lyx/youlu-openclaw-workspace
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

email processor

Browse Source
This commit is contained in:
Yanxin Lu
2026-02-26 20:54:07 -08:00
parent c5c9be3f00
commit b14a93866e
18 changed files with 1365 additions and 666 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
scripts/email_processor/.gitignore vendored Normal file
View File

@@ -0,0 +1,3 @@
__pycache__/
*.pyc
venv

233
scripts/email_processor/README.md Normal file
View File

@@ -0,0 +1,233 @@
# Email Processor
Learning-based mailbox cleanup using Himalaya (IMAP) + Ollama (local LLM). Classifies emails, learns from your decisions over time, and gradually automates common actions.
## Prerequisites
- **Himalaya** — CLI email client, handles IMAP connection and auth.
- **Ollama** — local LLM server.
- **Python 3.8+**
```bash
# Install himalaya (macOS)
brew install himalaya
# Configure himalaya for your IMAP account (first time only)
himalaya account list # should show your account after setup
# Install and start Ollama, pull the model
brew install ollama
ollama pull kamekichi128/qwen3-4b-instruct-2507:latest
# Set up Python venv and install dependencies
cd scripts/email_processor
python3 -m venv venv
source venv/bin/activate
pip install ollama
```
## How It Works
The system has two phases: a **learning phase** where it builds up knowledge from your decisions, and a **steady state** where it handles most emails automatically.
### Learning Phase (first ~20 decisions)
The confidence threshold is automatically raised to 95%. Most emails get queued.
1. **Cron runs `scan`.** For each unseen email, the classifier uses Qwen3's general knowledge (no history yet) to suggest an action. Most come back at 60-80% confidence — below the 95% threshold — so they get saved to `pending_emails.json` with the suggestion attached. A few obvious spam emails might hit 95%+ and get auto-deleted.
2. **You run `review list`.** It prints what's pending:
```
1. [msg_f1d43ea3] Subject: New jobs matching your profile
From: LinkedIn Suggested: delete (82%)
2. [msg_60c56a87] Subject: Your order shipped
From: Amazon Suggested: archive (78%)
3. [msg_ebd24205] Subject: Meeting tomorrow at 3pm
From: Coworker Suggested: keep (70%)
```
3. **You act on them.** Either individually or in bulk:
```bash
./email-processor.sh review 1 delete # agree with suggestion
./email-processor.sh review 2 archive # agree with suggestion
./email-processor.sh review accept # accept all suggestions at once
```
Each command executes via himalaya, appends to `decision_history.json`, and marks the pending entry as done.
4. **Next scan is smarter.** The classifier now has few-shot examples in the prompt:
```
History for linkedin.com: delete 2 times
--- Past decisions ---
From: LinkedIn | Subject: New jobs matching your profile -> delete
From: Amazon | Subject: Your package delivered -> archive
---
```
Confidence scores climb. You keep reviewing. History grows.
### Steady State (20+ decisions)
The threshold drops to the configured 75%. The classifier has rich context.
- **Repeat senders** (LinkedIn, Amazon, Uber) get auto-acted at 85-95% confidence during `scan`. They never touch the pending queue.
- **New or ambiguous senders** may fall below 75% and get queued.
- **You occasionally run `review list`** to handle stragglers — each decision further improves future classifications.
- **`stats` shows your automation rate** climbing: 60%, 70%, 80%+.
The pending queue shrinks over time. It's not a backlog — it's an ever-narrowing set of emails the system hasn't learned to handle yet.
## Usage
All commands are non-interactive — they take arguments, act, and exit. Compatible with cron/OpenClaw.
```bash
# Make the entry script executable (first time)
chmod +x email-processor.sh
# --- Scan ---
./email-processor.sh scan # classify unseen emails
./email-processor.sh scan --recent 30 # classify last 30 days
./email-processor.sh scan --dry-run # classify only, no changes
./email-processor.sh scan --recent 7 --dry-run # combine both
# --- Review ---
./email-processor.sh review list # show pending queue
./email-processor.sh review 1 delete # delete email #1
./email-processor.sh review msg_f1d43ea3 archive # archive by ID
./email-processor.sh review all delete # delete all pending
./email-processor.sh review accept # accept all suggestions
# --- Other ---
./email-processor.sh stats # show decision history
./email-processor.sh migrate # import old decisions
```
Or call Python directly: `python main.py scan --dry-run`
## Actions
| Action | Effect |
|---|---|
| `delete` | Move to Trash (`himalaya message delete`) |
| `archive` | Move to Archive folder |
| `keep` | Leave unread in inbox (no changes) |
| `mark_read` | Add `\Seen` flag, stays in inbox |
| `label:<name>` | Move to named folder (created if needed) |
## Auto-Action Criteria
Scan auto-acts when the classifier's confidence meets the threshold. During the learning phase (fewer than `bootstrap_min_decisions` total decisions, default 20), a higher threshold of 95% is used automatically. Once enough history accumulates, the configured `confidence_threshold` (default 75%) takes over.
This means on day one, only very obvious emails (spam, clear promotions) get auto-acted. As you review emails and build history, the system gradually handles more on its own.
## Configuration
`config.json` — only Ollama and automation settings. IMAP auth is managed by himalaya's own config.
```json
{
"ollama": {
"host": "http://localhost:11434",
"model": "kamekichi128/qwen3-4b-instruct-2507:latest"
},
"rules": {
"max_body_length": 1000
},
"automation": {
"confidence_threshold": 75,
"bootstrap_min_decisions": 20
}
}
```
| Key | Description |
|---|---|
| `ollama.host` | Ollama server URL. Default `http://localhost:11434`. |
| `ollama.model` | Ollama model to use for classification. |
| `rules.max_body_length` | Max characters of email body sent to the LLM. Longer bodies are truncated. Keeps prompt size and latency down. |
| `automation.confidence_threshold` | Minimum confidence (0-100) for auto-action in steady state. Emails below this get queued for review. Lower = more automation, higher = more manual review. |
| `automation.bootstrap_min_decisions` | Number of decisions needed before leaving the learning phase. During the learning phase, the threshold is raised to 95% regardless of `confidence_threshold`. Set to 0 to skip the learning phase entirely. |
## Testing
```bash
# 1. Verify himalaya can reach your mailbox
himalaya envelope list --page-size 3
# 2. Verify Ollama is running with the model
ollama list # should show kamekichi128/qwen3-4b-instruct-2507:latest
# 3. Dry run — classify recent emails without touching anything
./email-processor.sh scan --recent 7 --dry-run
# 4. Live run — classify and act (auto-act or queue)
./email-processor.sh scan --recent 7
# 5. Check what got queued
./email-processor.sh review list
# 6. Act on a queued email to seed decision history
./email-processor.sh review 1 delete
# 7. Check that the decision was recorded
./email-processor.sh stats
```
## File Structure
```
email_processor/
main.py # Entry point — scan/review/stats/migrate subcommands
classifier.py # LLM prompt builder + response parser
decision_store.py # Decision history storage + few-shot retrieval
config.json # Ollama + automation settings
email-processor.sh # Shell wrapper (activates venv, forwards args)
data/
pending_emails.json # Queue of emails awaiting review
decision_history.json # Past decisions (few-shot learning data)
logs/
YYYY-MM-DD.log # Daily processing logs
```
## Design Decisions
### Himalaya instead of raw IMAP
All IMAP operations go through the `himalaya` CLI via subprocess calls. This means:
- No IMAP credentials stored in config.json — himalaya manages its own auth.
- No connection management, reconnect logic, or SSL setup in Python.
- Each action is a single himalaya command (e.g., `himalaya message delete 42`).
The tradeoff is a subprocess spawn per operation, but for email volumes (tens per run, not thousands) this is negligible.
### Non-interactive design
Every command takes its full input as arguments, acts, and exits. No `input()` calls, no interactive loops. This makes the system compatible with cron/OpenClaw and composable with other scripts. The pending queue on disk (`pending_emails.json`) is the shared state between scan and review invocations.
### decision_history.json as the "database"
`data/decision_history.json` is the only persistent state that matters for learning. It's a flat JSON array — every decision (user or auto) is appended as an entry. The classifier reads the whole file on each email to find relevant few-shot examples via relevance scoring.
The pending queue (`pending_emails.json`) is transient — emails pass through it and get marked "done". Logs are for debugging. The decision history is what the system learns from.
A flat JSON file works fine for hundreds or low thousands of decisions. SQLite would make sense if the history grows past ~10k entries and the linear scan becomes noticeable, or if concurrent writes from multiple processes become necessary. Neither applies at current scale.
### Few-shot learning via relevance scoring
Rather than sending the entire decision history to the LLM, `decision_store.get_relevant_examples()` scores each past decision against the current email using three signals:
- Exact sender domain match (+3 points)
- Recipient address match (+2 points)
- Subject keyword overlap (+1 per shared word, stop-words excluded)
The top 5 most relevant examples are injected into the prompt as few-shot demonstrations. This keeps the prompt small while giving the model the most useful context.
### Conservative auto-action
Auto-action uses a single confidence threshold with an adaptive learning phase. When the decision history has fewer than `bootstrap_min_decisions` (default 20) entries, the threshold is raised to 95% — only very obvious classifications get auto-acted. Once enough history accumulates, the configured `confidence_threshold` (default 75%) takes over. This lets the system start working from day one while being cautious until it has enough examples to learn from.
### `keep` means unread
The `keep` action is a deliberate no-op — it leaves the email unread in the inbox, meaning it needs human attention. This is distinct from `mark_read`, which dismisses low-priority emails without moving them. During scan, queued emails are marked as read to prevent re-processing, but that's a scan-level concern separate from the `keep` action itself.
### Fail-safe classification
If the LLM call fails (Ollama down, model not loaded, timeout), the classifier returns `action="keep"` with `confidence=0`. This guarantees the email gets queued for manual review rather than being auto-acted upon. The system never auto-trashes an email it couldn't classify.

191
scripts/email_processor/classifier.py Normal file
View File

@@ -0,0 +1,191 @@
#!/usr/bin/env python3
"""
Classifier - LLM-based email classification with learning.
This module builds a rich prompt for the local Ollama model (Qwen3) that
includes few-shot examples from past user decisions, per-sender statistics,
and a list of known labels. The model returns a structured response with
an action, confidence score, summary, and reason.
The prompt structure:
1. System instructions (action definitions)
2. Known labels (so the model reuses them)
3. Sender statistics ("linkedin.com: deleted 8 times, kept 2 times")
4. Few-shot examples (top 5 most relevant past decisions)
5. The email to classify (subject, sender, recipient, body preview)
6. Output format specification
"""
import time
from datetime import datetime
from pathlib import Path
import decision_store
LOGS_DIR = Path(__file__).parent / "logs"
def _build_prompt(email_data, config):
"""Assemble the full classification prompt with learning context.
The prompt is built in sections, each providing different context to
help the model make better decisions. Sections are omitted when there
is no relevant data (e.g., no history yet for a new sender).
"""
max_body = config.get("rules", {}).get("max_body_length", 1000)
# Gather learning context from decision history
examples = decision_store.get_relevant_examples(email_data, n=10)
sender_domain = decision_store._extract_domain(email_data.get("sender", ""))
sender_stats = decision_store.get_sender_stats(sender_domain) if sender_domain else {}
known_labels = decision_store.get_known_labels()
# /no_think disables Qwen3's chain-of-thought, giving faster + shorter output
parts = ["/no_think\n"]
# Section 1: Action definitions
parts.append(
"You are an email classifier. Classify the email into one of these actions:\n"
"- delete: Spam, ads, promotions, unwanted notifications\n"
"- archive: Informational emails worth keeping but not needing attention "
"(receipts, shipping updates, automated confirmations)\n"
"- keep: Important emails that need attention or action (left unread in inbox)\n"
"- mark_read: Low-priority, leave in inbox but mark as read\n"
"- label:<name>: Categorize with a specific label\n"
)
# Section 2: Known labels (helps model reuse instead of inventing)
if known_labels:
parts.append(f"\nLabels used before: {', '.join(sorted(known_labels))}\n")
# Section 3: Sender statistics (strong signal for repeat senders)
if sender_stats:
stats_str = ", ".join(
f"{action} {count} times" for action, count in sender_stats.items()
)
parts.append(f"\nHistory for {sender_domain}: {stats_str}\n")
# Section 4: Few-shot examples (top 5 most relevant past decisions)
if examples:
parts.append("\n--- Past decisions (learn from these) ---")
for ex in examples[:5]:
parts.append(
f"From: {ex['sender'][:60]} | To: {ex['recipient'][:40]} | "
f"Subject: {ex['subject'][:60]} -> {ex['action']}"
)
parts.append("--- End examples ---\n")
# Section 5: The email being classified
body_preview = email_data.get("body", "")[:max_body]
parts.append(
f"Now classify this email:\n"
f"Subject: {email_data.get('subject', '(No Subject)')}\n"
f"From: {email_data.get('sender', '(Unknown)')}\n"
f"To: {email_data.get('recipient', '(Unknown)')}\n"
f"Body: {body_preview}\n"
)
# Section 6: Required output format
parts.append(
"Respond in this exact format (nothing else):\n"
"Action: [delete|archive|keep|mark_read|label:<name>]\n"
"Confidence: [0-100]\n"
"Summary: [one sentence summary of the email]\n"
"Reason: [brief explanation for your classification]"
)
return "\n".join(parts)
def _log_llm(prompt, output, email_data, action, confidence, duration):
"""Log the full LLM prompt and response to logs/llm_YYYY-MM-DD.log."""
LOGS_DIR.mkdir(exist_ok=True)
log_file = LOGS_DIR / f"llm_{datetime.now().strftime('%Y-%m-%d')}.log"
timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
subject = email_data.get("subject", "(No Subject)")[:60]
sender = email_data.get("sender", "(Unknown)")[:60]
with open(log_file, "a", encoding="utf-8") as f:
f.write(f"{'=' * 70}\n")
f.write(f"[{timestamp}] {subject}\n")
f.write(f"From: {sender} | Result: {action} @ {confidence}% | {duration:.1f}s\n")
f.write(f"{'-' * 70}\n")
f.write(f"PROMPT:\n{prompt}\n")
f.write(f"{'-' * 70}\n")
f.write(f"RESPONSE:\n{output}\n")
f.write(f"{'=' * 70}\n\n")
def _parse_response(output):
"""Parse the model's text response into structured fields.
Expected format (one per line):
Action: delete
Confidence: 92
Summary: Promotional offer from retailer
Reason: Clearly a marketing email with discount offer
Falls back to safe defaults (keep, 50% confidence) on parse failure.
"""
action = "keep"
confidence = 50
summary = "No summary"
reason = "Unknown"
for line in output.strip().split("\n"):
line = line.strip()
if line.startswith("Action:"):
raw_action = line.replace("Action:", "").strip().lower()
valid_actions = {"delete", "archive", "keep", "mark_read"}
if raw_action in valid_actions or raw_action.startswith("label:"):
action = raw_action
elif line.startswith("Confidence:"):
try:
confidence = int(line.replace("Confidence:", "").strip().rstrip("%"))
confidence = max(0, min(100, confidence)) # clamp to 0-100
except ValueError:
confidence = 50
elif line.startswith("Summary:"):
summary = line.replace("Summary:", "").strip()[:200]
elif line.startswith("Reason:"):
reason = line.replace("Reason:", "").strip()
return action, confidence, summary, reason
def classify_email(email_data, config):
"""Classify an email using the local LLM with few-shot learning context.
Connects to Ollama, sends the assembled prompt, and parses the response.
On any error, falls back to "keep" with 0% confidence so the email
gets queued for manual review rather than auto-acted upon.
Args:
email_data: dict with subject, sender, recipient, body keys.
config: full config dict (needs ollama.model and rules.max_body_length).
Returns:
Tuple of (action, confidence, summary, reason, duration_seconds).
"""
import ollama
prompt = _build_prompt(email_data, config)
model = config.get("ollama", {}).get("model", "kamekichi128/qwen3-4b-instruct-2507:latest")
start_time = time.time()
try:
# Low temperature for consistent classification
response = ollama.generate(model=model, prompt=prompt, options={"temperature": 0.1})
output = response["response"]
action, confidence, summary, reason = _parse_response(output)
except Exception as e:
# On failure, default to "keep" with 0 confidence -> always queued
output = f"ERROR: {e}"
action = "keep"
confidence = 0
summary = "Classification failed"
reason = f"error - {str(e)[:100]}"
duration = time.time() - start_time
_log_llm(prompt, output, email_data, action, confidence, duration)
return action, confidence, summary, reason, duration

12
scripts/email_processor/config.json
View File

@@ -1,16 +1,14 @@
{
"imap": {
"host": "imap.migadu.com",
"port": 993,
"email": "youlu@luyanxin.com",
"password": "kDkNau2r7m.hV!uk*D4Yr8mC7Dyjx9T"
},
"ollama": {
"host": "http://localhost:11434",
"model": "qwen3:4b"
"model": "kamekichi128/qwen3-4b-instruct-2507:latest"
},
"rules": {
"max_body_length": 1000,
"check_unseen_only": true
},
"automation": {
"confidence_threshold": 75,
"bootstrap_min_decisions": 20
}
}

52
scripts/email_processor/data/pending_emails.json
View File

@@ -1,52 +0,0 @@
{
"msg_f1d43ea3": {
"imap_uid": "2",
"subject": "Delivered: \"Voikinfo Bottom Gusset Bags...\"",
"sender": "\"Amazon.com - order-update(a)amazon.com\"\r\n <order-update_at_amazon_com_posyo@simplelogin.co>",
"recipient": "sho.amazon@ylu17.com",
"summary": "Your Amazon package (order #114-1496788-7649829) was delivered today to Argo, Los Angeles, CA and left near the front door or porch.",
"email_date": "Wed, 18 Feb 2026 04:15:24 +0000",
"status": "pending",
"found_at": "2026-02-18T16:18:42.347538"
},
"msg_60c56a87": {
"imap_uid": "3",
"subject": "=?UTF-8?b?5L2V5LiN5ruh6Laz6Ieq5bex55qE5Y+j6IW55LmL5qyy?=",
"sender": "\"Uber Eats - uber(a)uber.com\" <uber_at_uber_com_kjwzyhxn@simplelogin.co>",
"recipient": "uber@ylu17.com",
"summary": "Uber Eats has sent a notification that the user's order is ready for pickup.",
"email_date": "Wed, 18 Feb 2026 11:36:59 +0000",
"status": "pending",
"found_at": "2026-02-18T08:05:56.594842"
},
"msg_ebd24205": {
"imap_uid": "4",
"subject": "Your order has been shipped (or closed if combined/delivered).",
"sender": "\"cd(a)woodenswords.com\"\r\n <cd_at_woodenswords_com_xivwijojc@simplelogin.co>",
"recipient": "mail@luyx.org",
"summary": "This email confirms that your order has been shipped or closed (if combined/delivered).",
"email_date": "Wed, 18 Feb 2026 16:07:58 +0000",
"status": "pending",
"found_at": "2026-02-18T12:01:19.048091"
},
"msg_fa73b3bd": {
"imap_uid": "6",
"subject": "=?UTF-8?Q?Yanxin,_I=E2=80=99m_still_waiting_for_your_response?=",
"sender": "\"Arslan (via LinkedIn) - messages-noreply(a)linkedin.com\"\r\n <messages-noreply_at_linkedin_com_ajpnalmwp@simplelogin.co>",
"recipient": "Yanxin Lu <acc.linkedin@ylu17.com>",
"summary": "Arslan Ahmed, a Senior AI | ML | Full Stack Engineer from Ilford, invited you to connect on February 11, 2026 at 10:08 PM and is waiting for your response.",
"email_date": "Wed, 18 Feb 2026 18:53:45 +0000 (UTC)",
"status": "pending",
"found_at": "2026-02-18T12:04:34.602407"
},
"msg_59f23736": {
"imap_uid": "1",
"subject": "New Software Engineer jobs that match your profile",
"sender": "\"LinkedIn - jobs-noreply(a)linkedin.com\"\r\n <jobs-noreply_at_linkedin_com_zuwggfxh@simplelogin.co>",
"recipient": "Yanxin Lu <acc.linkedin@ylu17.com>",
"summary": "LinkedIn has notified the user of new software engineering jobs that match their profile and includes a link to update their top card.",
"email_date": "Wed, 18 Feb 2026 02:07:12 +0000 (UTC)",
"status": "pending",
"found_at": "2026-02-18T16:16:00.784822"
}
}

253
scripts/email_processor/decision_store.py Normal file
View File

@@ -0,0 +1,253 @@
#!/usr/bin/env python3
"""
Decision Store - Manages decision history for learning-based email classification.
This module persists every user and auto-made decision to a flat JSON file
(data/decision_history.json). Past decisions serve as few-shot examples
that are injected into the LLM prompt by classifier.py, enabling the
system to learn from user behavior over time.
Storage format: a JSON array of decision entries, each containing sender,
recipient, subject, summary, action taken, and whether it was a user or
auto decision.
"""
import json
import re
from datetime import datetime
from pathlib import Path
from collections import Counter
# ---------------------------------------------------------------------------
# Paths
# ---------------------------------------------------------------------------
SCRIPT_DIR = Path(__file__).parent
DATA_DIR = SCRIPT_DIR / "data"
HISTORY_FILE = DATA_DIR / "decision_history.json"
PENDING_FILE = DATA_DIR / "pending_emails.json"
# Stop-words excluded from subject keyword matching to reduce noise.
_STOP_WORDS = {"re", "fwd", "the", "a", "an", "is", "to", "for", "and", "or", "your", "you"}
# ---------------------------------------------------------------------------
# Internal helpers
# ---------------------------------------------------------------------------
def _load_history():
"""Load the full decision history list from disk."""
if not HISTORY_FILE.exists():
return []
with open(HISTORY_FILE, "r", encoding="utf-8") as f:
return json.load(f)
def _save_history(history):
"""Write the full decision history list to disk."""
DATA_DIR.mkdir(exist_ok=True)
with open(HISTORY_FILE, "w", encoding="utf-8") as f:
json.dump(history, f, indent=2, ensure_ascii=False)
def _extract_domain(sender):
"""Extract the domain part from a sender string.
Handles formats like:
"Display Name <user@example.com>"
user@example.com
"""
match = re.search(r"[\w.+-]+@([\w.-]+)", sender)
return match.group(1).lower() if match else ""
def _extract_email_address(sender):
"""Extract the full email address from a sender string."""
match = re.search(r"([\w.+-]+@[\w.-]+)", sender)
return match.group(1).lower() if match else sender.lower()
# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------
def record_decision(email_data, action, source="user"):
"""Append a decision to the history file.
Args:
email_data: dict with keys: sender, recipient, subject, summary.
action: one of "delete", "archive", "keep", "mark_read",
or "label:<name>".
source: "user" (manual review) or "auto" (high-confidence).
"""
history = _load_history()
entry = {
"timestamp": datetime.now().isoformat(timespec="seconds"),
"sender": email_data.get("sender", ""),
"sender_domain": _extract_domain(email_data.get("sender", "")),
"recipient": email_data.get("recipient", ""),
"subject": email_data.get("subject", ""),
"summary": email_data.get("summary", ""),
"action": action,
"source": source,
}
history.append(entry)
_save_history(history)
return entry
def get_relevant_examples(email_data, n=10):
"""Find the N most relevant past decisions for a given email.
Relevance is scored by three signals:
- Exact sender domain match: +3 points
- Recipient string match: +2 points
- Subject keyword overlap: +1 point per shared word
Only entries with score > 0 are considered. Results are returned
sorted by descending relevance.
"""
history = _load_history()
if not history:
return []
target_domain = _extract_domain(email_data.get("sender", ""))
target_recipient = email_data.get("recipient", "").lower()
target_words = (
set(re.findall(r"\w+", email_data.get("subject", "").lower())) - _STOP_WORDS
)
scored = []
for entry in history:
score = 0
# Signal 1: sender domain match
if target_domain and entry.get("sender_domain", "") == target_domain:
score += 3
# Signal 2: recipient substring match
if target_recipient and target_recipient in entry.get("recipient", "").lower():
score += 2
# Signal 3: subject keyword overlap
entry_words = (
set(re.findall(r"\w+", entry.get("subject", "").lower())) - _STOP_WORDS
)
score += len(target_words & entry_words)
if score > 0:
scored.append((score, entry))
scored.sort(key=lambda x: x[0], reverse=True)
return [entry for _, entry in scored[:n]]
def get_sender_stats(sender_domain):
"""Get action distribution for a sender domain.
Returns a dict like {"delete": 5, "keep": 2, "archive": 1}.
"""
history = _load_history()
actions = Counter()
for entry in history:
if entry.get("sender_domain", "") == sender_domain:
actions[entry["action"]] += 1
return dict(actions)
def get_sender_history_count(sender_domain):
"""Count total past decisions for a sender domain.
Used by the scan command to decide whether there is enough history
to trust auto-actions for this sender.
"""
history = _load_history()
return sum(1 for e in history if e.get("sender_domain", "") == sender_domain)
def get_known_labels():
"""Return the set of all label names used in past "label:<name>" decisions.
These are offered to the LLM so it can reuse existing labels rather
than inventing new ones.
"""
history = _load_history()
labels = set()
for entry in history:
action = entry.get("action", "")
if action.startswith("label:"):
labels.add(action[6:])
return labels
def get_all_stats():
"""Compute aggregate statistics across the full decision history.
Returns a dict with keys: total, by_action, by_source, top_domains.
Returns None if history is empty.
"""
history = _load_history()
if not history:
return None
total = len(history)
by_action = Counter(e["action"] for e in history)
by_source = Counter(e["source"] for e in history)
# Top 10 sender domains by decision count
domain_counts = Counter(e.get("sender_domain", "") for e in history)
top_domains = domain_counts.most_common(10)
return {
"total": total,
"by_action": dict(by_action),
"by_source": dict(by_source),
"top_domains": top_domains,
}
# ---------------------------------------------------------------------------
# Migration
# ---------------------------------------------------------------------------
def migrate_pending():
"""One-time migration: import 'done' entries from pending_emails.json.
Converts old-style action names ("archived" -> "archive", etc.) and
records them as user decisions in the history file. Safe to run
multiple times (will create duplicates though, so run once only).
"""
if not PENDING_FILE.exists():
print("No pending_emails.json found, nothing to migrate.")
return 0
with open(PENDING_FILE, "r", encoding="utf-8") as f:
pending = json.load(f)
# Map old action names to new ones
action_map = {
"archived": "archive",
"kept": "keep",
"deleted": "delete",
}
migrated = 0
for msg_id, data in pending.items():
if data.get("status") != "done":
continue
old_action = data.get("action", "")
action = action_map.get(old_action, old_action)
if not action:
continue
email_data = {
"sender": data.get("sender", ""),
"recipient": data.get("recipient", ""),
"subject": data.get("subject", ""),
"summary": data.get("summary", ""),
}
record_decision(email_data, action, source="user")
migrated += 1
print(f"Migrated {migrated} decisions from pending_emails.json")
return migrated

27
scripts/email_processor/email-processor.sh Executable file
View File

@@ -0,0 +1,27 @@
#!/usr/bin/env bash
# email-processor — wrapper script for the email processor.
#
# Usage:
# ./email-processor.sh scan # classify unseen emails
# ./email-processor.sh scan --recent 30 # last 30 days
# ./email-processor.sh scan --dry-run # classify only, no changes
# ./email-processor.sh scan --recent 7 --dry-run # combine both
# ./email-processor.sh review list # show pending queue
# ./email-processor.sh review 1 delete # act on email #1
# ./email-processor.sh review all delete # act on all pending
# ./email-processor.sh review accept # accept all suggestions
# ./email-processor.sh stats # show history stats
# ./email-processor.sh migrate # import old decisions
#
# Requires: Python 3.8+, himalaya, Ollama running with model.
set -euo pipefail
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
# Activate the virtualenv if it exists
if [ -d "$SCRIPT_DIR/venv" ]; then
source "$SCRIPT_DIR/venv/bin/activate"
fi
exec python3 "$SCRIPT_DIR/main.py" "$@"

50
scripts/email_processor/logs/2026-02-15.log
View File

@@ -1,50 +0,0 @@
[2026-02-15 21:14:02] KEPT: Please confirm your mailbox youlu@luyanxin.com
From: "noreply@simplelogin.io" <noreply@simplelogin.io>
Analysis: KEEP: Legitimate service confirmation email for mailbox addition (not promotional)
[2026-02-15 21:15:04] KEPT: =?utf-8?B?RndkOiBHZXQgMTAlIG9mZiB5b3VyIG5leHQgb3JkZXIg4pyF?=
From: "Yanxin Lu - crac1017(a)hotmail.com"
<crac1017_at_hotmail_com_fndbbu@simplelogin.co>
Analysis: KEEP: error - HTTPConnectionPool(host='localhost', port=11434): Read timed out. (read timeout=60)
[2026-02-15 21:15:37] KEPT:
=?utf-8?B?RndkOiDigJxzb2Z0d2FyZSBlbmdpbmVlcuKAnTogTWljcm9
From: "Yanxin Lu - crac1017(a)hotmail.com"
<crac1017_at_hotmail_com_fndbbu@simplelogin.co>
Analysis: KEEP: LinkedIn job alert notification for subscribed job search (not promotional)
[2026-02-15 21:15:52] KEPT: Fwd: Your receipt from OpenRouter, Inc #2231-9732
From: "Yanxin Lu - crac1017(a)hotmail.com"
<crac1017_at_hotmail_com_fndbbu@simplelogin.co>
Analysis: KEEP: This is a legitimate receipt for a payment made to OpenRouter, Inc (a known AI service provider), not promotional content.
[2026-02-15 21:16:10] KEPT: Fwd: Your ChatGPT code is 217237
From: "Yanxin Lu - crac1017(a)hotmail.com"
<crac1017_at_hotmail_com_fndbbu@simplelogin.co>
Analysis: KEEP: Legitimate security verification code from OpenAI (standard login confirmation)
[2026-02-15 22:49:44] KEPT (69.0s): =?UTF-8?B?5rWL6K+V6YKu5Lu2?=
From: Yanxin Lu <lyx@luyanxin.com>
Analysis: KEEP: Test email for delivery verification
From: Yanxin Lu <lyx@luyanxin.com>
Analysis: KEEP: Test email for delivery verification
[2026-02-15 22:57:03] MOVED_TO_TRASH (68.5s): =?utf-8?B?RndkOiBHZXQgMTAlIG9mZiB5b3VyIG5leHQgb3JkZXIg4pyF?=
From: "Yanxin Lu - crac1017(a)hotmail.com"
<crac1017_at_hotmail_com_fndbbu@simplelogin.co>
Analysis: AD: Forwarded Uber promotional offer
From: "Yanxin Lu - crac1017(a)hotmail.com"
<crac1017_at_hotmail_com_fndbbu@simplelogin.co>
Analysis: AD: Forwarded Uber promotional offer
[2026-02-15 23:00:09] KEPT (120.1s): Fwd: Your ChatGPT code is 217237
From: "Yanxin Lu - crac1017(a)hotmail.com"
<crac1017_at_hotmail_com_fndbbu@simplelogin.co>
Analysis: KEEP: error - HTTPConnectionPool(host='localhost', port=11434): Read timed out. (read timeout=120)
From: "Yanxin Lu - crac1017(a)hotmail.com"
<crac1017_at_hotmail_com_fndbbu@simplelogin.co>
Analysis: KEEP: error - HTTPConnectionPool(host='localhost', port=11434): Read timed out. (read timeout=120)

29
scripts/email_processor/logs/2026-02-18.log
View File

@@ -1,29 +0,0 @@
[2026-02-18 08:04:26] ADDED_TO_PENDING (msg_f1d43ea3) (108.6s): Delivered: "Voikinfo Bottom Gusset Bags..."
From: "Amazon.com - order-update(a)amazon.com"
<order-update_at_amazon_com_posyo@simplelogin.co>
Analysis: KEEP: Standard delivery confirmation from Amazon
[2026-02-18 08:05:56] ADDED_TO_PENDING (msg_60c56a87) (88.0s): =?UTF-8?b?5L2V5LiN5ruh6Laz6Ieq5bex55qE5Y+j6IW55LmL5qyy?=
From: "Uber Eats - uber(a)uber.com" <uber_at_uber_com_kjwzyhxn@simplelogin.co>
Analysis: KEEP: The decoded subject line "Your Uber Eats order is ready!" indicates a transactional order update, not an advertisement.
[2026-02-18 12:01:19] ADDED_TO_PENDING (msg_ebd24205) (66.7s): Your order has been shipped (or closed if combined/delivered
From: "cd(a)woodenswords.com"
<cd_at_woodenswords_com_xivwijojc@simplelogin.co>
Analysis: KEEP: System-generated shipping update notification from an e-commerce store, not promotional content.
[2026-02-18 12:03:36] MOVED_TO_TRASH (133.4s): =?UTF-8?Q?=E2=80=9Csoftware_engineer=E2=80=9D:_Snap_Inc._-_S
From: "LinkedIn Job Alerts - jobalerts-noreply(a)linkedin.com"
<jobalerts-noreply_at_linkedin_com_cnrlhok@simplelogin.co>
Analysis: AD: This email is a promotional job alert notification from LinkedIn's service for users who have set up job preferences.
[2026-02-18 12:04:34] ADDED_TO_PENDING (msg_fa73b3bd) (57.3s): =?UTF-8?Q?Yanxin,_I=E2=80=99m_still_waiting_for_your_respons
From: "Arslan (via LinkedIn) - messages-noreply(a)linkedin.com"
<messages-noreply_at_linkedin_com_ajpnalmwp@simplelogin.co>
Analysis: KEEP: This is a standard LinkedIn connection request notification with no promotional content, discounts, or advertisements—only a reminder of an existing invitation.
[2026-02-18 16:18:42] ADDED_TO_PENDING (msg_f1d43ea3) (102.1s): Delivered: "Voikinfo Bottom Gusset Bags..."
From: "Amazon.com - order-update(a)amazon.com"
<order-update_at_amazon_com_posyo@simplelogin.co>
Analysis: KEEP: Standard delivery confirmation from Amazon, not a promotional message.

