Escalation after 5× REJECTED may lose accumulated 🟡/🟢 findings

[pedrofuentes]

Issue

When an agent hits 5 REJECTED Sentinel cycles and escalates to the user, the accumulated 🟡/🟢 findings from those cycles have not been filed as GitHub issues (filing is deferred to the final verdict report per v0.11.1).

Since escalation has no \inal verdict, these findings exist only in the Sentinel report text — they could be overlooked.

Current mitigation

Escalation hands off to a human who has access to all Sentinel reports and the PR diff. The findings aren't lost — they're in the reports. Filing issues automatically during escalation could create noise for a PR that might be abandoned.

When to revisit

If we observe a pattern of 🟡 findings getting lost or forgotten after escalation in a downstream project.

Proposed fix (if needed)

Add to the AGENTS.md escalation table row:

\Sentinel rejects 5× | Escalate to user with full finding list (🔴 + 🟡 + 🟢) — do not retry same approach\

Or add to the REJECTED After Sentinel row:

\If escalating, file all accumulated 🟡/🟢 as issues before handoff.\

Source

Multi-model panel review (Opus 4.5) during v0.11.1 development.

[pedrofuentes]

Triage review (v0.15.0): still valid — remains sentinel:deferred.

Re-verified: AGENTS.md §After Sentinel still defers filing 🟡/🟢 to the "final verdict report," and the 5× escalation path has no final verdict — so the concern stands in principle. No finding-loss-after-escalation has been observed in practice.

Note: the v0.16.0 report-persistence work (#5) materially strengthens the existing mitigation. Once every Sentinel report is persisted to a durable location, the accumulated 🟡/🟢 findings survive an escalation inside the persisted reports rather than living only in ephemeral agent text — which is most of what this issue was worried about. That further reduces the need for a dedicated fix here.

Keeping open and deferred until observed in practice, per the repo's deferred-issue policy.