cost-booster-edit

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Cost Booster Edit

Direct wrapper around

agent-booster.apply()

(npm

agent-booster

v0.2.x, exposed via

agentic-flow/agent-booster

). Use when a transform is already classified as Tier 1 eligible —

cost-booster-route

recommends whether; this skill executes.

直接封装

agent-booster.apply()

（npm包

agent-booster

v0.2.x，通过

agentic-flow/agent-booster

暴露）。当转换已被归类为符合Tier 1条件时使用——

cost-booster-route

负责判断是否适用，本技能负责执行转换。

When to use

使用场景

Bulk transforms across many files (

var → const

add-types

remove-console

add-error-handling

async-await

add-logging

Any simple, structural edit where an LLM would otherwise be called and billed.
Inside CI pipelines where determinism + zero-cost matter more than naturalness.

Do NOT use when the transform requires reasoning about intent, naming, or cross-file context — those are Tier 2/3 jobs.

多文件批量转换（

var → const

、

add-types

、

remove-console

、

add-error-handling

、

async-await

、

add-logging

）。

任何原本需要调用LLM并产生费用的简单结构化编辑场景。
在CI流水线中，确定性与零成本比自然语义更重要的场景。

请勿使用场景：转换需要推理意图、命名或跨文件上下文——这些属于Tier 2/3任务。

Steps

步骤

Take inputs —
```
intent
```
(one of the 6 booster intents) and
```
file
```
path.
Read the source to a variable, derive the intended
```
edit
```
text from the intent (caller supplies).

Invoke — run from anywhere under

v3/

agent-booster

resolves:

bash

node --input-type=module -e '
  import("agent-booster")
    .then(async ({ AgentBooster }) => {
      const booster = new AgentBooster();
      const r = await booster.apply({
        code: process.argv[1],
        edit: process.argv[2],
        language: process.argv[3] || "javascript",
      });
      console.log(JSON.stringify({
        success: r.success, output: r.output, latency: r.latency,
        confidence: r.confidence, strategy: r.strategy,
        tokens: r.tokens,
      }));
    })
    .catch(e => console.log(JSON.stringify({ success: false, error: String(e.message) })));
' -- "$CODE" "$EDIT" "$LANG"

Check confidence — default threshold is
```
0.5
```
. Below that, fail closed: do NOT write the file; report and escalate to Tier 2/3.
Write back the
```
output
```
field if
```
success && confidence >= 0.5
```
.

Persist outcome —

memory_store --namespace cost-tracking --key "booster-edit-..." --value '{"intent":..., "latency":..., "confidence":..., "strategy":..., "applied":true}'

. Feed the routing learner via

hooks_model-outcome

(use the

cost-optimize

skill's step 8).

接收输入——
```
intent
```
（6种booster意图之一）和
```
file
```
路径。
读取源码到变量，根据intent推导目标
```
edit
```
文本（由调用方提供）。

调用执行——在

v3/

目录下任意位置运行，确保

agent-booster

可以被解析：

bash

node --input-type=module -e '
  import("agent-booster")
    .then(async ({ AgentBooster }) => {
      const booster = new AgentBooster();
      const r = await booster.apply({
        code: process.argv[1],
        edit: process.argv[2],
        language: process.argv[3] || "javascript",
      });
      console.log(JSON.stringify({
        success: r.success, output: r.output, latency: r.latency,
        confidence: r.confidence, strategy: r.strategy,
        tokens: r.tokens,
      }));
    })
    .catch(e => console.log(JSON.stringify({ success: false, error: String(e.message) })));
' -- "$CODE" "$EDIT" "$LANG"

检查置信度——默认阈值为
```
0.5
```
。低于该阈值时，终止操作：请勿写入文件，上报并升级至Tier 2/3任务。
写入结果——若
```
success && confidence >= 0.5
```
，将
```
output
```
字段内容写回文件。

保存结果——

memory_store --namespace cost-tracking --key "booster-edit-..." --value '{"intent":..., "latency":..., "confidence":..., "strategy":..., "applied":true}'

。通过

hooks_model-outcome

将结果反馈给路由学习器（参考

cost-optimize

技能的步骤8）。

Measured benchmark (2026-05-04, this checkout)

实测基准（2026-05-04，当前版本）

5 representative intents run through

AgentBooster.apply()

intent	latency (ms)	wall (ms)	confidence	strategy	success
var-to-const	5	5	0.65	fuzzy_replace	true
add-types	1	1	0.64	fuzzy_replace	true
remove-console	0	0	0.70	fuzzy_replace	true
add-error-handling	0	0	0.85	exact_replace	true
async-await	0	0	0.85	exact_replace	true

Avg measured latency ≈ 1.2 ms. All 5 above the default 0.5 confidence threshold. See

docs/benchmarks/0002-baseline.md

for the LLM-baseline comparison.

