Anything-as-API

Anything-as-API 是一个面向浏览器场景的最小闭环原型：先录制一次明确的网页操作场景，再从录制结果里提取业务 API，最终基于显式 skill_spec_source.json 生成一个可交付的原生 skill_package。当前主线交付形态是带 SKILL.md、scripts/、lib/ 和 assets/ 的 skill package。

当前已经打通这条最小链路：

capture：录制场景并输出结构化 session_bundle
extract：从 bundle 提取业务 API artifacts
spec：把 extract 产物编译成 generated_skill_spec.json
builder：基于 spec 生成原生 skill_package
verify：对阶段产物和生成的 skill package 做 smoke 验证

Current Status

当前已经完成的阶段：

Plan A：采集入口、事件模型、session_bundle 基础契约
Plan B：manifest、分级 network evidence、storage snapshot、action trace
Plan C：API 提取、endpoint catalog、OpenAPI 基线
Plan D：闭环验证与阶段归因报告
Plan E：generated_skill_spec.json 契约与 spec 编译
Plan F：原生 skill package 生成器
Plan G：生成脚本的 runtime/session loop 基线
Plan H：原生 skill package smoke verify

当前不是生产可用版本，而是一个已经可运行、可验证的最小技术闭环。generic_items_api 只是当前回归 fixture，用来验证生成框架的端到端能力，不是框架内置目标；真正的浏览器登录刷新和长期 cookie 持久化仍在后续阶段。

Repository Layout

capture/   Browser capture and session bundle generation
extract/   Bundle-to-API extraction and artifact generation
spec/      Extract artifacts to generated skill spec
builder/   Native skill package generation
verify/    End-to-end verification and stage attribution
tests/     Tests for capture, extract, spec, builder, and verify
docs/      Design notes and implementation plans
output/    Sample outputs generated during local verification

Prerequisites

Python 3.11+
Poetry
Chromium for Playwright

Setup

Install dependencies into a project-local virtualenv:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry install

Install the Playwright Chromium browser:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run playwright install chromium

Minimal Flow

One-Command Fixture Workflow

For local verification, run the full extract, refine, build, and verify chain with one command:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m workflow.cli \
  --bundle-dir tests/fixtures/session_bundles/plan_b_with_bodies \
  --output-dir output/workflow-sample \
  --refine-provider fake

Expected output includes:

output/workflow-sample/extract/
output/workflow-sample/refined/
output/workflow-sample/build/skill_package/
output/workflow-sample/verify/verification_report.json
output/workflow-sample/workflow_report.json

workflow_report.json records each stage status, exit code, and compact nested reports from refine, build smoke, and verification stages so failures are diagnosable without opening every output directory.

1. Capture a Scenario Bundle

The capture CLI records one browser entry URL and writes a structured session_bundle.

Example:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m capture.cli \
  --url https://example.com/login \
  --output-dir output/plan-b-sample \
  --manifest tests/fixtures/manifests/login_submit.json \
  --capture-mode guided \
  --headless

Expected output includes:

output/plan-b-sample/session_meta.json
output/plan-b-sample/network/requests.ndjson
output/plan-b-sample/network/responses.ndjson
output/plan-b-sample/browser/navigation_trace.json
output/plan-b-sample/browser/cookies_timeline.json
output/plan-b-sample/browser/storage_snapshots.json
output/plan-b-sample/user_actions/action_trace.json
output/plan-b-sample/labels/scenario_manifest.json

2. Extract API Artifacts

The extraction CLI consumes a captured bundle and emits API artifacts.

Example:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m extract.cli \
  --bundle-dir tests/fixtures/session_bundles/plan_b_minimal \
  --output-dir output/plan-c-sample

Expected output includes:

output/plan-c-sample/api_calls.json
output/plan-c-sample/api_catalog.json
output/plan-c-sample/openapi_spec.yaml
output/plan-c-sample/replay_context.json
output/plan-c-sample/skill_spec_source.json

If captured response bodies are available for the selected endpoint, extract also emits:

output/plan-c-sample/recorded_sample.json

3. Optionally Refine The Skill Spec Source

The refine CLI rewrites the baseline skill_spec_source.json through a provider and copies the required runtime assets into a builder-ready output directory. The default fake provider is deterministic and offline-only. The openai provider calls the OpenAI Responses API when OPENAI_API_KEY is set.

Refinement is contract-guarded: providers may improve names, summaries, descriptions, schemas, and parser hints, but they cannot change tool names, tool kinds, execution request templates, runtime asset paths, or build script filenames.

Example:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m spec.refine \
  --extract-dir tests/fixtures/extracted_scenarios/generic_items_api \
  --output-dir output/refined-skill-source \
  --provider fake

OpenAI-backed refinement:

OPENAI_API_KEY=... POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m spec.refine \
  --extract-dir output/plan-c-sample \
  --output-dir output/refined-skill-source \
  --provider openai \
  --model gpt-5.4

Expected output includes:

output/refined-skill-source/skill_spec_source.json
output/refined-skill-source/replay_context.json
output/refined-skill-source/recorded_sample.json
output/refined-skill-source/refine_report.json

4. Build And Verify A Native Skill Package

The builder CLI compiles an explicit skill_spec_source.json, assembles the package, and smoke-verifies it in one command. The Example directory below is a fixture that contains a spec source; it is not a framework special case.

Example:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m builder.cli \
  --extract-dir output/refined-skill-source \
  --output-dir output/native-skill-sample \
  --verify

Expected output includes:

output/native-skill-sample/generated_skill_spec.json
output/native-skill-sample/skill_package/SKILL.md
output/native-skill-sample/skill_package/scripts/
output/native-skill-sample/skill_package/lib/
output/native-skill-sample/skill_package/assets/
output/native-skill-sample/skill_smoke_report.json

SKILL.md includes skill loader frontmatter (name and description), and agents/openai.yaml includes UI-facing interface metadata so the package can be handed to another agent as a native skill artifact.

The smoke report should include:

{
  "status": "pass",
  "package_type": "native_skill",
  "checks": [
    "assets_complete",
    "scripts_importable",
    "recorded_tool_runs",
    "live_tool_auth_required",
    "refresh_session_runs",
    "live_tool_session_ready",
    "live_tool_request_context"
  ],
  "failures": []
}

The older two-step builder flow still works if a spec has already been generated:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m spec.cli \
  --extract-dir output/plan-c-sample \
  --output-dir output/spec-sample

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m builder.cli \
  --spec-file output/spec-sample/generated_skill_spec.json \
  --output-dir output/native-skill-sample \
  --verify

4. Verify the End-to-End Flow

The verification CLI checks whether the expected artifacts exist and reports which stage failed if the flow is incomplete.

Example:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m verify.cli \
  --bundle-dir tests/fixtures/session_bundles/plan_b_minimal \
  --artifact-dir output/plan-c-sample \
  --refined-artifact-dir output/refined-skill-source \
  --package-dir output/native-skill-sample \
  --output-dir output/plan-d-sample

Expected output:

output/plan-d-sample/verification_report.json

The current report shape is:

{
  "overall_status": "pass",
  "stages": {
    "capture": {"status": "pass"},
    "extraction": {"status": "pass"},
    "generation": {"status": "pass"}
  }
}

5. Smoke Test A Generated Skill Package

The skill smoke CLI validates native skill packages.

Example:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m verify.skill_smoke_cli \
  --package-dir output/native-skill-sample/skill_package \
  --output-dir output/plan-e-sample

Expected output:

output/plan-e-sample/skill_smoke_report.json

Test Commands

Run all current project tests:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m pytest tests/capture tests/extract tests/spec tests/builder tests/verify tests/workflows -v

Current expected result:

all tests pass

Run by area:

POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m pytest tests/capture -v
POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m pytest tests/extract -v
POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m pytest tests/spec -v
POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m pytest tests/builder -v
POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m pytest tests/verify -v
POETRY_VIRTUALENVS_IN_PROJECT=true poetry run python -m pytest tests/workflows -v

Example Inputs

Current bundled fixtures:

tests/fixtures/manifests/login_submit.json
tests/fixtures/session_bundles/plan_b_minimal/

Current generated sample outputs:

output/plan-b-sample/
output/plan-c-sample/
output/plan-d-sample/
output/native-skill-sample/

What Is Implemented vs Not Yet Implemented

Implemented now:

manifest-aware browser capture
graded request / response evidence capture
storage snapshots and action traces
bundle-to-API extraction baseline
endpoint catalog baseline
baseline OpenAPI generation
response-body-aware recorded sample extraction when body evidence is available
baseline input schema inference from captured request body and query values
optional fake-provider and OpenAI-provider skill spec refinement stage with builder-ready artifact output
refine_report.json success/failure reporting for spec refinement
refinement contract guards for runtime assets, request templates, tool identity, and generated script filenames
end-to-end verification with stage attribution
generated native skill_package output
native skill smoke verification
recorded sample and session-state runtime script baseline
one-command extract-dir -> skill_package -> smoke report builder flow
generated session guidance docs and refresh_session.py JSON instructions
generated live scripts can build authenticated HTTP request context from replay and session assets
generated live scripts can apply simple input overrides to captured request body or query values
generated live scripts support dry-run request inspection plus injected/default HTTP execution and baseline response parsing

Not implemented yet:

robust schema inference from multiple observations
dependency graph generation
stronger LLM-assisted field mapping and workflow generation across arbitrary sites
browser-driven auth refresh and long-lived session recovery
production-ready multi-site skill packaging
support for complex real-world multi-step authenticated apps

Design Docs

Project design and plan docs live in docs/:

Name		Name	Last commit message	Last commit date
Latest commit History 90 Commits
builder		builder
capture		capture
docs		docs
extract		extract
spec		spec
templates/skill_v1/lib		templates/skill_v1/lib
tests		tests
verify		verify
workflow		workflow
.gitignore		.gitignore
ARCHITECTURE.md		ARCHITECTURE.md
CONTRIBUTING.md		CONTRIBUTING.md
README.md		README.md
poetry.lock		poetry.lock
pyproject.toml		pyproject.toml

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Anything-as-API

Current Status

Repository Layout

Prerequisites

Setup

Minimal Flow

One-Command Fixture Workflow

1. Capture a Scenario Bundle

2. Extract API Artifacts

3. Optionally Refine The Skill Spec Source

4. Build And Verify A Native Skill Package

4. Verify the End-to-End Flow

5. Smoke Test A Generated Skill Package

Test Commands

Example Inputs

What Is Implemented vs Not Yet Implemented

Design Docs

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

Anything-as-API

Current Status

Repository Layout

Prerequisites

Setup

Minimal Flow

One-Command Fixture Workflow

1. Capture a Scenario Bundle

2. Extract API Artifacts

3. Optionally Refine The Skill Spec Source

4. Build And Verify A Native Skill Package

4. Verify the End-to-End Flow

5. Smoke Test A Generated Skill Package

Test Commands

Example Inputs

What Is Implemented vs Not Yet Implemented

Design Docs

About

Resources

Contributing

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages