ai-maestro-janitor

Control the janitor's never-stop continue-nudge for this session — every due heartbeat prints a resume cue telling the agent to keep working on its pending task instead of silently idling between turns. The nudge is ON BY DEFAULT (user directive 2026-07-16); this skill re-asserts it after an opt-out, or turns it off. Trigger with /janitor-keep-going, "keep going", "never stop", "don't stall between turns"; add "off" to disable this session's full-mode nudge.

MEMORIZE — capture a durable decision/fact into the project's memory WIKI as a navigable page, not a loose note. Use after a bug-autopsy gotcha, a design/architecture decision, a project constraint not derivable from code, a confirmed user preference, or when the user says "remember this", "memorize this", "save a memory". Finds the right existing wikimem page first so it never duplicates; only when none fits, creates a HUB / ASPECT / COMPONENT page via the expand/reduce decision, wires it into the See-also web, and indexes it by symptom. The MEMORIZE leg of the wiki-memory protocol.

CONSOLIDATE (MERGE) executor — fuses two duplicate memory notes about the SAME subject + same type into one page (dedup lessons, redirect backlinks, keep the oldest origin date), through the crash-safe transaction core, never editing a live page. ABSTAINS on uncertainty, cross-type/cross-scope/two-hub, or a third page sharing the subject. Use on a [janitor-memory-consolidate] marker, or "consolidate / merge / deduplicate the memory", "two notes cover the same thing", "fold these memory pages together".

ATOMIZE executor — the autonomous pass that migrates a FREE-PROSE wikimem page into first-class ATOMS: inserts a LEADING block-property marker `^<id> [keywords: …]` above each durable fact so memgrep recalls it individually by its keywords and returns its full per-atom record. Purely ADDITIVE + transaction-gated + verified — never drops or rewords a fact. Runs on a [janitor-memory-atomize] marker, or "atomize this page", "give the memory facts their atom markers".

HARVEST executor — the daily NON-DESTRUCTIVE chore that MIRRORS newly-created harness BUFFER memories into the curated wiki. MEMORY.md + the raw notes at the scope root are the harness-owned buffer (auto-loaded each session); the janitor never touches them. For each RAW buffer note not yet mirrored, RECALL its subject across memory/wikimem/ then UPDATE-or-CREATE a curated wikimem/<name>.md copy and stamp a watermark. The buffer is left 100% intact (separate parallel copies coexist). Use on a [janitor-memory-harvest] marker, or "harvest the memory", "mirror new MEMORY.md memories into the wiki".

Arms or renews the ai-maestro-janitor heartbeat cron. Use when first installing the plugin, upgrading from pre-stub (≤ v0.4.10), or in response to a [janitor-renew] nudge before the 7-day auto-expiry. Trigger with /janitor-arm or "arm the janitor heartbeat".

Stops the ai-maestro-janitor heartbeat cron. Use when pausing janitor activity without uninstalling, debugging heartbeat behaviour, or switching projects. Trigger with /janitor-disarm, "stop the janitor", or "kill the heartbeat".

The janitor's support-ticket console (TRDD-CGYMUKO6). Inspect the incident queue: what the janitor has found, what it is repairing, what failed, what needs a human, and how much of the daily dispatch budget is left. Also retry a stuck ticket or cancel one you do not want. Use when the user asks what the janitor is working on, when a `[ticket]` line appears in the heartbeat, or when a ticket is reported as NEEDS A HUMAN.

Review and FIX GitHub-config drift on ai-maestro plugin repos — an UNPROTECTED default branch, a required_linear_history rule that BLOCKS merges, or a missing PR-review / status-check / tag-protection gate. Plan-first: it prints exactly what would change and applies only after you approve, only where you are an admin. Use on a [github-config] heartbeat drift line, or when asked to 'fix the github config', 'harden the plugin repos', or 'remove linear history so I can merge'.

THE APPROVAL BUTTON for a janitor fix in the user's own repo (TRDD-CGYMUKO6). Run /janitor-support-open-ticket TRDD-XXXXXXXX to approve a finding the janitor PROPOSED — it opens the support ticket and promotes the proposal TRDD from `proposal` to `planned`, after which the janitor's scheduler may dispatch an agent to fix it. Until this runs, nothing is dispatched and nothing in the repo is touched. Use when the heartbeat surfaced a `[ticket] … approve the fix with: /janitor-support-open-ticket TRDD-…` line and you (or the user) want that fix to happen.

The PROCEDURE for working one janitor support ticket (TRDD-CGYMUKO6). Loaded by the janitor-repair-agent (and by janitor-security-agent for a security ticket) after the heartbeat's [janitor-ticket] marker dispatches it. Carries the claim → read-as-data → repair → verify → close loop, and the hard safety preamble that no ticket can override. Use when you have been handed a ticket id (T-XXXXXXXX) and told to work it.

