Language: English | 한국어

Skill Olympus

Twelve Greek gods. One command. A working SaaS.

Speak the name of cloud-gathering Zeus, and the entire pantheon descends — Zephermine drafts the spec, Poseidon raises the fleet, Argos counts every plank, Minos judges every test, and Clio carves the whole story into bronze.

A production agent harness for Claude Code, Codex CLI, Gemini CLI, and Grok Build — a harness-engineering and loop-engineering stack, named after the Twelve Olympians, forged across 3 months of daily real-product builds.

/zeus "Build a shopping mall. React + Spring Boot + PostgreSQL"

One line. Twelve gods. Design → Implement → Inspect → Test → Ship. No questions asked. No blueprints needed. No human in the loop.

Built for the loop-engineering era. Agents don't run on a single prompt — they run on loops: act → observe → verify, repeated until an objectively verifiable completion criterion is met. Olympus was built that way from the start — Chronos (verification gate that runs real tests), Minos (fix-until-pass), Argos (AC cross-check), Clio (GO/NO-GO). Since v4.7.0 these loops run on top of CLI-native features (/goal stop gate, review engines, built-in subagents) — the harness is the foundation, the loop is the operating model. And since v4.8.0 the loops know how to stay running: a loop survives on structure, not willpower — heartbeat in the machine, state in the audit log, blocked issues parked with decision-ready briefs, and false completion claims rejected at the hook level.

What you actually get

99 public skill sources (default allowlist union: 17 = 11 entrypoint harnesses + 6 runtime adapters; 13 or 14 active per installed surface, 82 source-only internal/optional modules) · 42 agent source references (40 top-level + 2 skill-owned; 0 custom agents registered by default) · 9 hooks · 4 CLIs · 1 mythology

Harness and Loop Engineering

Zeus is the harness layer: one non-interactive request is kept alive, decomposed, implemented, inspected, deployed, tested, and reported without handing "next steps" back to the user. In this repo that means skills/zeus/SKILL.md bootstraps Chronos persistence, drives Zephermine planning, routes implementation through Poseidon (agent-team-codex on Codex), then requires Argos, Docker, Minos, and a final evidence report before SUCCESS is allowed.

Chronos is the loop layer under that harness. It prefers native /goal, falls back to the Codex notify chain (save-turn -> continue-loop -> codex exec --skip-git-repo-check resume --last -), and only stops on Chronos Complete or an exact matching <promise>. Exhaustion is reported as incomplete, blockers are parked with Owner Decision Briefs, and each cycle returns to READ -> FIND -> FIX -> VERIFY -> LOG.

Quick Start

First install

# Clone
git clone https://github.com/Dannykkh/skill-olympus.git
cd skill-olympus

# Windows
.\install.bat

# macOS/Linux
chmod +x install.sh && ./install.sh

Running without arguments is the default full installation for Claude, Codex, Gemini, and Grok. --all is an optional explicit spelling of the same selection. The installer prepares each selected CLI's files even when that CLI executable is not on PATH; only CLI-dependent commands such as MCP registration are skipped. Install the missing CLI and rerun the same installer to finish those commands.

Updating an existing installation

A normal update does not require an uninstall first. Running the installer again reconciles only Olympus-managed names with the current policy and leaves unrelated third-party skill names in place.

git pull
.\install.bat

Use ./install.sh on macOS/Linux. Uninstall and reinstall only when the installation is broken or when you deliberately want to rebuild Olympus hooks and MCP registrations from scratch.

.\install.bat --uninstall
.\install.bat

What does source-only mean?

Source-only does not mean that an old skill version is being retained. It means that the current Olympus SKILL.md and supporting files remain available without being registered in the CLI's automatic discovery directory.

• Current sources are copied under each CLI's .olympus/source-skills/ and recorded in SKILLS-CATALOG.md.
• An active Olympus harness or the LLM can read that catalog source when a task needs it.
• Use --include-source-only-skills to put every compatible source-only skill back into slash menus and automatic matching.
• That option activates current Olympus sources; it does not restore an older modified copy from _olympus-preserved.

