// PAGE — HOME

Home

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

Persistent BasicAbility optimization factory.

▶ INSTALL ▶ USER GUIDE ▶ ARCHITECTURE ▶ CHANGELOG ▶ LIVE DEMO

Status (v0.4.0): SHIP_ELIGIBLE at gate standard (= v2_tightened). FIRST Si-Chip release at v2_tightened (vs v0.2.0 / v0.3.0 at relaxed). 19 consecutive v1_baseline + 2 consecutive v2_tightened passes (Round 18 + Round 19). T2_pass_k unblocked via evals/si-chip/runners/real_llm_runner.py — Round 18 ran k=4 sampling against claude-haiku-4-5 + claude-sonnet-4-6 via Veil litellm-local at 127.0.0.1:8086 ($0.20 spend, 640 calls); Round 19 replayed at 100% cache hit ($0). v3_strict deferred to v0.4.x — single blocker is metadata_tokens = 94 vs ≤ 80 budget.

What is Si-Chip?

Si-Chip is a persistent BasicAbility optimization factory. Its first-class object is the BasicAbility, not any particular file shape. It improves abilities through metric-driven, multi-round dogfood loops:

profile -> evaluate -> diagnose -> improve -> router-test -> half-retire-review -> iterate -> package-register

Si-Chip ships as its own first user: the v0.4.0 release is the result of 19 complete dogfood rounds run against Si-Chip itself, building on the v0.1.0 ship (2 rounds), v0.2.0 ship (+11), v0.3.0 ship (+2), v0.4.0 ship (+4).

Headline numbers (v0.4.0 ship)

metric v0.1.0 v0.4.0 (Round 18-19) ceiling (v2_tightened)
pass_rate 0.85 1.0 best cell, 0.9719 mean ≥ 0.82
T2_pass_k 0.5478 (PROXY) 1.0 best cell ≥ 0.55
trigger_F1 0.89 1.0 ≥ 0.85
near_miss_FP_rate 0.05 0.0 ≤ 0.10
metadata_tokens 78 94 ≤ 100
per-invocation footprint 4071 4726 ≤ 7000
wall_clock_p95 (s) 1.47 17.5726 s ≤ 30 s
routing_latency_p95 n/a 0.16 ms ≤ 1200 ms
routing_token_overhead n/a 0.0233 ≤ 0.12
iteration_delta (task_quality) +0.31 (context_delta) +0.4522 (Round 18) ≥ +0.10

See the demo page for the full ship-report walkthrough and the changelog for the per-round v0.1.x → v0.4.0 story.

What’s new in v0.4.0

  • §18 Token-Tier Invariant — token_tier {C7_eager_per_session, C8_oncall_per_trigger, C9_lazy_avg_per_load} decomposition; OPTIONAL EAGER-weighted iteration_delta formula; lazy_manifest packaging gate.
  • §19 Real-Data Verification — 3-layer pattern (msw fixture provenance
    • user-install + post-recovery live verification); BLOCKER 13 REAL_DATA_FIXTURE_PROVENANCE.
  • §20 Stage Transitions & Promotion History — stage_transition_table forbids reverse transitions; promotion_history append-only; ship_decision.yaml is the 7th evidence file when round_kind == ship_prep.
  • §21 Health Smoke Check — 4-axis {read, write, auth, dependency} pre-ship probes; REQUIRED when live_backend: true; BLOCKER 14 HEALTH_SMOKE_DECLARED_WHEN_LIVE_BACKEND.
  • §22 Eval-Pack Curation — 40-prompt minimum @ v2_tightened; G1 _provenance REQUIRED; deterministic seeding hash(round_id + ability_id).
  • §23 Method-Tagged Metrics — <metric>_method companion fields, _ci_low / _ci_high 95% CI bands, U1_language_breakdown, U4_state.
  • §6.1 value_vector axes 7 → 8 — adds eager_token_delta (FIRST byte-identicality break since v0.1.0 → v0.2.0 prose-count alignment).
  • Real-LLM runner at evals/si-chip/runners/real_llm_runner.py (884 LoC + 27 tests) — first dogfood-side invocation against real models unblocks T2_pass_k v2_tightened.
  • Spec validator at 14 BLOCKER invariants (was 11 at v0.3.0); accepts v0.1.0 / v0.2.0 / v0.3.0 / v0.4.0-rc1 / v0.4.0 (backward-compat).
  • AGENTS.md §13 at 13 hard rules (was 10 at v0.3.0).
  • Deterministic tarball docs/skills/si-chip-0.4.0.tar.gz (SHA-256 2cfcce00f989faf2467014e638b0ea1fa67870b5a1ee6b0531942be5a4be21ab; 1 SKILL.md + 1 DESIGN.md + 14 references + 5 scripts; 21 files).

