LAYAL CAFE ($2.80 instead of $42.90):
- Add (?!\s*tax) lookahead to _TOTAL_RE so "Total Taxes $2.80" is never
confused with the receipt total when OCR drops the "Taxes" word
- Change Pass 1 from matches[-1] to max() so the largest labeled amount
always wins, regardless of line order in the OCR output
United Airlines (Subway/$0/wrong date):
- Add OSD-based rotation correction in receipt_parser.py: after EXIF
transpose, ask Tesseract's orientation-detection engine (--psm 0) what
angle to rotate; applies to receipts photographed lying sideways where
EXIF metadata cannot help
- Add month-name date patterns (DD MON YYYY / MON DD YYYY) to
_extract_date_from_text for airline/hotel receipts that print dates
like "05 MAY 2026" instead of "05/07/26"
85 tests, all passing.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Gas station receipts (Costco, Shell, etc.) print "Total Sale $X.XX" — the
word "Sale" between "Total" and the amount prevented _TOTAL_RE from matching,
causing the Costco receipt to fall through to the max-scan heuristic and
return a garbled OCR value instead of the correct total.
Also add "Net Sale" and "Sale Total" variants for broader coverage.
79 tests, all passing.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add _is_likely_bank_statement(): if OCR text has ≥10 lines with dollar
amounts it is almost certainly a bank/card statement screenshot, not a
single receipt. Return skip=True so _act() skips it and adds a note to
the escalations list instead of creating a $1,699 expense line.
- Fix default product selection in _act(): prefer "Meals" over whatever
happens to be first in Odoo's expense product list ("Communication"),
so unrecognised receipts get a sensible fallback category.
- Improve LLM category prompt: remove hardcoded product names (airline →
Transport) that don't exist in every Odoo install; describe business
types semantically so the model picks from the actual available list.
- Mention skipped statements in the final summary message.
- 77 tests, all passing.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Remove card brands (VISA/MC/Amex) from _SKIP_LINE_RE so card-terminal
lines like "VISA USD$ 36.78" are no longer skipped
- Replace bottom-50% scan with full-text max scan (Pass 2): scans every
line in the receipt and returns the largest dollar amount, correctly
handling display-style receipts that show the charge at the top with
no label (e.g. LAYAL CAFE $40.10 before the item list)
- Update vendor LLM prompt to ask the model to correct OCR garbling
(e.g. "NeDonald's" → "McDonald's") and detect bank statements
- Add 4 new tests covering top-amount, card-terminal, max-beats-items,
and change-exclusion scenarios (71 tests, all passing)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Logs per-receipt: OCR text length, first 120 chars of OCR output,
and final parsed vendor/amount/date/product_name.
This will show whether Tesseract is producing usable text.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Receipt quality: replace LLM amount/date extraction with regex.
LLM was hallucinating 2021/2022 dates and returning '198.40 USD' strings.
Amounts now use deterministic regex (Total:/Grand Total:/Amount Due:).
Dates: filename timestamp > OCR regex > today (no LLM date guessing).
LLM only asked for vendor name + product category.
Approval: fix GET /approval/pending 500 by using correct column
name 'started_at' instead of 'created_at' (which does not exist).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The llama3.2-vision model was producing unreliable structured data
(wrong vendors, amounts, dates) making expense reports worse than
Tesseract + LLM extraction. Removes _ocr_image_vision(), the
vision JSON fast path in _parse_receipt_text(), _match_category(),
and the vision_ocr_model config setting entirely.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Three receipts per batch were failing with JSONDecodeError (e.g.
"Expecting ':' delimiter: line 1 column 90") because activeblue-chat
(llama3.2-vision) occasionally outputs near-JSON with trailing commas,
single-quoted strings, or unquoted keys.
Two-layer fix:
1. Add format='json' to the Ollama chat call — Ollama JSON mode forces
syntactically valid output at the sampler level, eliminating most
structural errors.
2. Add _repair_json() fallback that runs on any remaining JSONDecodeError:
strips trailing commas, converts single→double quotes, and quotes
unquoted keys. If repair succeeds, the result is re-serialised as
canonical JSON before being returned.