.\install.bat --include-source-only-skills

That's it. The 99 public sources split into a 17-skill default allowlist union (11 user-facing harnesses + 6 runtime adapters) and 82 source-only internal or optional modules. Runtime compatibility then removes adapters meant for other CLIs: Codex and Gemini each expose 95 compatible entries (13 active + 82 source-only), while Claude exposes 96 (14 + 82). Grok's standalone policy is also 95 (13 + 82), but the installed Grok surface reads the shared Claude directory and therefore sees the same 14 active entries as Claude. The internal-only deploymonitor source remains local and is not part of the public distribution. Active harnesses resolve required source-only modules through the catalog and read them directly; those modules do not need independent registration. No Olympus custom agent is registered by default; all 42 agent references remain source-only, and each CLI keeps its native subagents. New skill and agent sources are default-denied until deliberately allowlisted.

A missing CLI does not suppress asset preparation. Its skill catalog, source library, hooks, and configuration files are prepared; only commands that require the executable are reported as skipped.

When upgrading an existing installation, unrelated third-party skill names are preserved. Modified same-name collisions are moved to _olympus-preserved instead of being discarded. See the skill registry migration guide for the lightweight default, full opt-in, collision recovery, and uninstall/reinstall procedure. Recovery is manual: copy the preserved skill back only after giving both its directory and frontmatter name a unique name. Otherwise the next sync will preserve it again. --uninstall never restores backups.

The Pantheon of Olympus

Beyond the wine-dark sea, where clouds part above the world, Mount Olympus rises. On its windswept summit dwell the Twelve, each crowned with their own dominion, each known by an ancient and many-named song. Speak the name of one, and that one alone shall descend the holy mountain. Speak the name of cloud-gathering Zeus, and the whole pantheon shall come down with him in golden procession.

This is no mere tool-chest. It is a small mythology of work, a council of immortals each shaped to a single craft. They work as the old singers tell us they have always worked — gentle Zephermine, breath of the west wind, whispers a blueprint into the ear of earth-shaking Poseidon; Poseidon stirs the deep and his fleet sails forth in waves; hundred-eyed Argos walks the shore at dusk to count every nail and every beam; stern Minos sits upon his marble throne to weigh each soul at the gate; and last of all, fair-tressed Clio takes up her stylus to carve the whole tale upon a tablet of bronze, that mortals yet unborn may read of it.

Below stand the immortals. Call upon one — or call upon all.

The Twelve Who Sit Upon the Mountain

Songs from the Mountain

Hear now the voices of the Twelve, as the old singers heard them.

🜲 Zeus, Cloud-Gatherer Upon the topmost peak he sits, and his nod is the law of mountains. When his voice rolls out across Olympus, the council rises as one and descends — designer, builder, watchman, judge, and chronicler — all at his single word. "Speak my name once, mortal, and the whole council shall walk beside you to the end."

🜂 Zephermine, Bringer of the West Wind She is the soft breath that wakes the seed beneath the soil. Six and twenty are her questions, and the breath of each one is gentle — yet none may pass her by, for the spec is sacred and half-told tales bear no fruit. "I ask, and ask, and ask again — until what was unspoken becomes a thing of stone."

🜄 Poseidon, Earth-Shaker He stands knee-deep in the wine-dark sea, his trident raised, and the waters listen. The fleet of teammates lies in the harbor, and at his bidding the wave gathers them all and bears them out together, each prow pointed where the dependency graph commands. "The sea does not yield to the swimmer. The swimmer who knows the tide — she yields to him."

🜔 Daedalus, Master Builder Before him there was no labyrinth in all of Crete. He took stone from the mountain and shaped it with his own hands, and the work was good. Where the blueprint is missing, where the architect has not spoken, send for him — he will research, he will draft, and if no other hand will rise, his alone shall raise the walls. "Give me the stone. The plan I shall make as I go."

