(bug) Pi loses earlier conversation turns after local resume

[mmabrouk]

When I continue a local Pi conversation after the execution environment has gone cold, I expect Pi to remember the earlier turns. Instead, continuity can report that the native session loaded while Pi sees only the newest follow-up message.

Steps to reproduce

1. Configure an agent that uses the local pi_core harness and an author skill.
2. Start a conversation and send: Remember marker ALPHA-7.
3. Let the warm environment expire or fully recreate the runner so the next turn must cold-start.
4. Continue the same Agenta conversation and ask: What marker did I ask you to remember?
5. Inspect the continuity logs and the Pi transcript and skill paths.

Expected

Pi resumes the same native conversation after the cold load, answers ALPHA-7, and can still read the configured skill from a stable session-workspace path.

Actual

The load path can report success while Pi is attached to a blank replacement session and cannot see the earlier turn. Skills can also point into a deleted per-run agent directory.

Observed logs:

[continuity] ... loaded=true

Environment

• Local EE deployment
• Harness: pi_core
• Pi ACP adapter: pi-acp@0.0.29
• Observed: 2026-07-13

Evidence

• Agenta conversation: 7e9ad207-17bd-4555-ae40-9bf4f485a8f2
• Recorded Pi session: 019f5b55-9a07-7337-8209-18b93fb1a167
• Replacement Pi session: 019f5b56-306e-7cb9-b969-6ce56b3d9a8e
• Local reproduction report: debug/session-7e9ad207-bug-report.md

Delivery

• Persist Pi native transcripts for the Agenta conversation lifetime without persisting credentials.
• Materialize Pi skills at a stable, explicit cwd path without destructive reconciliation.

Launch non-goal

Do not change warm versus cold selection, native session/load, token usage, prompt caching, or canonical replay behavior in these fixes. Transcript verification and replay fallback remain separate follow-up work.

[linear-code]

AGE-3949

[ashrafchowdury]

I think this is fixed, please reopen if not fixed