Automattic / Automattic/wp-codebox

Bridge approved parent tools for imported sandbox agents

Open
#1,783 0 comments 0 reactions 0 assignees View on GitHub
Dominant language
TypeScript
Stars
16
Forks
4
Avg merge
59m
Merged PRs (30d)
131

Description

## Problem

Imported native agents execute in WordPress Playground but receive no workspace/Git/GitHub tools. The reusable workflow correctly classifies those tools as `parent` / `control_plane`; `sandbox-tool-policy` therefore excludes them and `tool_contract.tools` is empty.

Evidence: https://github.com/Automattic/build-with-wordpress/actions/runs/29311239264

The technical bootstrap agent runs but cannot inspect or modify the target checkout and returns no-op.

## Required outcome

Implement the generic parent-tool bridge already modeled by `wp-codebox/parent-tool-bridge/v1`:

- Expose only approved parent tool schemas to the model as bridged client tools while retaining parent/control-plane execution classification.
- Pause the Agents API loop on a bridged call, return a bound pending request to the Node host, execute it through the policy-aware runner-workspace executor, submit the result, and resume the same runtime conversation until terminal.
- Bind request/result to run, session, tool call, allowed tool and allowed repository; enforce idempotency, turn/tool/time budgets, approval and redaction.
- Never expose host paths, tokens, shell access, or arbitrary command execution to the sandbox.
- Support workspace read/search/write/edit/patch and the declared Git/GitHub operations through existing typed executors.
- Add deterministic imported-package E2E: model reads and edits a temp workspace through the bridge, verification passes, patch captured; spoof/denied/malformed/replay cases fail.
- No Docs Agent-specific production code.
- Release and rerun both maintenance and bootstrap.

Contributor guide

No contributing guide indexed for this repository

Research direction

Start with the modeled wp-codebox/parent-tool-bridge/v1 contract, sandbox-tool-policy, the Agents API loop, and the existing typed executors. Verify the imported-package E2E covers workspace read/edit, patch capture, spoofing, denial, malformed requests, and replay; done also requires release and reruns of maintenance and bootstrap.

Written by the indexing model from the issue text.

Assessment

Tech stack
git, github, typescript, wordpress
Domain
backend-api-design, security, testing
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Quiet
Clarity
Mostly clear
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.