👁 Argos Panoptes, the Hundred-Eyed He paces the half-built city by night, and at no hour are all his eyes closed at once. The plank a mortal builder forgot to nail — he has already seen it. The line of code that fails to match the spec — he has already named it. "While fifty of my eyes rest, fifty more keep watch. Nothing passes Argos in the dark."

⚖ Minos, Judge of the Dead He sits upon a throne of cold marble at the gate where the souls of the dead must come. He raises his golden scales, and the work is weighed against itself. His verdict is two-fold and no other: it shall pass, or it shall return to the fire. "Stand before the scales, child of mortals. We shall see if your tests are honest."

📜 Clio, Muse of the Long Memory She comes last of all the gods, after the labor is laid down. Her stylus is bronze and her tablet is the years to come. What the heroes have done, she sets down — diagram, decree, manual, and song — that the children of the children of those mortals may know the deeds were real. "The work has ended. Now begins the telling, and the telling endures."

⏳ Chronos, the Unwearying He is older than memory, older than the gods themselves. He turns the great wheel of the hours and does not tire when mortals sleep. The bug shall die or the dawn shall come — and Chronos shall outlast them both. "Mortals close their eyes. I do not. The work shall be finished, by sunrise or by the next."

🪶 Hermes, Wing-Footed He moves between the worlds — the high palace and the low marketplace, both are his road. He reads the trade-winds of distant lands and the price of grain in city-gates yet unseen. Before a single coin is risked, before a single line of code is written, he speaks first. "Every market is a road, traveler. Every road has its toll. Bring silver, or bring nothing."

🦉 Athena, Gray-Eyed She was born full-grown from the skull of her father, helmeted, spear in hand. Her wisdom does not flatter; her counsel is the cold edge of the bronze. She will ask the question the mortal fears most — should this thing be made at all? "Wisdom, child, is to know which work must never be begun. I shall ask. You shall answer."

🌹 Aphrodite, Foam-Born She rose from the white foam of the sea, and the world has not been plain since. Nine palettes lie at her hand, seven and forty font pairings, four and eighty styles. What leaves her workshop is not merely useful — it is loved, and that is the difference. "Beauty is not the ornament of the work. Beauty is what makes the work survive its maker."

📚 Mnemo, Mother of the Muses Long before the nine sisters sang, Mnemosyne kept the long memory of the world. The conversation a mortal had three moons ago is the answer she carries to him today. Three layers she keeps — the index of names, the meaning of things, the tale itself — and her remembering crosses every session, every CLI, every dawn. "Forget nothing, child. The word you spoke long ago is the gift you needed now."

What's New

v5.1.0 — Aphrodite Evidence Gates · Browser Motion · Dual Video Engines (August 2026)

Aphrodite now treats Codex Product Design as an optional execution adapter whose exact marketplace selector must be verified before it can be recommended once; an empty marketplace or unverified selector remains UNKNOWN, so local design work continues without fabricated plugin output or installation pressure. Local and ready-adapter results use the same brief, Experience Contract, DESIGN.md, content, data, states, viewport, and theme. The contract now records Product Facts, Content Integrity, Asset Provenance, and a browser-only motion ladder that prefers CSS, View Transitions, and scroll-driven animations before justified GSAP Timeline, ScrollTrigger, Flip, SplitText, or SVG plugins. Separately, video-maker now routes each video project to exactly one engine: Remotion for React/TSX or HyperFrames for HTML/CSS/GSAP, with engine-specific runtime and license gates plus shared render QA. HyperFrames is not an Aphrodite runtime, and no video engine is globally installed or synchronized by the skill.

v5.0.0 — Entrypoint-Only Registry · Native Agents · Provider-Safe Runtime Routing (August 2026)

