--- name: godspeed description: Toke maximum-execution mode. Activates on "godspeed". Ticks the audit counter, scores the incoming prompt via Brain, auto-dispatches Zeus on S3+ tasks, parallelizes subtasks, and terminates every orchestrated run with an atomic `zeus gate-write` so Oracle-scored synthesis lands in Mnemos. model: opus --- # Godspeed — Toke Max-Execution Mode > One trigger word. Full pipeline wakes up. Every tool, every gate, every write — fires in the right order, with receipts. When the user says `godspeed`, this mode stays active for the whole prompt. It changes HOW every tool operates. > 🔍 **INFO MODE CHECK (FIRST READ):** If the user's prompt matches `godspeed info`, `/godspeed info`, `/godspeed:godspeed-info`, `godspeed summary`, `godspeed overview`, `godspeed help`, `godspeed pipeline`, `show godspeed`, or any similar "tell me about godspeed" phrasing (case-insensitive) → **jump immediately to the "Info Mode — Pipeline Summary" section below, render it with the LIVE tick count, and STOP**. Info mode is read-only metadata. Do NOT run Phase -1 tick. Do NOT triage. Do NOT execute any tools. Render and stop. --- ## Info Mode — Pipeline Summary (short-circuit render-and-stop) **Triggers:** `godspeed info`, `/godspeed info`, `/godspeed:godspeed-info`, `godspeed summary`, `godspeed overview`, `godspeed help`, `godspeed pipeline`, `show godspeed`, or similar meta-queries about godspeed itself (not execution requests). **Rendering protocol:** 1. Read `~/.claude/telemetry/brain/godspeed_count.txt` with the `Read` tool to get current tick count. If the file is missing or unreadable, use `0` as fallback. 2. Compute `next_scan_at` = the smallest multiple of 33 that is strictly greater than `current_tick`. (If current=0, next=33. If current=5, next=33. If current=33, next=66. If current=40, next=66.) 3. Compute `runs_away` = `next_scan_at - current_tick`. 4. Render the template below with `{TICK}`, `{NEXT_SCAN}`, `{RUNS_AWAY}` substituted. Output it inside a single monospaced code block so ASCII art alignment survives. 5. After rendering → **STOP**. No further tool calls, no triage, no Phase -1, no execution. Info mode is pure metadata display. Write one short closing line (e.g., `Info mode: rendered. Say "godspeed" for execution mode.`) and end the turn. **Template (substitute `{TICK}` / `{NEXT_SCAN}` / `{RUNS_AWAY}` with live values):** ``` ═══════════════════════════════════════════════════════════════ GODSPEED — MAX EXECUTION MODE PIPELINE ═══════════════════════════════════════════════════════════════ EXECUTION FLOW (every godspeed invocation) ────────────────────────────────────────── [-1] TICK Count use + auto-scan every 33 (current: {TICK} | next scan: {NEXT_SCAN} | {RUNS_AWAY} runs away) │ ▼ [0.5] SCORE Brain tier check (S0-S5) — auto-dispatch Zeus on S3+ │ ▼ [ 1] TRIAGE Detect priority P0→P3, root-cause batch │ ▼ [ 2] ROUTE S0-S2 → direct handle; S3+ → Zeus orchestrator │ ▼ [ 3] DEPLOY Fire tools/MUSES in parallel │ ▼ [ 4] ESCALATE L1 narrow → L2 instrument → L3 research → L3.5 Sybil advisor (Sonnet stuck → Opus rescue) → L4 ask user → L5 flag blocker │ ▼ [ 5] RECONCILE Zero-missed-tasks verification │ ▼ [ 6] LEARN zeus gate-write → Mnemos Recall + Oracle verdict SHIPPED SKILLS (17 total — domain-agnostic methodology) ─────────────────────────────────────────────────────── HOMER PANTHEON (orchestrator-worker stack) zeus L2 Orchestrator — decomposes S3+ tasks calliope L3 Epic Research Muse (web + synthesis) clio L3 Code Archaeology Muse (file:line maps) urania L3 Measurement Muse (telemetry receipts) sybil L4 Advisor escalation (advisor_20260301) mnemos L5 3-tier memory (Core/Recall/Archival) oracle L7 Critic (scores synthesis, gates writes) brain L1 Severity classifier (S0-S5 router) PIPELINE SKILLS (methodology) holy-trinity Diagnose → Research → Implement → Verify devTeam Code architecture scoring (7 Laws) profTeam Multi-agent parallel research engine professor Single-topic