Quick install

Cursor and Claude Code auto-discover the included Skill mirrors. See Install for full instructions.

git clone https://github.com/YoRHa-Agents/Si-Chip.git
cd Si-Chip
python tools/spec_validator.py --json    # 14 invariants — verdict PASS

Out of scope (per spec section 11.1)

  • Skill / plugin marketplace
  • Router model training
  • Markdown-to-CLI converter
  • Generic IDE compatibility layer

PRs introducing any of these will be closed without review.

Install v0.4.0

curl -fsSL https://raw.githubusercontent.com/YoRHa-Agents/Si-Chip/main/install.sh | bash -s -- --yes --target cursor --scope repo

or

wget https://github.com/YoRHa-Agents/Si-Chip/releases/download/v0.4.0/si-chip-0.4.0.tar.gz
tar xzf si-chip-0.4.0.tar.gz -C .agents/skills/

持久化的 BasicAbility 优化工厂。

▶ 安装 ▶ 用户指南 ▶ 架构 ▶ 更新日志 ▶ 演示

状态（v0.4.0）： 在 gate standard（即 v2_tightened）达到 SHIP_ELIGIBLE。 Si-Chip 首次在 v2_tightened 档位发版（v0.2.0 / v0.3.0 仍在 relaxed）。 19 轮连续 v1_baseline + 2 轮连续 v2_tightened（Round 18 + Round 19）。 T2_pass_k 由 evals/si-chip/runners/real_llm_runner.py 解锁—— Round 18 通过 Veil litellm-local（127.0.0.1:8086）在 claude-haiku-4-5 + claude-sonnet-4-6 上跑 k=4 采样（$0.20、640 次调用）；Round 19 用 100% 命中的 cache replay（$0）完成第二轮通过。v3_strict 延后到 v0.4.x——唯一 阻塞项是 metadata_tokens = 94 vs ≤ 80 预算。

什么是 Si-Chip？

Si-Chip 是一个持久化的 BasicAbility 优化工厂。它的第一类对象是 BasicAbility，而不是任何特定的文件形态。它通过指标驱动的多轮 dogfood 循环来改进能力：

profile -> evaluate -> diagnose -> improve -> router-test -> half-retire-review -> iterate -> package-register

Si-Chip 以自身作为第一个用户进行交付：v0.4.0 是对 Si-Chip 自身运行 19 轮 完整 dogfood 后产出的结果（v0.1.0 完成 2 轮、v0.2.0 加 11 轮、v0.3.0 加 2 轮、 v0.4.0 加 4 轮）。

关键指标（v0.4.0 交付）

指标 v0.1.0 v0.4.0（Round 18-19） 上限（v2_tightened）
pass_rate 0.85 1.0 best cell，0.9719 mean ≥ 0.82
T2_pass_k 0.5478（PROXY） 1.0 best cell ≥ 0.55
trigger_F1 0.89 1.0 ≥ 0.85
near_miss_FP_rate 0.05 0.0 ≤ 0.10
metadata_tokens 78 94 ≤ 100
per-invocation footprint 4071 4726 ≤ 7000
wall_clock_p95（秒） 1.47 17.5726 s ≤ 30 s
routing_latency_p95 n/a 0.16 ms ≤ 1200 ms
routing_token_overhead n/a 0.0233 ≤ 0.12
iteration_delta（task_quality） +0.31（context_delta） +0.4522（Round 18） ≥ +0.10