Olympus now keeps only user-facing harnesses and the current CLI's runtime adapters in automatic discovery. The 99 public skill sources split into an allowlist union of 17 (11 common entrypoints plus 6 runtime adapters) and 82 source-only modules: Claude and the Grok-shared Claude surface expose 14 active skills, while Codex and Gemini expose 13; all three homes retain the same 82-module source library through exact catalog paths. The internal-only deploymonitor source remains local and is excluded from public release counts. Source-only means current code is available for direct, on-demand reading without consuming slash-menu or startup-description budget; --include-source-only-skills remains the explicit full-registry opt-in. Olympus custom agents now default to zero: orchestration uses each CLI's semantic read-only and write-capable native roles, with Main owning shared state and sequential fallback when delegation is unavailable. Runtime-specific adapters are selected by provider so Claude/Grok, Codex, and Gemini do not load one another's incompatible mnemo or agent-team entrypoints. The installer now treats no arguments as all four CLIs, prepares assets even when a CLI executable is absent, skips only executable-dependent registration, preserves unrelated third-party names, and moves modified same-name collisions to _olympus-preserved for manual recovery.

v4.21.0 — Aphrodite Experience-Led Redesign · Stitch as Execution Adapter · Zeus Scope Gate (August 2026)

Aphrodite no longer starts from color tokens. The pipeline now runs source-mode routing → site-benchmark dissection (evidence-backed Adopt/Adapt/Avoid verdicts on header, message, section order, CTA, trust, and mobile transformations) → three actually-rendered directions → an Experience Contract capturing user tasks, messaging, CTAs, trust, and responsive behavior → implementation → rendered UX/accessibility/performance gates → a learning handoff, with a validation script enforcing the contract shape. Stitch is repositioned as an execution adapter: it compiles Aphrodite's decisions (DESIGN.md, design-refs) into Stitch MCP operations — work-contract, state-management, and file-transfer patterns selectively adopted from google-labs-code/stitch-skills — and never invents a visual direction of its own. Zeus gained a scope gate: a zero-interaction run has no user to approve scope creep, so any task, feature, or dependency not in the fixed plan must answer three questions (does it serve the one-line goal? what observed evidence demands it now? would something smaller do?) and record a pass/reduce/hold verdict in the Decision Ledger — held items surface as Deferred for post-hoc promotion (principle absorbed from hosioobo/track).

v4.20.2 — Themis: Official-Guideline Pinning · Runtime Freshness Check (August 2026)

(spans v4.20.1–v4.20.2) The Korean policy path is now anchored to the PIPC's official Privacy Policy Writing Guideline (Apr 2025): the guideline edition is pinned in the skill, and at generation time Themis fetches the official notice board to detect newer revisions — adopting them and telling the user when one exists (environments without web tools fall back to the pinned edition with an explicit "check latest" note). The ko template gained the 2025.4 specifics: a grievance-handling contact separate from the privacy officer, behavioral-advertising collection/refusal clauses, and a children's-data clause (Art. 22-2). Statute citations are verified against law.go.kr by direct fetch; the briefly-added optional law-MCP integration was removed as an unnecessary dependency — the "guidelines keep changing" problem is solved by runtime checks, not by owning scraper infrastructure.

v4.20.0 — Themis: Privacy-Policy Generator · mnemo Opt-Out (August 2026)

Themis, Titaness of law and order, joins as skill #100. She audits a codebase for every point where personal data is received, stored, sent out, or deleted (5 audit questions + grep hints + a trap checklist covering soft deletes, append-only logs, uninstall scope, masking coverage, silent outbound requests, and git ls-files-verified ignore rules), interviews the operator for the 14 things code cannot tell you (privacy officer, retention, children, server region, marketing use…), then generates per-country policy drafts (Korean PIPA Art. 30 / US CCPA-CPRA / EU GDPR Art. 13) from audit facts only — unverified items stay explicit fill-in blanks, legal judgments are flagged for counsel, and every draft carries a not-legal-advice notice. Dogfooded on this very repo: the three resulting drafts ship as docs/privacy-policy-draft-*. That audit also exposed our own gap — recording hooks with no off switch — so mnemo gained MNEMO_DISABLE (all 15 hook entry points across 4 CLIs exit immediately; existing data untouched) and the session-start version check gained OLYMPUS_UPDATE_CHECK_DISABLE, both verified by on/off comparison runs.

