c0ecf2fa3bebd46bdfd0169efd73b89653d1dfed
3 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Jakob Wennberg
|
c7a75d069d |
feat(ai): job-shaped AI service with OpenAI-compatible backend, extraction-first; stop extracting every inbox document twice (#1740)
* feat(ai): job-shaped AI service with OpenAI-compatible backend, extraction-first; stop extracting every inbox document twice Sovereign plan WS1 PR1 (#1406 Tier 2, extraction-first, aligned with the AI surface audit). lib/ai grows a job-shaped service (generateText / generateStructured / extractFromDocument; no streaming members yet, see plan rule R3): - services/anthropic-family delegates to the existing createAiClient() and sends the exact request literals the inbox extractor sent before (request-shape tests deep-equal them), so hosted Bedrock stays byte-identical. - services/openai-compatible talks to any chat-completions endpoint (BYO Swedish provider) via Vercel AI SDK 6.x, exact-pinned and guarded: images as parts, PDFs rasterized with poppler (AI_PDF_MODE) or sent natively, AI_VISION / AI_STRICT_JSON declared, honest skips (ai_no_vision, pdf_rasterizer_missing) instead of fake failures. - config.ts: AI_PROVIDER/AI_BASE_URL/AI_API_KEY/AI_MODEL and per-tier AI_*_MODEL with the legacy BEDROCK_* names kept as the same overrides; getAiStatus() is the single source of truth for "is AI wired up". - provider.ts: openai-compatible in the auto-detect chain (after Bedrock and the direct API); createAiClient() refuses it loudly. Document extraction moves onto the service and gets the audit's fixes: - Inbox documents were extracted TWICE (pipeline A ran inside uploadDocument() before the inbox row existed, so its dedupe branch never fired; 3 707 + 1 666 calls / 30 d). The inbox now declares extractionOwner on the upload, the extension yields, and the inbox mirrors its single outcome onto document_attachments from every writer (sync, deferred, attach, retry, MCP). - Every "no extraction will ever happen" outcome is stamped (skipped:no_ai_entitlement / ai_unconfigured / system_generated / ...); the status route maps the quiet ones to 'disabled' on the first poll instead of a 30 s client timeout. Prod showed 309 of the 327 never-extracted uploads were the paywall working silently. - Self-generated documents (our own invoice PDFs, payout files) are no longer OCR'd. - Agent invoke answers 503 ai_unconfigured when the deployment has no assistant backend, distinct from the paywall. Guard: new direct-ai-client antipattern check (shrink-only allowlist of the pre-abstraction SDK callers) plus exact pins for @anthropic-ai/sdk, ai and @ai-sdk/openai-compatible. Verified: 15 958 unit tests green, guards, lint ratchet, typecheck, and a live smoke against hosted Bedrock through the new service (ping, streamed tool turn, thinking+cache, PDF extraction). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(ai): make AI_API_KEY optional for OpenAI-compatible endpoints (keyless local model servers) A local model server (llama.cpp's server, Ollama /v1, LM Studio, vLLM) usually has no auth. Before, the OpenAI-compatible backend required both AI_BASE_URL and AI_API_KEY to count as configured, so running Accounted on a local model meant setting a meaningless placeholder key. - resolveAiProvider / hasAiCredentials: a base URL alone is now enough. - services/openai-compatible: only send Authorization: Bearer when AI_API_KEY is set, so a keyless server is never handed an empty bearer; a hosted provider that needs a key still sets it. - Docs (SELF-HOSTING Option 3: local-model example, key marked optional), DECISIONS. Verified: with no AI_API_KEY, just AI_BASE_URL + AI_MODEL, getAiStatus() reports configured=true / provider=openai-compatible (live). lib/ai suite 71 green; tsc, guards, lint clean. Bedrock/Anthropic logic unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
f0f3050f54 |
fix(assistant): stop the chat loading in stages (#1210)
* fix(assistant): stop the chat loading in stages PR2 of the assistant UI makeover (dev_docs/assistant_redesign_plan.md section 7). No redesign; this is the "it loads in different stages" complaint, traced to four separate staging points and one dead link. Resumed conversations rendered a column of EMPTY bordered cards until the markdown chunk arrived, then filled in all at once and reflowed the thread. The chunk was deferred with a null fallback, which is invisible while a reply streams (nobody reads that fast) but very visible on hydrate, where every assistant bubble is already text. The chunk is now prefetched as soon as any chat surface mounts, and until it resolves the raw text renders instead of nothing, so a bubble is never blank. Clicking the assistant launcher showed NOTHING until the sheet chunk loaded: the dynamic import had no loading state at all. It now renders a skeleton in the same geometry, and the chunk is warmed on idle so the click usually hits an already-loaded module. /chat's route skeleton drew a 320px sidebar while ChatSidebar mounts collapsed as a 48px rail, so every load snapped one to the other. The skeleton now matches what actually mounts, per breakpoint. The first turn read agent_profiles twice: once in the route to build the intent's prompt template, once again in run-turn for the system prompt. The route now hands its result over. Ranked memory is deliberately NOT shared: the two queries differ (the route's selects fewer columns and orders without is_pinned, and run-turn needs ids to stamp last_accessed_at), so reusing it would silently change both the prompt and memory touch. Command palette's "Fråga Anna: ..." pointed at /chat?prompt=, but only /chat/new reads ?prompt=, so the typed question was silently dropped and the user landed on an empty state. Verified: 9526 unit tests pass, lint clean and tsc clean on every touched file, guards pass. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(assistant): make the prefetches fail-safe and bounded Review follow-ups on the staged-loading batch. A rejected markdown import left the cached promise permanently rejected, so every bubble for the rest of the session stayed on the plain-text fallback and the rejection went unhandled. The cache is now cleared on failure so a later surface retries, and the rejection is swallowed. requestIdleCallback can defer indefinitely on a page that never goes idle; the 2s fallback only applied where the API is missing. The idle request now carries a 2s timeout, and the warm import cannot produce an unhandled rejection either. Adds the first-turn test for the profile-summary handover: it asserts the value read for the prompt template is what reaches the turn, so a regression that re-introduces the second read (or drops the template) fails. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
ee8ddb3849 |
fix(assistant): stop cross-user conversation access, bricked threads and lost sessions (#1209)
* fix(assistant): stop cross-user conversation access, bricked threads and lost sessions Hotfix batch (PR1 of the assistant UI makeover, dev_docs/assistant_redesign_plan.md section 7). No visual change; each of these is wrong today regardless of which design lands, and three are unrecoverable per incident. /api/agent/invoke never checked who owns a resumed conversation_id. RLS on agent_conversations/agent_messages is company-scoped, not user-scoped (20260517204000), so a member could post a colleague's conversation id, have their history loaded into the prompt and read it back, while their own turns were appended to that thread. The conversations list route filters on user_id for exactly this reason. Also pins company and intent: resuming a thread from another company would mix ledgers, and resuming under a different intent would swap the tool whitelist under history the model has already seen. A turn persists the assistant message carrying tool_use blocks before the tools run, and their results only after the batch finishes. Dying in between (client disconnect terminating the function, a deploy, a slow tool) left history ending on an unanswered tool_use, which the Messages API rejects on replay: every later turn 400s, and agent_messages is append-only for the BFL trail, so nothing could repair it. History is now patched on read by synthesizing is_error tool_results, leaving the stored trail untouched. check_and_increment_agent_quota is SECURITY DEFINER in public with a caller-chosen p_user_id, so any authenticated user could drain a colleague's minute/day budget and lock them out of every agent endpoint. A plain REVOKE would break the limiter (all three callers use the user's RLS client) and, as it fails open, silently remove the spend cap: the function now refuses to act for anyone but the caller, while service-role connections keep passing an explicit id. The single reject route re-read status and then wrote unguarded, so losing the race with commit's atomic pending -> committing claim stamped `rejected` over an operation that had already posted a verifikat, invisible to the committing-state recovery sweep. Guarded on status like bulk-reject already is; a lost race is now a 409. The sheet's Escape handler listened on window with no defaultPrevented or target check while the sheet is deliberately non-modal, so pressing Esc to dismiss the reject-reason Select inside an approval card, the command palette or any dialog unmounted the sheet and discarded the conversation, the streaming turn and the un-actioned proposal. It now yields to open overlays and to focus outside the sheet. Verified: 9526 unit tests pass, lint clean on touched files, guards pass, and the new pg-real test proves the quota guard against real Postgres (attacker raises 42501, victim counters stay at 0). The four unrelated pg-real failures on this machine reproduce identically with these changes stashed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(assistant): close anon path on the quota RPC, order the ownership check ahead of writes Review follow-ups on the hotfix batch. The caller guard used auth.uid() alone, which is NULL for the `anon` role just as it is for backend roles, so an unauthenticated caller holding the public anon key (it ships in the browser bundle) could still pick any p_user_id and drain that user's quota. The guard now keys on the request role: anon and authenticated may only ever spend their own quota, backend roles keep passing an explicit id. The default PUBLIC execute grant is revoked as a second layer, with execute granted only to authenticated and service_role. Covered by a new pg test for the anon path. The ownership check ran after the onboarding.intake stamp, so a request that was about to be rejected could still write intake_completed_at. It now sits directly after the capability gate, ahead of every side effect and ahead of the company and profile reads, which also makes a rejected request cheaper. The tool-result repair matched ids anywhere in the history, but the API needs results in the message IMMEDIATELY after the tool_use. A result persisted after an intervening turn (two turns racing on one conversation) left a shape that still 400s. The repair is now positional, and orphaned or late-duplicate tool_results are dropped, since an unmatched tool_result is rejected just as an unanswered tool_use is. The Escape guard matched the Radix popper wrapper, which stays mounted when a popper is force-mounted; it now requires data-state="open" so a closed popper cannot block Escape for the rest of the session. Both new route errors are Swedish, per the user-facing error rule. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |