fix(agent): chat console keeps its thread across turns and reloads (#1859)

Three user-reported failures in the assistant panel, one root cause each:

1. "The chat asks what I'm referring to" when continuing a thread. The
   single-call console (general.help, AskConsole -> /api/agent/ask) was
   stateless since the 08-20 model-agnostic cutover: conversationId was only
   the tool actor id, so every turn was answered blind, reload or not.
   The provider-agnostic GenerateTextRequest gains an optional `history`
   (real message turns before the prompt, in both the Anthropic-family and
   the OpenAI-compatible adapter; absent/empty leaves the request
   byte-identical to the single-turn call). The route loads the thread's
   earlier turns server-side (loadChatHistory: text only, hidden and tool
   rows dropped, alternation repaired, newest 16 rows / 10k chars) before
   writing the new question, and hands them to the model.

2. A full page reload (the deploy prompt's "Ladda om") closed the docked
   panel and dropped the thread from view. The panel now remembers its open
   thread per tab in sessionStorage (lib/agent-panel/session-restore) and
   the provider reopens it on mount; the sheet loads it exactly like a pick
   from "Tidigare konversationer". Close and "Ny konversation" forget it; a
   thread that no longer opens is dropped instead of retried on every reload.

3. "Can't type any more" once the update banner shows. DeployReloadPrompt's
   full-width wrapper sits at z-[60] after the panel in DOM order and
   swallowed clicks on the panel's composer; only the card takes input now.


Claude-Session: https://claude.ai/code/session_01VjoXN3xdNZrHZeYA6qMi3g

Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

DECISIONS.md

@@ -1188,3 +1188,4 @@ One line per decision: `[YYYY-MM-DD] <decision>: <why>`. Appended by agents and
 [2026-08-24] Declared currency/voucher_series nullable in three MCP listing schemas on column-nullability alone (no traced null producer): loosening an output schema can only stop false validation failures, never cause one, and legacy rows predate the columns' defaults. Declined (for now) a full Ajv execute-vs-schema round-trip harness in output-schema.test.ts: right long-term answer to this bug class, but a session-sized project of its own; the audit's seven confirmed sites are pinned by a targeted declaration test instead.
 [2026-08-24] Manual matching (PR 6b) ships N:1 only (many outside rows -> one verifikat): bank links are independent per transaction (the engine allows it by design), skattekonto groups are all-or-nothing with the sum settling the verifikat (one guarded UPDATE, partial hit rolled back). 1:M (one row over several verifikat) and residual booking wait for a link table in 6c: the single journal_entry_id pointer on both row kinds cannot express them, and faking it (pointing the row at the residual verifikat) would break the bridge. The worksheet therefore enables Koppla only when the selection nets to zero and says so otherwise.
 [2026-08-24] Skattekonto payment file gets pain.001 through the supplier-payment generator (generateSupplierPain001), not the salary pain001 generator: the payment is a plain BG+OCR giro transfer (no SALA CtgyPurp), and the supplier dialect is the Validex-validated shape for exactly that; the LB path stays the default so nothing changes for banks still on LB.
+[2026-08-24] Single-call chat console (general.help, AskConsole → /api/agent/ask) now carries the thread's earlier turns into every model call, via a new optional `history` on the provider-agnostic GenerateTextRequest (real message turns before the prompt in BOTH adapters: Anthropic-family messages array, OpenAI-compatible via AI SDK `messages`; an absent/empty history leaves the request byte-identical to the single-turn call, so hosted extraction and every other caller are untouched). The 08-20 RIP-3 cutover made each turn stateless (conversationId was only the tool actor id), so a follow-up in a resumed thread was answered blind (user report: "frågar vad jag refererar till"). History is loaded server-side from agent_messages (loadChatHistory: text only, hidden + tool rows dropped, alternation repaired, newest 16 rows / 10k chars) rather than sent by the client, so the client cannot forge earlier turns and old streaming threads replay cleanly. Rejected: inlining a transcript into the prompt (works everywhere but weaker turn semantics and blurs data vs instructions) and loading history in AskConsole (client-trusted history). Separately: the docked assistant panel now remembers its open thread per tab in sessionStorage (lib/agent-panel/session-restore) and reopens it after a full reload (the deploy prompt's "Ladda om" wiped it); sessionStorage, not user_preferences, because this is this-tab-this-session state that must not follow the user to other devices or tabs. And DeployReloadPrompt's full-width wrapper gets pointer-events-none: at z-[60] after the panel in DOM order it swallowed clicks on the panel's composer ("går ej att skriva").