feat(screen): 화면을 만들고 재고 고친 뒤에만 보여주는 screen skill · screen-designer (v0.7.8)

.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
     {
       "name": "goax",
       "description": "AI 에이전트 결과 일관성을 환경으로 통제하는 4계층 하네스 (Triage·Constitution·Module·Spec/ADR) + Spirit·Mistake Loop. bash+markdown only, 자연어로 도입. Claude Code·OpenCode 지원.",
-      "version": "0.7.7",
+      "version": "0.7.8",
       "author": {
         "name": "bluecheat",
         "email": "itsinil@gmail.com"
@@ -30,5 +30,5 @@
       ]
     }
   ],
-  "version": "0.7.7"
+  "version": "0.7.8"
 }

.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "goax",
-  "version": "0.7.7",
+  "version": "0.7.8",
   "description": "AX 4-Layer harness — Triage / Constitution / Module / Spec·ADR + Spirit·Mistake Loop. Bash + Markdown only, with Claude-driven onboarding. Multi-CLI: Claude Code (native) + OpenCode (Hybrid compat via AGENTS.md SSOT + opencode.json).",
   "author": {
     "name": "bluecheat",

CLAUDE.md

@@ -20,13 +20,13 @@ There is no `npm`/`make`/`gradle`. Anything more complex is invoked through Clau
 ## Repository layout (what each top-level dir does)
 