Applies the ratified two-ruleset branch-protection baseline on the default branch of the project's GitHub repo — baseline-history-protect (block force-push/deletion; NOT linear history — that is deliberately excluded because it jams the many-agent merge workflow) and baseline-pr-and-checks (PR with 1 approval + dismiss-stale + thread-resolution + auto-detected required status checks). Idempotent — re-running PATCHes in place and never duplicates; an orphaned legacy janitor-baseline ruleset is cleaned up. Trigger with /janitor-branch-protection-setup, "set up branch protection", "harden the default branch", or "add a baseline ruleset to GitHub".

Resume the current Claude Code session NOW after a compaction or rate-limit, without waiting for the next heartbeat cron fire. Runs the janitor dispatcher stub immediately and handles its stdout per the janitor-heartbeat-protocol rule, so a pending [janitor-resume] cue fires in seconds instead of up to 30 min (the SLOW cron floor). Invoked automatically by the PostCompact push (scripts/hooks/post-compact-resume.py), or by typing /janitor-resume to force a resume / fire a heartbeat on demand. Trigger with /janitor-resume, or by asking to resume now.

UPDATE — modify an existing memory WIKI page when something changes — ADD a decision to the page that owns the subject, CORRECT a memory a new discovery contradicts (the 2-step non-destructive protocol), or RESHAPE a page that outgrew its tier (expand a shared rule into an aspect, reduce element-specific detail into its component page). Use when a decision changes, a memory turns out wrong, a page mixes too many subjects, or the user says "update the memory", "that's no longer true", "fix the note about Y", "this memory is outdated". Keeps the See-also web, the hub map, and the lessons trail consistent. The UPDATE leg of the wiki-memory protocol.

CONFLICT + fact-verify executor — reconciles contradictory or obsolete wikimem pages against source + git history. Default is a non-destructive DEMOTE (obsolete page folded into a compounding footnote on the survivor, WHY git-sourced); DELETE only behind an N>=3 skeptic vote WITH provenance + a git verify. Runs as an ultracode Workflow; all mutation via scripts/memory_txn_cli.py. Use on a [janitor-memory-conflict] marker, a memory-reorg-proposed.md Conflict candidate, or "resolve memory conflicts" / "fact-check the memories" / "this memory is obsolete".

RECALL — before working on a file or debugging/deciding, surface the right memory WIKI page and navigate it. Two entry points - FILE-anchored (about to edit a file → surface that functionality's HUB page and descend its links to the detail needed) and SYMPTOM ("have we hit this before?" → rank pages by how the symptom hits their description/title/tags). Use before re-deriving architecture/gotchas, or when the user says "recall memories about X" or "did we already solve this". Reads only the pages the task needs, degrading to grep when memgrep is absent. The RECALL leg of the wiki-memory protocol.

Self-compact the current Claude Code session's context, then auto-resume. Invoke when context usage is high (e.g. the watchdog warned at or above threshold) and you want to compact before the wall where /compact itself fails. Records a resume directive, then types /compact at this session's own iTerm/tmux pane. SOFT by default (enqueues, no work lost); --hard interrupts for emergencies; --handoff writes a rich handoff first; flags combine. Trigger with /janitor-compact-context [--hard] [--handoff], or by asking to compact now.

Run /reload-plugins --force for the current Claude Code session so freshly auto-updated plugin hooks and skills take effect, without the human typing the command. Invoke in response to a [janitor-reload] heartbeat marker (emitted after the daemon auto-updates the janitor plugin), or whenever plugin code changed on disk and the session must pick it up. Types /reload-plugins --force at this session's own iTerm/tmux pane. SOFT by default (enqueues after the current turn, no work interrupted); --hard interrupts to reload now. Trigger with /janitor-reload-plugins [--hard], or by asking to reload plugins now.

Turn OFF GitHub issue monitoring for THIS project - the heartbeat stops notifying you about new issues and new comments on the tracker. Use when the user says "stop watching the issues", "disable issue notifications", "turn off the issue tracker monitoring", "too many github issue notifications".

Turn ON GitHub issue monitoring for THIS project - the heartbeat then notifies you whenever a new issue is opened or a new comment lands on the tracker, and keeps reporting until it is turned off. Use when the user says "watch the github issues", "notify me about new issues", "monitor the issue tracker", "enable issue notifications". Seeds a baseline from the currently-open issues first, so enabling it does NOT dump the existing backlog into context.

