English | 中文
backtrader-skills is an offline, independently installable author/review/test product for this
Backtrader fork. It turns a registered local dataset and a typed StrategySpec v1 into either a
collected pytest strategy or a three-file Python bundle, reviews the candidate without importing it,
and runs approved candidates in separate runonce/runnext child processes.
It does not import or start sibling MCP or Agent products. The bundled catalog snapshot contains metadata for 1,152 functional strategy tests and 1,035 three-file packages, with 1,032 mapped IDs, so normal operation does not require either source corpus.
Online documentation is published at https://cloudquant.github.io/backtrader-skills/
(GitHub Pages, live once Pages is enabled — see
docs/iterations/iteration-30-bilingual-docs-site/acceptance.md) and at
https://backtrader-skills.readthedocs.io/ (ReadTheDocs, live once the project is imported on
readthedocs.org). Both URLs follow the standard <owner>.github.io/<repo> and
<repo>.readthedocs.io patterns.
This repository follows Semantic Versioning; the current release is 0.2.0. The change history is in CHANGELOG.md and the P1 backlog is in docs/roadmap.en.md. Agents working on this repository must read CLAUDE.md (non-Claude hosts: AGENTS.md), which defines the maintainer environment, the verification commands, and the distribution manifest rule.
The only Backtrader source accepted for strategy execution and acceptance is
cloudQuant/backtrader. A matching package name or
version number is not sufficient: the tool verifies the Git remote, or PEP 610 installation
metadata that leads back to that remote.
Run doctor before using the product. If the active Python environment has no backtrader,
doctor installs git+https://github.com/cloudQuant/backtrader.git with that same interpreter and
verifies it again. If a backtrader package already exists but cannot be proven to be the
cloudQuant fork, doctor returns a BACKTRADER_SOURCE_WARNING; it does not silently replace the
existing package. The run command uses the same preflight and writes that warning to stderr.
Every --target and source-checkout --repository is also required to be a cloudQuant Git
checkout. A valid-looking package from another fork is rejected with
BACKTRADER_SOURCE_MISMATCH.
From the backtrader-skills checkout, activate any supported Python 3.10–3.13 environment and
install the distribution. Conda is optional; for example, conda activate base may be used before
these commands:
python -m pip install .
backtrader-skills --target /path/to/backtrader doctorRuntime state is always <target>/.backtrader-skills/. Dataset objects, manifests, draft bytes,
approval-token digests, run evidence, and install manifests remain there. The 256-bit token handle
is returned once to the caller and is never persisted in plaintext. doctor records the actual
interpreter and environment used by the installed command; no local machine path is part of the
distribution interface.
The same distribution supports four project-level layouts:
| Host | Destination |
|---|---|
| Claude Code | .claude/skills/backtrader-* |
| Codex | .agents/skills/backtrader-* |
| OpenCode | .opencode/skills/backtrader-* |
| OpenClaw | <workspace>/skills/backtrader-* |
Preview, approve, and apply:
BT_TARGET=/path/to/backtrader
backtrader-skills --target "$BT_TARGET" \
install preview --host codex
backtrader-skills --target "$BT_TARGET" \
approval approve --token-id tok_...
backtrader-skills --target "$BT_TARGET" \
install apply --plan-id install_codex_... --token-id tok_...Use claude, opencode, or openclaw for the other native locations. Installation is create-only.
For OpenClaw, set BT_TARGET to the actual agent workspace root because its native skill directory
is <workspace>/skills; the installer does not register an OpenClaw agent for you.
Uninstall is also preview/approval/apply; files whose hash changed after installation are preserved:
backtrader-skills --target "$BT_TARGET" \
install uninstall-preview --host codex
backtrader-skills --target "$BT_TARGET" \
approval approve --token-id tok_...
backtrader-skills --target "$BT_TARGET" \
install uninstall-apply --plan-id uninstall_codex_... --token-id tok_...Each installed skill has one thin scripts/backtrader_skills.py forwarder. Deterministic behavior
lives only in src/backtrader_skills/.