v4.19.0 — Native-First Realignment: 4-CLI Native Multi-Agent · impeccable Detection Hook (August 2026)

The Aphrodite interference experiment (real defects: native-alone 1 ≪ custom 20 ≈ mixed 25) proved the native-first principle — "engines upstream, extensions local" — and this release applies it across the skill roster. Design axis: the frontend-design fork now yields auto-invocation to the official plugin (with the v2.9.1 core grafted in), and prohibition rules moved from prompts to a linter — impeccable's 59 deterministic rules wired as a PostToolUse hook. Shipping that surfaced a real defect: PostToolUse exit-0 plain stdout never reaches the model, so the hook was rewritten to emit hookSpecificOutput.additionalContext JSON (without this fix it would have run silently forever). Multi-agent axis: with all four CLIs now shipping native multi-agent, Poseidon and Daedalus delegate execution to each CLI's native primitives (Claude Agent Teams / Codex spawn_agent / Gemini subagents / Grok spawn_subagent), repositioning the orchestrator MCP as a policy layer for hard file locks and external task boards plus a legacy fallback. Zeus gained an ultracode-session-only Workflow verification fan-out branch, Chronos got its Gemini AfterAgent wiring corrected (plus removal of stale deprecated /loop help text), and Argos now delegates its generic code-quality layer to the native review engines while keeping spec-vs-implementation audit as its own domain. Humanizer re-absorbed both upstreams (blader v2.9.1 + im-not-ai v2.3). The Gemini paths are designed from official docs but not yet field-tested — the verification checklist is on record.

v4.15.0 — Grok Build Support: grok-mnemo Adapter (July 2026)

Grok Build (xAI's CLI) joins the roster with near-zero integration cost: grok inspect --json confirmed that Grok reads ~/.claude/ directly via its [compat.claude] defaults — 99 skills, 41 agents, MCP servers, and the global CLAUDE.md rules all load without any sync script. The only real parity gap was mnemo auto-save: Grok's hook envelope is camelCase (no transcript_path), so Claude's Stop hooks silently no-op'd and Grok conversations were never recorded. grok-mnemo closes that gap — one script dispatching on two events (UserPromptSubmit strips Grok's <user_query> wrapper; Stop saves lastAssistantMessage directly, no transcript parsing, filtering out the session-end observe re-fire by reason == "end_turn"), writing to conversations/YYYY-MM-DD-grok.md. Since Grok also loads ~/.claude/settings.json hooks, the five Claude mnemo hook pairs gained a GROK_HOOK_EVENT guard to prevent double/mislabeled saves (Grok prompts were landing in -claude.md). Shipping this surfaced two environment gotchas now on record: Korean-commented .ps1 files must be UTF-8 with BOM (PowerShell 5.1 reads BOM-less files as CP949, and a misdecoded comment can swallow the newline and absorb the next code line — the <user_query> strip silently never ran), and the installer test had been failing on NoDefaultCurrentDirectoryInExePath=1 machines because it invoked install.bat by relative name. Verified end to end with a real grok -p headless session.

v4.14.0 — Layout Block Anatomy · Atmosphere Recipe · Prompt Consumption Gate (July 2026)

