Higher-effort items from the upstream triage (ao-ops/docs/UPSTREAM-CHERRY-PICK-TRIAGE.md), deferred from the Tier-1/Tier-2 folding (#1) because they need real implementation work, not a clean cherry-pick.
Security (Go-only upstream → re-implement in TS)
Dependency upgrade (test carefully)
Read-path rearchitecture (the 'platform-down' root cause; upstream's wholesale fix never merged)
Pre-existing test gap noticed during folding
🤖 Generated with Claude Code
Higher-effort items from the upstream triage (
ao-ops/docs/UPSTREAM-CHERRY-PICK-TRIAGE.md), deferred from the Tier-1/Tier-2 folding (#1) because they need real implementation work, not a clean cherry-pick.Security (Go-only upstream → re-implement in TS)
bypassPermissions; read-only is prompt-only. Re-implement scoped AllowedTools/DisallowedTools through the claude-code plugin + reviewer path in core. Highest-value security item for an autonomous orchestrator.Dependency upgrade (test carefully)
Read-path rearchitecture (the 'platform-down' root cause; upstream's wholesale fix never merged)
ghcalls —ghis the dominant 20-40s bottleneck (benchmarked), not tmux/ps; #1858 is scoped to the wrong probes AgentWrapper/agent-orchestrator#1885 dashboard read path makes per-sessionghcalls (20-40s bottleneck)ao startPre-existing test gap noticed during folding
readLastJsonlEntry returns a recognized message typefails on hosts with current Claude JSONL — the recognized-message-type set is stale. Low-pri, env-dependent.🤖 Generated with Claude Code