[NarryG]

ELI5

After a restart, Orca may not know whether OMP received your message. It can now check that chat's saved conversation for reliable evidence, without sending the message again. If the conversation cannot reopen, Orca warns that earlier context was lost instead of silently treating an uncertain chat as unused.

What Changed

• Connect OMP's saved user messages to existing restart reconciliation, using the execution host's sealed account and workspace.
• Require a uniquely owned conversation and complete active-branch history. Old branches, incomplete text, partial images and already-claimed copies cannot confirm delivery.
• Preserve uncertainty for pending, unknown or accepted nonqueued sends without a saved turn. Queued/rejected sends keep existing empty-chat behavior.
• Reuse the existing restore warning; add boundary regressions and documentation. No new recovery engine, controls or automatic replay.

Why

Saved conversation evidence is safer than guessing from terminal output or retrying a message that might already have arrived. Missing evidence never establishes non-delivery. This is separate from waiting for terminal writers to exit.

Linked Issue

Fixes stablyai#30542.

Visual Proof

Windows: before at 51530bb, after at 7d6bfef. A controlled interrupted send remains uncertain and is not replayed; a failed reopen shows the existing lost-context warning.

Before

After

Testing

• Agent-driven check in the rebuilt Windows app: uncertain-send preservation, failed reopen, warning and no automatic resend.
• Behavior regressions retained for ownership, branches, duplicate claims, timestamps, malformed/bounded history and queued/rejected boundaries.
• Windows pnpm typecheck, pnpm build and changed-code quality passed.
• History focus: 65 passed, 2 skipped. Restore focus: 21 failed on SQLite fixture cleanup (EBUSY); no assertions or application behavior were changed to hide them.
• Linux full lint, typecheck and all four pnpm test shards passed, alongside untouched upstream: run.

AI Disclosure

OpenAI Codex GPT-6.1 Sol integrated this candidate. Earlier independent reviews used Codex and Claude Fable 5.1. Agent-driven checks are not human testing.

Review

Independent source review found no blocker. Windows/macOS/Linux paths, folder workspaces, execution-host ownership, supported agents/integrations, Git providers and backwards compatibility were checked. Only OMP gains a history reader; no new wire protocol or UI control. Reads reject symlinks, malformed UTF-8 and ambiguous ownership, and are limited to 16 MiB/100,000 records. Windows app verified; macOS, real SSH/WSL and mobile-device checks were not run.

Agent skill upstream boundary

• Not applicable; no upstream skill-installer material copied.

Notes

No dependency or release-version changes.

Author

GitHub: @NarryG. X: @NarryOnTheNet.

Checklist

• Focused change, mechanism, alternative and issue explained.
• Genuine before/after attached with their source revisions.
• Correctness, security, performance and compatibility considered.
• Required contribution checks completed; Windows-specific cleanup failures disclosed above.