Loveca db02b77837
feat(plugins): add evidence-gated skill self-evolution example (#7984)
* feat(plugins): add evidence-gated skill self-evolution example

Demonstrates the #7884 roadmap deliverable: a proposed change, the regression
case it fixes, the old and new results, and a rollback to a known-good version,
driven only through the public SkillMutationService contract merged in #5645.

The example stages a candidate against a sealed completed-run evidence anchor,
evaluates recorded JSON regression cases against both frozen bundles with a
deterministic static evaluator (never executing package content), and commits
only a "justified" verdict with an AssessmentRef binding the report to the
exact candidate hash and base revision. Unjustified or failed proposals are
discarded; revert stays append-only. Ships an HTTP walkthrough surface, an
offline lifecycle test against a contract-faithful fake host, and a README
section proposing a concrete evidence standard for the automatic-publication
discussion.

Co-Authored-By: Claude Code <noreply@anthropic.com>

* fix(skill-evolution): resolve request models and honor APPLIED publication

The walkthrough routes had never been exercised through FastAPI. With
`from __future__ import annotations`, request models defined inside
`build_router` are unresolvable forward references, so FastAPI read `body`
as a query parameter and every POST answered 422. Move the models to module
scope and cover the routes through TestClient.

A committed operation reports `publication == "APPLIED"`; `READY` is a
`views` state. The committed check listed READY, so a successful apply never
flipped `record.committed` and the ALREADY_COMMITTED guard never fired. Fix
the check, mirror the real literals in the lifecycle fake host, and assert
both the flag and the second-apply refusal.

Co-Authored-By: Claude Code <noreply@anthropic.com>

* fix(skill-evolution): address review feedback and register the example catalog

- Bind walkthrough routes to the authenticated principal: owner checks on propose/apply/revert, owner-scoped listing, and non-disclosing 404s.
- Map unknown skills, proposals and operations to 404 for both NOT_FOUND and the real host's NOT_FOUND_OR_FORBIDDEN; map broken case files to 400.
- Ignore URL paths and trim sentence punctuation when extracting in-package references.
- Validate config.enabled's type instead of treating any non-False value as on.
- Soften the README claim about report persistence and add CHANGELOG entries.
- Register the example in the frontend extension catalog, fixing frontend-unit-tests.

Co-Authored-By: Claude Code <noreply@anthropic.com>

* fix(skill-evolution): narrow case-file error mapping and cover run visibility

- Label ValueError as CASE_INVALID only on POST /proposals, the route that actually loads case files; apply and revert now surface host bugs as 500 instead of pointing at the case directory.
- Cover the request-scoped run-evidence reader the real Gateway installs: hidden anchor returns 403 SOURCE_NOT_VISIBLE before the host is touched, a denied resolver returns 403, and a visible anchor still proceeds.
- Add regressions pinning host ValueError on apply/revert to 500.

Co-Authored-By: Claude Code <noreply@anthropic.com>

* docs(changelog): drop contributor changelog entry per policy

Contributor PRs no longer modify CHANGELOG.md or CHANGELOG_zh.md (#8202);
maintainers add the entry at merge time.

---------

Co-authored-by: Claude Code <noreply@anthropic.com>
Co-authored-by: Nan Gao <thunderggnn@gmail.com>
2026-11-16 16:04:42 +08:00
..

Develop an extension from an existing example

Start with the smallest example that covers your extension. Each directory is an independent Python package using the public deerflow-extension-api contract.

Start here What it demonstrates
Extension example Middleware, task lifecycle, model observations, services, and HTTP routes. Start here for backend extensions.
Conversation bookmarks A browser page, conversation action, backend actions, packaged resources, and a model tool. Start here for full-stack plugins.
Agent teams Full Agent orchestration. Use this after the basic contribution and lifecycle patterns are familiar.

Additional examples cover Office file previews, text-list classification, context pruning, advisory screening of fetched content, sectioned long-form report writing, and evidence-gated skill self-evolution.

Describe the extension

Package deerflow-extension.json beside the entrypoint's Python module (beside __init__.py for a package). Capability Center renders this public description without requiring a browser module. Use the full-stack plugin namespace as id, or a unique stable ID for a backend-only extension. A v1 document requires an ID, at least one locale, and a nonempty title, description (purpose) and usage list in every supplied locale. requirements and docs_url are optional; omit requirements when none apply. Keep secrets and deployment configuration out of this file. Names, purposes and instructions are authored by the package; loaded/enabled state and registered tools come from the host.

{
  "schema_version": 1,
  "id": "company.helper",
  "locales": {
    "en-US": {
      "title": "Task statistics",
      "description": "Collect tool-call statistics during tasks.",
      "usage": [
        {"title": "Run a task", "description": "Statistics are collected automatically after activation."}
      ]
    }
  }
}

The packaging checker validates this file when present and detects omitted wheel resources. Older packages without it remain compatible. Invalid optional copy produces a startup diagnostic while the working extension remains available. Gateway reads copy at startup only, after successful registration. Disabled and failed packages are not imported for documentation; a catalog entry alone does not prove activation. A documented backend extension with actual registrations appears in discovery even if it supplies no full-stack plugin.

For bundled examples, update their package files and run python scripts/sync_extension_documentation.py from the repository root. The generated frontend snapshot supplies descriptions before installation; installed package copy takes precedence. Backend tests check that the snapshot is current.

Edit, test, check, then install

  1. Copy the selected example outside the DeerFlow checkout. Rename its Python package, distribution and deerflow.extensions entry point. For a full-stack plugin, also choose its own namespace and browser module name.

  2. Follow the example's README to configure its dependencies and run its tests. Keep package behavior tests alongside the extension. For a bundled example with a tests/ directory, CI installs the package with its dev extra next to this checkout's deerflow-extension-api, then runs pytest, ruff check and ruff format --check.

  3. From the DeerFlow checkout root, prepare the backend environment with make -C backend install, then check the package:

    uv run --project backend --no-sync python scripts/check_extension.py /absolute/path/to/your-extension
    

    To try the checker on an unchanged example:

    uv run --project backend --no-sync python scripts/check_extension.py examples/deerflow-extension-bookmarks
    
  4. Follow that example's deployment and try-it instructions. Restart Gateway, refresh the browser when applicable, and exercise the installed feature. Use make extension-upgrade SOURCE=/absolute/path/to/your-extension to apply later source edits: local installs are snapshots.

What the checker covers

The script builds a wheel in a temporary directory, checks exactly one extension entry point and its packaged Python module, parses packaged Python syntax, and checks that the API dependency range accepts this checkout's contract version. It validates conventional ui_manifest.json resources with the same validator used by Gateway, including detecting manifests left out of the wheel in flat or src/ layouts. A custom manifest filename needs an explicit wheel-relative path:

uv run --project backend --no-sync python scripts/check_extension.py /absolute/path/to/your-extension --manifest your_package/resources.json

The build executes the package's build backend, so check trusted sources. The checker does not import the extension, call install(), install the extension, or change deployment configuration. Registration/configuration behavior and the @extension version marker are verified when loading the plugin; continue to run the package tests and the installed feature after packaging checks pass.