feat(conversation): persist plan snapshots and expose them for rehydration

[kaizhou-lab]

Closes #915

What

Makes the agent's plan / to-do snapshot survive a conversation switch, and gives the frontend a way to rehydrate it mid-turn.

Changes

persist_plan + a dedicated relay arm — AgentStreamEvent::Plan was in the _ => catch-all of stream_relay.rs: forwarded to the WebSocket, never written. It now persists one row per turn (id = plan:{msg_id}), upserted, because a plan is a full-replacement snapshot per the ACP spec.

The arm deliberately does not set saw_tool_or_side_effect (that would make an otherwise-replayable turn look unsafe to retry) and does not close the active text segment (a plan refresh lands mid-reply and would shatter it into a fresh bubble) — same reasoning as the existing WorkflowProgress arm.

msg_id stores the bare turn msg_id, not the plan: form: the live WS frame carries the bare id and the renderer dedupes history against live frames on ${type}:${msg_id}. Storing the prefixed form would show one live card plus one history card after reload.

turn_id rides inside the content JSON so the frontend can gate the bar on the running turn without a schema change.

latest_message_of_type + GET /api/conversations/{id}/messages/latest?type=plan — required, not a nicety: upsert_message does not refresh created_at, so a plan row stays anchored at the START of its turn and a turn with many tool calls buries it outside the default 50-message page. A just-ended turn would otherwise rehydrate with an empty bar.

Deliberately not a type filter on list_messages_page: that paginator has four SQL variants and cursor semantics (has_more_before / has_more_after) that a type filter would muddy. One extra single-row query, hitting the existing idx_messages_type_created, is the smaller blast radius.

adapter/claude.rs: TodoWrite → SessionEvent::Plan — with the paired tool_result suppressed. That suppression is mandatory: session_agent.rs translates SessionEvent::ToolResult into a terminal AgentStreamEvent::ToolCall, and a terminal frame with no card to settle makes the renderer append a nameless junk card. A malformed payload (no todos array) falls back to the ordinary tool path — losing the plan is acceptable, losing the card is not.

map_plan_status is re-exported pub(crate) from backend/mod.rs rather than copied a third time. codex_conn.rs's existing private copy is left alone.

No migration

messages.type has no CHECK constraint, content is free-form JSON, MessageType::Plan already exists, and idx_messages_type_created already indexes (type, created_at DESC). Plan rows have never been written, so there is no legacy data to migrate.

Version boundary on the claude lane

Bisected 2026-08-21 over the npm line: claude removed TodoWrite from its headless tool set in 2.1.142 — 2.1.141 still advertises it in the system:init frame, 2.1.142 does not. (The init frame's tools array is the ground truth; it is emitted before the model sees the prompt. The todoFeatureEnabled setting is unrelated — it enables the interactive TUI panel and does nothing headless.)

So on 2.1.142+ this translation is dormant: the model has no such tool to call. It is kept because the logic is correct and cheap, and the CLI surface moves. The test fixture is wire-pinned to a real 2.1.141 frame (input.todos[{content, activeForm, status}], caller:{type:"direct"}) rather than a hand-written shape. VERIFIED_CLAUDE_VERSION (2.1.235) is already past the removal.

codex (update_plan) and ACP agents emitting sessionUpdate:"plan" are unaffected — both verified live.

Verification

• cargo test -p aionui-db --lib — 362 passed
• cargo test -p aionui-conversation --lib — 425 passed
• cargo test -p aionui-session --lib — 598 passed
• cargo clippy -p aionui-db -p aionui-conversation -p aionui-session --all-targets -- -D warnings — clean
• Live: plan rows confirmed written for a codex turn and for a grok (ACP) turn.

New tests cover: first-write then same-turn upsert collapsing to one row; turn_id present in persisted content; latest_message_of_type reaching a plan row buried under 60 later messages; the TodoWrite translation emitting exactly one Plan and zero ToolCall; the paired tool_result producing zero terminal frames; and a malformed TodoWrite falling back to ToolCall.

There is also a small guard test that the new /messages/latest route can be registered alongside the {messageId} wildcard — axum builds its route trie eagerly, so an overlap panics at construction and cargo check would never catch it.