 - `.claude-plugin/{plugin.json,marketplace.json}` — Claude Code plugin manifest. Both files plus `VERSION` must hold the **same version string** (smoke test fails otherwise).
-- `skills/<name>/SKILL.md` — 15 skills (`up`, `onboarding`, `zero`, `doctor`, `spec`, `spec-{tasks,implement,validate}`, `lane`, `audit`, `mistake`, `hud`, `triage`, `adr`, `vendor`). Each starts with frontmatter `name:` + `description:` where description is the trigger-keyword bag. Note: `install` was renamed to `up` (0.1.18), `spec-plan` was retired (0.1.16), `mistake` was added as the capture-only counterpart to `audit`. **`zero` and `lane` were added in 0.5.0** (`lane` was born as `parallel` and renamed in 0.5.1 — the unit of the skill is the lane, and the agents `lane-*` and scripts `lanes-*` already carried that name) — `zero` is the 0→1 entry point (`up` §6 hands greenfield to it; `onboarding` stays brownfield-only), and `lane` sits between `spec-tasks` and `spec-implement` to decide whether the work can be split at all. **`rules` and `spirit` were retired in 0.5.3**: both were prose telling the model to grep; they became `rules-index.sh` and `spirit-lint.sh`, and `doctor` owns their triggers ("rules 보여줘", "spirit 점검"). Big skills keep `SKILL.md` under 500 lines and push box mockups / incident write-ups / mode-specific loops into `<skill>/references/*.md` (`onboarding`, `lane`, `spec-implement`, `triage`, `zero` do this) — `tests/smoke.sh` enforces the line cap, the 1,536-char description cap, and that every referenced `references/<x>.md` exists.
+- `skills/<name>/SKILL.md` — 16 skills (`up`, `onboarding`, `zero`, `doctor`, `spec`, `spec-{tasks,implement,validate}`, `lane`, `screen`, `audit`, `mistake`, `hud`, `triage`, `adr`, `vendor`). Each starts with frontmatter `name:` + `description:` where description is the trigger-keyword bag. Note: `install` was renamed to `up` (0.1.18), `spec-plan` was retired (0.1.16), `mistake` was added as the capture-only counterpart to `audit`. **`zero` and `lane` were added in 0.5.0** (`lane` was born as `parallel` and renamed in 0.5.1 — the unit of the skill is the lane, and the agents `lane-*` and scripts `lanes-*` already carried that name) — `zero` is the 0→1 entry point (`up` §6 hands greenfield to it; `onboarding` stays brownfield-only), and `lane` sits between `spec-tasks` and `spec-implement` to decide whether the work can be split at all. **`rules` and `spirit` were retired in 0.5.3**: both were prose telling the model to grep; they became `rules-index.sh` and `spirit-lint.sh`, and `doctor` owns their triggers ("rules 보여줘", "spirit 점검"). Big skills keep `SKILL.md` under 500 lines and push box mockups / incident write-ups / mode-specific loops into `<skill>/references/*.md` (`onboarding`, `lane`, `spec-implement`, `triage`, `zero` do this) — `tests/smoke.sh` enforces the line cap, the 1,536-char description cap, and that every referenced `references/<x>.md` exists.
 - `commands/goax.md` — single slash-command index that delegates to skills (other thin per-skill wrappers were retired).
-- `agents/{architect,evaluator,lane-scout,lane-worker}.md` — sub-agents invoked by skills. `evaluator` is the Generator/Evaluator separation: `spec-implement` §8 spawns it in a fresh context and it writes `<spec>/review.md` whose first line is `verdict: 진행|보강 필요|재논의 필요`; `tasks-gate.sh` G6 reads that line. `lane-scout`/`lane-worker` are the read-only investigation lane and the file-ownership execution lane that `spec-implement` runs in lane mode.
+- `agents/{architect,evaluator,lane-scout,lane-worker,screen-designer}.md` — sub-agents invoked by skills. `screen-designer` is the screen executor: the `screen` skill does intake (`design-caps.sh`), mode and spec-first tables, then spawns it in a fresh context to build → render → measure (`screen-measure.sh`) → fix (≤3) and write a report whose first line is `verdict: 통과|보강 필요|실측 아님`. The pair is **self-contained** — all design knowledge lives in `skills/screen/references/`; MCPs, browsers and simulators are optional accelerators, never requirements. `evaluator` is the Generator/Evaluator separation: `spec-implement` §8 spawns it in a fresh context and it writes `<spec>/review.md` whose first line is `verdict: 진행|보강 필요|재논의 필요`; `tasks-gate.sh` G6 reads that line. `lane-scout`/`lane-worker` are the read-only investigation lane and the file-ownership execution lane that `spec-implement` runs in lane mode.
 - `scripts/provision.sh` — the installer. `up` calls it with `--json`; it reads MANIFEST and copies `templates/default/` into the user project. Lives at the plugin root (not `templates/default/.ax/scripts/`) because it *creates* `.ax/` — it cannot live inside what it installs. Root `install.sh` is an unrelated legacy stub that only prints install instructions.
 - `templates/default/` — **everything that gets copied into a user project** when the `up` skill runs.
 - `templates/default/MANIFEST` — installer SSOT. `up` reads this line-by-line; lines ending in `/` are recursive directory copies, `src -> dst` lines do rename copies. **When you add anything under `templates/default/`, update MANIFEST or the installer will not copy it.** Conditional copies (CLAUDE.md, settings.json, config.yml, mistakes/README.md, .gitignore) are intentionally outside MANIFEST and live in `skills/up/SKILL.md` §5–§6.
-- `templates/default/.ax/scripts/bash/` — deterministic shell tooling (42 scripts). All follow the standard in `templates/default/.ax/scripts/bash/README.md`: `set -euo pipefail`, `--json --dry-run --help` options, `[goax]` stderr prefix, exit codes `0` ok / `1` error / `2` skipped. Their `--json` shape is `{status, result, next_step, warnings, errors}`. SKILL.md files invoke these and parse the JSON; they do not reimplement the logic inline.
+- `templates/default/.ax/scripts/bash/` — deterministic shell tooling (45 scripts). All follow the standard in `templates/default/.ax/scripts/bash/README.md`: `set -euo pipefail`, `--json --dry-run --help` options, `[goax]` stderr prefix, exit codes `0` ok / `1` error / `2` skipped. Their `--json` shape is `{status, result, next_step, warnings, errors}`. SKILL.md files invoke these and parse the JSON; they do not reimplement the logic inline.
 - `templates/default/.ax/hooks/{session-start,pre-compact,user-prompt,pre-bash,pre-edit,post-edit,pre-commit,subagent-start,stop}/*.sh` — sensors. The companion `.claude/settings.json.template` registers them; `doctor` cross-checks installed hooks against this template (it is the SSOT for which hooks *and which event keys* should be registered). `subagent-start/harness-pointer.sh` hands Constitution/Spirit/active-spec/handoff **pointer** to every non-goax subagent (the harness used to stop at the main session); `stop/spec-gate.sh` blocks the end of a turn once when an `implementing`/`review` spec fails `tasks-gate.sh`, unless `current-task.json.handoff.now` records it within 24h (`now_at`) (bounded by `stop_hook_active` and an 8-per-session counter). The two pre-edit injection hooks dedupe per session via `goax_inject_fresh` (`.ax/.session/<sid>/`, 4h TTL) — measured 1,200 repeat injections / 529KB in one real session before this. Four hooks are **fact-forcing gates** that block even in `warning` mode because passing is cheap: `pre-edit/rule-read-gate.sh` denies an edit until every path-matched spirit/module rule file has been Read this session (Read marker, not the model's word; `@`-imported rules are exempt), `pre-bash/destructive-facts.sh` denies an in-project destructive command once and passes the identical retry after the model lists targets/rollback/user instruction, `pre-bash/block-hook-bypass.sh` denies `--no-verify`/`core.hooksPath`/`HUSKY=0`, and `pre-edit/quality-config-gate.sh` denies the first edit per session of an *existing* lint/format/type-check/coverage/git-hook config (cross-ecosystem default file names plus `sensors.quality_configs` — files that mix other settings, like `package.json`/`pyproject.toml`, are deliberately not defaults) until the model writes loosen-or-tighten and the user's instruction. Shell commands are judged by `goax_shell_scan` (shlex tokens; heredoc bodies, here-strings and word-initial `#` comments handled like the shell — shlex's own `commenters` is off because it swallowed the rest of a joined multi-line command) — never by a one-line regex, which reads `-m "fix -n"` as a flag. Python availability is `goax_py_ok` (actually imports; `command -v python3` passes for a broken macOS xcrun shim) and a failed scan returns 3 so the hook degrades to a coarse check instead of passing. Regenerable directories are a cross-ecosystem default plus the project's `sensors.regenerable_paths` — do not grow the built-in list per stack. Every registered hook goes through `goax_hook_enabled <id> <minimal|standard>` (`sensors.disabled_hooks`, `sensors.hook_profile`, `GOAX_DISABLED_HOOKS`); `smoke §53` fails a hook that doesn't, and block-destructive's CATASTROPHIC tier must stay above that switch. `session-start/session-brief.sh` only relays `session-brief.sh --json` (handoff · version lag · overdue audit · same-category mistake recurrence); `pre-compact/snapshot.sh` writes `session-brief.sh --snapshot` facts (branch/HEAD, uncommitted files, active spec's tasks progress) to `.ax/.session/<sid>/precompact.txt` and the `source=compact` SessionStart prepends them once. `post-edit/lint-changed.sh` runs only `commands.lint_file` (`"<glob> => <cmd with {file}>"`, first match) and hands failures to the model as `additionalContext` — it never guesses a linter from the extension; `detect-stack.sh` proposes values from what the project declares (package.json scripts, Makefile/justfile/Taskfile targets, wrappers, manifests) and `config-set.sh` writes them under the `config.yml` lock. In a monorepo the session usually opens at the repo root while `.ax/` lives in `projects/<app>/`, so hooks registered only in the project's settings never fire; `ade-settings.sh --apply` generates the root `.claude/settings.json` entries from the template as `env CLAUDE_PROJECT_DIR="${CLAUDE_PROJECT_DIR}/<rel>" bash "…/<rel>/.ax/hooks/…"` (replacing only that project's goax entries), and doctor §3.14 checks it — never hand-write that file (a real monorepo's hand-written copy was missing four hooks). The same script in a single repo (ADE root = project root) is how `.ax/settings.json.suggested` gets merged: never `jq -s '.[0] * .[1]'`, whose `*` replaces the whole `PreToolUse` array and drops the user's own hooks. It writes no backup file — git is the undo. Model-facing hook text goes through `goax_cap_context [max_bytes] [where]` (default `GOAX_CONTEXT_MAX`=8000 bytes): Claude Code moves any `additionalContext`/stdout over 10,000 chars to a file and shows only a 2,000-char preview without asking the model to read it, so the hook cuts first — on a byte budget (bytes ≥ chars in any locale), never mid-UTF-8, with a `… N바이트 생략 — <where>` tail; `smoke §61` feeds each injecting hook a worst-case input and fails a hook that emits `additionalContext` without the helper. Hooks stay synchronous: `async`/`asyncRewake` deliver on the next turn (or wake only on exit 2 via stderr), which would let the model keep editing on top of a lint failure it hasn't seen.
 - `templates/default/.ax/spirit/{values.md,tone.md,README.md}` — cross-cut Spirit (shared agent personality), the trio the plugin ships. `spirit/rules/` is user-curated (no plugin-shipped instances) — categories get added via onboarding/audit, seeded from `_templates/spirit/ops.md` as an opt-in baseline. Spirit rule headers must match `## SP-<CAT>-<NNN>: text` — `doctor` lints this.
 - `templates/default/.ax/_templates/{spec,adr,module,spirit,mistakes}/` — SDD templates the user copies into their own work. `_templates/spec/.origin` is a sha snapshot the installer writes; `check-templates-drift.sh` diffs current vs `.origin` to detect user edits vs plugin updates.

