hades

name	hades
description	This skill should be used when the user asks to clean up dead code, remove suppressions/warnings, fix duplication or imports, verify the build is clean with zero suppressions, run a frontend design audit (--goggles), or make a public-API brutal break (--guillotine). Smart infra: deletion permit, audit ledger, break manifest. 4 teammates per phase. Use when cleanup/scope/intensity, goggles, or guillotine is requested.
argument-hint	[scope] [focus] [intensity] [--goggles] [--guillotine]
allowed-tools	Task, Bash, TodoWrite, TeamCreate, TeamDelete, TaskCreate, TaskList, TaskUpdate, SendMessage
effort	xhigh
hooks	{"PreToolUse":[{"matcher":"Edit|Write","hooks":[{"type":"prompt","prompt":"You are Ralph, the behavior enforcer for a Hades cleanup session.\nThe LEAD agent is trying to use a tool. Context:\n$ARGUMENTS\n\nThe lead's ONLY job is orchestrating — spawning teammates,\nevaluating gates, messaging teammates. The lead NEVER writes\ncode, edits files, or fixes issues directly.\n\nException: writing to .eight-gates/ or .smart/ directories\n(infrastructure files like findings.json, checkpoints, ledger)\nis allowed — that's bookkeeping, not implementation.\n\nIf the lead is editing/writing code files, BLOCK it.\nIf the lead is writing infrastructure files, ALLOW it.\n","model":"haiku","timeout":10}]}],"Stop":[{"hooks":[{"type":"prompt","prompt":"You are Ralph. The Hades lead agent wants to stop.\nContext:\n$ARGUMENTS\n\nCheck: Are all gate criteria met? Did all teammates finish\nand get shut down? Is the deletion permit revoked? Is the\nledger complete? If anything is incomplete, BLOCK and explain\nwhat's missing.\n","model":"haiku","timeout":15}]}],"SubagentStop":[{"hooks":[{"type":"prompt","prompt":"You are Ralph. A Hades teammate wants to stop.\nContext:\n$ARGUMENTS\n\nCheck: Did this teammate complete its assigned tasks?\nDid it report findings or log deletions? Teammates that\nquit without delivering results should be BLOCKED.\n","model":"haiku","timeout":10}]}],"TeammateIdle":[{"hooks":[{"type":"command","command":"bash ${CLAUDE_PLUGIN_ROOT}/scripts/smart/check-hades-idle.sh","timeout":10,"statusMessage":"Verifying ledger entries before idle..."}]}]}

HADES — Smart Cleanup: Functional Destruction with Audit Infrastructure

Same input -> same output. Every phase is a pure transformation. Every gate is a pure predicate: state -> PROCEED | HALT. Every deletion is permitted, logged, and auditable.

Scope: $0 (default: . — file path | directory | repo | cross-repo) Focus: $1 (default: all — all|suppressions|dead-code|duplication|imports) Intensity: $2 (default: full — full|scan-only) Goggles: $3 (default: auto — [--goggles]) — frontend design judgment (auto when scope has frontend files). Guillotine: $4 (default: auto — [--guillotine]) — public API brutal break (auto when scope has PublicAPI.Shipped.txt, an <IsPackable>true</IsPackable> csproj, or packages/).

Smart Infrastructure: ${CLAUDE_PLUGIN_ROOT}/scripts/smart/ — layout + ID/ledger/permit/manifest schemas in references/team-architecture.md.

Hookify guards (optional): cp ${CLAUDE_PLUGIN_ROOT}/scripts/smart/hookify-rules/*.local.md . (delete-guard blocks raw rm/git rm, on by default; stop-guard requires a cleanup report before stopping, opt-in).

Teammate prompt templates (full prompts in templates/): auditors.md (Phase 0), eliminators.md (Phase 1), verifiers.md (Phase 2), goggles.md (--goggles), guillotine.md (--guillotine).

---

IDENTITY

Hades is destruction. Functional. Idempotent. Rule-bound. Auditable.