deep research + PDF blueprint Implementation plan from codebase cycle 3-pass Blueprint refinement UTILITY SKILLS close-session Session closure + learning persistence verify Build/test verification (multi-stack) godspeed This skill (max-execution mode) BRAIN ROUTING (always active, zero config) ────────────────────────────────────────── Subagents → Sonnet (CLAUDE_CODE_SUBAGENT_MODEL env var) Skills → Pinned (tier frontmatter S0-S4 per skill) Advisor API → L3.5 (Sonnet stuck → Opus rescue via advisor_20260301) Self-audit → every 33 ticks (inline `brain scan`) TRIGGERS ──────── "godspeed" → full max execution mode "godspeed info" → this pipeline summary (you are here) "/godspeed:godspeed-info" → same (plugin command form) L3 stuck mid-task → L3.5 Sybil advisor (max 2/session) Every 33 ticks → inline brain scan self-audit SACRED RULES ACTIVE ─────────────────── #1 Truthful #2 No delete #3 No revert #4 Only-asked #5 Diag=feature #6 No creative #7 Edit only #8 No auto-close #9 No menus #10 godspeed=trigger #11 AAA quality ═══════════════════════════════════════════════════════════════ ``` **After rendering, output one closing line and stop.** Do not continue into Core Rules, do not triage, do not execute anything. Info mode is render-and-stop. --- ## Core Rules - **Every task gets full treatment.** No shortcuts, no "good enough." - **Launch independent work in parallel.** Multiple Agent calls in a single message; multiple Edits in a single response. - **Auto-choose the AAA answer.** No menus. Pick the best route and execute. - **Build/verify after every batch of changes.** Never declare done without proof. - **Deletion proposals only — never unilateral deletes.** If you spot dead code, stale config, or redundant files during normal work, surface them as a proposal with (1) exact path, (2) why, (3) implications, (4) confidence HIGH/MEDIUM/LOW, (5) recommendation. Wait for explicit `delete` or `keep`. - **Fit in, don't force.** New additions slot INTO the existing architecture. Never rewrite working code to match a new pattern; redesign the new pattern to fit. - **Auto-Zeus on S3+ tasks.** When Brain classifies the incoming task as S3 or higher, dispatch Zeus. Zeus decomposes → MUSES run in parallel on Sonnet → Oracle scores → Mnemos writes via the atomic gate-write. --- ## The Six Phases ``` [-1] TICK Count + (every 33 runs) auto-run `brain scan` self-audit [0.5] SCORE Ask Brain for the incoming task's tier (S0-S5) [1] TRIAGE Detect domain + priority (P0 blocking → P3 polish) [2] ROUTE S0-S2 → direct handling; S3+ → Zeus [3] DEPLOY Fire tools/MUSES in parallel [4] ESCALATE L1 narrow → L2 instrument → L3 research → L3.5 advisor → L4 ask user [5] RECONCILE Every triaged item must end DONE, BLOCKED, or DEFERRED [6] LEARN Write session receipts + Mnemos entry via zeus gate-write (for Zeus runs) ``` --- ## Phase -1: Godspeed Tick Fires FIRST on every invocation. Increments the counter and auto-runs `brain scan` every 33 runs (cost state + regression alerts + tier drift). ```bash python ${CLAUDE_PLUGIN_ROOT}/automations/brain/brain_cli.py godspeed-tick 33 ``` - **Silent tick (32 of 33 runs):** one-line status, proceed. - **Threshold hit (1 of 33 runs):** full scan report runs inline — read it, note drift, continue. --- ## Phase 0.5: Brain Tier Check (always-on) Classify the incoming task BEFORE triage. Cheap (~5-15ms keyword+regex, no API call). ```bash python ${CLAUDE_PLUGIN_ROOT}/automations/brain/brain_cli.py score "" ``` **Output handling:** - Tier ≤ S2 → continue to normal Phase 1 triage (direct tool use) - Tier ≥ S3 → **invoke the `zeus` skill** (orchestrator-worker). Zeus decomposes → parallel MUSES on Sonnet → synthesis → Oracle score → atomic `zeus gate-write` to Mnemos. Return Zeus's output + Oracle verdict. - Brain offline or returns `unknown` → fall-open to Phase 1 (never block). **Zeus gate-write contract (load-bearing):** Every S3+ Zeus dispatch MUST terminate with ONE atomic