Applying an install plan proves that the canonical files reached the native directory; it does not
prove that an external model session discovered them. Reload the project or start a new host session
after installation. Use the following read-only first request, replacing /path/to/backtrader with
the Backtrader project root:
Without writing any files, use the backtrader-strategy-author skill. Run:
backtrader-skills --target /path/to/backtrader doctor
Return the doctor pass/fail result, the no-sibling-product-imports check, and the catalog counts.
The expected smoke result has passed=true, a passing no-sibling-product-imports check, and the
verified catalog baseline 1,152/1,035/1,032. A host that cannot name or load the skill has not
completed discovery, even if the files exist.
Confirm .claude/skills/backtrader-strategy-author/SKILL.md exists, restart Claude Code in the
project, and send the first request above. Prefixing the request with “use the
backtrader-strategy-author skill” is the explicit trigger; keep the returned transcript or host
tool trace as discovery evidence.
Confirm .agents/skills/backtrader-strategy-author/SKILL.md exists and start a new task in the
project. Invoke the skill explicitly with:
$backtrader-strategy-author Perform the read-only doctor smoke described above.
Record the resolved skill name and command output. A filesystem-only check is not a Codex discovery test.
Confirm .opencode/skills/backtrader-strategy-author/SKILL.md exists, reload the project, and ask
OpenCode to “load and use the backtrader-strategy-author skill” before sending the first request.
Retain the skill/tool trace and the doctor JSON result as evidence.
Set BT_TARGET to an existing, explicitly registered OpenClaw agent workspace before installation,
then confirm skills/backtrader-strategy-author/SKILL.md exists below that workspace. Reload the
registered agent and ask it to “use the workspace skill backtrader-strategy-author” for the first
request. This installer does not create or register the agent itself.
OpenClaw was not installed in the environment used for the current acceptance snapshot. Its layout, metadata, forwarders, conflict handling, and protected uninstall are statically tested; live discovery must remain unchecked until an installed OpenClaw agent completes the smoke above.
The evals/ suite measures whether a real host agent can use the three installed skills to
produce reviewed, runnable strategies, and whether the review gates hold. Ten prompts live in
evals/prompts/: seven golden prompts (one per archetype) plus three adversarial/cross-skill
probes. The runbook is evals/README.md.
Score the artifact a prompt run produced with the shipped mechanical scorer:
python scripts/record_eval.py \
--target <target> \
--artifact <target>/strategies/generated/<archetype>/<artifact_id>_<slug>/strategy.py \
--dataset-id 'ds_<64hex>' \
--out evals/results/<prompt-file>.jsonThe scorer invokes the installed CLI for review and run prepare only. run execute
requires a human-approved token inside the host session, so live LLM runs are manual; the
runbook's rubric rows cover what the scorer cannot.
P0 accepts only offline local files inside explicitly registered, read-only roots. Portable manifests contain an opaque root ID and relative path, never the local absolute path.
backtrader-skills --target "$BT_TARGET" \
data root-add --directory /path/to/fixtures --root-id prices
backtrader-skills --target "$BT_TARGET" \
data inspect --feed-spec feed.json
backtrader-skills --target "$BT_TARGET" \
data register --spec data-spec.json
backtrader-skills --target "$BT_TARGET" \
data preview --dataset-id 'ds_<64hex>' --rows 5DataSpec supports multiple named feeds, roles, timeframe/compression, timezone, explicit column
mapping, deterministic transforms, and intersection|left|explicit_asof declarations. Registration
normalizes header-based CSV/tabular inputs to UTF-8 canonical CSV, validates timestamps, finite
OHLC, ordering and duplicates, and stores content-addressed objects. Formats are
generic_csv, backtrader_csv, yahoo_csv, mt5_csv, pandas, and
pandas_custom_lines; Pandas profiles consume a safely materialized CSV, never pickle or a callable.
Any source-byte change invalidates the manifest and its approvals.
Search the shipped catalog and create a scaffold:
backtrader-skills --target "$BT_TARGET" \
catalog search --query "multi timeframe momentum" --archetype multi_timeframe
backtrader-skills --target "$BT_TARGET" \
spec scaffold --archetype multi_timeframe --output-profile python_bundle \
--dataset-id 'ds_<64hex>' --feed-count 2 > strategy-spec.jsonValidate the JSON after removing any surrounding CLI presentation, then use the two-phase writer:
backtrader-skills --target "$BT_TARGET" \
spec validate --spec strategy-spec.json
backtrader-skills --target "$BT_TARGET" \
render preview --spec strategy-spec.json
backtrader-skills --target "$BT_TARGET" \
render validate --draft-id draft_...
backtrader-skills --target "$BT_TARGET" \
approval approve --token-id tok_...
backtrader-skills --target "$BT_TARGET" \
render apply --draft-id draft_... --token-id tok_...Bundles are created under strategies/generated/. Collected generated tests are created under
tests/functional/strategies/generated/. Existing files require an explicit expected hash.
Multi-file apply stages every byte first and uses a journal plus rollback, so a later-file failure
does not leave a partially applied bundle.
All seven archetypes-single-data indicator, multi-indicator, multi-asset allocation,
multi-timeframe, pairs/spread, order/risk, and precomputed/ML signal-use the same restricted
Expression/Action/StateRule IR for both output profiles. Direct bt.Strategy templates intentionally
do not call super().__init__() in this fork.
backtrader-skills --target "$BT_TARGET" \
review --file "$BT_TARGET/strategies/generated/.../strategy.py"
backtrader-skills --target "$BT_TARGET" \
repair --draft-id draft_...
backtrader-skills --target "$BT_TARGET" \
run prepare --candidate "$BT_TARGET/strategies/generated/.../strategy.py" \
--dataset-id 'ds_<64hex>'
backtrader-skills --target "$BT_TARGET" \
approval approve --token-id tok_...
backtrader-skills --target "$BT_TARGET" \
run execute --run-id run_... --token-id tok_...When a diagnostic requires a semantic change, revise the typed spec and bind it to the failed ValidationReport:
backtrader-skills --target "$BT_TARGET" \
repair --spec revised-strategy-spec.json --validation-report failed-validation.jsonThe controller never imports candidate code. It proves the candidate is an unchanged artifact from
an approved render/apply, recomputes candidate, dataset, source-data, and environment hashes,
consumes a separate execution approval, and invokes the distribution's active Python interpreter
with -I for each mode.
Approval capabilities expire after 15 minutes by default and end in
CONSUMED, REVOKED, or EXPIRED. Reports are stored as JSON and Markdown.
The 11 metrics have frozen units: six integer bar/trade counts; account-currency final_value;
nullable dimensionless sharpe_ratio; nullable ratio annual_return; percent max_drawdown; and
percent return_rate. Integers and normalized events compare exactly. Floats use rel_tol=1e-7,
abs_tol=1e-9, with a documented amount override. Null equals only null; missing, NaN, and Infinity
fail.
- P0 runs only product-generated and explicitly approved candidates. Unknown code may receive a static review but cannot receive a run token.
- Generated candidates may import only top-level
backtrader. AST gates reject controller and filesystem imports, dynamic execution/import, subprocess, known network clients, sockets, live stores, absolute paths, traversal, and positive line offsets. python -Ichild isolation is not a complete OS sandbox. P0 has no network namespace, container, seccomp, or resource cgroup.- Data is offline and header-based. No download, database, API key, pickle, live feed, or arbitrary loader/callable is accepted.
- Alignment (
intersection,left, orexplicit_asof), resample, and replay intent is frozen in the DatasetManifest and validated before feed assembly. The P0 runner delegates bar-clock advancement to Backtrader and does not silently fill missing bars or change calendars. - The automated P0 runner proves runonce/runnext parity. A separately checked-out, human-approved master/dev financial baseline remains an explicit release workflow, not an inferred expected return.
- Host-client UI discovery cannot be emulated without each client binary. Product tests verify all four native paths, canonical skill metadata, forwarders, conflicts, and protected uninstall.