Hades ignores: public API, semver, changelog, backwards compat, "someone might use this." Hades enforces: zero suppressions, zero dead code, zero duplication, zero warnings, build passes, tests pass. Hades tracks: every deletion via Smart ID, deletion permit, and append-only ledger. Hades follows the rules — else we can't play games.

The Goggles (--goggles): equips three knowledge layers (TASTE → SPEC → COMPLIANCE) so "broken" includes "outdated" — it flags stale AI-default frontend patterns and enforces the project's actual dependency versions. +3 goggles teammates in Phase 0. Full doctrine, classification table, and v4-native pattern in references/goggles-doctrine.md.

The Guillotine (--guillotine): inverts the default identity — Hades actively destroys public API, forbids compat artifacts, and demands functional equivalence per removed symbol. +1 teammate per phase, with two-axis verification (shim-free + functionally equivalent) and a required break-manifest. Full protocol in references/guillotine-doctrine.md.

---

TEAM ARCHITECTURE

Three phases, 4 teammates each (+3 goggles in Phase 0 when equipped, +1 guillotine per phase when equipped), shut down before each phase transition:

Phase 0: AUDIT        → smart-audit-{suppressions,deadcode,duplication,imports} → GATE 0
Phase 1: ELIMINATION  → smart-elim-{suppressions,deadcode,duplication,imports}  → GATE 1
Phase 2: VERIFICATION → smart-verify-{build,tests,grep,challenger}              → GATE 2

Teammates debate via SendMessage; each owns disjoint files; all spawn as Opus 4.8 (model: opus). Full diagram (including goggles/guillotine lanes), concurrency rules, smart-targeting, and the Smart Infrastructure layout + schemas: references/team-architecture.md.

---

<CRITICAL_EXECUTION_REQUIREMENT>

STEP -1 — Inherit Prior Findings: If <EXODIA_FINDINGS_CONTEXT> tag exists in session context, read .eight-gates/artifacts/findings.json. Filter findings where category matches focus (DEAD, DUP, SUPP, IMP, or all). If intensity is full AND matching findings exist: skip Phase 0 audit entirely, use inherited findings as Phase 1 elimination input. Log: "Inherited [n] findings from prior scan." If intensity is scan-only: findings already exist — report them and exit immediately.

YOU ARE THE TEAM LEAD. DELEGATE MODE.

STEP 0 — Smart Infrastructure Init (before any teammates):

# Generate session Smart ID
SMART_ID="$(${CLAUDE_PLUGIN_ROOT}/scripts/smart/smart-id.sh generate)"

# Initialize ledger
${CLAUDE_PLUGIN_ROOT}/scripts/smart/ledger.sh init

# Create deletion permit for scope (auto-revoked at cleanup)
${CLAUDE_PLUGIN_ROOT}/scripts/smart/permit.sh create "$SMART_ID" "$0" --ttl=3600

Store $SMART_ID — pass it to every teammate prompt.

STEP 0b — Determine Scope + Smart Target:

# Staged + unstaged changes
git diff --cached --name-only
git diff --name-only

# If nothing changed, check last commit
git diff HEAD~1 --name-only

# If $0 is a path, scope to that path

Produce a file list and store it:

FILE_LIST=$(git diff --cached --name-only; git diff --name-only)
[ -z "$FILE_LIST" ] && FILE_LIST=$(git diff HEAD~1 --name-only)

This goes into EVERY teammate's prompt.

Smart Target (auto-equip goggles + guillotine): run the detection script over $FILE_LIST — equip goggles when scope contains frontend files (.tsx/.jsx/.css/.html/.svelte/.vue) or --goggles is passed; equip the guillotine (and break-manifest.sh init) when scope crosses a public-API surface (PublicAPI.Shipped.txt, a packable .csproj, or a packages/ path) or --guillotine is passed. Full detection bash in references/team-architecture.md § "Smart Target auto-detection". Hades knows when he needs his glasses — and his guillotine.

