Files
accounted/lib/agent/ask/persist.ts
T
Jakob Wennberg 3edbf0a2e3 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>
2026-08-24 16:37:30 +02:00

228 lines
8.5 KiB
TypeScript

import type { SupabaseClient } from '@supabase/supabase-js'
import type { AiChatTurn } from '@/lib/ai'
/**
* Thread persistence for the single-call chat console (/chat).
*
* The console answers in one call via answerAssistantQuestion (no tool loop,
* no streaming: runs on a local model). But the founder wants the /chat
* sidebar and "resume a thread" to keep working, so each console turn is
* written to the SAME durable tables the streaming runtime uses:
* agent_conversations + agent_messages. That way old and new threads live in
* one list and one schema.
*
* Content is stored as the canonical Anthropic content array
* ([{ type: 'text', text }]) so normalizeStoredMessages() renders these rows
* identically to run-turn's, and the BFL append-only audit invariant on
* agent_messages holds (no UPDATE/DELETE policy exists on that table).
*
* Only the general.help intent uses this path. Tool-loop intents
* (categorization, invoice draft, supplier review) still go through
* run-turn.ts because they stage operations and need the tool loop.
*/
/** The only intent the single-call console persists under. */
export const CHAT_INTENT_ID = 'general.help'
// How much of a thread the model sees on each turn. Bounded because every
// resumed turn replays it: a long-lived thread must not grow the prompt
// without limit, and a local model may have a small context window. The
// newest turns win; the tail that fits is what the model gets.
export const HISTORY_MAX_MESSAGES = 16
export const HISTORY_MAX_CHARS = 10_000
// One turn is clamped too so a single pasted wall of text cannot eat the
// whole budget on its own.
const HISTORY_MAX_TURN_CHARS = 3_000
// The sidebar caches a one-line preview per row; the title is the sidebar
// label. Both are bounded so a long first question can't bloat the row.
const PREVIEW_MAX = 200
const TITLE_MAX = 80
/** Truncate on a whole-grapheme boundary is overkill here; a hard slice with an ellipsis is fine for a preview. */
function clamp(text: string, max: number): string {
const t = text.trim().replace(/\s+/g, ' ')
if (t.length <= max) return t
return `${t.slice(0, max - 1).trimEnd()}…`
}
function textBlocks(text: string): { type: 'text'; text: string }[] {
return [{ type: 'text', text }]
}
export type ResolveConversationResult =
| { ok: true; conversationId: string; created: boolean }
// The client passed a conversation id that isn't this user's general.help
// thread in this company. Same outcome as "doesn't exist": a 404, never a
// 403 that would confirm someone else's id is real.
| { ok: false; reason: 'not_found' }
/**
* Resolve the conversation to append this turn to.
*
* With no id, create a fresh general.help conversation titled from the first
* question. With an id, verify ownership the same way /api/agent/invoke does:
* RLS on agent_conversations is COMPANY-scoped, so a colleague's thread would
* otherwise load; the user_id + company_id + intent_id checks close that.
*/
export async function resolveChatConversation(
supabase: SupabaseClient,
userId: string,
companyId: string,
conversationId: string | null | undefined,
firstQuestion: string,
contextRef?: string | null,
): Promise<ResolveConversationResult> {
if (conversationId) {
const { data: conv } = await supabase
.from('agent_conversations')
.select('id, user_id, company_id, intent_id')
.eq('id', conversationId)
.maybeSingle()
if (
!conv ||
conv.user_id !== userId ||
conv.company_id !== companyId ||
conv.intent_id !== CHAT_INTENT_ID
) {
return { ok: false, reason: 'not_found' }
}
return { ok: true, conversationId: conv.id as string, created: false }
}
const { data: created, error } = await supabase
.from('agent_conversations')
.insert({
company_id: companyId,
user_id: userId,
intent_id: CHAT_INTENT_ID,
context_ref: contextRef ?? null,
title: clamp(firstQuestion, TITLE_MAX) || 'Fråga din assistent',
})
.select('id')
.single()
if (error || !created) throw error ?? new Error('Failed to create conversation')
return { ok: true, conversationId: created.id as string, created: true }
}
/** Append the user's question as a persisted turn (append-only). */
export async function persistUserTurn(
supabase: SupabaseClient,
conversationId: string,
question: string,
): Promise<void> {
const { error } = await supabase.from('agent_messages').insert({
conversation_id: conversationId,
role: 'user',
content: textBlocks(question),
})
if (error) throw error
}
/**
* Append the assistant's answer and roll the conversation row forward so the
* sidebar shows this thread at the top with a fresh preview. run-turn updates
* the same two columns on every assistant turn; we mirror that exactly.
