Repository navigation
Summary: On current Claude Code versions the final "ready to merge" verdict was always rejected, even after every quality check had run, so a fully finished unattended run ended on a false blocker. The check that guards that verdict now recognises how current versions record the specialist agents' work.
Details: The guard looked for subagent records inside the main session transcript. Claude Code 2.1.280 writes them to separate files, so it saw none and refused every verdict: a live run with 27 agents, 9/9 gates and 226/226 tests ended blocked. It now accepts two more tamper-proof signals written by Claude Code itself. A verdict with no agent activity at all is still refused. Version 4.11.2.
Technical details: Evidence sources are now the legacy sidechain records, the harness-written tool result carrying the subagent's id, and the per-session subagent transcript folder; all fail open and none can be written by the model. Replayed against the live run's real transcript: old handler refuses, new handler allows. Finding recorded in docs/findings. No screens, endpoints, schema or deployment impact.
Update an installed project: claude plugin update claude-code-sdlc@claude-code-sdlc --scope project