// PAGE — INSTALL

Install

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Quick Install (one-line)

The fastest path. Picks Cursor or Claude Code (or both) and installs to a global location (~/.cursor/skills/si-chip/ and/or ~/.claude/skills/si-chip/) or to a single repo (<repo>/.cursor/skills/si-chip/ etc.).

# Interactive (TTY): prompts for target and scope
curl -fsSL https://yorha-agents.github.io/Si-Chip/install.sh | bash

# Non-interactive: install Cursor globally
curl -fsSL https://yorha-agents.github.io/Si-Chip/install.sh | bash -s -- \
  --target cursor --scope global --yes

# Non-interactive: install Claude Code into a specific repo
curl -fsSL https://yorha-agents.github.io/Si-Chip/install.sh | bash -s -- \
  --target claude --scope repo --repo-root ~/code/myrepo --yes

# Install for both Cursor and Claude Code, globally
curl -fsSL https://yorha-agents.github.io/Si-Chip/install.sh | bash -s -- \
  --target both --scope global --yes

Installer flags

Flag	Values	Default	Required
`--target`	`cursor` / `claude` / `both`	(interactive prompt)	when `--yes`
`--scope`	`global` / `repo`	(interactive prompt)	when `--yes`
`--repo-root`	path	`$PWD`	when `--scope repo --yes`
`--version`	tag	`v0.4.0`	no
`--source-url`	URL	`https://yorha-agents.github.io/Si-Chip`	no (mostly for testing)
`--yes` / `-y`	flag	`false`	no
`--dry-run`	flag	`false`	no
`--force`	flag	`false`	no
`--uninstall`	flag	`false`	no
`--help`	flag	`false`	no

What gets installed (21 files via tarball, ~115 KB)

The HTTPS installer downloads docs/skills/si-chip-0.4.0.tar.gz (SHA-256 2cfcce00f989faf2467014e638b0ea1fa67870b5a1ee6b0531942be5a4be21ab; 83060 bytes; deterministic and reproducible) and extracts 21 files (1 SKILL.md + 1 DESIGN.md + 14 references + 5 scripts):

<install-dir>/
  SKILL.md                                              (metadata 94 / body 4646 tokens)
  DESIGN.md                                             (internal architecture notes)
  references/basic-ability-profile.md                   (§2)
  references/self-dogfood-protocol.md                   (§8)
  references/metrics-r6-summary.md                      (§3 — 7 dim / 37 sub-metrics)
  references/router-test-r8-summary.md                  (§5 — 8-cell MVP / 96-cell Full)
  references/half-retirement-r9-summary.md              (§6 — 8-axis value vector)
  references/core-goal-invariant-r11-summary.md         (§14 — C0 invariant; v0.3.0)
  references/round-kind-r11-summary.md                  (§15 — round_kind enum; v0.3.0)
  references/multi-ability-layout-r11-summary.md        (§16 — Informative; v0.3.0)
  references/token-tier-invariant-r12-summary.md        (§18 — C7/C8/C9; v0.4.0)
  references/real-data-verification-r12-summary.md      (§19 — fixture provenance; v0.4.0)
  references/lifecycle-state-machine-r12-summary.md     (§20 — promotion history; v0.4.0)
  references/health-smoke-check-r12-summary.md          (§21 — 4-axis probes; v0.4.0)
  references/eval-pack-curation-r12-summary.md          (§22 — 40-prompt v2 minimum; v0.4.0)
  references/method-tagged-metrics-r12-summary.md       (§23 — _method companions; v0.4.0)
  scripts/profile_static.py                             (§8 step 1)
  scripts/count_tokens.py                               (packaging gate)
  scripts/aggregate_eval.py                             (§8 step 2)
  scripts/eval_skill_quickstart.md                      (CLI cheat-sheet; v0.3.0)
  scripts/real_llm_runner_quickstart.md                 (CLI cheat-sheet; v0.4.0)

DESIGN.md carries internal architecture notes and is included in the tarball / file:// install but is not mirrored into .cursor/skills/si-chip/ or .claude/skills/si-chip/ (those mirror the 20-file SKILL.md + references + scripts set per the cross-tree drift contract — see CONTRIBUTING.md §9).

