[idea] preserve agent permissions, model, and compute settings on restart

[lsisoft]

idea / problem

Herdr correctly restores supported agent conversations after a server restart, but it does not preserve the complete operating profile of the agent session. The resumed agent can have different permissions, model selection, or compute/reasoning settings from the session that was running before the restart.

For example, a Codex pane started with --yolo, an explicit model, and a reasoning effort can resume with only codex resume <session-id>. This can reintroduce approval prompts or silently change model behavior. The same problem applies to the equivalent documented settings in Claude Code, GitHub Copilot CLI, and OpenCode.

requested change

Persist a provider-specific, allow-listed runtime profile with each native agent session reference and reapply the supported settings when Herdr performs native session restore. The profile should cover:
… +5 lines (ctrl + t to view transcript)

Capture these settings only from explicit launch arguments, provider-owned session events, and official Herdr integrations. Apply only options documented by that provider for session resume. A setting may be recorded for visibility without being reapplied if the provider does not expose a matching resume option.

The current implementation covers these runtime fields:

Agent	Runtime settings
Codex	Model, model provider, reasoning effort/summary, verbosity, and service tier
Claude Code	Model, effort, and agent profile
GitHub Copilot CLI	Model, reasoning effort, context tier, custom agent, and reasoning summaries
OpenCode	Provider/model and agent profile; variant is observed but not reapplied because OpenCode does not document a TUI resume flag for it

The persisted profile must exclude arbitrary arguments, prompts, config overrides, environment variables, credentials, and secrets. Unknown, invalid, or unsupported values should fail closed to the provider's normal safe behavior. The effective profile should also be exposed read-only through agent_session.runtime so restore behavior is auditable.

why you want this

Long-running multi-agent workspaces become unreliable after planned restarts, updates, host reboots, or reconnects: agents return to the right conversation but no longer have the same operating mode. Preserving this allow-listed operating profile would make restart behavior deterministic while retaining a narrow, auditable security boundary.

I have a working fork implementation on branch feature/preserve-agent-resume-access. It captures the allow-listed profile from launch arguments, provider session logs, and managed integrations, persists it in the Herdr session snapshot, exposes the runtime profile through the API, and reapplies documented options during native restore.

I will only open a PR if this direction is approved and an issue/approval path is provided.