887
scripts/email_processor/main.py
View File

@@ -1,297 +1,704 @@
#!/usr/bin/env python3
"""
Email Processor - Auto filter ads using local Qwen3
Moves ad emails to Trash folder (not permanently deleted)
Email Processor - Learning-based mailbox cleanup using Himalaya + Ollama.
Uses himalaya CLI for all IMAP operations (no raw imaplib, no stored
credentials). Uses a local Qwen3 model via Ollama for classification,
with few-shot learning from past user decisions.
All commands are non-interactive — they take arguments, mutate files on
disk, and exit. Suitable for cron (OpenClaw) and scripting.
Subcommands:
python main.py scan # classify unseen emails
python main.py scan --recent 30 # classify last 30 days
python main.py scan --dry-run # classify only, no changes
python main.py scan --recent 7 --dry-run # combine both
python main.py review list # print pending queue
python main.py review <num-or-id> <action> # act on one email
python main.py review all <action> # act on all pending
python main.py review accept # accept all suggestions
python main.py stats # show decision history
python main.py migrate # import old decisions
Action mapping (what each classification does to the email):
delete -> himalaya message delete <id> (moves to Trash)
archive -> himalaya message move Archive <id>
keep -> no-op (leave unread in inbox)
mark_read -> himalaya flag add <id> seen
label:X -> himalaya message move <X> <id>
"""
import json
import imaplib
import email
import os
import subprocess
import hashlib
import sys
from datetime import datetime
from datetime import datetime, timedelta
from pathlib import Path
# Config
import classifier
import decision_store
# ---------------------------------------------------------------------------
# Paths — all relative to the script's own directory
# ---------------------------------------------------------------------------
SCRIPT_DIR = Path(__file__).parent
CONFIG_FILE = SCRIPT_DIR / "config.json"
LOGS_DIR = SCRIPT_DIR / "logs"
DATA_DIR = SCRIPT_DIR / "data"
PENDING_FILE = DATA_DIR / "pending_emails.json"
# ---------------------------------------------------------------------------
# Config
# ---------------------------------------------------------------------------
def load_config():
"""Load configuration"""
"""Load config.json from the script directory.
Only ollama, rules, and automation settings are needed — himalaya
manages its own IMAP config separately.
"""
with open(CONFIG_FILE) as f:
return json.load(f)
def connect_imap(config):
"""Connect to IMAP server"""
imap_config = config['imap']
mail = imaplib.IMAP4_SSL(imap_config['host'], imap_config['port'])
mail.login(imap_config['email'], imap_config['password'])
return mail
def get_unseen_emails(mail):
"""Get list of unseen email IDs"""
mail.select('INBOX')
_, search_data = mail.search(None, 'UNSEEN')
email_ids = search_data[0].split()
return email_ids
# ---------------------------------------------------------------------------
# Himalaya CLI wrappers
#
# All IMAP operations go through himalaya, which handles connection,
# auth, and protocol details. We call it as a subprocess and parse
# its JSON output.
# ---------------------------------------------------------------------------
def fetch_email(mail, email_id):
"""Fetch email content"""
_, msg_data = mail.fetch(email_id, '(RFC822)')
raw_email = msg_data[0][1]
msg = email.message_from_bytes(raw_email)
# Extract subject
subject = msg['Subject'] or '(No Subject)'
# Extract sender
sender = msg['From'] or '(Unknown)'
# Extract recipient
recipient = msg['To'] or '(Unknown)'
# Extract date
date = msg['Date'] or datetime.now().isoformat()
# Extract body
body = ""
if msg.is_multipart():
for part in msg.walk():
if part.get_content_type() == "text/plain":
try:
body = part.get_payload(decode=True).decode('utf-8', errors='ignore')
break
except:
pass
def _himalaya(*args):
"""Run a himalaya command and return its stdout.
Raises subprocess.CalledProcessError on failure.
"""
result = subprocess.run(
["himalaya", *args],
capture_output=True, text=True, check=True,
)
return result.stdout
def _himalaya_json(*args):
"""Run a himalaya command with JSON output and return parsed result."""
return json.loads(_himalaya("-o", "json", *args))
# ---------------------------------------------------------------------------
# Email fetching via himalaya
# ---------------------------------------------------------------------------
def get_unseen_envelopes():
"""Fetch envelope metadata for all unseen emails in INBOX.
Returns a list of envelope dicts from himalaya's JSON output.
Each has keys like: id, subject, from, to, date, flags.
"""
return _himalaya_json("envelope", "list", "not", "flag", "seen")
def get_recent_envelopes(days):
"""Fetch envelope metadata for all emails from the last N days.
Includes both read and unread emails — useful for testing and
bulk-classifying historical mail.
"""
since = (datetime.now() - timedelta(days=days)).strftime("%Y-%m-%d")
return _himalaya_json("envelope", "list", "after", since)
def read_message(envelope_id):
"""Read the full message body without marking it as seen.
The --preview flag prevents himalaya from adding the \\Seen flag,
so the email stays unread for the actual action to handle.
"""
# Read plain text, no headers, without marking as seen
return _himalaya("message", "read", "--preview", "--no-headers", str(envelope_id))
def build_email_data(envelope, body, config):
"""Build the email_data dict expected by classifier and decision_store.
Combines envelope metadata (from himalaya envelope list) with the
message body (from himalaya message read).
"""
max_body = config.get("rules", {}).get("max_body_length", 1000)
# himalaya envelope JSON uses "from" as a nested object or string
sender = envelope.get("from", {})
if isinstance(sender, dict):
# Format: {"name": "Display Name", "addr": "user@example.com"}
name = sender.get("name", "")
addr = sender.get("addr", "")
sender_str = f"{name} <{addr}>" if name else addr
elif isinstance(sender, list) and sender:
first = sender[0]
name = first.get("name", "")
addr = first.get("addr", "")
sender_str = f"{name} <{addr}>" if name else addr
else:
try:
body = msg.get_payload(decode=True).decode('utf-8', errors='ignore')
except:
pass
sender_str = str(sender)
# Same for "to"
to = envelope.get("to", {})
if isinstance(to, dict):
name = to.get("name", "")
addr = to.get("addr", "")
to_str = f"{name} <{addr}>" if name else addr
elif isinstance(to, list) and to:
first = to[0]
name = first.get("name", "")
addr = first.get("addr", "")
to_str = f"{name} <{addr}>" if name else addr
else:
to_str = str(to)
return {
'id': email_id,
'subject': subject,
'sender': sender,
'recipient': recipient,
'date': date,
'body': body[:300] # Limit body length
"id": str(envelope.get("id", "")),
"subject": envelope.get("subject", "(No Subject)"),
"sender": sender_str,
"recipient": to_str,
"date": envelope.get("date", ""),
"body": body[:max_body],
}
def analyze_with_qwen3(email_data, config):
"""Analyze email with local Qwen3 using official library"""
import ollama
import time
prompt = f"""/no_think
Analyze this email and provide two pieces of information:
# ---------------------------------------------------------------------------
# IMAP actions via himalaya
#
# Each function executes one himalaya command. Returns True on success.
# On failure, prints the error and returns False.
# ---------------------------------------------------------------------------
1. Is this an advertisement/promotional email?
2. Summarize the email in one sentence
def execute_action(envelope_id, action):
"""Dispatch an action string to the appropriate himalaya command.
Email details:
Subject: {email_data['subject']}
Sender: {email_data['sender']}
Body: {email_data['body'][:300]}
Action mapping:
"delete" -> himalaya message delete <id>
"archive" -> himalaya message move Archive <id>
"keep" -> no-op (leave unread in inbox)
"mark_read" -> himalaya flag add <id> seen
"label:X" -> himalaya message move <X> <id>
Respond in this exact format:
IsAD: [YES or NO]
Summary: [one sentence summary]
Reason: [brief explanation]
"""
start_time = time.time()
model = config['ollama'].get('model', 'qwen3:4b')
Returns True on success, False on failure.
"""
eid = str(envelope_id)
try:
response = ollama.generate(model=model, prompt=prompt, options={'temperature': 0.1})
output = response['response']
# Parse output
is_ad = False
summary = "No summary"
reason = "Unknown"
for line in output.strip().split('\n'):
if line.startswith('IsAD:'):
is_ad = 'YES' in line.upper()
elif line.startswith('Summary:'):
summary = line.replace('Summary:', '').strip()[:200]
elif line.startswith('Reason:'):
reason = line.replace('Reason:', '').strip()
if is_ad:
result = f"AD: {reason}"
if action == "delete":
_himalaya("message", "delete", eid)
elif action == "archive":
_himalaya("message", "move", "Archive", eid)
elif action == "keep":
pass # leave unread in inbox — no IMAP changes
elif action == "mark_read":
_himalaya("flag", "add", eid, "seen")
elif action.startswith("label:"):
folder = action[6:]
_himalaya("message", "move", folder, eid)
else:
result = f"KEEP: {reason}"
except Exception as e:
result = f"KEEP: error - {str(e)[:100]}"
summary = "Analysis failed"
is_ad = False
duration = time.time() - start_time
return result, summary, is_ad, duration
def move_to_trash(mail, email_id):
"""Move email to Trash folder"""
# Copy to Trash
result = mail.copy(email_id, 'Trash')
if result[0] == 'OK':
# Mark original as deleted
mail.store(email_id, '+FLAGS', '\\Deleted')
print(f" Unknown action: {action}")
return False
return True
return False
except subprocess.CalledProcessError as e:
print(f" Himalaya error: {e.stderr.strip()}")
return False
def log_result(log_file, email_data, analysis, action, duration=None):
"""Log processing result with Qwen3 duration"""
timestamp = datetime.now().strftime('%Y-%m-%d %H:%M:%S')
duration_str = f" ({duration:.1f}s)" if duration else ""
with open(log_file, 'a') as f:
f.write(f"[{timestamp}] {action}{duration_str}: {email_data['subject'][:60]}\n")
f.write(f" From: {email_data['sender']}\n")
f.write(f" Analysis: {analysis}\n\n")
# ---------------------------------------------------------------------------
# Pending queue — emails awaiting manual review
#
# Stored as a JSON dict in data/pending_emails.json, keyed by msg_id.
# Each entry tracks the envelope ID (for himalaya), classifier suggestion,
# and status (pending/done).
# ---------------------------------------------------------------------------
def load_pending():
"""Load pending emails from JSON file"""
"""Load the pending queue from disk."""
if not PENDING_FILE.exists():
return {}
with open(PENDING_FILE, 'r', encoding='utf-8') as f:
with open(PENDING_FILE, "r", encoding="utf-8") as f:
return json.load(f)
def save_pending(pending):
"""Save pending emails to JSON file"""
"""Write the pending queue to disk."""
DATA_DIR.mkdir(exist_ok=True)
with open(PENDING_FILE, 'w', encoding='utf-8') as f:
with open(PENDING_FILE, "w", encoding="utf-8") as f:
json.dump(pending, f, indent=2, ensure_ascii=False)
def add_to_pending(email_data, summary, imap_uid, recipient):
"""Add email to pending queue"""
def add_to_pending(email_data, summary, reason, action_suggestion, confidence):
"""Add an email to the pending queue for manual review.
Stores the classifier's suggestion and confidence alongside the
email metadata so the user can see what the model thought.
"""
pending = load_pending()
# Generate unique ID
import hashlib
msg_id = f"msg_{hashlib.md5(f'{imap_uid}_{email_data['subject']}'.encode()).hexdigest()[:8]}"
# Extract date from email
email_date = email_data.get('date', datetime.now().isoformat())
# Generate a stable ID from envelope ID + subject
eid = str(email_data["id"])
key = f"{eid}_{email_data['subject']}"
msg_id = f"msg_{hashlib.md5(key.encode()).hexdigest()[:8]}"
pending[msg_id] = {
"imap_uid": str(imap_uid),
"subject": email_data['subject'],
"sender": email_data['sender'],
"recipient": recipient,
"envelope_id": eid,
"subject": email_data["subject"],
"sender": email_data["sender"],
"recipient": email_data.get("recipient", ""),
"summary": summary,
"email_date": email_date,
"reason": reason,
"suggested_action": action_suggestion,
"confidence": confidence,
"email_date": email_data.get("date", ""),
"status": "pending",
"found_at": datetime.now().isoformat()
"found_at": datetime.now().isoformat(),
}
save_pending(pending)
return msg_id
def main():
"""Main processing function"""
print("📧 Email Processor Starting...")
# Load config
config = load_config()
# Setup logging
# ---------------------------------------------------------------------------
# Logging
# ---------------------------------------------------------------------------
def log_result(log_file, email_data, action, detail, duration=None):
"""Append a one-line log entry for a processed email."""
timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
dur = f" ({duration:.1f}s)" if duration else ""
with open(log_file, "a") as f:
f.write(f"[{timestamp}] {action}{dur}: {email_data['subject'][:60]}\n")
f.write(f" From: {email_data['sender']}\n")
f.write(f" Detail: {detail}\n\n")
# ---------------------------------------------------------------------------
# Subcommand: scan
# ---------------------------------------------------------------------------
def cmd_scan(config, recent=None, dry_run=False):
"""Fetch emails, classify each one, then auto-act or queue.
Auto-action is based on a single confidence threshold. When the
decision history has fewer than 20 entries, a higher threshold (95%)
is used to be conservative during the learning phase. Once enough
history accumulates, the configured threshold takes over.
Args:
config: full config dict.
recent: if set, fetch emails from last N days (not just unseen).
dry_run: if True, classify and print but skip all actions.
"""
mode = "DRY RUN" if dry_run else "Scan"
print(f"Email Processor - {mode}")
print("=" * 50)
LOGS_DIR.mkdir(exist_ok=True)
log_file = LOGS_DIR / f"{datetime.now().strftime('%Y-%m-%d')}.log"
try:
# Connect to IMAP
print("Connecting to IMAP...")
mail = connect_imap(config)
print("✅ Connected")
# Get unseen emails
email_ids = get_unseen_emails(mail)
print(f"Found {len(email_ids)} unread emails")
if not email_ids:
print("No new emails to process")
mail.logout()
return
# Process each email
processed = 0
moved_to_trash = 0
added_to_pending = 0
for email_id in email_ids:
print(f"\nProcessing email {email_id.decode()}...")
# Fetch email
email_data = fetch_email(mail, email_id)
print(f" Subject: {email_data['subject'][:50]}")
# Analyze with Qwen3 (one call for both ad detection and summary)
analysis, summary, is_ad, duration = analyze_with_qwen3(email_data, config)
print(f" Analysis: {analysis[:100]}")
print(f" Summary: {summary[:60]}...")
print(f" Qwen3 time: {duration:.1f}s")
# Check if analysis was successful (not an error)
if 'error -' in analysis.lower():
# Analysis failed - keep email unread for retry
print(f" -> Analysis failed, keeping unread for retry")
log_result(log_file, email_data, analysis, "FAILED_RETRY", duration)
# Don't increment processed count - will retry next time
continue
# Analysis successful - determine action
if is_ad:
print(" -> Moving to Trash")
if move_to_trash(mail, email_id):
log_result(log_file, email_data, analysis, "MOVED_TO_TRASH", duration)
moved_to_trash += 1
else:
log_result(log_file, email_data, analysis, "MOVE_FAILED", duration)
# Load automation threshold
automation = config.get("automation", {})
configured_threshold = automation.get("confidence_threshold", 75)
# Adaptive threshold: be conservative when history is thin
stats = decision_store.get_all_stats()
total_decisions = stats["total"] if stats else 0
bootstrap_min = automation.get("bootstrap_min_decisions", 20)
if total_decisions < bootstrap_min:
confidence_threshold = 95
print(f"Learning phase ({total_decisions}/{bootstrap_min} decisions) — threshold: 95%\n")
else:
confidence_threshold = configured_threshold
# Fetch envelopes via himalaya
if recent:
envelopes = get_recent_envelopes(recent)
print(f"Found {len(envelopes)} emails from last {recent} days\n")
else:
envelopes = get_unseen_envelopes()
print(f"Found {len(envelopes)} unread emails\n")
if not envelopes:
print("No new emails to process.")
return
auto_acted = 0
queued = 0
for envelope in envelopes:
eid = envelope.get("id", "?")
print(f"[{eid}] ", end="", flush=True)
# Read message body without marking as seen
try:
body = read_message(eid)
except subprocess.CalledProcessError:
body = ""
email_data = build_email_data(envelope, body, config)
print(f"{email_data['subject'][:55]}")
# Run the LLM classifier (includes few-shot examples from history)
action, confidence, summary, reason, duration = classifier.classify_email(
email_data, config
)
print(f" -> {action} (confidence: {confidence}%, {duration:.1f}s)")
print(f" {reason[:80]}")
# Auto-act if confidence meets threshold
can_auto = confidence >= confidence_threshold
if dry_run:
# Dry run: log what would happen, touch nothing
log_result(log_file, email_data, f"DRYRUN:{action}@{confidence}%", reason, duration)
if can_auto:
print(f" -> Would AUTO-execute: {action}")
auto_acted += 1
else:
# Non-ad email - add to pending queue
print(" -> Adding to pending queue")
# Add to pending
msg_internal_id = add_to_pending(
email_data,
summary,
email_id.decode(),
email_data.get('recipient', 'youlu@luyanxin.com')
print(f" -> Would queue for review")
queued += 1
elif can_auto:
# Auto-execute the action via himalaya
success = execute_action(eid, action)
if success:
decision_store.record_decision(
{**email_data, "summary": summary}, action, source="auto"
)
# Mark as read (so it won't be processed again)
mail.store(email_id, '+FLAGS', '\\Seen')
log_result(log_file, email_data, analysis, f"ADDED_TO_PENDING ({msg_internal_id})", duration)
added_to_pending += 1
processed += 1
# Expunge deleted emails
mail.expunge()
mail.logout()
# Summary
print(f"\n{'='*50}")
print(f"Total emails checked: {len(email_ids)}")
print(f"Successfully processed: {processed} emails")
print(f" - Moved to trash (ads): {moved_to_trash}")
print(f" - Added to pending queue: {added_to_pending}")
print(f"Failed (will retry next time): {len(email_ids) - processed}")
print(f"\n📁 Pending queue: {PENDING_FILE}")
print(f"📝 Log: {log_file}")
print(f"\n💡 Run 'python process_queue.py' to view and process pending emails")
except Exception as e:
print(f"❌ Error: {e}")
log_result(log_file, email_data, f"AUTO:{action}", reason, duration)
print(f" ** AUTO-executed: {action}")
auto_acted += 1
else:
# Himalaya action failed — fall back to queuing
log_result(log_file, email_data, "AUTO_FAILED", reason, duration)
print(f" !! Auto-action failed, queuing instead")
add_to_pending(email_data, summary, reason, action, confidence)
queued += 1
else:
# Not enough confidence or history — queue for manual review
add_to_pending(email_data, summary, reason, action, confidence)
# Mark as read to prevent re-processing on next scan
if not dry_run:
try:
_himalaya("flag", "add", str(eid), "seen")
except subprocess.CalledProcessError:
pass
log_result(log_file, email_data, f"QUEUED:{action}@{confidence}%", reason, duration)
print(f" -> Queued (confidence {confidence}% < {confidence_threshold}%)")
queued += 1
# Print run summary
print(f"\n{'=' * 50}")
print(f"Processed: {len(envelopes)} emails")
print(f" Auto-acted: {auto_acted}")
print(f" Queued for review: {queued}")
print(f"\nRun 'python main.py review list' to see pending emails")
# ---------------------------------------------------------------------------
# Subcommand: review
#
# Non-interactive: each invocation takes arguments, acts, and exits.
# No input() calls. Compatible with cron and scripting.
# ---------------------------------------------------------------------------
def _get_pending_items():
"""Return only pending (not done) items, sorted by found_at."""
pending = load_pending()
items = {k: v for k, v in pending.items() if v.get("status") == "pending"}
sorted_items = sorted(items.items(), key=lambda x: x[1].get("found_at", ""))
return sorted_items
def cmd_review_list():
"""Print the pending queue and exit.
Shows each email with its number, ID, subject, sender, summary,
and the classifier's suggested action with confidence.
"""
sorted_items = _get_pending_items()
if not sorted_items:
print("No pending emails to review.")
return
print(f"Pending emails: {len(sorted_items)}")
print("=" * 60)
for i, (msg_id, data) in enumerate(sorted_items, 1):
suggested = data.get("suggested_action", "?")
conf = data.get("confidence", "?")
print(f"\n {i}. [{msg_id}]")
print(f" Subject: {data.get('subject', 'N/A')[:55]}")
print(f" From: {data.get('sender', 'N/A')[:55]}")
print(f" To: {data.get('recipient', 'N/A')[:40]}")
print(f" Summary: {data.get('summary', 'N/A')[:70]}")
print(f" Suggested: {suggested} ({conf}% confidence)")
print(f"\n{'=' * 60}")
print("Usage:")
print(" python main.py review <number> <action>")
print(" python main.py review all <action>")
print(" python main.py review accept")
print("Actions: delete / archive / keep / mark_read / label:<name>")
def cmd_review_act(selector, action):
"""Execute an action on one or more pending emails.
Args:
selector: a 1-based number, a msg_id string, or "all".
action: one of delete/archive/keep/mark_read/label:<name>.
"""
# Validate action
valid_actions = {"delete", "archive", "keep", "mark_read"}
if action not in valid_actions and not action.startswith("label:"):
print(f"Invalid action: {action}")
print(f"Valid: {', '.join(sorted(valid_actions))}, label:<name>")
sys.exit(1)
sorted_items = _get_pending_items()
if not sorted_items:
print("No pending emails to review.")
return
# Resolve targets
if selector == "all":
targets = sorted_items
else:
target = _resolve_target(selector, sorted_items)
if target is None:
sys.exit(1)
targets = [target]
LOGS_DIR.mkdir(exist_ok=True)
log_file = LOGS_DIR / f"{datetime.now().strftime('%Y-%m-%d')}.log"
# Execute action on each target
for msg_id, data in targets:
eid = data.get("envelope_id") or data.get("imap_uid")
if not eid:
print(f" {msg_id}: No envelope ID, skipping")
continue
success = execute_action(eid, action)
if success:
# Record decision for future learning
decision_store.record_decision(data, action, source="user")
# Mark as done in pending queue
pending = load_pending()
pending[msg_id]["status"] = "done"
pending[msg_id]["action"] = action
pending[msg_id]["processed_at"] = datetime.now().isoformat()
save_pending(pending)
log_result(log_file, data, f"REVIEW:{action}", data.get("reason", ""))
print(f" {msg_id}: {action} -> OK ({data['subject'][:40]})")
else:
log_result(log_file, data, f"REVIEW_FAILED:{action}", data.get("reason", ""))
print(f" {msg_id}: {action} -> FAILED")
def cmd_review_accept():
"""Accept all classifier suggestions for pending emails.
For each pending email, executes the suggested_action that the
classifier assigned during scan. Records each as a "user" decision
since the user explicitly chose to accept.
"""
sorted_items = _get_pending_items()
if not sorted_items:
print("No pending emails to review.")
return
LOGS_DIR.mkdir(exist_ok=True)
log_file = LOGS_DIR / f"{datetime.now().strftime('%Y-%m-%d')}.log"
for msg_id, data in sorted_items:
action = data.get("suggested_action")
if not action:
print(f" {msg_id}: No suggestion, skipping")
continue
eid = data.get("envelope_id") or data.get("imap_uid")
if not eid:
print(f" {msg_id}: No envelope ID, skipping")
continue
success = execute_action(eid, action)
if success:
decision_store.record_decision(data, action, source="user")
pending = load_pending()
pending[msg_id]["status"] = "done"
pending[msg_id]["action"] = action
pending[msg_id]["processed_at"] = datetime.now().isoformat()
save_pending(pending)
log_result(log_file, data, f"ACCEPT:{action}", data.get("reason", ""))
print(f" {msg_id}: {action} -> OK ({data['subject'][:40]})")
else:
log_result(log_file, data, f"ACCEPT_FAILED:{action}", data.get("reason", ""))
print(f" {msg_id}: {action} -> FAILED")
def _resolve_target(selector, sorted_items):
"""Resolve a selector (number or msg_id) to a (msg_id, data) tuple.
Returns None and prints an error if the selector is invalid.
"""
# Try as 1-based index
try:
idx = int(selector) - 1
if 0 <= idx < len(sorted_items):
return sorted_items[idx]
else:
print(f"Invalid number. Range: 1-{len(sorted_items)}")
return None
except ValueError:
pass
# Try as msg_id
for msg_id, data in sorted_items:
if msg_id == selector:
return (msg_id, data)
print(f"Not found: {selector}")
return None
# ---------------------------------------------------------------------------
# Subcommand: stats
# ---------------------------------------------------------------------------
def cmd_stats():
"""Print a summary of the decision history.
Shows total decisions, user vs. auto breakdown, action distribution,
top sender domains, and custom labels.
"""
stats = decision_store.get_all_stats()
if not stats:
print("No decision history yet.")
print("Run 'python main.py scan' and 'python main.py review' to build history.")
return
print("Decision History Stats")
print("=" * 50)
print(f"Total decisions: {stats['total']}")
# User vs. auto breakdown
print(f"\nBy source:")
for source, count in sorted(stats["by_source"].items()):
pct = count / stats["total"] * 100
print(f" {source}: {count} ({pct:.0f}%)")
auto = stats["by_source"].get("auto", 0)
if stats["total"] > 0:
print(f" Automation rate: {auto / stats['total'] * 100:.0f}%")
# Action distribution
print(f"\nBy action:")
for action, count in sorted(stats["by_action"].items(), key=lambda x: -x[1]):
print(f" {action}: {count}")
# Top sender domains with per-domain action counts
print(f"\nTop sender domains:")
for domain, count in stats["top_domains"]:
domain_stats = decision_store.get_sender_stats(domain)
detail = ", ".join(
f"{a}:{c}" for a, c in sorted(domain_stats.items(), key=lambda x: -x[1])
)
print(f" {domain}: {count} ({detail})")
# Custom labels
labels = decision_store.get_known_labels()
if labels:
print(f"\nKnown labels: {', '.join(sorted(labels))}")
# ---------------------------------------------------------------------------
# Subcommand: migrate
# ---------------------------------------------------------------------------
def cmd_migrate():
"""Import old pending_emails.json 'done' entries into decision history.
Run once after upgrading from the old system. Converts old action
names (archived/kept/deleted) to new ones (archive/keep/delete).
"""
decision_store.migrate_pending()
# ---------------------------------------------------------------------------
# Entry point & argument parsing
#
# Simple hand-rolled parser — no external dependencies. Supports:
# main.py [subcommand] [--recent N] [--dry-run] [review-args...]
# ---------------------------------------------------------------------------
if __name__ == "__main__":
main()
args = sys.argv[1:]
subcommand = "scan"
recent = None
dry_run = False
extra_args = [] # for review subcommand arguments
# Parse args
i = 0
while i < len(args):
if args[i] == "--recent" and i + 1 < len(args):
recent = int(args[i + 1])
i += 2
elif args[i] == "--dry-run":
dry_run = True
i += 1
elif not args[i].startswith("--") and subcommand == "scan" and not extra_args:
# First positional arg is the subcommand
subcommand = args[i]
i += 1
elif not args[i].startswith("--"):
# Remaining positional args go to the subcommand
extra_args.append(args[i])
i += 1
else:
print(f"Unknown flag: {args[i]}")
sys.exit(1)
config = load_config()
if subcommand == "scan":
cmd_scan(config, recent=recent, dry_run=dry_run)
elif subcommand == "review":
if not extra_args or extra_args[0] == "list":
cmd_review_list()
elif extra_args[0] == "accept":
cmd_review_accept()
elif len(extra_args) == 2:
cmd_review_act(extra_args[0], extra_args[1])
else:
print("Usage:")
print(" python main.py review list")
print(" python main.py review <number-or-id> <action>")
print(" python main.py review all <action>")
print(" python main.py review accept")
sys.exit(1)
elif subcommand == "stats":
cmd_stats()
elif subcommand == "migrate":
cmd_migrate()
else:
print(f"Unknown subcommand: {subcommand}")
print("Usage: python main.py [scan|review|stats|migrate] [--recent N] [--dry-run]")
sys.exit(1)