完整的 ship-report 走查详见 演示页面，每轮 v0.1.x → v0.4.0 故事详见 更新日志。

v0.4.0 新增亮点

  • §18 Token-Tier Invariant —— token_tier {C7_eager_per_session, C8_oncall_per_trigger, C9_lazy_avg_per_load} 三层分解；可选 EAGER 加权 iteration_delta 公式；lazy_manifest 打包闸门。
  • §19 Real-Data Verification —— 三层模式（msw fixture 溯源 + user-install
    • post-recovery live verification）；BLOCKER 13 REAL_DATA_FIXTURE_PROVENANCE。
  • §20 Stage Transitions & Promotion History —— stage_transition_table 禁止反向迁移；promotion_history 仅可追加；当 round_kind == ship_prep 时新增 第 7 件证据文件 ship_decision.yaml。
  • §21 Health Smoke Check —— 4 维 {read, write, auth, dependency} pre-ship probe；当 live_backend: true 时 REQUIRED；BLOCKER 14 HEALTH_SMOKE_DECLARED_WHEN_LIVE_BACKEND。
  • §22 Eval-Pack Curation —— v2_tightened 最少 40 prompt；G1 _provenance REQUIRED；确定性种子 hash(round_id + ability_id)。
  • §23 Method-Tagged Metrics —— <metric>_method 伴随字段、_ci_low / _ci_high 95% CI 区间、U1_language_breakdown、U4_state。
  • §6.1 value_vector 7 → 8 维 —— 新增 eager_token_delta（自 v0.1.0 → v0.2.0 prose-count alignment 之后首次破坏字节一致性）。
  • Real-LLM runner evals/si-chip/runners/real_llm_runner.py（884 LoC + 27 tests）—— dogfood 侧首次对接真实模型，解锁 T2_pass_k v2_tightened。
  • Spec validator 升至 14 个 BLOCKER（v0.3.0 时 11 个）；接受 v0.1.0 / v0.2.0 / v0.3.0 / v0.4.0-rc1 / v0.4.0（向后兼容）。
  • AGENTS.md §13 升至 13 条 hard rules（v0.3.0 时 10 条）。
  • Deterministic tarball docs/skills/si-chip-0.4.0.tar.gz（SHA-256 2cfcce00f989faf2467014e638b0ea1fa67870b5a1ee6b0531942be5a4be21ab； 1 SKILL.md + 1 DESIGN.md + 14 references + 5 scripts；21 个文件）。

快速安装

Cursor 与 Claude Code 会自动发现本仓库内置的 Skill 镜像。完整说明见 安装。

git clone https://github.com/YoRHa-Agents/Si-Chip.git
cd Si-Chip
python tools/spec_validator.py --json    # 14 项不变量 — verdict PASS

范围之外（依据规范 §11.1）

  • Skill / plugin marketplace（技能 / 插件市场）
  • Router model training（路由模型训练）
  • Markdown-to-CLI converter（Markdown 到 CLI 转换器）
  • Generic IDE compatibility layer（通用 IDE 兼容层）

任何引入上述方向的 PR 将被直接关闭，不进入评审。

安装 v0.4.0

curl -fsSL https://raw.githubusercontent.com/YoRHa-Agents/Si-Chip/main/install.sh | bash -s -- --yes --target cursor --scope repo

或

wget https://github.com/YoRHa-Agents/Si-Chip/releases/download/v0.4.0/si-chip-0.4.0.tar.gz
tar xzf si-chip-0.4.0.tar.gz -C .agents/skills/