Triage Agent
ready
Run queued
Bug report normalized and waiting for a local workspace.
Bug reports into verified patches.
A local Codex repair loop that reproduces a defect, writes the regression test, patches the source, verifies the suite, and presents a reviewable diff.
Report
->Reproduce
->Patch
->Verify
->PR
Triage Agent
Test Agent
Patch Agent
Runner
Review Agent
demo-repo
fix/roi-zero-equity
idle
not-started
Reproduced
pending
Patched
pending
Verified
pending
Ready to Review
pending
Bug Report
Select a seeded failure and stage the repair prompt.
live
Selected target
ROI calculator crashes when equity is zero.
Agent Timeline
Triage, test, patch, run, and review are separated into visible handoffs.
live path
Test Agent
next
Regression path armed
The repair agent will add a failing test before patching source.
Runner
next
Verification pending
Final Vitest output and diff will replace the preview after the API returns.
Test Output
Regression first, full suite after the patch.
queued
Before patch
1 failed
->
After patch
12 passed
$ pnpm test RUN v2.1.9 .workspaces/preview waiting for repair run...
Diff Viewer
Reviewable unified diff from the isolated workspace.
0 files
unifiedreview ready
01diff preview
02
03Select a seeded bug and start a repair to collect the real unified diff.
PR Summary
Judge-ready explanation of the verified patch.
ready
Repair run waiting
Start a repair to generate a Codex-authored PR title, summary, changed files, test command, and risk level.
0 files changed
pending
medium risk
Awaiting verified patch
Run the repair loop to unlock the merge-ready state.
Evidence Mode
Judge-facing proof that the agent reproduced, patched, verified, and packaged the run.
golden fixture
captured
Report
captured
Reproduce
captured
Patch
captured
Verify
captured
PR
Bug report
ROI calculator crashes when equity is zero.
When a user enters equity = 0, the ROI calculator returns Infinity. Expected behavior: return 0 and keep the report stable.
Runtime
25s
Changed files
2
Retry count
0
Evals Dashboard
Local signals that make the coding-agent run measurable for Loops House and judges.
eval-ready
Repair success
100%
3/3 golden runs passed
Avg time to patch
25s
end-to-end evidence runtime
Bug coverage
3/3
seeded bugs have golden evidence
Selected diff
25 lines
low risk
Judging signal map
Agentic Coding
strong
Codex performs the test-first repair loop inside an isolated workspace.
Building Evals
strong
Runs expose pass/fail, runtime, diff size, changed files, retry count, and risk.
UX
strong
Timeline, evidence, diff, tests, and PR summary are visible in one local cockpit.
Selected run evaluation
Fix ROI zero-equity handling
Testspnpm test (passed)
Files changed2
Diff size25 lines
Risklow