- General concurrent CLI invocations against the same
--targetremain unsupported. State files outside approval tokens are not globally serialized; run one command at a time per target. - A single approval token is protected across local processes: render apply and install or uninstall hold a per-token lock through their protected writes, while run atomically consumes its token before launching child processes. At most one operation can consume the same token. This does not serialize unrelated tokens, data-root registration, draft previews, or arbitrary target writes.
Run these commands from the backtrader-skills checkout with the intended environment activated.
Repository maintainers use the Anaconda base environment required by the repository's AGENTS.md,
but that machine-specific executable path is not part of the public commands.
The source-checkout helpers automatically locate a Backtrader repository only when this product is
either nested below that repository or next to a sibling directory named backtrader. They validate
that the selected root contains backtrader/version.py and that its Git remote is
cloudQuant/backtrader. They return SOURCE_CHECKOUT_NOT_FOUND when neither layout exists and
BACKTRADER_SOURCE_MISMATCH when a candidate is another fork, rather than guessing or continuing
with an incompatible implementation.
After changing a distribution-included file, rebuild the tracked manifest once with
python scripts/build_manifest.py. For routine, read-only validation use
python scripts/build_manifest.py --check; both --check and --help leave
manifest.json unchanged.
# Automatic discovery for a nested or sibling Backtrader checkout
python scripts/doctor.py
python scripts/build_manifest.py --check
python scripts/build_catalog.py --check
python -m pytest tests -q
python scripts/run_acceptance.py \
--matrix all --require-no-mcp --require-no-agent
# A Backtrader checkout in any other location
python scripts/doctor.py --target /path/to/backtrader
python scripts/run_acceptance.py --repository /path/to/backtrader \
--matrix all --require-no-mcp --require-no-agentdoctor, the acceptance matrix, and the execution tests run strategies through the cloudQuant
Backtrader source package. The test suite also honors BT_BACKTRADER_DIR pointing at that
checkout's backtrader package directory, and skips source-backed tests when the package is absent.
Continuous integration enforces this on every push to master: a dedicated acceptance job checks
out the cloudQuant Backtrader fork, runs the full suite with the source package present (execution,
doctor, and acceptance tests included), enforces at least 80% coverage over
src/backtrader_skills (except the two python -I child modules, child_runner.py and
isolate_entry.py, which the coverage config omits), and then runs the complete 7×2 acceptance
matrix. Pull requests keep the lighter quality and supported-Python jobs. Emulating each host client binary and the
human-approved master/dev financial baseline remain local release steps.
The acceptance command builds a wheel, installs it into an isolated directory, exposes only the
Backtrader source package to a clean fixture repository, and runs the full 7×2 matrix from that
installed distribution with the source checkout absent from sys.path. The seven archetypes use
seven distinct DatasetManifests covering all six declared adapters; multi-data, resample, and
precomputed custom-line semantics are recorded per cell. Every cell records independent runonce
and runnext hashes plus comparison results. Multi-data, multi-timeframe, and precomputed/ML
representative cells must also pass a structured failure -> typed-IR repair -> revalidation ->
approved dual-mode backtest gate.
The wheel contains seven named JSON Schemas, comparison-profile-v1.json, the full metadata
snapshot, four host adapter manifests, and all three canonical skills. manifest.json records every
published file hash and compatibility range.