28
scripts/email_processor/move_ad_to_trash.py
View File

@@ -1,28 +0,0 @@
#!/usr/bin/env python3
"""Move specific email to trash"""
import imaplib
import email
# Connect
mail = imaplib.IMAP4_SSL('imap.migadu.com', 993)
mail.login('youlu@luyanxin.com', 'kDkNau2r7m.hV!uk*D4Yr8mC7Dyjx9T')
mail.select('INBOX')
# Search for the email with "10% off" in subject
_, search_data = mail.search(None, 'SUBJECT', '"10% off"')
email_ids = search_data[0].split()
print(f"Found {len(email_ids)} emails with '10% off' in subject")
for email_id in email_ids:
# Copy to Trash
result = mail.copy(email_id, 'Trash')
if result[0] == 'OK':
mail.store(email_id, '+FLAGS', '\\Deleted')
print(f"✅ Moved email {email_id.decode()} to Trash")
else:
print(f"❌ Failed to move email {email_id.decode()}")
mail.expunge()
mail.logout()
print("Done!")

214
scripts/email_processor/process_queue.py
View File

@@ -1,214 +0,0 @@
#!/usr/bin/env python3
"""
Email Queue Processor - Handle user commands for pending emails
Reads pending_emails.json and executes user commands (archive/keep/reply)
"""
import json
import imaplib
import os
import sys
from datetime import datetime
from pathlib import Path
SCRIPT_DIR = Path(__file__).parent
DATA_FILE = SCRIPT_DIR / "data" / "pending_emails.json"
def load_pending():
"""Load pending emails from JSON file"""
if not DATA_FILE.exists():
return {}
with open(DATA_FILE, 'r', encoding='utf-8') as f:
return json.load(f)
def save_pending(pending):
"""Save pending emails to JSON file"""
DATA_FILE.parent.mkdir(exist_ok=True)
with open(DATA_FILE, 'w', encoding='utf-8') as f:
json.dump(pending, f, indent=2, ensure_ascii=False)
def connect_imap(config):
"""Connect to IMAP server"""
mail = imaplib.IMAP4_SSL(config['imap']['host'], config['imap']['port'])
mail.login(config['imap']['email'], config['imap']['password'])
return mail
def show_pending_list():
"""Display all pending emails"""
pending = load_pending()
if not pending:
print("📭 没有待处理的邮件")
return
print(f"\n📧 待处理邮件列表 ({len(pending)} 封)")
print("=" * 60)
# Sort by email_date
sorted_items = sorted(
pending.items(),
key=lambda x: x[1].get('email_date', '')
)
for msg_id, data in sorted_items:
if data.get('status') == 'pending':
print(f"\n🆔 {msg_id}")
print(f" 主题: {data.get('subject', 'N/A')[:50]}")
print(f" 发件人: {data.get('sender', 'N/A')}")
print(f" 收件人: {data.get('recipient', 'N/A')}")
print(f" 时间: {data.get('email_date', 'N/A')}")
print(f" 摘要: {data.get('summary', 'N/A')[:80]}")
print("\n" + "=" * 60)
print("\n可用指令:")
print(" • 归档 [ID] - 移动到 Archive 文件夹")
print(" • 保留 [ID] - 标记已读，留在收件箱")
print(" • 删除 [ID] - 移动到 Trash")
print(" • 全部处理 - 列出所有并批量操作")
def archive_email(config, msg_id):
"""Archive a specific email by ID"""
pending = load_pending()
if msg_id not in pending:
print(f"❌ 未找到邮件 ID: {msg_id}")
return False
email_data = pending[msg_id]
uid = email_data.get('imap_uid')
if not uid:
print(f"❌ 邮件 {msg_id} 没有 UID")
return False
try:
mail = connect_imap(config)
mail.select('INBOX')
# Copy to Archive
result = mail.copy(uid, 'Archive')
if result[0] == 'OK':
# Mark original as deleted
mail.store(uid, '+FLAGS', '\\Deleted')
mail.expunge()
# Update status
pending[msg_id]['status'] = 'done'
pending[msg_id]['action'] = 'archived'
pending[msg_id]['processed_at'] = datetime.now().isoformat()
save_pending(pending)
print(f"✅ 已归档: {email_data.get('subject', 'N/A')[:40]}")
return True
else:
print(f"❌ 归档失败: {result}")
return False
except Exception as e:
print(f"❌ 错误: {e}")
return False
finally:
try:
mail.logout()
except:
pass
def keep_email(config, msg_id):
"""Keep email in inbox, mark as read"""
pending = load_pending()
if msg_id not in pending:
print(f"❌ 未找到邮件 ID: {msg_id}")
return False
email_data = pending[msg_id]
uid = email_data.get('imap_uid')
if not uid:
print(f"❌ 邮件 {msg_id} 没有 UID")
return False
try:
mail = connect_imap(config)
mail.select('INBOX')
# Mark as read (Seen)
mail.store(uid, '+FLAGS', '\\Seen')
# Update status
pending[msg_id]['status'] = 'done'
pending[msg_id]['action'] = 'kept'
pending[msg_id]['processed_at'] = datetime.now().isoformat()
save_pending(pending)
print(f"✅ 已保留: {email_data.get('subject', 'N/A')[:40]}")
return True
except Exception as e:
print(f"❌ 错误: {e}")
return False
finally:
try:
mail.logout()
except:
pass
def delete_email(config, msg_id):
"""Move email to Trash"""
pending = load_pending()
if msg_id not in pending:
print(f"❌ 未找到邮件 ID: {msg_id}")
return False
email_data = pending[msg_id]
uid = email_data.get('imap_uid')
if not uid:
print(f"❌ 邮件 {msg_id} 没有 UID")
return False
try:
mail = connect_imap(config)
mail.select('INBOX')
# Copy to Trash
result = mail.copy(uid, 'Trash')
if result[0] == 'OK':
mail.store(uid, '+FLAGS', '\\Deleted')
mail.expunge()
# Update status
pending[msg_id]['status'] = 'done'
pending[msg_id]['action'] = 'deleted'
pending[msg_id]['processed_at'] = datetime.now().isoformat()
save_pending(pending)
print(f"✅ 已删除: {email_data.get('subject', 'N/A')[:40]}")
return True
else:
print(f"❌ 删除失败: {result}")
return False
except Exception as e:
print(f"❌ 错误: {e}")
return False
finally:
try:
mail.logout()
except:
pass
def main():
"""Main function - show pending list"""
import json
# Load config
config_file = Path(__file__).parent / "config.json"
with open(config_file) as f:
config = json.load(f)
show_pending_list()
if __name__ == "__main__":
main()

38
scripts/email_processor/test_single.py
View File

@@ -1,38 +0,0 @@
#!/usr/bin/env python3
"""Test single email analysis"""
import requests
import json
email_data = {
"subject": "Fwd: Get 10% off your next order 🎉",
"sender": "crac1017@hotmail.com",
"body": "Get 10% off your next order! Limited time offer. Shop now and save!"
}
prompt = f"""Analyze this email and determine if it's an advertisement/promotional email.
Subject: {email_data['subject']}
Sender: {email_data['sender']}
Body preview: {email_data['body'][:200]}
Is this an advertisement or promotional email? Answer with ONLY:
- "AD: [brief reason]" if it's an ad/promo
- "KEEP: [brief reason]" if it's important/legitimate
Be conservative - only mark as AD if clearly promotional."""
print("Sending to Qwen3...")
try:
response = requests.post(
"http://localhost:11434/api/generate",
json={
"model": "qwen3:4b",
"prompt": prompt,
"stream": False
},
timeout=120
)
result = response.json()
print(f"Result: {result.get('response', 'error')}")
except Exception as e:
print(f"Error: {e}")

2
scripts/email_processor/venv/bin/python
View File

@@ -1 +1 @@
python3
python3.13

2
scripts/email_processor/venv/bin/python3
View File

@@ -1 +1 @@
/usr/bin/python3
python3.13

1
scripts/email_processor/venv/bin/python3.12
View File

@@ -1 +0,0 @@
python3

1
scripts/email_processor/venv/lib64
View File

@@ -1 +0,0 @@
lib

8
scripts/email_processor/venv/pyvenv.cfg
View File

@@ -1,5 +1,5 @@
home = /usr/bin
home = /opt/homebrew/opt/python@3.13/bin
include-system-site-packages = false
version = 3.12.3
executable = /usr/bin/python3.12
command = /usr/bin/python3 -m venv /home/lyx/.openclaw/workspace/scripts/email_processor/venv
version = 3.13.0
executable = /opt/homebrew/Cellar/python@3.13/3.13.0_1/Frameworks/Python.framework/Versions/3.13/bin/python3.13
command = /opt/homebrew/opt/python@3.13/bin/python3.13 -m venv /Users/ylu/Documents/me/youlu-openclaw-workspace/scripts/email_processor/venv