Evidence
Gray run 3 (34693008936, publish=false): both upsert-entities.md pages (guides + byoc trees) were written by the agent and iterated through all four repair rounds, then failed the external gate with:
Protected content mismatch for inline_code: source .../upsert-entities.md: line 50, column 10, offset 2500, token "`autoID`" sha256 2364854702b6; ta[rget ...]
Every other stuck page failed for the runner-level sandbox reason (fixed by #601/#606 + the sandbox-mode/preflight change); these two are a distinct, content-level problem.
Working hypothesis
The draft alters the inline-code span around autoID (dropped backticks or changed span text), and repair sessions without in-sandbox validator access could not localize the exact span from the bounded violation text alone. The protected-span-aware terminology work (#596) already prevents deterministic repairs from causing this class; the residual case is agent-authored edits.
Direction (separate change, after the sandbox fix lands)
- Re-run these two pages with the sandbox fix in place — repair sessions regain validator access, which may resolve it outright.
- If it persists: enrich the violation payload for inline_code mismatches with the exact source span text and surrounding line so the repair prompt can pinpoint the edit, or add a deterministic restore pass for dropped inline-code formatting.
- Both pages remain recoverable through the per-file recovery artifact from the retained run.
Tracked as the last item of the external review on the agentic gray rollout.
Evidence
Gray run 3 (34693008936, publish=false): both
upsert-entities.mdpages (guides + byoc trees) were written by the agent and iterated through all four repair rounds, then failed the external gate with:Every other stuck page failed for the runner-level sandbox reason (fixed by #601/#606 + the sandbox-mode/preflight change); these two are a distinct, content-level problem.
Working hypothesis
The draft alters the inline-code span around
autoID(dropped backticks or changed span text), and repair sessions without in-sandbox validator access could not localize the exact span from the bounded violation text alone. The protected-span-aware terminology work (#596) already prevents deterministic repairs from causing this class; the residual case is agent-authored edits.Direction (separate change, after the sandbox fix lands)
Tracked as the last item of the external review on the agentic gray rollout.