Feature request: multi-repo coordinator workflow for cross-repo fixes

[danshapiro]

Summary

Low priority suggestion: consider a stock workflow for multi-repo incident/feature work that coordinates across repositories, then delegates mutation to per-repo workers.

This came up while fixing a real issue that spanned two repos:

• one repo owned a platform adapter and its reconnect behavior
• another repo owned the gateway recovery path that needed to react when the adapter receive task died

The existing coding-loop was useful for per-repo checks, but it was not the right shape for the broader question: "does this cross-repo incident fix hold together as a system?"

General Problem

Some tasks are not naturally single-worktree tasks. They involve a contract between two or more repositories:

• an adapter raises a specific exception, while a gateway has to notice and recover from it
• a plugin changes runtime behavior, while a host repo owns deployment/retry/status logic
• one repo owns client behavior, while another repo owns server behavior and tests

The stock coding-loop currently works best as a single-worktree loop:

1. choose one small subtask
2. implement it in that worktree
3. review the resulting diff
4. repeat until done

That is good for isolated repo work. For cross-repo incidents, it tends to narrow too early. A separate run per repo can miss integration-level questions unless a human coordinates the contract manually.

What Might Serve This Better

A useful stock workflow could be a multi-repo coordinator loop, not one giant mutating agent.

Rough shape:

1. Accept a spec plus multiple repo/worktree paths.
2. Run a coordinator stage that may read all listed repos and identify cross-repo contracts, risks, and expected behavior.
3. Produce per-repo task specs with explicit ownership boundaries.
4. Launch isolated per-repo implementation/review workers, each mutating only its own worktree.
5. Run a final coordinator review that checks the system-level contract across the resulting branches:
   • exception/status names line up
   • retry budgets match
   • logging and failure paths are observable
   • tests cover both sides of the integration
   • no worker solved its local piece by creating a conflicting behavior elsewhere

The coordinator could remain read-only except for writing run artifacts and subtask specs. Actual code changes would still happen in normal isolated worktrees with normal per-repo commits.

Why This Would Be Better

For this class of task, a coordinator workflow would likely catch issues that a per-repo coding-loop may not prioritize:

• mismatched assumptions between producer and consumer repos
• successful local fixes that do not compose into a working incident fix
• missing tests for the cross-repo handoff
• duplicated or conflicting recovery paths
• unclear ownership of retry/deadline/status behavior

It would also give the human operator a clearer artifact: one final cross-repo report explaining whether the overall incident/feature contract is satisfied, plus the per-repo diffs that implement it.

Non-Goals

• Not asking for one agent to freely edit multiple repos in one working tree.
• Not asking to replace the existing coding-loop.
• Not urgent; current per-repo workflows are workable with manual coordination.

Priority

Low. This is a workflow quality improvement suggestion for better support of cross-repo tasks, not a blocker.