Agent Bridge — Supervised Maintenance

Inspiration

An AI-generated patch is only the start of maintenance. A maintainer still needs to reproduce the report, inspect the diff, check tests at the exact commit, and decide whether to merge. Agent Bridge makes that chain visible for one real output-pagination defect in an existing MIT project.

What it does

The public workbench replays a UTF-16 boundary bug and compares it with the genuine Duo-fixed production core. In the historical implementation, paging A🧭B at a two-unit limit splits the emoji across malformed pages. The fixed core returns well-formed A🧭 and B pages while retaining the UTF-16 offset API. Visitors can enter another public sample and download a focused issue brief; samples are processed locally in their browser.

The supervised workflow connects issue #2 to GitLab Duo Developer Flow session 9666771, a narrow production fix, unchanged regression coverage, exact-commit CI, and a Draft merge request. Merge and deployment remain human decisions. The workbench links the authentic issue, session, MR, and pipeline records. Its local evidence checklist cannot approve or merge code.

How we built it

We extracted the public session/output core from the original MIT Agent Bridge project into a separate 40-file contest edition. We added synthetic fixtures, Unicode regression cases, a read-only evidence collector, the browser workbench, GitLab CI, foundational Duo instructions, and a manual GitLab Pages job.

Issue #2 provided a reproduction, acceptance criteria, an isolated source branch, and a human review stop. Developer Flow session 9666771 inspected the source, created its fix branch, changed only src/render.ts, pushed the result, and opened MR !2 as Draft with auto-merge off. Its net production change was +14/−1 lines. The fix extends a page boundary by one UTF-16 unit when a surrogate pair would otherwise split, and rejects offsets inside a pair.

The historical baseline failed 13 regression cases. The genuine fix's hosted pipeline passed the original 51 tests at 010b79fbdb99cdbae0484e89bcebf8d1288ab406. After review and explicit user approval, the fix and audited foundation were manually merged and deployed. A later presentation change added one replay-versus-fixed test. Final main a7a5e3e3e592c9ce9aa01e15f9cc9fbf973bcfb7 passed 52 tests before the manual Pages deployment. Fresh October 26 checks of that unchanged source in a separate cloud checkout also passed all 52 tests.

Challenges we ran into

The Duo runner had Node 20 and an npm registry 403, so it could not finish the normal local build. Its temporary hand-stripped JavaScript proxy was weaker evidence, disclosed in the MR. The genuine Node 24 hosted pipeline supplied the exact-commit verification. The agent restored original line endings to keep the production diff narrow.

We distinguish local checks from hosted execution. A local report does not establish a Duo run; CI at an older commit does not validate a changed HEAD; and the browser checklist is not an enforced GitLab approval policy. Approvals are optional in this provisioned project. Draft status and the instructed manual decision supplied the actual review stop.

Accomplishments that we're proud of

The project connects a reproducible defect to a genuine agent fix and inspectable verification. The public contest edition contains 40 files, preserves regression coverage, and demonstrates lossless, well-formed pagination in an actual running browser. The edition excludes unrelated private work, credentials, and real session data. No new paid service was added.

What we learned

Small reproductions and explicit outcome criteria make agent work easier to assess. Disclosing agent-environment limitations prevents plausible-looking local results from being overstated. Binding evidence to a commit makes the final human decision clearer.

What's next

Extend the approach to more session/output defects, improve the evidence collector, and add stronger approval policy where administrators authorize it. The current edition is a focused supervised maintenance workflow with a browser replay. New Duo execution requires an authorized GitLab account; anonymous visitors cannot start agents.

Built With

• gitlab
• typescript