Problem
A session today operates against a single repository working directory. Real changes frequently span more than one repository: a service and its client, a shared package and its consumers, or a target repository plus a separate repository that holds reusable instructions, prompts, and skills.
The only workaround is to start a separate session per repository. Each session has its own context, its own history, and its own result, so the agent cannot reason across the set, validate a change end to end, or produce a coordinated outcome.
What is missing
- No way to declare more than one repository when creating a session.
- No per-repository permission model, so a caller cannot grant read-only access to one repository and write access to another.
- No way to produce linked pull requests, or a combined status and rollback plan, across repositories.
Proposed behavior
- Session creation accepts a list of repositories or workspace roots rather than a single working directory.
- Each entry carries an explicit access level (for example
read or write).
- The agent can read, edit, and run validation across all declared roots within one session and one shared context.
- The session can open linked pull requests and report a combined result, identifying which repositories changed and which were read only.
- Repositories that fail to resolve are reported explicitly rather than silently omitted.
Example scenario
A caller asks the agent to trace a failing API call from a consuming application into the library that implements it, fix the library, update the consumer, and run both test suites. This requires both repositories present in one session.
Acceptance criteria
- A session can be created with two or more repositories.
- Access level is configurable per repository and enforced.
- Shared context is retained across all roots for the lifetime of the session.
- The result identifies per-repository changes and any linked pull requests.
Problem
A session today operates against a single repository working directory. Real changes frequently span more than one repository: a service and its client, a shared package and its consumers, or a target repository plus a separate repository that holds reusable instructions, prompts, and skills.
The only workaround is to start a separate session per repository. Each session has its own context, its own history, and its own result, so the agent cannot reason across the set, validate a change end to end, or produce a coordinated outcome.
What is missing
Proposed behavior
readorwrite).Example scenario
A caller asks the agent to trace a failing API call from a consuming application into the library that implements it, fix the library, update the consumer, and run both test suites. This requires both repositories present in one session.
Acceptance criteria