ForgeLoopRecorded run · October 9, 2026
Watch an issue become a merged PR.
A real run on a public repository, replayed step by step from its recorded evidence. No staged screenshots, no invented numbers.
Replay: Cover non-string file search queries with regression tests
Intake
A labeled GitHub issue opens a run
Issue #10 asked for regression tests proving that null, undefined, numeric, and object queries return an empty array. It restricted changes to one test file and named the exact commands that must pass.
webhook accepted · repository policy revision captured · budget USD 5Plan
The planner produces a bounded task graph
The request becomes validated tasks with owned paths and acceptance criteria. Only tests/search.test.js is writable; runtime code, dependencies, and workflows are out of scope.
owned path: tests/search.test.js · criteria stored server-sideExecute
A runner claims the lease and edits an isolated worktree
The enrolled runner checks out the repository on its own machine, calls the model with its locally stored provider key, and commits the change in a dedicated Git worktree.
lease acknowledged · heartbeat streaming · commit recordedVerify
The repository gate runs in a pinned container
The required unit gate ran Node tests and the production build in a digest-pinned Node 22 image. An agent saying “done” does not count; only the gate result does.
unit gate PASSED · every acceptance criterion coveredReview
An independent review checks the integrated diff
A separate review stage receives the bounded, redacted diff against the same remote-tracking base used for execution. It assesses each stored criterion exactly once.
review approved · criteria identities match the server listDeliver
Policy allowed auto-merge, so the verified commit merged
Human approval was disabled for this repository and auto-merge was enabled. ForgeLoop opened PR #11 as its GitHub App, merged the verified commit, and closed the source issue.
PR #11 merged · approved_by: null · issue #10 closed
$ npm ci && node --test tests/search.test.js && npm run build
✓ unit gate PASSED
merged fbae689 · October 9, 2026 at 7:09 PM UTC
What this demo does not claim
The cost figure is recorded provider telemetry, not a verified total; some telemetry was unavailable. Two earlier attempts on the same repository blocked safely and are documented as debugging history rather than hidden. Unattended merge is opt-in: by default, a human approves before anything reaches GitHub.
Inspect the evidence yourself.
The issue, the bot-authored pull request, and the merge commit are public. The validation record in the source repository covers the blocked attempts and the fixes that followed.