Also re-serialise with json.dumps() on success so the fast path in
_parse_receipt_text always receives clean, canonical JSON regardless of
whitespace or key ordering in the model's original output.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two sources of hallucinated values in receipt parsing:
1. The LLM extraction prompt had no explicit "don't guess" constraint, so
when Tesseract produced garbled OCR text the LLM substituted plausible-
looking values (wrong vendor names, wrong totals) instead of returning
safe defaults.
2. The date field asked the LLM to extract the date from the OCR text even
when date_hint (from the filename timestamp, e.g. 20260509_180857.jpg)
was already available — a reliable signal that was being ignored.
expenses_agent._parse_receipt_text:
- LLM path: new prompt leads with "copy values EXACTLY, do NOT guess or
infer"; adds "if OCR looks corrupted, return safe default rather than
a more logical value"; injects date_hint directly as an authoritative
value when available so the LLM never needs to extract the date.
- Vision fast path: normalise "null" string for date the same way as time;
prefer date_hint over a null date returned by the vision model.
receipt_parser._ocr_image_vision:
- Vision prompt now leads with the same "copy exactly, do not guess"
constraint and explicitly accepts null for date/time when not clearly
visible, matching the conservative tone of the LLM extraction prompt.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When a zip/image arrives via /upload, the LLM was classifying the
message as needs_clarification=True (because the chat body was just a
filename like "download (8).zip", not an instruction), and the early
return on line 91 fired before the receipts safety guard on line 106,
so the guard never executed.
master_agent: move the receipts safety guard to BEFORE the
needs_clarification early-return. If extra_context contains receipts,
unconditionally set needs_clarification=False and ensure expenses_agent
is in the agents list — the LLM cannot veto an upload with a question.
upload router: normalize empty or filename-only messages (e.g. when the
user drops a file in Discuss chat with no text) to
"Create an expense report from these uploaded receipts." so the LLM
intent classification also has a sensible string to work with.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
/approval/pending was returning 500 UndefinedColumnError because the
approval router and MCP get_pending_approvals tool both query columns
(agent_name, action_type, description, context_data, approver_id,
approval_note, updated_at) that were never added in the initial schema
migration 001.
Adds migration 002 to ALTER TABLE ab_directive_log with all seven
missing columns (all nullable so existing rows are unaffected) and an
index on updated_at for efficient polling.
Deploy: after pulling on miaai, run:
cd /root/odoo/odoo-ai && docker compose exec agent-service \
alembic -c agent_service/migrations/alembic.ini upgrade head
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
receipt_parser: change _ocr_image_vision() to extract structured JSON
{vendor,amount,date,time,category} directly from the image instead of
transcribing raw text, so the downstream LLM extraction step is
unnecessary and the two-step error-compounding is eliminated.
expenses_agent: add _match_category() helper to map vision category
labels to expense product names via substring/fuzzy match; add fast
path in _parse_receipt_text() that detects pre-extracted vision JSON
(text starts with '{') and skips the second LLM submit call entirely.
Fix text[:2000] truncation that discarded receipt totals — now keeps
first 1500 + last 1500 chars of long receipts so the grand total at
the bottom is always included.
tests: fix stale test_act_enters_awaiting_confirmation_on_first_pass
(confirmation gate was removed); add TestMatchCategory and three new
tests for the vision JSON fast path and LLM fallthrough.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
1. OdooClient missing self._timeout — every _xmlrpc_call raised
AttributeError, making the odoo health check permanently fail.
Fix: set self._timeout = XMLRPC_TIMEOUT in __init__.
2. action_ping only accepted ollama=='ok' but health.py now returns
'warming' when the model is not yet hot in VRAM. Fix: treat
warming as passing so the bot goes online and the model loads
on the first real request.
3. /ai/approval/pending declared methods=['GET'] on a type='json'
route — Odoo JSON-RPC always POSTs, so every browser call got
405 METHOD NOT ALLOWED. Fix: change to methods=['POST'].
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ping() was calling ollama.AsyncClient.list() which parses /api/tags with
ollama==0.3.3 pydantic models. Vision models carry metadata fields that 0.3.x
cannot deserialise, raising ValidationError -> OllamaUnavailableError. This
made the /health/detailed ollama field 'error: ...' instead of 'ok', so
ab_ai_bot.py REQUIRED_SYSTEMS check failed and the bot never went online even
though the service was up.
Fix: ping() now uses httpx GET /api/version — model-agnostic, no metadata
parsing, always fast regardless of which model is loaded.
Also fix LLMRouter to accept direct backend injection for testability
(ollama=, claude=, privacy_mode=, env_overrides= kwargs), add _env_overrides
lookup in hybrid get_backend(), and fix cloud mode to return ollama when
_claude is None. All 6 test_llm_router tests now pass.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- OllamaBackend enforces _MIN_TIMEOUT=300s (overrides OLLAMA_TIMEOUT env var)
- warm_model() background task loads activeblue-chat into VRAM at startup
- health/detailed reports "warming" vs "ok" via Ollama ps() API
- README updated with May 2026 changes and test coverage details
- All specialist agents: handle_peer_request(request_type, params, directive_id)
replaces handle_peer_request(request: dict) so callers pass structured args
- ab_ai_bot: force-write bus_presence.status via SQL so Odoo 18 WebSocket presence
shows the correct colour immediately (ORM compute does not trigger on last_poll writes)
- odoo_client: wrap XML-RPC executor calls in asyncio.wait_for to enforce timeout
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously ab_ai_mail.py intercepted file uploads before reaching the
LLM and responded with a hardcoded clarification template. The LLM had
no involvement in the file upload response.
Changes:
- ab_ai_mail.py: remove _post_file_clarification, _find_pending_attachments,
_describe_zip, and the two-step pending-attachment lookup. All messages
(text, files, or both) are dispatched to the agent service immediately.
Files with no text pass an empty message — the LLM decides what to do.
- upload.py: default message changed from hardcoded receipt instruction
to '' so the LLM determines intent from file content.
- master_agent._synthesize: always runs through the LLM for both single
and multi-agent cases — no raw templates reach the user.
- master_system.txt: add FILE UPLOADS routing rule so the LLM knows to
route receipts to expenses_agent without asking for clarification.
New flow: upload → parse → LLM classifies → agent acts → LLM synthesizes
natural response → user sees it. Zero scripted intercepts.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The old flow required a "confirm" reply after showing a parsed-receipt
table, but that follow-up dispatch call carries no receipts (they only
exist in the /upload context). The confirmation gate was architecturally
broken: the second turn would always create nothing.
Fix: create the expense sheet immediately when receipts are present.
Byte-exact and semantic duplicates are auto-skipped; the count of
skipped items is reported in the success message. The report is always
created in Odoo as a draft so users can review amounts and submit
manually via Odoo > Expenses.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds a new specialist agent that gives the AI system control over its
own infrastructure:
- sysops_tools.py: docker SDK (ps/logs/restart) + git CLI (pull/status/log)
+ Odoo channel notifier for autonomous action broadcasts
- sysops_agent.py: BaseAgent subclass handling on-demand chat requests,
auto_heal() triggered by health failures, and sweep() for audits
- Background auto-heal loop (main.py): runs every 2 minutes, calls
_get_failing_systems() and triggers auto_heal() when degraded
- health.py: extracted _get_failing_systems() helper reused by both
the /health/detailed endpoint and the auto-heal loop
- docker-compose.yml: mount docker socket + /root/odoo workspace +
SSH keys for git authentication
- Dockerfile: add git to apt-get
- requirements.txt: add docker==7.1.0 Python SDK
Auto-heal behavior:
- Detects failing containers, restarts them, notifies all bot DM channels
- Ollama (192.168.2.9) is flagged as external and skipped
- On-demand via chat: "restart agent", "check logs", "pull latest code"
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- expenses_tools: remove 'date' from hr.expense.sheet field lists (Odoo 18
uses accounting_date; querying 'date' raised ValueError at runtime)
- master_system.txt: add few-shot routing examples so Llama 3.1 8B correctly
outputs agents=[] for general questions instead of defaulting to expenses_agent
- ab_ai_bot.py: increase bot presence last_poll offset from 90s to 10min so
the green dot stays on between cron runs (cron fires every ~5min in practice,
not every 20s as configured)
- ARCHITECTURE.md: full system documentation covering component layout, request
flow, LLM routing, agent registry, access control, health/presence mechanism,
known issues fixed today, and future self-healing concept
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Introduces VISION_OCR_MODEL setting. When set (e.g. llama3.2-vision:11b),
receipt images are transcribed by the Ollama vision model before falling
back to Tesseract. Also improves Tesseract preprocessing with adaptive
binarisation (autocontrast + threshold at 140) for better accuracy on
thermal receipts.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Before writing any expense records the bot now posts a numbered table
of parsed vendor/amount/date for every receipt, with duplicate entries
flagged inline. User replies 'confirm' (skips dups) or 'confirm, keep
all'. This catches OCR amount misreads before they land in Odoo.
Also removes the separate awaiting_dup_approval step; duplicate review
is now part of the single confirmation table.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Pass 1 unchanged: same date + amount within 0.05 + vendor similarity 60%.
Pass 2 (new): same vendor (>= 80% similarity) + same date, regardless
of amount, to catch receipts where OCR misread the total.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replaced 'pick the largest one' guidance with 'bottom-most total' and
'return 0 if no clear total found' to avoid picking line items or tips.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
DispatchResponse declared actions_taken as list[dict] but agents return
list[str], causing a 422 on every successful upload.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The /registry/agents endpoint was 500 on every call because
AgentRegistry.get_all() is async but was called without await.
Also aligns get_all() dict keys (name, domain) with what the router reads.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- master_agent: thread raw user message into extra_context and peer_data so
expenses_agent can check it directly without relying on LLM intent_summary
- master_agent: when receipts are in extra_context always route to expenses_agent,
so replies like 'skip duplicates' still trigger expense processing
- expenses_agent: _plan() checks peer_data raw_message alongside task so
skip/keep keywords are detected even when master rewrites the intent
- ab_ai_mail: wrap clarification message HTML in Markup() so Odoo does not
re-escape the tags; use <br> instead of <br/>
- ab_ai_mail: convert agent plain-text replies newlines to <br> for proper
line-break rendering in Discuss
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- expenses_agent: extract transaction time (HH:MM) from OCR receipt text
- expenses_agent: _find_semantic_duplicate uses time to rule out false positives (>30 min apart = different receipts)
- expenses_agent: pause when duplicates found, set mode=awaiting_dup_approval, ask user before creating sheet
- expenses_agent: _report formats approval message listing each dup pair with vendor/amount/date/times/filenames
- ab_ai_mail: _find_pending_attachments recognises dup-approval bot message so ZIP re-attaches on user reply
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
After parsing all receipts, identify photos that are different shots of
the same physical receipt by comparing amount + date + vendor similarity
(difflib ratio >= 0.6). When a duplicate is found, keep whichever photo
produced the most OCR text (clearest shot) and report the skipped ones.
Zero-amount receipts (OCR failed entirely) are excluded from semantic
dedup to avoid false positives.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Dockerfile: add tesseract-ocr-osd for orientation detection data
- receipt_parser: resize large phone photos to 1800px, convert to
grayscale, sharpen before OCR; use psm 1 (auto + OSD) so rotated
receipts are correctly oriented before text extraction
- expenses_agent: tighten amount extraction prompt to pick the FINAL
total, not subtotal or tax line, reducing misreads like 42.90->409.00
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- ab_ai_bot: raise requests.post timeout 120s -> 600s so long OCR+LLM
runs don't silently drop the reply in Discuss
- upload: run parse_upload in ThreadPoolExecutor so tesseract OCR
doesn't block the FastAPI event loop
- expenses_agent: parse all receipts concurrently with asyncio.gather
(Ollama semaphore caps parallelism at 2); reduces 13-receipt LLM
time from ~39s sequential to ~20s parallel
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Dockerfile: install tesseract-ocr so Pillow+pytesseract can OCR receipt images
- operational_store: JSON-serialize raw_data before passing to asyncpg JSONB
- receipt_parser: add SHA256 hash + date extracted from filename timestamps
- expenses_agent: deduplicate receipts by hash before creating expense records
- expenses_agent: fetch all expensable Odoo products, pass list to LLM for
category selection (Meals, Flights, etc.) per receipt
- expenses_agent: pass date_hint from filename (e.g. 20260509_180857.jpg -> 2026-05-09)
as fallback when OCR text is unavailable
- expenses_tools: add get_expense_products() to fetch all expensable products
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- _synthesize: short-circuit on any single-agent report (avoids extra
Ollama call that can timeout); wrap multi-agent LLM call in try/except
- _update_memory: catch exceptions so DB/memory failures don't kill reply
- _log_directive_start: use 0 instead of NULL for channel_id (NOT NULL col)
- create_expense: drop 'description' field (not valid on hr.expense in Odoo 18)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
AGENT_ACCESS_GROUPS uses XML IDs (e.g. hr_expense.group_hr_expense_user)
but the check compared them against res.groups.full_name strings which
never matched, denying every user access to all restricted agents.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The master agent was routing expense/receipt requests to finance_agent
instead of expenses_agent because only DB-registered agents appeared
in get_active_agents(). This adds auto-activation of all in-memory
registered agents with precise capability summaries so the LLM picks
the right specialist.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- ollama_backend: add format='json' for 'master' and receipt_parser
callers so llama3.1:8b returns valid JSON instead of plain English
- ab_ai_mail: add debug logging to trace attachment_ids from Discuss;
handle file-only messages and clarification look-back flow
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Discuss bot now reads ir.attachment from incoming messages; file-only
messages no longer silently dropped
- ZIP files are described (contents listed) and bot asks clarifying
question before acting; user's follow-up reply looks back for pending
attachments so files don't need to be re-uploaded
- receipt_parser: extracts text from ZIP (recursive), JPG/PNG/etc (OCR),
PDF (pdfplumber), HTML, TXT
- expenses_agent: full rewrite fixing broken method signatures; adds
create_expense_sheet / create_expense / attach_receipt flow driven by
LLM receipt parsing (Ollama, HIPAA-locked)
- master_agent: extra_context threads receipts + user_id into directives
- FastAPI /upload multipart endpoint; registered in main.py
- Odoo /ai/upload controller proxies files to agent service
- ab_ai_bot: dispatch_message_with_files() for multipart uploads
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- ElearningTools: add create_course, update_course, publish_course,
add_section, create_slide, enroll_user write methods using OdooClient
- ElearningAgent: fix all BaseAgent method signatures (_plan/_gather/
_reason/_act/_report no longer take wrong positional args)
- Replace dead _dispatch_tool pattern with _tool_<name> methods so
BaseAgent._run_tool() can drive them via LLM tool calls in _loop()
- Add LLM-driven course creation in _reason(): when intent is create,
_loop() is called with a course-building system prompt and all tools;
the LLM calls create_course → add_section → create_slide → publish
- Fix handle_peer_request signature to match BaseAgent interface
- Fix AgentReport missing directive_id; fix SweepReport invalid kwargs
- Extend ELEARNING_TOOLS list with all new write-side tools
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
BaseAgent._lookup_odoo_context() calls odoo_doc_agent via PeerBus before
_plan() runs on every directive. The RAG answer is stored in
self._gathered['odoo_context'] and injected into every _loop() LLM call
so agents reason with correct Odoo 18 workflow steps automatically.
No changes required to individual agents. odoo_doc_agent opts out via
auto_rag=False to prevent self-referential calls.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Wraps odootrain RAG API (http://192.168.2.9:8000) as a BaseAgent so any
specialist agent can query Odoo 18 docs mid-execution via PeerBus
request_type=query_docs. Participates in sweep health checks.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Health endpoint called .ping() on both but neither implemented it,
causing ollama/odoo to always show as error and the bot to stay offline.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Routers calling /registry/agents raised AttributeError because
get_all() was not defined. Added method returning all registered
agents with active status, capabilities and instance flags.
The classifier was silently falling back to a clarification prompt every
time the LLM wrapped its JSON in markdown fences, prefixed it with
'json', or added surrounding prose. The bot then asked 'Could you
clarify what you need?' to every message regardless of clarity.
Now: strip code fences, slice to the first {...} block, and on parse
failure log the raw content (truncated) and treat the message as 'no
specialist agent' so the direct-answer fallback responds instead of
looping on clarification.
Previously when the LLM classified a message as needing no specialist
agent, the dispatcher built zero directives and _synthesize returned
'No agent responses received.' Greetings, follow-up clarifications,
and general questions all fell into this dead end.
Now when intent.agents is empty and no clarification is needed, the
master makes a second LLM call with the recent conversation as context
and answers directly. Updated master_system.txt to steer the classifier
toward agents=[] for chitchat instead of forcing a clarification loop.
The f-string only spanned the first fragment ('You don') so the
{chr(44).join(...)} placeholder leaked into chat output as literal
text. Build the message with plain string concat.