A full round-trip: build a real page with Aphrodite, notice it's still generic, fix the pipeline, rebuild, repeat — validated end-to-end on this project's own landing page (dannykkh.github.io/skill-olympus). Layout block anatomy: bag-ui's (MIT) block wireframe catalog translated into a structure grammar — 30 block anatomy contracts (Marketing/App/Ecommerce) with shared rules (three-level ink hierarchy, exactly-one emphasis per block, CTA syntax, sustained-asymmetry ratio reuse across sections) wired in as design-plan Phase 2.5, so structure is locked before style. Prompt consumption gate: DESIGN.md's prose contracts and docs/design-refs/ superprompts were being produced but never mandatorily read at implementation time — Phase 3 now requires reading them first, and direction cards must be saved to file instead of living only in chat. Atmosphere/texture recipe: SKILL.md told implementers to add grain/mesh/noise "atmosphere" with no verified CSS values to reach for, so it evaporated in practice — a new technique-recipes.md §11 supplies a real grain overlay, a restrained mesh gradient, an emphasis-break rule for repeated blocks, and a "boring test." The rebuild also surfaced two ground-truth bugs only visible by actually shipping: a font-loading false negative (document.fonts reports "unloaded" for subsetted Korean webfonts even after the network fetch succeeds — verified by inspecting the actual woff2 requests) and a mobile CSS-selector bug (.pipeline-grid > div deleted the whole timeline, not just the empty gutter) caught by screenshot, not static review.

v4.13.0 — Data Visualization Skill (July 2026)

Vendored Anthropic's official data-visualization skill (from anthropics/knowledge-work-plugins, Apache-2.0, body unmodified) — chart-selection guidance by data relationship (trend/comparison/ranking/distribution/correlation/flow), chart anti-patterns (pie charts over 6 slices, 3D, dual axes), Python code patterns (matplotlib/seaborn/plotly), design principles, and an accessibility checklist.

v4.12.0 — Style Recipes · Reference Capture · Chronos Heartbeat (July 2026)

Aphrodite absorbed the best of MengTo/Skills (MIT) after a full read-through. 12 style recipes: each aesthetic capsuled as "identity boundary + hex tokens + Korean/Latin font stacks + tuning knobs + avoid list" — the original's value-less prose was fixed by binding to our CSV DBs, and every color pair was verified by WCAG computation (3 sub-4.5:1 pairs fixed before shipping). 9 technique recipes: copy-paste-ready values for shadows, progressive blur, border gradients, text reveals, and the GSAP+Lenis motion system. Reference capture (Phase 2 rework): screenshots/URLs/videos/HTML become section-anatomy superprompts (docs/design-refs/) — prompts as versioned assets. Chronos heartbeat: skipping /goal setup used to leave the loop engineless and stall mid-run; the new 1.5-tier engine re-enters via native /loop intervals (--heartbeat), and score-based completion (e.g., "90+ points") is now formalized as a 3-part --completion-promise (threshold, measurement, printed evidence). The pipeline was validated end-to-end with an editorial-tech sample page (Korean fallback rendering, motion, responsive, contrast), surfacing and fixing gotcha 045 (document.fonts.check() false-negatives on unicode-range subset fonts).

v4.11.0 — Unknowns-First Planning · Lean Zephermine Flow · Implementation Learning Loop (July 2026)

Zephermine now plans around what is unknown instead of stopping for broad preference interviews. Unknowns-first discovery: Step 4 auto-selects codebase/web/GitHub/academic/competitor research from the spec and risk profile, writes research-decision.md, then Step 5A writes unknowns.md with known knowns, known unknowns, unknown knowns, unknown unknowns, and architecture-changing questions. Lean interview: Step 6 asks only critical blockers that could change architecture, data model, security boundary, UX flow, rollout, or compliance; otherwise it writes inferred assumptions and continues. No-stop domain flow: domain-dictionary now auto-seeds clear global terms, auto-merges low-risk ADD/REFINE/MERGE updates, and asks only on blocking dictionary conflicts. Implementation learning: workpm and agent-team prompts now keep implementation-notes.md deviations; Clio can generate a CHANGE-QUIZ.md; frontend-design can use divergent static prototypes to surface "you know it when you see it" preferences before implementation.

