feat(sessions): opt-in clear history and continue from plan

[igennova]

Summary

• Adds an opt-in plan-approval option: Approve, clear history, and continue from plan (alongside normal approve + reject).
• Normal Approve and proceed keeps planning history unchanged.

Why

Issue #704: after planning, clearing context before implementation avoids auto-compaction mid-implementation. This mirrors Claude’s clear-on-plan-accept flow as an explicit choice.

How it works

1. User is in plan mode → agent proposes a plan
2. Plan card offers:
   • Approve and proceed (keep context)
   • Approve, clear history, and continue from plan (fresh context)
   • Reject with feedback
3. Clear path: resolve permission UI locally → cancelPrompt → createNewLocalSession with plan seed
4. Same task / title; new task run (chat thread resets by design)

Video

Screen.Recording.2026-07-15.at.12.49.28.PM.mov

Test plan

• Enter plan mode → create a plan → confirm new clear option appears
• Approve and proceed → planning history remains; agent continues in-place
• Clear history and continue → chat resets; agent starts from plan only; no ACP/error toast
• [x]Mode selector still applies on clear path (e.g. auto / acceptEdits)
• Reject with feedback still works
• Error-state New Session still reconnects in place (resetSession unchanged)
• Unit tests: planContinuation, sessionServicePlanContinuation, permission-options, PlanApprovalSelector

[trunk-io]

Merging to main in this repository is managed by Trunk.

• To merge this pull request, check the box to the left or comment /trunk merge below.

After your PR is submitted to the merge queue, this comment will be automatically updated with its status. If the PR fails, failure details will also be posted here

[igennova]

The option id is duplicated here and in the Claude permission-options module on purpose — @posthog/agent can't import @posthog/core (import-direction rule), so there's no shared home. The doc comment above flags they must stay in sync. Open to a better spot (e.g. @posthog/shared).

[igennova]

Key decision: for clear-and-continue I deliberately don't call respondToPermission. Answering "allow" makes the old planning session start streaming the implementation reply in the stale (full-planning) context, and cancelling that mid-stream is what closed the ACP connection for ~2s (Upstream fetch aborted after client disconnect). Instead I resolve the pending-permission card locally and let continueFromApprovedPlan interrupt the turn. Normal approve/reject paths are untouched — they still go through respondToPermission below.

[igennova]

Intentionally not calling teardownSession here — it does removeSession, which leaves a window with no session for this task. The connect effect fills that window by resuming the stale run (the Resuming session race I saw in the logs). Keeping the store entry and letting createNewLocalSession → setSession swap it atomically (store auto-deletes the old taskRunId for the same taskId) means there's no dead gap and no resume race.