English | 中文
backtrader-skills 是面向本 Backtrader fork 的离线、可独立安装的编写 / 审查 / 测试产
品。它把已登记的本地数据集和 typed StrategySpec v1 转换为一个收集型 pytest 策略或
三文件 Python bundle，在不导入候选项的前提下审查它，并在独立的 runonce/runnext 子进
程中运行已批准的候选项。
它不导入或启动 sibling 的 MCP 或 Agent 产品。内置的 catalog 快照含 1,152 个功能策略 测试和 1,035 个三文件包的元数据，以及 1,032 个已映射 ID，因此正常使用不需要任一源语 料。
在线文档发布在 https://cloudquant.github.io/backtrader-skills/（GitHub Pages，启用
Pages 后才会上线——见 docs/iterations/iteration-30-bilingual-docs-site/acceptance.md）和
https://backtrader-skills.readthedocs.io/（ReadTheDocs，在 readthedocs.org 上导入本项目
后才会上线）。两个 URL 都遵循标准的 <owner>.github.io/<repo> 与
<repo>.readthedocs.io 模式。
本仓库遵循语义化版本（SemVer）；当前发布版本为 0.2.0。变更历史见 CHANGELOG.md，P1 backlog 见 docs/roadmap.zh.md。在本仓库 工作的 Agent 必须阅读 CLAUDE.md（非 Claude 宿主：AGENTS.md）， 其中定义了维护者环境、验证命令和分发 manifest 规则。
策略执行和验收只接受
cloudQuant/backtrader。仅凭同名包或版本号不能
证明兼容性：工具会校验 Git remote，或校验可回溯至该 remote 的 PEP 610 安装元数据。
使用产品前先运行 doctor。若当前 Python 环境没有 backtrader，doctor 会使用同一解释器
安装 git+https://github.com/cloudQuant/backtrader.git，随后再次验证来源。若已存在
backtrader 但无法证明来自 cloudQuant fork，doctor 会返回
BACKTRADER_SOURCE_WARNING；它不会静默替换已有包。run 命令使用同一预检，并将该警告
写入 stderr。
每个 --target 和源码检出 --repository 也必须是 cloudQuant Git 检出。来自其他 fork 的
目录即使包含看似有效的包，也会以 BACKTRADER_SOURCE_MISMATCH 拒绝。
在 backtrader-skills 检出目录下，激活任意受支持的 Python 3.10–3.13 环境并安装本分
发。Conda 是可选的；例如，可在这些命令前先 conda activate base：
python -m pip install .
backtrader-skills --target /path/to/backtrader doctor运行时状态始终位于 <target>/.backtrader-skills/。数据集对象、manifest、草稿字节、
审批 token 摘要、运行证据和安装 manifest 都留在那里。256 位 token 句柄只返回一次给调
用方，绝不以明文持久化。doctor 记录已安装命令实际使用的解释器和环境；任何本地机器
路径都不是分发接口的一部分。
同一分发支持四种项目级布局：
| 宿主 | 目标位置 |
|---|---|
| Claude Code | .claude/skills/backtrader-* |
| Codex | .agents/skills/backtrader-* |
| OpenCode | .opencode/skills/backtrader-* |
| OpenClaw | <workspace>/skills/backtrader-* |
preview、approve、apply：
BT_TARGET=/path/to/backtrader
backtrader-skills --target "$BT_TARGET" \
install preview --host codex
backtrader-skills --target "$BT_TARGET" \
approval approve --token-id tok_...
backtrader-skills --target "$BT_TARGET" \
install apply --plan-id install_codex_... --token-id tok_...其他原生位置用 claude、opencode 或 openclaw。安装是 create-only 的。对
OpenClaw，把 BT_TARGET 设为实际的 agent 工作区根目录，因为其原生 skill 目录是
<workspace>/skills；安装器不会替你注册 OpenClaw agent。卸载同样是
preview/approval/apply；安装后哈希变化过的文件会被保留：
backtrader-skills --target "$BT_TARGET" \
install uninstall-preview --host codex
backtrader-skills --target "$BT_TARGET" \
approval approve --token-id tok_...
backtrader-skills --target "$BT_TARGET" \
install uninstall-apply --plan-id uninstall_codex_... --token-id tok_...每个已安装 skill 都有一个薄薄的 scripts/backtrader_skills.py 转发器。确定性逻辑只存
在于 src/backtrader_skills/。
应用安装计划只证明规范文件到达了原生目录，并不能证明外部 model 会话发现了它们。安装
后请重载项目或开启新的宿主会话。使用下面的只读首个请求，把 /path/to/backtrader 替
换为 Backtrader 项目根目录：
Without writing any files, use the backtrader-strategy-author skill. Run:
backtrader-skills --target /path/to/backtrader doctor
Return the doctor pass/fail result, the no-sibling-product-imports check, and the catalog counts.