v4.10.0 — DESIGN.md as Canonical Source · Korean Font Real-Load Guardrails · Hook Timeout & Schema Fixes (July 2026)

Driven by running /aphrodite end-to-end on a real Korean-language UI. DESIGN.md canonical source: Google's @google/design.md format (YAML tokens + two-layer prose) is adopted as Aphrodite's design source of truth, with the agent wiring, single-source, and verification Google omits — a new design-md-guide.md (schema + lint/export + legacy migration), CSV palettes pinned into DESIGN.md to stop re-invocation drift, and a "read DESIGN.md first" rule injected into the 3-CLI always-on guardrails. Korean font real-load guardrail: picking a Latin-only pairing let Korean silently fall back to a system font (Space Grotesk/DM Sans have no Hangul glyphs) — fonts must now be actually loaded (@import/<link> + document.fonts.check), and Korean UIs prefer a Korean-only pairing (Hahmlet/Noto Serif KR + Pretendard/Noto Sans KR) that covers both scripts, structurally avoiding the fallback trap (gotcha 041); the Latin+Pretendard stack remains the fallback path. Dark/light theme rule: the three commonly-missed spots (container div backgrounds, text-color inversion, <select> option color-scheme) are now checked by the mnemo guardrail and ui-ux-auditor. Fixes: hook 30s-timeout prevention (60s timeout + powershell -NoProfile across all four hook-command generators, closing "hook timed out after 30s — output discarded"), and AskUserQuestion schema violations (≤4 options, ≤12-char headers) in the Aphrodite presets and zephermine templates.

v4.9.0 — Always-on Design Guardrail · Consult-Before-Implement · Handoff Feature Map (June 2026)

Three always-on capabilities, plus a compatibility audit. Design guardrail injection: frontend-design's anti-slop guidance only fired when the skill was explicitly invoked (the auto_apply field is read by no hook — a no-op), so casual design requests defaulted to "internet-average" output. The condensed guardrail is now injected into the 3-CLI always-on context (CLAUDE.md/AGENTS.md), updated to the 2026 web platform — native scroll-driven animation-timeline first (GSAP·Lenis only for pin/snap/WebGL), View Transitions API, container queries, :has(), OKLCH — plus a Korean/Latin font-pairing principle (Pretendard + noonnu, matched by weight/contrast DNA). Condensed ~24% and browser-verified on three distinct directions (dopamine SaaS / editorial / brutalist). Consult-before-implement guard: "implement X" requests now check existing work first (codemap → README/handoffs → grep), classify new/improve/duplicate, and compare against adjacent pipelines (zeus·zephermine·agent-team) — closing the "re-implementing what already exists" loop. Handoff feature map: every handoff records a Feature/Flow/Decision Snapshot + a Menu/Screen Map (screen-level features with done/partial/planned status), with diagrams required only for feature-bearing sessions. agent-team·chronos gain a static boundary-coherence cross-check ("build passes ≠ correct" — TS generics hide API↔hook contract mismatches) and chronos grounds its FIND in tool signals before model intuition. Compatibility audit: Pydantic v2, Next.js 15 async params, MySQL 8.4, docker compose v2, MUI v7, OpenAPI 3.1, Tailwind v4, and stale tool/model labels across the utility skills.

v4.8.7 — Loop Honesty: Completion Contract, Independent Verify, Exhaustion Surfacing (June 2026)

Loop Library 028/034 patterns propagated across the loop stack. A completion contract (each requirement mapped to reproducible evidence, scored proved/weak/missing/contradicted, and exhaustion never reported as success) now governs termination in Chronos, Zeus, minos, autoresearch, argos, agent-team, and workpm. Independent cross-model verification is added where a single actor would otherwise grade its own work — autoresearch re-scores the final champion with a different model family, and Zeus's argos gate runs cross-model only on risk triggers (deterministic build/test/static gate first, the other model in the background) so the implementation path stays fast. At the hook layer, loop-stop and continue-loop now surface EXHAUSTED on max-iterations/stale termination and inject a final-allowed-turn honest-report warning (dry-run verified, ps1 5/5; sh syntax-checked). Hermes gains a grounding guard: TAM/SAM/SOM figures require a cited source or an explicit [확인 필요].