5种代表性意图通过

AgentBooster.apply()

运行的结果：

intent	延迟 (ms)	实际耗时 (ms)	置信度	策略	成功
var-to-const	5	5	0.65	fuzzy_replace	true
add-types	1	1	0.64	fuzzy_replace	true
remove-console	0	0	0.70	fuzzy_replace	true
add-error-handling	0	0	0.85	exact_replace	true
async-await	0	0	0.85	exact_replace	true

平均实测延迟≈1.2毫秒。上述5种意图的置信度均高于默认阈值0.5。LLM基准对比请参考

docs/benchmarks/0002-baseline.md

。

What's verified locally

本地验证项

Claim	Status here
100% win rate	Verified — 12/12 on `bench/booster-corpus.json` (see `runs/latest.json` ). Booster AND Gemini 2.0 Flash both score 12/12 — this is a structural-correctness corpus, not a hard adversarial one.
Sub-millisecond latency	Verified — avg 0.67 ms, p50 0 ms, p99 6 ms, max 6 ms.
$0 per edit	Verified structurally — no API call, no token billing.
Deterministic AST-based merge	Verified — same inputs reproduce the same `output` and `strategy` .
Confidence ≥ 0.5 ⇒ correct	Verified on this corpus — 12/12 above 0.5 (min 0.551), all correct.
`350×` speedup vs. LLM	Verified — exceeded against every tier: 1000.9× vs Gemini 2.0 Flash, 1838.7× vs Claude Sonnet 4.6, 2634.1× vs Claude Opus 4.7. Run `BENCH_LLM_BASELINE=1 BENCH_ANTHROPIC=1 node scripts/bench.mjs` to refresh.
Cost saved per edit	Measured: $0.000020 vs Gemini, $0.000722 vs Sonnet 4.6, $0.004720 vs Opus 4.7 (the booster side is $0 in all cases).
Win parity with frontier LLMs	Verified — Booster, Gemini 2.0 Flash, Sonnet 4.6, Opus 4.7 all scored 12/12 on this corpus. Booster matches LLM accuracy structurally for deterministic transforms.

To extend: add cases to

bench/booster-corpus.json

, run

( cd v3 && node ../plugins/ruflo-cost-tracker/scripts/bench.mjs )

(or with

BENCH_LLM_BASELINE=1

), commit

runs/latest.json

. Smoke step 23 fails the build if win rate drops below 0.80.

Override the LLM model:

BENCH_LLM_MODEL='claude-sonnet-4'

(when wired against

api.anthropic.com

) or

BENCH_LLM_MODEL='models/gemini-2.5-flash'

for a reasoning-model comparison. Pricing flags:

BENCH_LLM_PRICE_IN

BENCH_LLM_PRICE_OUT

fuzzy_replace

is best-effort; for production transforms prefer cases that route to

exact_replace

(≥0.85 confidence in our sample).

声明	当前验证状态
100%成功率	已验证 — 在 `bench/booster-corpus.json` 上12/12通过（见 `runs/latest.json` ）。Booster与Gemini 2.0 Flash均获得12/12的成绩——这是一个结构正确性测试集，而非高难度对抗测试集。
亚毫秒级延迟	已验证 — 平均0.67毫秒，p50为0毫秒，p99为6毫秒，最大值6毫秒。
每次编辑零成本	结构验证通过 — 无API调用，无token计费。
基于AST的确定性合并	已验证 — 相同输入会生成相同的 `output` 和 `strategy` 。
置信度≥0.5即正确	在当前测试集上已验证 — 12项测试的置信度均高于0.5（最低0.551），全部正确。
比LLM快350倍	已验证——远超所有LLM层级：比Gemini 2.0 Flash快1000.9倍，比Claude Sonnet 4.6快1838.7倍，比Claude Opus 4.7快2634.1倍。运行 `BENCH_LLM_BASELINE=1 BENCH_ANTHROPIC=1 node scripts/bench.mjs` 可刷新基准数据。
每次编辑节省的成本	已测算：对比Gemini节省$0.000020，对比Sonnet 4.6节省$0.000722，对比Opus 4.7节省$0.004720（Booster侧均为$0）。
与前沿LLM准确率持平	已验证 — Booster、Gemini 2.0 Flash、Sonnet 4.6、Opus 4.7在该测试集上均获得12/12的成绩。对于确定性转换，Booster在结构准确性上与LLM持平。

扩展方法：在

bench/booster-corpus.json

中添加测试用例，运行

( cd v3 && node ../plugins/ruflo-cost-tracker/scripts/bench.mjs )

（或添加

BENCH_LLM_BASELINE=1

参数），提交

runs/latest.json

。若成功率低于0.80，冒烟测试步骤23会导致构建失败。

覆盖LLM模型：

BENCH_LLM_MODEL='claude-sonnet-4'

（对接

api.anthropic.com

时）或

BENCH_LLM_MODEL='models/gemini-2.5-flash'

，用于推理模型对比。定价参数：

BENCH_LLM_PRICE_IN

、

BENCH_LLM_PRICE_OUT

。

fuzzy_replace

为尽力而为策略；生产环境转换优先选择路由到

exact_replace

的场景（在我们的样本中置信度≥0.85）。

Cross-references

交叉引用

ADR-0002 §"Decision 1" (route classifier) and §"Riskiest assumption" (Bash-shelled invocation) ·

cost-booster-route

(classifier-side companion) ·

agent-booster

npm README (3-mode install, MCP / npm / HTTP).

ADR-0002 §"Decision 1"（路由分类器）和§"Riskiest assumption"（Bash shell调用） ·

cost-booster-route

（分类器侧配套工具） ·

agent-booster

npm README（三种安装模式：MCP / npm / HTTP）。