预期 smoke 结果为 passed=true、通过的 no-sibling-product-imports 检查，以及已验证的
catalog 基线 1,152/1,035/1,032。无法命名或加载该 skill 的宿主即便文件存在，也未完
成发现。
确认 .claude/skills/backtrader-strategy-author/SKILL.md 存在，在项目中重启 Claude
Code，并发送上文首个请求。在请求前加“use the backtrader-strategy-author skill”是显
式触发；保留返回的 transcript 或宿主工具 trace 作为发现证据。
确认 .agents/skills/backtrader-strategy-author/SKILL.md 存在，并在项目中开启新任
务。用如下方式显式调用 skill：
$backtrader-strategy-author Perform the read-only doctor smoke described above.
记录解析出的 skill 名和命令输出。仅文件系统检查不算 Codex 发现测试。
确认 .opencode/skills/backtrader-strategy-author/SKILL.md 存在，重载项目，并在发送
首个请求前要求 OpenCode “load and use the backtrader-strategy-author skill”。保留
skill/tool trace 和 doctor JSON 结果作为证据。
安装前把 BT_TARGET 设为一个已存在、已显式注册的 OpenClaw agent 工作区，然后确认该
工作区下存在 skills/backtrader-strategy-author/SKILL.md。重载已注册的 agent，并要
求它对首个请求“use the workspace skill backtrader-strategy-author”。本安装器不创建
或注册 agent 本身。
当前验收快照所用环境未安装 OpenClaw。其布局、元数据、转发器、冲突处理和受保护卸载经 过静态测试；在已安装的 OpenClaw agent 完成上述 smoke 之前，live 发现保持未检查状态。
evals/ 套件衡量真实宿主 agent 能否用三个已安装 skill 产出通过审查、可运行的策略，以及
审查门禁能否挡住对抗性请求。十个 prompt 位于 evals/prompts/：七个 golden prompt（每个
archetype 一个）加三个对抗性 / 跨 skill 探针。运行手册是 evals/README.md。
用内置的机械评分器对一次 prompt 运行产出的 artifact 评分：
python scripts/record_eval.py \
--target <target> \
--artifact <target>/strategies/generated/<archetype>/<artifact_id>_<slug>/strategy.py \
--dataset-id 'ds_<64hex>' \
--out evals/results/<prompt-file>.json评分器只调用已安装 CLI 的 review 和 run prepare。run execute 需要宿主会话内人工批
准的 token，因此 live LLM 运行是手动的；评分器无法覆盖的部分由运行手册中的 rubric 行负责。
P0 只接受显式注册的只读 root 内的离线本地文件。可移植 manifest 含不透明 root ID 和相 对路径，绝不含本地绝对路径。
backtrader-skills --target "$BT_TARGET" \
data root-add --directory /path/to/fixtures --root-id prices
backtrader-skills --target "$BT_TARGET" \
data inspect --feed-spec feed.json
backtrader-skills --target "$BT_TARGET" \
data register --spec data-spec.json
backtrader-skills --target "$BT_TARGET" \
data preview --dataset-id 'ds_<64hex>' --rows 5DataSpec 支持多个具名 feed、角色、timeframe/compression、时区、显式列映射、确定性
transform，以及 intersection|left|explicit_asof 声明。注册把基于表头的 CSV / 表格输
入归一化为 UTF-8 规范 CSV，校验时间戳、有限 OHLC、顺序和重复，并存储内容寻址对象。格
式为 generic_csv、backtrader_csv、yahoo_csv、mt5_csv、pandas 和
pandas_custom_lines；Pandas profile 消费安全物化的 CSV，绝不接受 pickle 或
callable。任何源字节变化都会使 manifest 及其审批失效。
搜索内置 catalog 并创建脚手架：
backtrader-skills --target "$BT_TARGET" \
catalog search --query "multi timeframe momentum" --archetype multi_timeframe
backtrader-skills --target "$BT_TARGET" \
spec scaffold --archetype multi_timeframe --output-profile python_bundle \
--dataset-id 'ds_<64hex>' --feed-count 2 > strategy-spec.json去掉任何外围 CLI 展示后校验 JSON，然后用两阶段 writer：
backtrader-skills --target "$BT_TARGET" \
spec validate --spec strategy-spec.json
backtrader-skills --target "$BT_TARGET" \
render preview --spec strategy-spec.json
backtrader-skills --target "$BT_TARGET" \
render validate --draft-id draft_...