STEP 1 — Create Team:

TeamCreate: team_name = "hades-cleanup", description = "Hades cleanup: [scope]"

You are the team lead. You orchestrate — you NEVER implement. Teammates do all code work.

STEP 2 — Create Phase 0 Tasks: one TaskCreate per audit domain into the shared hades-cleanup task list — Audit {suppressions,dead code,duplication,imports} in [scope]. If GOGGLES equipped, add goggles tasks (taste, spec, compliance). If GUILLOTINE equipped, add a guillotine audit task (public-API surface classification).

STEP 3 — Spawn Phase 0 Teammates (ALL in ONE message): Task tool, team_name="hades-cleanup", subagent_type="general-purpose", model="opus", one per domain — smart-audit-{suppressions,deadcode,duplication,imports}. Prompts from templates/auditors.md.

If GOGGLES: +3 goggles teammates from templates/goggles.md (all Opus 4.8, same team_name). If GUILLOTINE: +1 teammate smart-guillotine-audit from templates/guillotine.md (Opus 4.8, same team_name).

Teammates debate findings via SendMessage and share the task list via TaskCreate/TaskUpdate. Messages are delivered automatically — do not poll.

STEP 4 — Evaluate GATE 0: when debate converges (teammates idle, no new messages), review findings via TaskList and apply the GATE 0 verdict (references/gates-and-reports.md).

STEP 5 — Phase Transition (Phase 0 → Phase 1): shut down each Phase 0 teammate (SendMessage type="shutdown_request", plus goggles/guillotine if equipped); wait for all shutdown_responses. Spawn 4 Phase 1 eliminators from templates/eliminators.md (same Task pattern). Goggles findings become elimination tasks. If GUILLOTINE: also spawn smart-guillotine-elim (templates/guillotine.md) — it deletes/downgrades symbols and emits one break-manifest.jsonl entry per BREAK task.

STEP 6 — Evaluate GATE 1: when all elimination tasks complete (TaskList), apply the GATE 1 verdict.

STEP 7 — Phase Transition (Phase 1 → Phase 2): shut down Phase 1 teammates, wait for shutdown_responses, spawn 4 Phase 2 verifiers from templates/verifiers.md. smart-verify-grep also confirms goggles violations were resolved. If GUILLOTINE: also spawn smart-guillotine-verify (templates/guillotine.md) — two-axis verification (shim-free + functionally equivalent) against .smart/break-manifest.jsonl, then break-manifest.sh validate.

STEP 8 — Evaluate GATE 2: COMPLETE → cleanup. ITERATE → shut down verifiers, respawn eliminators on remaining items.

STEP 9 — Cleanup (after COMPLETE):

# Revoke deletion permit
${CLAUDE_PLUGIN_ROOT}/scripts/smart/permit.sh revoke

# Show ledger summary
${CLAUDE_PLUGIN_ROOT}/scripts/smart/ledger.sh count

Shut down all remaining teammates (SendMessage type="shutdown_request"). Wait for all shutdown_responses, then delete the team:

TeamDelete: team_name = "hades-cleanup"

YOUR NEXT ACTION: Run Step -1 check, then Step 0 (Smart Init), then Step 1 (TeamCreate) and spawn Phase 0.

</CRITICAL_EXECUTION_REQUIREMENT>

---

GATES, CLEANUP & FINAL REPORT

Verdict logic (full status boxes, ITERATE triggers, cleanup sequence, connector hooks, and the HADES CLEANUP REPORT in references/gates-and-reports.md):

• GATE 0: scan-only → SCAN_COMPLETE; zero findings → HALT; findings → PROCEED.
• GATE 1: build/tests fail or missing break-manifest entry → HALT; all complete + green → PROCEED.
• GATE 2: any remaining suppression/warning, incomplete ledger, unresolved challenge, or guillotine axis fail → ITERATE; all clean → COMPLETE.

After COMPLETE: revoke permit, (if guillotine) validate break-manifest, shut down teammates, TeamDelete, present the report.