Parallel agent work produces a strange kind of confidence. Every lane can be clean in isolation while the system as a whole has moved. One worktree contains the live performance fix. Another contains the new page architecture. A third has the generated files. Testing any one of them proves only that one story is coherent.
§1What failed
The assumption was simple: the nearest clean checkout was the source. It built, linked, and rendered correctly. But production contained newer material from another lane. A release from the clean tree would have deleted live content while passing every local check.
§2What broke the confidence
- The live homepage contained an entry missing from the candidate tree.
- Utility routes existed locally but returned 404 in production.
- Multiple branches changed the same asset versions and page shell.
- The active deploy could not be explained by the candidate branch alone.
§3The protocol now
- Observe the live result and record hashes, routes, and asset versions.
- Identify the active deploy artifact.
- Match the artifact to a revision.
- Inspect every worktree touching the same surfaces.
- Create one clean integration lane.
- Re-run the full build and browser checks there.
If this clean tree shipped now, which live facts would disappear?
§4The guardrail now
Tests follow orientation. A green suite is accepted only after the tested tree is matched to the live artifact and the concurrent-work inventory is recorded. Generated files and shared cache versions are treated as release state, not branch-local trivia.
Steal this
- Test ownership before correctness.
- Compare live content against the candidate source.
- Inventory concurrent work before editing shared surfaces.
- Integrate first; weaken no test to make a stale tree pass.
- Download the live-truth checklist.