Multi-user team agent with cost management MCP server, document extraction, and shared knowledge base. Supports Telegram integration, guest demo mode, and per-user Durable Objects.
Mixed quality across 19 tools. Cost-core tools (record_expense, update_expense, etc.) have detailed Turkish descriptions and proper Zod schemas with type constraints, but lack English descriptions and output schema documentation. Standard tools (calculator, get_weather, get_datetime) have minimal descriptions (<50 chars). Tool composition is reasonable (separate record/list/get/update/delete), but parameter naming inconsistencies and missing output documentation significantly reduce usability. No tool annotations (readOnlyHint, destructiveHint) despite clear WRITE/DESTRUCTIVE/READ_ONLY risk classifications. Error handling guidance is absent.
Harcamaya resmi belge (fis/fatura/dekont) ekler ve CIFT-SENKRONIZASYON yapar: belgeden cozulen veri ile kaydi karsilastirir. Celiski varsa (orn. fisteki adres farkli sube) durum 'disputed' olur ve kullaniciya sorulmasi istenir.
Matematik islemi yapar.
Kullanicinin KISISEL/bireysel bilgisini kaydeder: tercihleri, teknik seviyesi, iletisim tarzi, egitimi, meslegi.
Harcamayi siler (soft-delete, geri alinabilir). ZORUNLU: once get_expense ile version al, expectedVersion gonder.
Bir notu siler.
R2'deki bir belge gorselini (fis/fatura/dekont) okur ve structured JSON dondurur: docType, amountTRY (TL), addressText (sube adresi), vendor, dateISO, confidence. Sonucu cost_attach_document'a tasi (extractedAmountTRY/extractedUnitId/extractedDate). addressText'i cost_list_units adresleriyle eslestirip extractedUnitId'yi sen belirle.
Output schemas undocumented. No tool returns a documented schema, LLMs cannot predict response structure for downstream chaining. E.g., record_expense returns 'describeCostResult()' but no schema is visible; list_expenses returns 'table + count + expenses' but structure is opaque.
Minimal descriptions on utility tools. get_weather ('Bir sehrin hava durumunu getirir' = 'Gets a city weather'), calculator ('Matematik islemi yapar' = 'Does math'), get_datetime ('Anlik tarih ve saat bilgisini verir' = 'Gives current date/time') are under 50 chars and lack context for LLM selection. No explanation of when to use vs alternatives.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Anlik tarih ve saat bilgisini verir.
Tek bir harcamayi getirir (GUNCEL version dahil — update/delete/attach oncesi bunu cagir).
Bir hisse senedinin guncel fiyatini getirir.
Bir sehrin hava durumunu getirir.
Secilebilir kategori/alan etiketlerini listeler.
Harcamalari listeler (birim/durum/etiket/silinmis filtreli). Sabit kolonlu tablo + ham veri doner.
Kaydedilmis notlari listeler.
Tum maliyet birimlerini (4 restoran + 1 merkez ofis) listeler.
Yeni harcama kaydeder. Maliyet buyuklugu (demirbas/orta/carcur) tutardan OTOMATIK atanir. En az 1 kategori/alan etiketi ZORUNLU (yoksa kayit reddedilir). Kategoriden emin degilsen categoryConfident=false ver.
Cift-senkron uyusmazligini cozer: kullanicidan alinan dogru bilgiyi (birim/tutar/tarih) uygular, belgeye gore tekrar kontrol eder; temizse 'verified' yapar. Once get_expense ile version al.
Bir not kaydeder.
Internette arama yapar ve sonuclari dondurur.
Harcamayi gunceller (OCC). ZORUNLU AKIS: once get_expense ile guncel 'version'i al, expectedVersion olarak gonder. Tutar degisirse buyukluk etiketi otomatik yeniden hesaplanir.
No tool annotations despite explicit risk classification. Tools marked WRITE (record_expense, update_expense, attach_document, resolve_discrepancy, save_note, create_knowledge, delete_note) and DESTRUCTIVE (delete_expense, delete_note) lack readOnlyHint/destructiveHint/idempotentHint annotations. Agents cannot distinguish safe from unsafe operations.
Error handling guidance absent. No tool description explains what to do on failure. E.g., record_expense says 'En az 1 kategori/alan etiketi ZORUNLU (yoksa kayit reddedilir)' (At least 1 category tag REQUIRED or record rejected) but does not guide recovery. No pattern:recovery-guide implementation.
Inconsistent parameter naming across tools. OCC (Optimistic Concurrency Control) uses 'expectedVersion' in cost tools but no other tools follow this pattern. 'actor' parameter appears in cost tools but not in note/knowledge tools. Inconsistency forces LLM to reason about field mappings.