backtrader-skills --target "$BT_TARGET" \
approval approve --token-id tok_...
backtrader-skills --target "$BT_TARGET" \
render apply --draft-id draft_... --token-id tok_...bundle 创建在 strategies/generated/ 下。收集型生成测试创建在
tests/functional/strategies/generated/ 下。既有文件需要显式预期哈希。多文件 apply 会
先暂存每个字节，并使用日志加回滚，因此后续文件失败不会留下部分应用的 bundle。
全部七个 archetype--单数据指标、多指标、多资产配置、多 timeframe、配对 / 价差、订单
/ 风险、precomputed/ML 信号--对两种输出 profile 都使用同一套受限的
Expression/Action/StateRule IR。本 fork 中直接的 bt.Strategy 模板刻意不调用
super().__init__()。
backtrader-skills --target "$BT_TARGET" \
review --file "$BT_TARGET/strategies/generated/.../strategy.py"
backtrader-skills --target "$BT_TARGET" \
repair --draft-id draft_...
backtrader-skills --target "$BT_TARGET" \
run prepare --candidate "$BT_TARGET/strategies/generated/.../strategy.py" \
--dataset-id 'ds_<64hex>'
backtrader-skills --target "$BT_TARGET" \
approval approve --token-id tok_...
backtrader-skills --target "$BT_TARGET" \
run execute --run-id run_... --token-id tok_...当诊断需要语义修改时，修订 typed spec 并把它绑定到失败的 ValidationReport：
backtrader-skills --target "$BT_TARGET" \
repair --spec revised-strategy-spec.json --validation-report failed-validation.json控制器绝不导入候选代码。它证明候选项是来自已批准 render/apply 的未变 artifact，重算
candidate、dataset、source-data 和环境哈希，消费一次单独的执行审批，并对每个 mode 用
分发的活动 Python 解释器加 -I 调用。审批能力默认 15 分钟后过期，并以
CONSUMED、REVOKED 或 EXPIRED 结束。报告以 JSON 和 Markdown 存储。
11 个指标有冻结的单位：六个整数 bar/trade 计数；账户币种 final_value；可空无量纲
sharpe_ratio；可空比率 annual_return；百分比 max_drawdown；百分比
return_rate。整数和归一化事件精确比较。浮点数用 rel_tol=1e-7、abs_tol=1e-9，并
有文档化的 amount override。null 仅等于 null；缺失、NaN 和 Infinity 失败。
- P0 只运行产品生成且已显式批准的候选项。未知代码可获静态审查，但不能拿到 run token。
- 生成的候选项只能 import 顶层
backtrader。AST 门禁拒绝控制器和文件系统 import、动 态执行 / import、subprocess、已知网络客户端、socket、实盘 store、绝对路径、穿越和正 行偏移。 python -I子进程隔离不是完整 OS 沙箱。P0 无网络 namespace、容器、seccomp 或资源 cgroup。- 数据为离线、基于表头。不接受下载、数据库、API key、pickle、实盘 feed 或任意 loader/callable。
- 对齐（
intersection、left或explicit_asof）、resample 和 replay 意图冻结在 DatasetManifest 中，并在 feed 装配前校验。P0 runner 把 bar 时钟推进委托给 Backtrader，不静默填充缺失 bar 或更改日历。 - 自动化 P0 runner 证明 runonce/runnext 一致性。单独检出、人工批准的 master/dev 财务 基线仍是显式发布工作流，而非推断出的预期收益。
