeline. Iterative worker + reviewer loop, returns SHIP verdict plus a committed implementation diff on the worktree branch. Long-tail phase (45-90 min internally).
Retrospective phase for the per-phase /task-v2 pipeline. Reads the full agent history (phasetransitions, failures, PR URL), produces zero-to-many retro issue bodies for the daemon to file.
Summary phase for the per-phase /task-v2 pipeline. Reads the issue body and the git diff, produces a process-summary issue comment (with AC mapping), a conventional commit message, a PR title, and a PR body.
Verify phase for the per-phase /task-v2 pipeline. Reads the merged PR plus deploy status plus acceptance criteria, produces a verification-evidence comment with per-criterion proof against dev.