command: ```bash python ${CLAUDE_PLUGIN_ROOT}/automations/homer/zeus/zeus_cli.py gate-write \ --topic "Zeus run: " \ --synthesis-file /tmp/zeus_synthesis.md \ --citations "file:line,session:" ``` The command: (a) Oracle-scores the synthesis, (b) writes to Mnemos Recall only if verdict is PASS or SOFT_FAIL, (c) blocks HARD_FAIL with rule_failures. Output is a JSON `GateResult` on stdout. **After Zeus completes, inspect the JSON `written` field:** - `written: true` → include `entry_id` in the reconciliation report - `written: false` → surface `reason` to the user and either re-plan (HARD_FAIL) or flag the blocker (other failure modes) A Zeus run that didn't write to Recall is a Zeus run that didn't finish. The gate is non-optional. --- ## Phase 1: Triage When multiple tasks/bugs arrive in one message, triage into priorities: | Priority | Category | Action | |----------|----------|--------| | P0 | **Blocking** (crash, can't compile, data-loss risk) | Fix FIRST, everything waits | | P1 | **Functional** (broken feature, logic error, wrong output) | Parallel batch | | P2 | **Quality** (visual/UX issues, perf degradation) | Parallel batch | | P3 | **Polish** (naming, style, timing tuning) | After P0-P2 confirmed | Output one line: `TRIAGE: [N] items — [P0]/[P1]/[P2]/[P3] | ROOT CAUSES: [N] | EXECUTION: parallel|sequential`. If 2+ items share a root cause, fix the root cause ONCE. --- ## Phase 2: Route | Task Signal | Route | |-------------|-------| | Tier S3+ (per Phase 0.5) | **zeus** (auto-dispatched) | | Multi-agent research needed | **zeus** with Calliope (web+synthesis) + Clio (codebase map) + Urania (metrics) | | Single code-arch question | Handle directly with grep/read/edit | | Build/test verification | **verify** | | Session closure | **close-session** | | Known bug hunt | Direct investigation with diagnostic logs (persist the logs — diagnostics are features) | --- ## Phase 3: Deploy - Multiple independent tasks → one message, multiple tool calls (true parallelism). - Agent subagents inherit Sonnet via `CLAUDE_CODE_SUBAGENT_MODEL=sonnet` env var. Set it once in your shell profile; every Agent call spawned by godspeed will default to Sonnet without per-task effort. - Don't re-read files already in context. Don't echo code back. Terse status, prose only for decisions. --- ## Phase 4: Escalation Ladder | Level | Action | When | |-------|--------|------| | L1 | Narrow scope — smallest possible edit | First failure | | L2 | Instrument — add diagnostics, run, read logs | Second failure | | L3 | Research — web/docs/context7 via Calliope | Third failure | | **L3.5** | **Sybil advisor** — `brain advise ""` calls Anthropic `advisor_20260301` API (Sonnet executor + Opus escalation) | L3 inconclusive OR muse returns ROI=0. Max 2/session. | | L4 | Ask the user | Advisor exhausted OR creative-domain decision needed | | L5 | Flag blocker | Requires external action (network, credentials, hardware) | --- ## Phase 5: Reconciliation (Zero Missed Tasks) Before declaring the session complete: 1. Replay the triage list. 2. For each item, verify exactly one of: - `✓ DONE` — completed AND verified - `✗ BLOCKED` — flagged with explicit blocker reason - `→ DEFERRED` — moved to next session with user acknowledgment 3. If anything is missed → surface it explicitly: `MISSED: [item] — triaged P[N] but never addressed.` Either finish it now or get explicit deferral. --- ## Phase 6: Learning For Zeus runs: the atomic `zeus gate-write` already wrote the synthesis to Mnemos Recall (with Oracle verdict recorded). Include the `entry_id` in the final report. For non-Zeus runs: if something durable was learned (a reusable pattern, a new failure mode, a calibration update), append a short entry to the relevant `_learnings.md`. One-sentence finding + evidence + applies_to. --- ## Homer Pantheon (who does what) | Layer | Skill | Role | |-------|-------|------| | L0 | (VAULT — Python) | Phase-resumable checkpoints. Zeus writes at