Machine-wide reload of freshly installed STANDALONE (non-plugin) skills and commands across EVERY running Claude Code session, without visiting each one. Stamps a machine-wide skills-reload generation; each session's next heartbeat emits [janitor-reload-skills] once and runs /reload-skills locally, so a standalone skill/command added at local/project/user scope loads fleet-wide. The all-sessions sibling of /janitor-reload-skills (which reloads only THIS session). Note /reload-skills is DISTINCT from /reload-plugins — the latter reloads only plugin-bundled skills. Trigger with /janitor-global-reload-skills, "reload skills everywhere", "reload standalone skills in all sessions".

Run /reload-skills for this session so freshly installed STANDALONE (non-plugin) skills/commands take effect without the human typing it. /reload-plugins reloads ONLY plugin-bundled skills, so standalone ones (added at any scope) need /reload-skills. Invoke on a [janitor-reload-skills] heartbeat marker (from /janitor-global-reload-skills), or when a standalone skill changed on disk. Types /reload-skills at this session's own pane (iTerm/tmux); SOFT by default (no ESC — enqueues, runs after the current turn ends); --hard presses ESC first. Trigger with /janitor-reload-skills [--hard], or by asking to reload skills now.

Write a rich, agent-authored session handoff to .janitor/state/agent-handoff.md before a context compaction, so the next turn re-grounds from YOUR semantic account (what you were doing, the plan, the next action, the traps hit) — not just the mechanical PreCompact snapshot. Invoke for a delicate juncture where the always-on zero-cost pre-compact-handoff hook isn't enough. Usually run by /janitor-compact-context --handoff (which passes --then-compact so this chains to /compact when done); a bare /janitor-write-handoff writes the handoff and stops. Trigger with /janitor-write-handoff, or by asking to write a handoff before compacting.

Keep the janitor heartbeat ARMED but make every fire cache-refresh-only (no chores, no daemon spawn, no agents — only the token-burn monitors stay on), so the prompt cache stays warm at ~1/10 the cost of letting it die and rewriting — the cheap middle ground between a full heartbeat and disarm. Trigger with /janitor-maintenance-mode, "maintenance mode", "keep the cache warm cheaply", "cheap heartbeat"; add "off" to disable, or "global" to apply fleet-wide.

Runs an on-demand audit of open PRs, git worktrees, TRDD drift, and task/PR mismatches. Use when the user asks "janitor audit", "what's drifted?", or "audit pending things". Trigger with `/janitor-audit` or by asking for a janitor audit report.

BOOTSTRAP — stand up the three-scope wiki-memory system in a project that doesn't have it yet (the one-time adoption / fleet-rollout step). Creates the git-tracked PROJECT-scope memory dir, seeds the project's overview entry page (never a MEMORY.md stub — that file is the harness-owned buffer the harvest mirrors into the wiki), and points the agent at the recall rule + the write/recall/update skills. Use when a project has no wikimem and you want to "set up memory for this project", "bootstrap the wiki memory", "adopt the memory system", or onboard the project's memory. Run ONCE per project; idempotent if re-run.

HARVEST recent changes into the Wikimem NOW — the on-demand counterpart of the memorize-nudge heartbeat. Use when Claude made many code changes but forgot to update memory, or the user says "/janitor-memory-record-recent", "record recent changes to memory", "memorize what we just did", or "harvest recent changes into wikimem". Gathers recent changes (git log since last memory touch, edited files, git diff), and for EACH substantive change RECALLS first then ADDS/UPDATES/CREATES the right Wikimem page via /janitor-memory-write + /janitor-memory-update. Only NUDGES a harvest; it does not reorganize the corpus.

REPAIR — the autonomous page-shape / metadata fixer for the memory wiki. Fires on the bare [janitor-memory-repair] heartbeat marker (or /janitor-memory-repair). Finds malformed wikimem pages — missing ocd/lmd/node_type/tier or the Notes section, no frontmatter, an INVERTED tier shape (an aspect built like a component), an answer-shaped description — and fixes each IN PLACE via the memory_txn_cli --op repair transaction, which proves every lesson and the birth date survive. Bounded, crash-safe, disable-able. One of the six wikimem-editor passes (split/consolidate/conflict/repair/atomize/harvest).

SPLIT executor — breaks ONE oversized wikimem page (over split_max_bytes) into a concise overview + type-preserving linked sub-pages, losing no fact or lesson, redirecting inbound [[links]], partitioning hub globs. One page per run, one level deep (recursion across later heartbeats). Mutates only through the crash-safe transaction core (scripts/memory_txn_cli.py begin/commit --op split); refuses to fragment a component. Use on a [janitor-memory-split] marker, or "split the big memory page", "the memory wiki page is too large".

