Add agent feedback behavior evals (#98887)
### What?
Adds eval coverage for the experimental agent feedback workflow:
- Routine debugging should not produce a report
- Qualifying friction should produce an anonymized structured report
- Distinct issues should produce separate review forms
Also adds repeat-run and variant controls so trigger frequency can be
measured across multiple runs.
Local-Skill and agent-feedback/privacy evals are marked `publish:
false`, with scoped `evals/AGENTS.md` instructions preventing their
fixtures, transcripts, or scores from being exported to the public
benchmark.
### Why?
We need to measure trigger precision, anonymization, issue splitting,
and duplicate prevention before expanding the experiment.
### How?
Depends on #98582.
The `agent-feedback` treatment uses the managed block and bundled
reporting protocol from the parent PR. Only the remote rollout gate is
forced on inside the eval sandbox so runs are deterministic.
| Eval | Baseline | Agent rules | Agent feedback | 10-run treatment |
| --- | --- | --- | --- | --- |
| Routine debugging | Pending | Pending | 1/1 passed | Pending |
| Anonymization | Pending | Pending | 2/2 reporting checks passed |
Pending |
| Distinct issues | Pending | Pending | 1/1 passed | Pending |
The first attempted run did not reach the agent because the local
sandbox was not linked to a Vercel project. It is infrastructure setup
and is not included in the results above.
The first distinct-issues run produced the expected two separate
payloads. Its scorer rejected them because the parser did not allow the
existing `token` query parameter and the browser criterion required an
open attempt even when the agent environment exposed no browser
capability. After correcting those assertions and clarifying the
stopping-point wording, the scored rerun passed.
Both anonymization treatments produced one valid payload with none of
the seeded customer, project, route, local-path, internal-URL, or secret
values. The corrected rerun also selected `misleading-error` and passed
every reporting assertion. An inherited `.next`-preservation assertion
was removed from this fixture because it measures `next-dev-loop`
behavior, not agent-feedback anonymization; Skill queue coordination
belongs in a separate focused eval.