README.md

@@ -123,6 +123,8 @@ The `spec` skill doesn't generate every artifact upfront. The triage result deci
                                   tasks carry lane assignments (lanes-dispatch.sh ledger)
                                   → completion gate G1–G6 (+ fresh-context evaluator for L / M×L3)
 "split into lanes"              → lane — first asks whether the work can be split at all
+"build this screen"             → screen — spec tables first → screen-designer (fresh context) builds,
+                                  renders, measures (screen-measure.sh) and fixes before showing
 # Design decisions go to ADR (.ax/docs/adr/<id>-*.md)
 ```
 
@@ -147,6 +149,7 @@ Natural-language tier overrides: `"simple"` / `"spec + tasks"` → standard · `
 | "split into lanes" / "run in parallel" | lane — decides first whether the work can be split at all |
 | "start implementation" / "run tasks" | spec-implement |
 | "validate spec" | spec-validate — NEEDS CLARIFICATION gating |
+| "build this screen" / "review this mockup" / `/screen` | screen — spec tables first (tokens · components · copy · exceptions), then the `screen-designer` agent builds, renders, measures in px and fixes (≤3) before showing; report starts with `verdict:`. Self-contained: all design knowledge ships in `skills/screen/references/`; MCPs · browsers · simulators are optional |
 | "plan the <task>" / "fix" / "refactor" | triage — Size × Risk classify in 30s |
 | "write ADR" / "record decision" | adr — new ADR file |
 | "spirit check" | doctor → `spirit-lint.sh` — header format · token duplicates · placeholders |

VERSION

@@ -1 +1 @@
-0.7.7
+0.7.8