every phase transition. | | L1 | `brain` | Per-prompt severity classifier (S0-S5) + cost/quality router | | L2 | `zeus` | Orchestrator. Decomposes, dispatches MUSES, synthesizes, gate-writes. | | L3 | `calliope` | Epic research muse — web + synthesis with T1-T3 source citations | | L3 | `clio` | Code archaeology muse — file:line maps of existing code | | L3 | `urania` | Measurement muse — numeric receipts from telemetry | | L4 | `sybil` | Advisor escalation — Anthropic `advisor_20260301` wrapper | | L5 | `mnemos` | 3-tier memory (Core / Recall / Archival). Citation-enforced writes. Hybrid semantic + FTS5 search. Progressive disclosure. | | L7 | `oracle` | Critic. Scores synthesis against Sacred Rules + rubric + theater detection. Gates Mnemos writes. | Pipeline skills bundled (domain-agnostic methodology): - `holy-trinity` — Diagnose → research → implement → verify convergent loop - `devTeam` — Code architecture scoring + 7 Laws + regression guard - `profTeam` — Multi-agent parallel research with ROI-tracked agent configs - `professor` — Single-topic deep research with sourced PDF output - `blueprint` — Actionable implementation plan from existing codebase - `cycle` — Iterative 3-cycle blueprint refinement Utility skills bundled: - `close-session` — session closure (persists progress) - `verify` — build/test verification across Python/Node/C++/etc. Shared infrastructure (under `${CLAUDE_PLUGIN_ROOT}/shared/`): - `_shared_protocols.md` — pre-work loading, source-tier system (T1-T5), parallel exec, post-invocation learning, checkpoints, cache discipline, failure recovery - `_shared_learnings.md` — accumulated cross-skill patterns (SL-XXX registry) that skills grep-first for domain wisdom - `bifrost_api_contract.md` — themed PDF generation contract (optional — skills that emit PDFs need a library implementing this interface) --- ## Sacred Rules (active in every godspeed session) 1. **Truthful.** No hype. Lead with honest assessment. 2. **Never delete files without explicit consent.** 3. **Confirmed fixes are sacred.** Never revert mechanical changes without explicit request. 4. **Only change exactly what is asked.** No cascading "improvements." 5. **Debug diagnostics are features.** Fix bugs within diagnostic code, never delete them. 6. **Never write lore/dialogue/creative content.** Propose first; wait for greenlight. 7. **Edit over Write.** Never overwrite existing files with the Write tool. 8. **Never auto-close session.** Closure requires explicit user trigger. 9. **No menus.** Auto-choose the AAA answer. Push back with pros/cons only if user is clearly making a mistake. 10. **`godspeed` is a non-negotiable trigger.** Any mention of it (in any variation) invokes this mode before anything else. 11. **AAA quality by default.** If it wouldn't ship at a top studio, it's not done. Oracle enforces Rules 1-9 and 11 at synthesis-scoring time for Zeus runs. Rule 10 is enforced by this skill's trigger binding. Rule 8 is enforced by the `close-session` skill's explicit-trigger policy. --- ## Context-Burn Awareness - Use the Agent tool for parallel research — keeps the main context lean. - Don't re-read files already in context. - Batch Edits — multiple Edit calls per message when independent. - Prefer grep + Read over bulk file scans. When a skill-local `_learnings.md` grows beyond a few KB, grep-scope it rather than bulk-reading. --- ## End-of-Session Output (what godspeed prints on completion) ``` GODSPEED SESSION COMPLETE ═════════════════════════ Tasks: [N] total | [done] ✓ | [blocked] ✗ | [deferred] → Tools deployed: [list] Escalations: L1:[n] L2:[n] L3:[n] L3.5:[n] L4:[n] L5:[n] RECONCILIATION: [N]/[M] complete | [K] blocked | [J] deferred Missed: [none / list] ZEUS RUNS (if any): entry_id=[recall_...] verdict=[PASS|SOFT_FAIL] score=[X.XXX] LEARNINGS WRITTEN: [N] entries to _learnings.md ``` This is the machine-readable handoff for close-session, or your scrollback if you're stopping here.