feat(agents): persist per-session plan state as a typed projection

[david-engelmann]

Summary

• Adds a backend projection of the latest agent plan per session, so a future pinned-plan UI can survive reloads without rescanning the chat transcript.
• Codex turn/plan/updated events and Claude ExitPlanMode tool calls get normalised into the same typed Plan shape and upserted into a new session_plan_state table.
• New getSessionPlanState Tauri command + UiMutationEvent::SessionPlanChanged invalidation channel. The pipeline accumulator/adapter — and therefore the existing inline plan card / todo list rendering — are untouched; this is a parallel side-table write.

Why

Closes #410. The bug is that Helmor already knows what the active plan is at runtime, but the only place it lives is the chat scrollback. After a reload the frontend has to rescan messages to reconstruct the plan, which is fragile and forces every consumer to re-implement the projection.

Scope boundary

Included:

• session_plan_state table with idempotent migration + snapshot test.
• agents::session_plan module — typed Plan, PlanItem, PlanItemStatus, PlanStatus, PlanSource, SessionPlanState shapes, plus pure parsers for both providers.
• Streaming-bridge hooks at the two existing plan-event sites:
  • Codex turn/plan/updated in the default arm.
  • ExitPlanMode next to the existing persist_exit_plan_message call.
• UiMutationEvent::SessionPlanChanged { session_id } with frontend type mirror, bridge handler, query key + sessionPlanStateQueryOptions.
• get_session_plan_state(session_id) Tauri command + getSessionPlanState TS wrapper.
• Tests:
  • 9 projection unit tests (Codex status normalisation, current-item resolution, ExitPlanMode bullet extraction with nested-line guard, raw-text fallback, empty/missing inputs, upsert round-trip, in-place replacement).
  • Schema migration snapshot for legacy DBs.
  • UI sync serialization test for the new variant + camelCase regression gate extended.

Deferred (intentionally out of scope; subsequent PRs in track C):

• Pinned-plan UI (Track C / PR C2).
• Continue / revise prompt builders + buttons (Track C / PR C3).
• Per-provider plan-completion / cancellation signals (the enum stays Active for now; new variants land when there's a clear provider signal to drive them).

Design notes

• The pipeline accumulator emits the inline plan card / todo-list exactly as before — projection is a separate observation. That keeps the existing pipeline snapshot suites stable and means the storage-shape "MUST have snapshot test coverage" rule from AGENTS.md applies only to the new table, not to the message rows.
• Codex events lack stable per-step ids; the projection derives codex-{idx} so the pinned UI has a key and replays land in lockstep when a subsequent turn/plan/updated arrives.
• ExitPlanMode is free-text markdown. The parser pulls top-level bullets into structured items (rejecting indented sub-items by checking for leading whitespace before the bullet) and keeps the raw markdown as a rawText fallback so the panel can render prose-only or nested plans even though the structured list is shallow.
• load_session_plan_state logs and returns None on JSON validation failure rather than bubbling — a stale row from a hypothetical breaking shape change won't crash the panel.

Test plan

• cd src-tauri && cargo nextest run — 1144 / 1144 pass (9 new projection tests + 1 new migration test + 1 new ui_sync test).
• cd src-tauri && cargo test --doc — clean.
• cd src-tauri && cargo fmt --all -- --check — clean.
• bun run lint — biome + clippy -D warnings clean.
• bun run typecheck — clean.
• bun run test:frontend — 1092 / 1092 pass.

Coverage added

New tests:

• agents::session_plan::tests::codex_plan_maps_statuses_and_picks_current_item
• agents::session_plan::tests::codex_plan_falls_back_to_first_pending_when_no_in_progress
• agents::session_plan::tests::codex_plan_returns_none_for_missing_or_empty_plan_array
• agents::session_plan::tests::codex_plan_treats_unknown_status_as_pending
• agents::session_plan::tests::exit_plan_mode_extracts_bullet_items_and_preserves_raw_text
• agents::session_plan::tests::exit_plan_mode_returns_none_for_empty_or_missing_plan
• agents::session_plan::tests::exit_plan_mode_handles_unbulleted_prose_as_raw_text_only
• agents::session_plan::tests::upsert_and_load_round_trips_typed_shape
• agents::session_plan::tests::upsert_replaces_existing_row_in_place
• ui_sync::events::tests::session_plan_changed_uses_camel_case_type_and_field
• tests/schema_migrations.rs::session_plan_state_migration_creates_table_on_legacy_dbs

Snapshot updates:

• tests/snapshots/schema_migrations__session_plan_state_migration.snap (new).

Follow-up

This is PR C1 in the durable-plan-state track. The natural follow-up slices:

• C2: pinned-plan UI rendering getSessionPlanState with sessionPlanChanged invalidation.
• C3: continue / revise prompt builders + composer actions.

The backend API is intended to be stable enough that either of those can land independently.

[vercel]

@david-engelmann is attempting to deploy a commit to the Caspian's Team Team on Vercel.

A member of the Team first needs to authorize it.