*/
export async function persistAssistantTurn(
supabase: SupabaseClient,
conversationId: string,
answer: string,
): Promise<void> {
const { error: msgErr } = await supabase.from('agent_messages').insert({
conversation_id: conversationId,
role: 'assistant',
content: textBlocks(answer),
})
if (msgErr) throw msgErr
// Best-effort: a failed roll-forward only mis-sorts the sidebar row, it does
// not lose the answer (already persisted above). Do not fail the request on it.
await supabase
.from('agent_conversations')
.update({
last_message_at: new Date().toISOString(),
last_message_preview: clamp(answer, PREVIEW_MAX),
})
.eq('id', conversationId)
}
/** Plain text of a stored Anthropic-shaped content array; empty for tool-only rows. */
function textOfContent(content: unknown): string {
if (typeof content === 'string') return content.trim()
if (!Array.isArray(content)) return ''
return content
.filter(
(b): b is { type: 'text'; text: string } =>
!!b && typeof b === 'object' && (b as { type?: unknown }).type === 'text' && typeof (b as { text?: unknown }).text === 'string',
)
.map((b) => b.text)
.join('\n')
.trim()
}
function clampTurn(text: string): string {
if (text.length <= HISTORY_MAX_TURN_CHARS) return text
return `${text.slice(0, HISTORY_MAX_TURN_CHARS).trimEnd()}…`
}
/**
* The earlier turns of a thread, as the model should see them.
*
* This is what makes a resumed thread a conversation rather than a series of
* unrelated questions: without it every turn was answered blind, so a
* follow-up ("och förra månaden?") got "vad syftar du på?". Only the
* conversation's own rows are read (the caller has already proven ownership
* via resolveChatConversation).
*
* Shape rules, so both AI adapters can send the result as-is:
* - text only: tool_use / tool_result rows from the old streaming runtime
* carry no prose, and hidden rows are prompt-template scaffolding, never
* something the user said; both are dropped.
* - alternating, starting with a user turn: consecutive same-role turns are
* merged (a question whose answer failed sits next to its retry), and a
* leading assistant turn (its question fell off the window) is dropped.
* - bounded: newest HISTORY_MAX_MESSAGES rows, then trimmed from the oldest
* end until the text fits HISTORY_MAX_CHARS.
*/
export async function loadChatHistory(
supabase: SupabaseClient,
conversationId: string,
): Promise<AiChatTurn[]> {
const { data, error } = await supabase
.from('agent_messages')
.select('role, content, hidden, created_at')
.eq('conversation_id', conversationId)
.order('created_at', { ascending: false })
.order('id', { ascending: false })
.limit(HISTORY_MAX_MESSAGES)
if (error) throw error
const rows = (data ?? []) as { role: string; content: unknown; hidden?: boolean | null }[]
const turns: AiChatTurn[] = []
// Oldest first from here on.
for (const row of rows.slice().reverse()) {
if (row.hidden) continue
if (row.role !== 'user' && row.role !== 'assistant') continue
const text = textOfContent(row.content)
if (!text) continue
const clamped = clampTurn(text)
const last = turns[turns.length - 1]
if (last && last.role === row.role) {
last.text = clampTurn(`${last.text}\n\n${clamped}`)
} else {
turns.push({ role: row.role, text: clamped })
}
}
// Trim from the oldest end until the total fits, then make sure what is
// left opens with the user.
let total = turns.reduce((n, t) => n + t.text.length, 0)
while (turns.length > 0 && total > HISTORY_MAX_CHARS) {
total -= turns[0].text.length
turns.shift()
}
while (turns.length > 0 && turns[0].role !== 'user') turns.shift()
return turns
}