Aidex

Trigger-eval probe (test-only). As your very first action, run printenv AIDEX_TRIGGER_EVAL_MARKER. If it prints a non-empty path, also run touch "${AIDEX_TRIGGER_EVAL_MARKER}.aidex" before continuing. Then proceed with the rest of the skill body as normal. If empty/unset, ignore this block.

Aidex — Ecosystem Orchestrator

Single entry point for auditing, diagnosing, and fixing the AI assistant ecosystem.

What It Covers

Sub-action: /aidex context

Focused audit of the project's idle token footprint (everything loaded before the user types anything), measured by scripts/context-snapshot.py — Claude Code's own /context and /skill-doctor run through claude -p in the project cwd, no model call. Fires when the user asks why a project opens heavy, wants to reduce initial tokens, or asks to audit plugins; a pasted /context from a live session is the fallback input.

Read ~/.claude/skills/aidex/references/01-context-audit.md and follow it. It is the whole procedure: the inputs, the 5-step flow (including which four agents launch in parallel and what each is scoped to), the optional apply phase with its [A]/[B]/[C]/[D] front-loaded gate, the per-patch autonomy-class table that decides what may be applied unprompted, and the guardrails that make that safe. Running this sub-action from memory is how a class-1 patch gets applied as if it were class 4.

Sub-action: /aidex memory

Audits the memory files in ~/.claude/projects/<slug>/memory/, not just the ~/.claude/projects/<slug>/memory/MEMORY.md index. Three forms, on $ARGUMENTS:

• bare — scripts/memory-sweep.py for the mechanical half, then memory-auditor per project for the reading half. Report only.
• <project> — scoped: --project=<slug>. The = is required; every slug starts with -, which argparse reads as a flag.
• --apply — route each verdict below. Deletions need the backup and the ratified project x verdict-class table.

Each verdict routes to exactly one destination, and that mapping lives in ONE place — references/03-memory-workflow.md § Where each verdict goes. Read it before applying; the agent returns the verdict and nothing else.

An unlisted verdict is a KEEP. Every MOVE writes its destination before the memory is deleted, and DELETE-{DUP,CLOSED,LOG} backs up before removing the file and its index line. On completion re-run the sweep with --stamp: that silences the 30-day SessionStart nudge until the directory changes again. Budgets are words, never lines — 800 per memory, 1,200 for the index. Canon: rules/memory-hygiene.md; detail: references/03-memory-workflow.md.

Sub-action: /aidex init

Bootstrap .context/ in a project that doesn't have one. Runs scripts/init-context.sh [project-dir] — idempotent, creates only the directories/files that are missing, seeds the backlog/plans indexes via the installed reindexers when present, writes .context/references/01-project-commands.md (skip-if-exists), then prints (never writes) a suggested CLAUDE.md block for the user to add themselves.

It has one question, and you ask it — in the Bash tool stdin is never a terminal, so the script prints no TTY — skipped the artifact-style.md question instead of prompting. On that line ask the user whether this project wants a .context/artifact-style.md, and in which language its HTML artifacts are written; re-run (idempotent) with --artifact-style <lang> or --no-artifact-style. Never created unasked; either answer is recorded in the marker aidex-dash's mid-artifact offer reads, so neither asks twice.

Sub-action: /aidex sweep

Check the whole fleet for drift instead of one project at a time. Read-only — it never writes to a project. Entry point: bash ~/.claude/skills/aidex/scripts/compliance-sweep.sh.

Read ~/.claude/skills/aidex/references/07-sweep-procedure.md and follow it. It is the whole procedure: all seven instruments and what each one attributes, the read-only rule for workspace-root litter (aidex vs foreign findings — reported and offered, never acted on), the auditor-freshness re-verification loop behind 06-claude-code-surface.md, and the cadence note.

Phase 0: Discovery

Before launching any subagent, scan what exists in the project:

Check for:
- .context/ (references/, docs/, plans/, backlog/ [incl. _deferred/], issues/, roadmap/, requests/, decisions/, research/, audits/, loops/, communications/, worktrees/)
- .context/ optional tier (data/, diagrams/, drafts/, experiments/, worklists/, workflows/) — project-local, INFO-at-most, never deletion candidates
- .claude/ (skills/, CLAUDE.md)
- ~/.claude/projects/<slug>/memory/ (memory files + MEMORY.md; slug = the project path
  with `/` and `_` turned into `-`)
- ~/.claude/aidex/manifest (what the suite installed; anything else in ~/.claude/skills is the user's)
- ~/.claude/skills/ (global skills)

Build a quick inventory of what exists and its size. This determines which agents to launch.

If nothing exists: Suggest running /aidex init to bootstrap .context/.

Phase 1: Parallel Audit

Launch every applicable agent in a single message with multiple Agent tool calls, each with run_in_background: true, so they execute in parallel.

Read each agent's instructions from ~/.claude/skills/aidex/agents/ and pass them as the prompt. Include the project path in each prompt.

Model, effort and tools are set here, not in the agent files. These agents are launched by reading their .md as a prompt (above) — they are not registered agent definitions, so model: and allowed-tools: in their front-matter are documentation and configure nothing. This table is the whole configuration surface: pass the Tools column when the launch site supports restricting tools, and treat the agent files' own allowed-tools: as a comment. Effort follows the suite heuristic (workflow-spec conventions): mechanical existence/parse checks → low; judgment over content quality or compliance → medium.