v4.8.1 — mnemo Root-Resolution Fix (June 2026)

Auto-save hooks used to key off the latest cwd, so a cd into a subfolder (e.g. reference/1week) in a non-git project scattered conversations/·memory/ into that subfolder. Root resolution is now a 2-pass candidate evaluation — Pass 1 takes the first candidate with a git root (a git repo normalizes from any subfolder), Pass 2 falls back to the session-launch cwd for non-git projects, so a persisted cd can't move the anchor. Also guards an environment where HOME itself is a git repo (HOME excluded from candidates; a git root equal to HOME is treated as a dotfiles repo and skipped). Applied across all 8 hooks (save-response/save-conversation/save-tool-use/reconcile-conversations × ps1·sh) with installed copies synced; PS 7/7 · SH 7/7 scenario tests pass.

v4.8.0 — Loop Programming: Park, Brief, Re-entry (June 2026)

A loop is sustained by structure, not by the model's will to continue. Five failure modes, five structural cures: heartbeat in the machine, state outside the context window, blockers parked instead of blocking, idle work defined, and the human taken off the critical path with decision-ready briefs.

The loop state lives in the audit log, not the context window; each cycle re-enters through READ, verifies with objective checks, parks valid blockers with decision-ready briefs, and lets hooks reject false completion claims.

• Chronos PARK rule — one blocked issue can no longer stall the whole loop. Four valid park reasons (verification 3-fail / permission boundary / missing external access / product decision); declaring "blocked" without naming a reason counts as evasion; the worker must go as far as it can (reproduce, root-cause, in-permission fixes) before parking
• Escalation ladder before PARK — a verification failure is not parked to the owner until the model raises its own capability once (higher reasoning effort, a stronger model, or a focused review pass when neither is available). Same-approach retries alone cannot trigger a park; escalation runs at most once per issue and its outcome is recorded in the Owner Decision Brief evidence
• Owner Decision Brief — parked issues are reported decision-ready, never as raw questions: what / why now / evidence / trade-offs / recommendation (mandatory — never offload analysis to the owner) / exact choices. The owner's job becomes a 4-way choice: approve as recommended, reject, grant exactly one access, or pick a documented alternative
• Re-entry protocol (READ step) — every cycle re-reads docs/chronos/chronos-log.md before FIND; when the model's memory and the log disagree, the log wins. A new session resumes the loop from the audit log alone — loop state lives in files, not in the context window
• Deadlock guard — goal statements now include the park clause: when only parked issues remain, the loop ends with a Brief-carrying Chronos Complete report. Outputting an untrue <promise> is forbidden — and the hooks stop rewarding it: the "any <promise> tag = done" branch is removed from loop-stop/continue-loop (a mismatched promise re-injects instead of terminating), multiline promises now match, and re-injection nudges teach the park rule. loop-stop.ps1 functionally tested (4/4 cases: mismatch-rejected / exact-match / parked-only complete / no-marker re-inject); .sh mirrored + syntax-verified
• Zeus Decision Ledger — [ZEUS-AUTO:taste] decisions now record rationale + rejected alternatives + how to reverse; zero-interaction stays intact — approval moves post-hoc instead of pre-hoc, with reversible defaults preferred
• codex-mnemo installer fix — notify-wrapper check order corrected: a wrapper that chains save-turn is preserved (refreshed) before the IDE-notification removal heuristic runs. External tools sharing Codex's single notify slot are no longer silently disconnected by reinstall

v4.7.1 — Design Visual Verification + Clio Gate Hardening (June 2026)

• ui-ux-auditor visual verification — after the Grep static scan (fi