- 宿主客户端 UI 发现无法在缺少各客户端二进制的情况下模拟。产品测试验证全部四条原生路 径、规范 skill 元数据、转发器、冲突和受保护卸载。
- 对同一
--target的一般并发 CLI 调用仍不受支持。审批令牌以外的状态文件没有全局 串行化；每个 target 一次只运行一条命令。 - 同一审批 token 在本地跨进程间受保护：render apply、install 和 uninstall 在受保护写入期间 持有 per-token 锁；run 会在启动子进程前原子消费 token。同一 token 最多只能被一个操作消费。 这不串行化无关 token、数据 root 登记、草稿 preview 或任意 target 写入。
在 backtrader-skills 检出目录下激活目标环境后运行这些命令。仓库维护者使用仓库
AGENTS.md 要求的 Anaconda base 环境，但该机器特定的可执行路径不属于公开命令。
源码检出辅助脚本只会在两种布局下自动定位 Backtrader 仓库：本产品位于该仓库内，或本产品
与名为 backtrader 的仓库目录同级。它会校验目标根目录含有
backtrader/version.py，且 Git remote 为 cloudQuant/backtrader；两种布局都不满足时返回
结构化 SOURCE_CHECKOUT_NOT_FOUND 错误，候选目录属于其他 fork 时返回
BACKTRADER_SOURCE_MISMATCH，不会猜测其他目录或继续使用不兼容实现。
修改被分发的文件后，先用 python scripts/build_manifest.py 重建一次受跟踪的清单。
日常只读验证使用 python scripts/build_manifest.py --check；--check 和
--help 都不会修改 manifest.json。
# 产品嵌套在 Backtrader 仓库中，或两个仓库同级时自动发现
python scripts/doctor.py
python scripts/build_manifest.py --check
python scripts/build_catalog.py --check
python -m pytest tests -q
python scripts/run_acceptance.py \
--matrix all --require-no-mcp --require-no-agent
# Backtrader 位于其他位置时显式指定其仓库根目录
python scripts/doctor.py --target /path/to/backtrader
python scripts/run_acceptance.py --repository /path/to/backtrader \
--matrix all --require-no-mcp --require-no-agentdoctor、验收矩阵和执行测试都要通过 cloudQuant Backtrader 源码包来运行策略。测试套件也
支持用 BT_BACKTRADER_DIR 指向该 checkout 的 backtrader 包目录，并在缺少该包时自动跳过
这些测试。
持续集成在每次 push 到 master 时强制执行上述验证：专用 acceptance 任务检出 cloudQuant
Backtrader fork，在源码包存在的情况下运行完整测试套件（包含执行、doctor 和 acceptance 测试），
要求 src/backtrader_skills 覆盖率至少 80%（不含两个 python -I 子进程模块
child_runner.py 和 isolate_entry.py，coverage 配置已将其排除），随后运行完整 7×2 验收矩阵。
Pull request 保留较轻量的 quality 与受支持 Python 任务。模拟各宿主客户端二进制和人工批准的 master/dev 财务基线仍属
于本地发布步骤。
验收命令构建 wheel，安装到隔离目录，只把 Backtrader 源码包暴露给一个干净 fixture 仓
库，并从该已安装分发运行完整 7×2 矩阵，源码检出不在 sys.path 上。七个 archetype 使
用七个不同的 DatasetManifest，覆盖全部六个已声明 adapter；多数据、resample 和
precomputed 自定义 line 语义按单元记录。每个单元记录独立的 runonce 和 runnext 哈希加
比较结果。多数据、多 timeframe 和 precomputed/ML 代表单元还须通过结构化的
失败 -> typed-IR 修复 -> 重新校验 -> 已批准双模式回测门禁。
wheel 含七个具名 JSON Schema、comparison-profile-v1.json、完整元数据快照、四个宿主
adapter manifest 和全部三个规范 skill。manifest.json 记录每个已发布文件哈希和兼容
范围。