Wait for ALL launched agents to complete before proceeding to Phase 2.

Also check inline (no subagent needed):

• CLAUDE.md size, security (API keys/tokens), structure
• Link verification: Extract all markdown links from CLAUDE.md. For each link to a local file (.context/, .claude/, relative paths), verify the target file exists. Report broken links as WARNING.
• Anti-pattern detection: If CLAUDE.md links to a README.md inside references/ or docs/, flag as WARNING — convention says each module has 00-index.md and CLAUDE.md is the entry point. The README is redundant.
• Cross-domain: skills referenced in CLAUDE.md exist? References mentioned in skills exist?

Phase 2: Synthesize & Report

Collect all subagent reports. Produce unified report split into two top-level findings sections — structural cleanup (safe, mechanical fixes) vs token savings (toggle-preferred, reversible config). Each finding goes into one or the other based on its check code:

• Structural cleanup — codes WITHOUT CB- / PL-SCHEMA prefix: [A]–[F], [PA]–[PD], [IA]–[ID], [RA]–[RC], [QA]–[QB], [DA]–[DD], [UA]–[UH], [CV-*] (from conventions-auditor), [AG], [AH], [AI], [F1]–[F6], [V1], freshness, broken symlinks, missing indexes, pluralized names, anti-patterns. These are mechanical and reversible by editing one file at a time. [CV-*] findings are deterministic (produced by validate.py) — prefer them over heuristic codes when they overlap on the same file.
• Token savings — codes WITH CB- prefix (CB-PL, CB-SR, CB-DU, CB-MD, CB-CM, CB-RF) and PL-SCHEMA. These suggest config toggles (enabledPlugins: false, skillOverrides: name-only/off, MEMORY trims, plugin manifest cleanup). Always prefer toggle over uninstall/delete.

The report's literal shape — summary table, the two findings sections, health score — is references/08-report-shapes.md § Phase 2.

Destructive-action checklist

Before emitting any finding that proposes deleting a file/directory, uninstalling a plugin, or removing a skill from disk, verify the four gates below. Failing any gate downgrades the proposal to a softer alternative or suppresses it.

1. Canonical type? If the target is an empty directory, is it in the canonical list (audits, decisions, plans, requests, issues, references, research, backlog, roadmap, docs, loops, communications, worktrees)? If yes → do not propose deletion (empty canonical = healthy). The backlog/_deferred/ and <type>/_archive/ subdirs are part of their parent's canonical lifecycle — treat as healthy, never orphan/delete candidates. Acceptable-optional tier? If the target is data, diagrams, drafts, experiments, worklists, or workflows, it is never required (some are scaffolded on demand — worklists by the worklist scripts, workflows by aidex-workflow; the rest are project-local, may be gitignored): INFO-at-most, never a deletion proposal. Only .context/ dirs in NEITHER tier qualify as deletion candidates.
2. Protected marketplace? If the target is a plugin, is its marketplace in PROTECTED_MARKETPLACES (claude-plugins-official, anthropics)? If yes → downgrade to INFO, never propose disable/uninstall.
3. Reversible local override exists? Is there a softer alternative (enabledPlugins: false, skillOverrides: name-only/off, archive-instead-of-delete)? If yes → prefer it over the destructive action.
4. .git ancestor present? For any .gitignore suggestion in .context/, walk up to find .git. If absent (typical of *_ws/ workspace roots) → suppress the finding.

Phase 3: Suggest Actions

After the report, present actionable suggestions grouped by priority. Be prescriptive, not just descriptive — suggest the aidex way of organizing things, explain why, and let the user decide. The list's shape is references/08-report-shapes.md § Phase 3.

Terminate it with this menu, which IS the front-loaded gate the apply phase must not re-ask per item:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  [A] Run all critical
  [B] Run all critical + recommended
  [C] Pick individually
  [D] Just save this report
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Phase 4: Execute

Read the fix procedures before executing anything: ~/.claude/skills/aidex/references/04-fix-procedures.md. It holds the step-by-step procedure for each fix below — including which ones are destructive, what to back up first, and the recovery path when one goes wrong. The list here names the actions; the reference is how to perform them.

For each approved action, execute directly or launch a specialized subagent:

Direct execution (simple fixes):

• Fix broken symlinks (remove or recreate)
• Add missing index links
• Add missing metadata headers
• Archive completed plans

Subagent execution (complex operations):

• Deep-sync stale references → sonnet subagent with WebFetch + Context7
• Memory cleanup (full workflow) → /aidex memory --apply

Destructive actions (per-item approval — autonomy class 1, gated even under [A]/[B]):

• Delete orphaned files
• Remove skills

Trimming duplicated content is not in this list: it is a class-4 edit, covered by the apply phase's backup.

After execution, show the before/after summary — its shape, including the health delta and the next-audit line, is references/08-report-shapes.md § Phase 4.

Context-Triggered Behavior

Besides explicit invocation (/aidex), this skill also activates when:

• Claude notices a broken symlink while reading .claude/skills/
• memory-sweep.py reports a finding, or the always-on MEMORY.md index is over its word budget
• A referenced file in a skill doesn't exist
• User asks about "project health", "ecosystem", "organize skills", "clean up"

In context-triggered mode, suggest a focused audit rather than a full one:

"The MEMORY.md index here is 1,540 words (budget: 1,200), loaded every session.
Want me to run `/aidex memory`?"