Allow the Editor Agent to retrieve remote web page contents via cURL

[manuelkiessling]

[JOORVIS]

ProductBuilder Planning Agent — Implementation Plan

Applying rules: 00-rule-acknowledgment.mdc, 07-workflow.mdc, 02-code-standards.mdc

I’ll quickly scan the codebase for the Editor Agent implementation and any existing remote-fetch tooling so the plan is concrete and scoped correctly.I found the likely backend path (src/LlmContentEditor/...) and I’m now reading the agent + tests to anchor the plan to existing abstractions and tool-wiring points.I confirmed the extension point: the Editor Agent gets custom tools from ContentEditorAgent, which delegate to WorkspaceToolingFacade. Next I’ll check existing docs/plans for expected behavior and security constraints around remote fetch.I’ve mapped the main integration points and I’m now checking dependency wiring and existing tests for WorkspaceToolingFacade so the plan includes exact files/methods to extend.I’ve got enough context for a concrete plan. I’ll include the exact extension points (WorkspaceTooling*, ContentEditorAgent, template instructions, progress i18n, and tests), plus assumptions and risk controls for URL fetching.Applying rules: 00-rule-acknowledgment.mdc, 07-workflow.mdc, 02-code-standards.mdc

Implementation Plan

Scope and assumptions

• Add a new Editor Agent tool that fetches remote web page content using the curl CLI (not Symfony HttpClient).
• Keep behavior safe and predictable: http/https only, request timeout, bounded response size, and non-throwing error responses.
• Return tool output as JSON so the agent can reliably parse status and content.

1) Extend workspace tooling contract

• Update WorkspaceToolingServiceInterface with a method like fetchRemoteWebPage(string $url): string.
• Define return contract as JSON string:
  • success: { url, finalUrl, statusCode, contentType, content, truncated }
  • error: { error, url }

2) Implement cURL-backed fetch in tooling facade

• Implement method in WorkspaceToolingFacade:
  • Validate URL (parse_url, require scheme+host, allow only http/https).
  • Resolve execution path from current workspace context (fail safely if missing).
  • Execute curl through existing shell abstraction so it runs in isolated agent container.
  • Use conservative flags (-L, --max-time, --connect-timeout, -sS) and include metadata (http_code, content_type, effective URL).
  • Truncate large bodies to a max size (to protect LLM context and payload size), set truncated=true when cut.
  • Never throw; return structured error JSON on command/network/validation failure.

3) Register the tool in the Editor Agent

• In ContentEditorAgent::tools(), add a Tool::make(...) entry (e.g. fetch_remote_web_page):
  • Description clarifies it retrieves remote page content via cURL.
  • Required url property.
  • Callable delegates to WorkspaceToolingServiceInterface::fetchRemoteWebPage.

4) Teach the agent when to use it

• Update default prompt template in AgentConfigTemplate:
  • Add a short “REMOTE WEB PAGES” section explaining when to call the new tool (e.g., when user asks to inspect/copy/adapt content from an external page).
• Add/adjust related template unit assertions in AgentConfigTemplateTest.

5) Add progress message support

• Map new tool name in ProgressMessageResolver (e.g. fetching_remote_web_page).
• Add translation keys in:
  • translations/progress.en.yaml
  • translations/progress.de.yaml
• Extend ProgressMessageResolverTest with one case for new tool-call mapping.

6) Add/extend unit tests for facade behavior

• In WorkspaceToolingFacadeTest, add tests for:
  • valid URL success path (mock shell command output and assert JSON fields),
  • invalid URL/scheme (returns error JSON, no shell call),
  • curl failure path (returns error JSON),
  • response truncation behavior (large body => truncated flag).
• Add one ContentEditorAgentTest assertion that new tool is present in the tool list (reflection-based, consistent with current test style).

7) Verification plan

• Run project checks via mandated workflow:
  • mise run quality
  • mise run tests
• If test runtime is large, start with affected unit suite first, then full mise run tests.

8) Acceptance criteria

• Editor Agent can call fetch_remote_web_page and receive remote page content without using local file tools.
• Tool output is deterministic JSON and robust on failures.
• Progress UI shows meaningful “fetching remote page” feedback.
• Unit tests cover success, validation, failure, and truncation paths.

---

Review this plan. Add comments to refine it, or add the prdb:plan-approved label to start implementation.

[JOORVIS]

ProductBuilder Implementation Agent — PR Created

The implementation has been completed and a pull request has been created:

#141