Where <install-dir> is one of:

target	scope	install dir
cursor	global	`~/.cursor/skills/si-chip/`
cursor	repo	`<repo-root>/.cursor/skills/si-chip/`
claude	global	`~/.claude/skills/si-chip/`
claude	repo	`<repo-root>/.claude/skills/si-chip/`

Verify the install

# Replace <install-dir> with the path the installer printed.
python3 <install-dir>/scripts/count_tokens.py --file <install-dir>/SKILL.md --both
# Expected: metadata_tokens=94, body_tokens=4646, pass=true
#           (against the v0.4.0 v2_tightened budget: meta <= 100, body <= 5000)

Uninstall

curl -fsSL https://yorha-agents.github.io/Si-Chip/install.sh | bash -s -- \
  --target cursor --scope global --uninstall --yes

Manual install (clone the repo)

If you prefer to inspect everything first, or if you want the full source tree (templates, evals, dogfood evidence, spec, …), clone the repo. This path covers Cursor and Claude Code (the two priorities per spec §7.2), the Codex bridge (still bridge-only at v0.4.0 per spec §11.2), developer setup, and smoke tests.

Prerequisites

Python >= 3.10
git
Optional: tiktoken (for accurate token counting; otherwise count_tokens.py falls back to a deterministic whitespace splitter and reports backend=fallback).
Optional: devolaflow (R7 §1 upstream — pip install git+https://github.com/YoRHa-Agents/DevolaFlow.git).
Optional: nines CLI (legacy live-LLM runner; the included evals/si-chip/runners/real_llm_runner.py is the v0.4.0 production runner and does NOT depend on nines).
Optional: requests (only required if you actually call evals/si-chip/runners/real_llm_runner.py against a live Anthropic Messages endpoint; cache-replay mode does not need it).

1. Clone the Repository

git clone https://github.com/YoRHa-Agents/Si-Chip.git
cd Si-Chip

2. Cursor Install (priority 1)

The Skill is mirrored at .cursor/skills/si-chip/. Cursor auto-discovers it on workspace open. The optional bridge rule .cursor/rules/si-chip-bridge.mdc is included and points back at .cursor/skills/si-chip/SKILL.md plus AGENTS.md (which is itself compiled from .rules/si-chip-spec.mdc; AGENTS.md §13 carries 13 hard rules at v0.4.0).

Reload Cursor; the Skill should appear under the project’s local skills.

Verify:

python .agents/skills/si-chip/scripts/count_tokens.py \
  --file .cursor/skills/si-chip/SKILL.md --both

Expect metadata_tokens=94, body_tokens=4646, pass=true (matches the spec §7.3 v2_tightened packaging gate; identical to the canonical mirror per the cross-tree drift contract — see CONTRIBUTING.md §9).

3. Claude Code Install (priority 2)

The Skill is mirrored at .claude/skills/si-chip/. Claude Code auto-discovers it on session start.

Verify:

python .agents/skills/si-chip/scripts/count_tokens.py \
  --file .claude/skills/si-chip/SKILL.md --both

Same gate numbers as the Cursor mirror (drift = 0).

4. Developer Setup

pip install pyyaml                                              # required for scripts
pip install tiktoken                                            # optional; matches CI
pip install requests                                            # optional; only for live real_llm_runner runs
pip install git+https://github.com/YoRHa-Agents/DevolaFlow.git  # optional

pyyaml is the only hard dependency for the bundled scripts. tiktoken matches CI’s token counting backend; devolaflow is required only when you want to drive Si-Chip through the upstream template_engine / memory_router paths (spec §5.1, §9). requests is only needed for live Anthropic Messages calls from evals/si-chip/runners/real_llm_runner.py; the --seal-cache / cache-replay flow does not require it.

5. Smoke Tests

# 14 BLOCKER spec invariants — verdict PASS
python tools/spec_validator.py --json

# Generate self-profile
python .agents/skills/si-chip/scripts/profile_static.py \
  --ability si-chip --out /tmp/profile.yaml

# Deterministic seeded baseline runners (no LLM cost)
python evals/si-chip/runners/no_ability_runner.py \
  --cases-dir evals/si-chip/cases/ --out-dir /tmp/no_ability/ --seed 42

python evals/si-chip/runners/with_ability_runner.py \
  --cases-dir evals/si-chip/cases/ --out-dir /tmp/with_ability/ --seed 42

# Aggregate to MVP-8 + 29 explicit-null R6 keys
python .agents/skills/si-chip/scripts/aggregate_eval.py \
  --runs-dir /tmp/with_ability --baseline-dir /tmp/no_ability \
  --skill-md .agents/skills/si-chip/SKILL.md \
  --templates-dir templates --out /tmp/metrics_report.yaml

Expected: spec_validator exits 0 with verdict: PASS (14/14 BLOCKER invariants — the original 9 + REACTIVATION_DETECTOR_EXISTS + 2 v0.3.0 additive invariants CORE_GOAL_FIELD_PRESENT + ROUND_KIND_TEMPLATE_VALID

3 v0.4.0 additive invariants TOKEN_TIER_DECLARED_WHEN_REPORTED + REAL_DATA_FIXTURE_PROVENANCE + HEALTH_SMOKE_DECLARED_WHEN_LIVE_BACKEND); profile_static emits a BasicAbilityProfile YAML against the §2.1 schema ($schema_version: 0.3.0); the two runners populate per-case result.json files; aggregate_eval produces a metrics_report.yaml with the MVP-8 keys filled and the remaining 29 keys explicitly null (matches evals/si-chip/SMOKE_REPORT.md).

Optional — real-LLM runner cache replay (v0.4.0)

The Round 18 / Round 19 cache lives at .local/dogfood/2026-04-30/round_18/raw/real_llm_runner_cache/ (640 entries). To replay it without paying for live calls:

python evals/si-chip/runners/real_llm_runner.py --help
# See .agents/skills/si-chip/scripts/real_llm_runner_quickstart.md for the
# full Round 18 / Round 19 invocation; cache replay is $0 and ~20 ms.

6. Troubleshooting

count_tokens.py reports backend=fallback: install tiktoken for parity with CI; the fallback uses a deterministic whitespace splitter and may report different token counts.
aggregate_eval.py warns about a schema cross-check: expected. The templates are JSON-Schema-shaped (properties.basic_ability.properties.metrics.properties), not a direct basic_ability.metrics map. MVP-8 keys are still validated independently. The smoke report documents this as a non-blocking warning.
spec_validator.py --strict-prose-count exits 1 against spec_v0.1.0.md but PASS against v0.2.0+: expected. The legacy v0.1.0 prose contained “28 sub-metrics” / “21 threshold cells” while §3.1 / §4.1 TABLE counts were 37 / 30. v0.2.0+ §13.4 prose was reconciled to 37 / 30 and the validator now passes strict mode against any v0.2.0 / v0.3.0 / v0.4.0 spec; the v0.1.0 mode is preserved for historical regression.
The packaging gate fails with metadata_tokens=94 > 80: expected. v0.4.0 ships at v2_tightened (meta <= 100); v3_strict (meta <= 80) is deferred to v0.4.x. See README “Headline Numbers” and the v0.4.0 ship report under .local/dogfood/2026-04-30/v0.4.0_ship_report.md.

7. Uninstall

Cursor: delete .cursor/skills/si-chip/ and reload the workspace.
Claude Code: delete .claude/skills/si-chip/ and restart the session.
Repo: rm -rf Si-Chip/.

Codex (bridge-only at v0.4.0)

Si-Chip ships AGENTS.md, which is compiled from .rules/si-chip-spec.mdc. Codex reads AGENTS.md, so the Normative spec content (§3 / §4 / §5 / §6 / §7 / §8 / §11 / §14 / §15 / §17 / §18 / §19 / §20 / §21 / §22 / §23 plus the 13 hard rules in §13) is in front of Codex on every session.

Native .codex/profiles/si-chip.md plus .codex/instructions/si-chip-bridge.md remain deferred per spec §11.2 (“Codex native SKILL.md runtime support; v0.x is bridge-only”). This is re-affirmed in spec §14.6 + §18.7 + §19.6 + §20.6 + §21.6 + §22.7 + §23.7 across the v0.3.0 + v0.4.0 add-on chapters; native Codex SKILL.md runtime will be re-evaluated in a future spec bump once v3_strict is earned.

一键安装

最快的安装方式。选择 Cursor 或 Claude Code（或两者），安装到全局位置（~/.cursor/skills/si-chip/ 和/或 ~/.claude/skills/si-chip/）或指定的单个仓库（<repo>/.cursor/skills/si-chip/ 等）。

# Interactive (TTY): prompts for target and scope
curl -fsSL https://yorha-agents.github.io/Si-Chip/install.sh | bash

# Non-interactive: install Cursor globally
curl -fsSL https://yorha-agents.github.io/Si-Chip/install.sh | bash -s -- \
  --target cursor --scope global --yes

# Non-interactive: install Claude Code into a specific repo
curl -fsSL https://yorha-agents.github.io/Si-Chip/install.sh | bash -s -- \
  --target claude --scope repo --repo-root ~/code/myrepo --yes

# Install for both Cursor and Claude Code, globally
curl -fsSL https://yorha-agents.github.io/Si-Chip/install.sh | bash -s -- \
  --target both --scope global --yes

安装器选项

选项	值	默认值	必填
`--target`	`cursor` / `claude` / `both`	（交互式提示）	`--yes` 时必填
`--scope`	`global` / `repo`	（交互式提示）	`--yes` 时必填
`--repo-root`	路径	`$PWD`	`--scope repo --yes` 时必填
`--version`	版本标签	`v0.4.0`	否
`--source-url`	URL	`https://yorha-agents.github.io/Si-Chip`	否（主要用于测试）
`--yes` / `-y`	开关	`false`	否
`--dry-run`	开关	`false`	否
`--force`	开关	`false`	否
`--uninstall`	开关	`false`	否
`--help`	开关	`false`	否

安装内容（通过 tarball 安装 21 个文件，约 115 KB）

HTTPS 安装器会下载 docs/skills/si-chip-0.4.0.tar.gz（SHA-256 2cfcce00f989faf2467014e638b0ea1fa67870b5a1ee6b0531942be5a4be21ab；83060 字节；确定性可复现），并解压出 21 个文件（1 个 SKILL.md + 1 个 DESIGN.md + 14 个 references + 5 个 scripts）：

<install-dir>/
  SKILL.md                                              (metadata 94 / body 4646 tokens)
  DESIGN.md                                             (内部架构说明)
  references/basic-ability-profile.md                   (§2)
  references/self-dogfood-protocol.md                   (§8)
  references/metrics-r6-summary.md                      (§3 — 7 维 / 37 子指标)
  references/router-test-r8-summary.md                  (§5 — 8-cell MVP / 96-cell Full)
  references/half-retirement-r9-summary.md              (§6 — 8 维 value vector)
  references/core-goal-invariant-r11-summary.md         (§14 — C0 不变量；v0.3.0)
  references/round-kind-r11-summary.md                  (§15 — round_kind 枚举；v0.3.0)
  references/multi-ability-layout-r11-summary.md        (§16 — Informative；v0.3.0)
  references/token-tier-invariant-r12-summary.md        (§18 — C7/C8/C9；v0.4.0)
  references/real-data-verification-r12-summary.md      (§19 — fixture 溯源；v0.4.0)
  references/lifecycle-state-machine-r12-summary.md     (§20 — 升档历史；v0.4.0)
  references/health-smoke-check-r12-summary.md          (§21 — 4 维探针；v0.4.0)
  references/eval-pack-curation-r12-summary.md          (§22 — v2 最少 40 prompt；v0.4.0)
  references/method-tagged-metrics-r12-summary.md       (§23 — _method 伴随字段；v0.4.0)
  scripts/profile_static.py                             (§8 步骤 1)
  scripts/count_tokens.py                               (打包闸门)
  scripts/aggregate_eval.py                             (§8 步骤 2)
  scripts/eval_skill_quickstart.md                      (CLI 速查；v0.3.0)
  scripts/real_llm_runner_quickstart.md                 (CLI 速查；v0.4.0)

DESIGN.md 是内部架构说明，包含在 tarball / file:// 安装中，但不会被镜像到 .cursor/skills/si-chip/ 或 .claude/skills/si-chip/（这两个镜像只包含 20 个文件的公开 SKILL.md + references + scripts 集合，遵循跨树漂移契约——详见 CONTRIBUTING.md §9）。

<install-dir> 取值如下：

target	scope	安装目录
cursor	global	`~/.cursor/skills/si-chip/`
cursor	repo	`<repo-root>/.cursor/skills/si-chip/`
claude	global	`~/.claude/skills/si-chip/`
claude	repo	`<repo-root>/.claude/skills/si-chip/`

验证安装

# Replace <install-dir> with the path the installer printed.
python3 <install-dir>/scripts/count_tokens.py --file <install-dir>/SKILL.md --both
# Expected: metadata_tokens=94, body_tokens=4646, pass=true
#           （对应 v0.4.0 v2_tightened 预算：meta <= 100, body <= 5000）