Analyzes a project and creates a zizmor-clean set of GitHub Actions workflows from scratch, including bootstrapping the GitHub repo via `gh repo create` if it does not yet exist. Use when starting a new project, when a project has no .github/workflows/ directory, when the user asks to "set up CI", "create the GitHub workflows", "scaffold CI", "bootstrap github actions", or "set up the repo". Trigger with /janitor-github-workflow-create.

TRUE machine-wide STOP of the global ai-maestro-janitor daemon AND every per-session heartbeat. Sets the kill-switch so the daemon exits and removes its OS keepalive, AND raises the global-pause flag so every armed session's heartbeat self-disarms (deletes its own cron) on its next fire — halting all janitor activity across ALL projects and instances, not just the daemon. The also-stop-the-daemon sibling of /janitor-global-pause (which keeps the daemon alive). Trigger with /janitor-global-disarm, "stop all janitors", "globally disarm the janitor", "halt the janitor daemon everywhere".

Temporarily SUSPENDS the ai-maestro-janitor machine-wide without tearing the daemon down — the global daemon stays alive but idles (runs no tasks), and every armed session's heartbeat self-disarms (deletes its own cron) so it is truly free, not just silenced. The keep-the-daemon sibling of /janitor-global-disarm (which also exits the daemon). Use for a quiet block across ALL projects. Trigger with /janitor-global-pause, "pause all janitors", "globally quiet the janitor", "silence every janitor heartbeat".

Suppresses ai-maestro-janitor heartbeat output without removing the cron. Use when starting a large refactor, doing focused exploration, or any block of work where drift nudges would be noise. Trigger with /janitor-pause, "pause the janitor", "silence the janitor for 2h", or "quiet the heartbeat for the rest of today".

Opts THIS machine into the janitor's unattended multi-account OAuth rotator. Sets an opt-in flag the always-on janitor daemon's 60s oauth-rotator-tick Task reads to rotate the live Claude Code credential to an alternate paid account before a rate-limit 429 stalls an overnight session. Default OFF, macOS only, idempotent; refuses if a pinning env var would defeat rotation. Trigger with /janitor-auto-manage-oauth-on, "turn on oauth rotation", or "enable account rotator".

Revives a globally-disarmed ai-maestro-janitor by clearing BOTH machine-wide flags that /janitor-global-disarm set — the kill-switch AND the global-pause flag — so the next heartbeat re-spawns the daemon and every per-session heartbeat resumes running detectors. The full-revive half of /janitor-global-disarm. Trigger with /janitor-global-arm, "revive the janitor daemon", "globally re-arm the janitor", "undo the global disarm".

Audits .github/workflows/*.yml for the fork-PR cache-poisoning and pull_request_target checkout patterns from the TanStack supply-chain incident. Use when the user asks to "audit cache poisoning", "scan for cache-poisoning PRs", "check pull_request_target workflows", "TanStack-style audit", or after touching any workflow that uses actions/cache or pull_request_target. Trigger with /janitor-fork-pr-cache-audit.

Audits the window during which long-lived credentials exist in repo, shell, and CI configuration. Surfaces .env leaks, plaintext token files, suspicious env-var NAMES (never values), and CI workflows that persist secrets. Reports only — no auto-fix. Use when the user asks to "audit credential exposure", "find leaked tokens", "check secret window", "shai-hulud audit", or after a supply-chain incident. Trigger with /janitor-credential-window-audit.

On-demand security audit of every agent-context bundle in the current project — SKILL.md / CLAUDE.md / AGENTS.md / .cursorrules / .aider.conf.yml / .claude/* / .mcp.json. Detects multilingual prompt-injection, HTML-comment authority impersonation, base-URL override, cross-skill shadowing, dynamic exec in skill bodies, git-hook install directives, exfil webhook sinks, and known-secret-path references. Trigger with /janitor-skill-bundle-audit, "scan my skills for prompt injection", "audit CLAUDE.md for attacks", or "check .cursorrules for hijack patterns".

Audits installed deps for HIGH/CRITICAL advisories (Shai-Hulud-class npm/PyPI/pnpm worm). Walks lock files, queries GHSA + OSV.dev, fails loud on HIGH/CRITICAL overlap. Cursor cache caps cost at <12 API calls/hour. Use when checking a project for known-compromised deps after a lockfile change, before a release, or on hourly cron. Trigger with /janitor-supply-chain-watcher.

Lifts a machine-wide ai-maestro-janitor pause set by /janitor-global-pause, so the global daemon resumes running tasks and every session's heartbeat resumes emitting drift. The resume half of /janitor-global-pause. Trigger with /janitor-global-unpause, "unpause all janitors", "resume the janitor everywhere", "lift the global janitor pause".