卸载

curl -fsSL https://yorha-agents.github.io/Si-Chip/install.sh | bash -s -- \
  --target cursor --scope global --uninstall --yes

手动安装（克隆仓库）

如果你想先审视所有内容，或者需要完整的源码树（templates、evals、dogfood 证据、spec 等），请克隆仓库。这条路径覆盖 Cursor 与 Claude Code（按规范 §7.2，这是目前的两个优先平台）、依然处于 bridge-only 状态的 Codex（v0.4.0 仍按规范 §11.2 延后原生支持）、开发环境配置以及冒烟测试。

前置依赖

Python >= 3.10
git
可选：tiktoken（用于精确的 token 计数；否则 count_tokens.py 会回退到确定性的空白切分器，并报告 backend=fallback）。
可选：devolaflow（R7 §1 上游 — pip install git+https://github.com/YoRHa-Agents/DevolaFlow.git）。
可选：nines CLI（旧版 live-LLM runner；v0.4.0 内置的 evals/si-chip/runners/real_llm_runner.py 是新版生产 runner，不依赖 nines）。
可选：requests（仅当你确实要让 evals/si-chip/runners/real_llm_runner.py 调用 Anthropic Messages 端点时才需要；cache-replay 模式不需要它）。

1. 克隆仓库

git clone https://github.com/YoRHa-Agents/Si-Chip.git
cd Si-Chip

2. Cursor 安装（优先级 1）

Skill 镜像位于 .cursor/skills/si-chip/。Cursor 在打开工作区时会自动发现它。可选的 bridge 规则 .cursor/rules/si-chip-bridge.mdc 也已包含在内，它指回 .cursor/skills/si-chip/SKILL.md 与 AGENTS.md（后者由 .rules/si-chip-spec.mdc 编译生成；v0.4.0 时 AGENTS.md §13 共有 13 条 hard rules）。

重新加载 Cursor；该 Skill 应当出现在该项目的本地 skills 列表下。

验证：

python .agents/skills/si-chip/scripts/count_tokens.py \
  --file .cursor/skills/si-chip/SKILL.md --both

预期输出 metadata_tokens=94、body_tokens=4646、pass=true（满足规范 §7.3 的 v2_tightened packaging gate；按跨树漂移契约，与标准镜像完全一致——详见 CONTRIBUTING.md §9）。

3. Claude Code 安装（优先级 2）

Skill 镜像位于 .claude/skills/si-chip/。Claude Code 在 session 启动时会自动发现它。

验证：

python .agents/skills/si-chip/scripts/count_tokens.py \
  --file .claude/skills/si-chip/SKILL.md --both

各项门控数值与 Cursor 镜像相同（drift = 0）。

4. 开发环境配置

pip install pyyaml                                              # 内置脚本的硬依赖
pip install tiktoken                                            # 可选；与 CI 一致
pip install requests                                            # 可选；仅 live real_llm_runner 调用需要
pip install git+https://github.com/YoRHa-Agents/DevolaFlow.git  # 可选

pyyaml 是内置脚本的唯一硬依赖。tiktoken 与 CI 使用的 token 计数后端一致；devolaflow 仅在你希望通过上游 template_engine / memory_router 路径（规范 §5.1、§9）驱动 Si-Chip 时才需要。requests 仅在 evals/si-chip/runners/real_llm_runner.py 真正调用 Anthropic Messages 时需要；--seal-cache / cache-replay 流程无需安装。

5. 冒烟测试

# 14 个 BLOCKER 规范不变量 — verdict PASS
python tools/spec_validator.py --json

# 生成自身 profile
python .agents/skills/si-chip/scripts/profile_static.py \
  --ability si-chip --out /tmp/profile.yaml

# 确定性种子化的 baseline runner（不消耗 LLM）
python evals/si-chip/runners/no_ability_runner.py \
  --cases-dir evals/si-chip/cases/ --out-dir /tmp/no_ability/ --seed 42

python evals/si-chip/runners/with_ability_runner.py \
  --cases-dir evals/si-chip/cases/ --out-dir /tmp/with_ability/ --seed 42

# 聚合得到 MVP-8 + 29 个显式 null 的 R6 key
python .agents/skills/si-chip/scripts/aggregate_eval.py \
  --runs-dir /tmp/with_ability --baseline-dir /tmp/no_ability \
  --skill-md .agents/skills/si-chip/SKILL.md \
  --templates-dir templates --out /tmp/metrics_report.yaml

预期：spec_validator 以 verdict: PASS 退出码 0 退出（14/14 个 BLOCKER 不变量——最早的 9 个 + REACTIVATION_DETECTOR_EXISTS + v0.3.0 新增的 2 个 CORE_GOAL_FIELD_PRESENT + ROUND_KIND_TEMPLATE_VALID + v0.4.0 新增的 3 个 TOKEN_TIER_DECLARED_WHEN_REPORTED + REAL_DATA_FIXTURE_PROVENANCE + HEALTH_SMOKE_DECLARED_WHEN_LIVE_BACKEND）；profile_static 输出符合 §2.1 schema 的 BasicAbilityProfile YAML（$schema_version: 0.3.0）；两个 runner 为各 case 生成 result.json；aggregate_eval 生成 metrics_report.yaml，其中 MVP-8 keys 已填入数值，剩余 29 个 key 显式置 null（与 evals/si-chip/SMOKE_REPORT.md 一致）。

可选 — real-LLM runner cache replay（v0.4.0）

Round 18 / Round 19 的缓存位于 .local/dogfood/2026-04-30/round_18/raw/real_llm_runner_cache/（640 个条目）。在不消耗 live 调用的情况下回放：

python evals/si-chip/runners/real_llm_runner.py --help
# 完整的 Round 18 / Round 19 调用方法详见
# .agents/skills/si-chip/scripts/real_llm_runner_quickstart.md；
# cache replay 花费 $0、约 20 ms。

6. 常见问题

count_tokens.py 报告 backend=fallback：安装 tiktoken 以获得与 CI 一致的结果；fallback 使用确定性的空白切分器，可能产生不同的 token 计数。
aggregate_eval.py 提示 schema 交叉校验告警：属于预期行为。模板采用 JSON-Schema 形态（properties.basic_ability.properties.metrics.properties），并非直接的 basic_ability.metrics 映射；MVP-8 keys 仍会单独校验。冒烟报告将其记为非阻塞告警。
spec_validator.py --strict-prose-count 在 spec_v0.1.0.md 上以退出码 1 退出，但在 v0.2.0+ 上 PASS：属于预期行为。旧版 v0.1.0 的散文写的是 “28 sub-metrics” / “21 threshold cells”，而 §3.1 / §4.1 表格其实是 37 / 30。v0.2.0+ §13.4 散文已对齐到 37 / 30，校验器在 strict 模式下对 v0.2.0 / v0.3.0 / v0.4.0 任一 spec 都会 PASS；保留 v0.1.0 模式仅用于历史回归。
packaging gate 报错 metadata_tokens=94 > 80：属于预期行为。v0.4.0 在 v2_tightened（meta <= 100）档位发版；v3_strict（meta <= 80）已延后到 v0.4.x。详见 README 的 “Headline Numbers” 与 .local/dogfood/2026-04-30/v0.4.0_ship_report.md。

7. 卸载

Cursor：删除 .cursor/skills/si-chip/ 并重新加载工作区。
Claude Code：删除 .claude/skills/si-chip/ 并重启 session。
仓库：rm -rf Si-Chip/。

Codex（v0.4.0 仍为 bridge-only）

Si-Chip 同时分发 AGENTS.md，它由 .rules/si-chip-spec.mdc 编译生成。Codex 会读取 AGENTS.md，因此每次 session 都会看到 Normative 规范内容（§3 / §4 / §5 / §6 / §7 / §8 / §11 / §14 / §15 / §17 / §18 / §19 / §20 / §21 / §22 / §23，以及 §13 中的 13 条 hard rules）。

原生的 .codex/profiles/si-chip.md 与 .codex/instructions/si-chip-bridge.md 仍按规范 §11.2 延后（”Codex native SKILL.md runtime support；v0.x 仅 bridge”）。这一点在 v0.3.0 + v0.4.0 的 §14.6 + §18.7 + §19.6 + §20.6 + §21.6 + §22.7 + §23.7 中被反复重申；只有当 v3_strict 达成后，才会在新一轮 spec bump 中重新评估 Codex 原生 SKILL.md runtime。

The canonical install guide also lives at the repository root.

标准安装指南也存放于仓库根目录的 INSTALL.md 中。