How Reply-Watch Classifies Employer Email Replies and Updates the Tracker in Career-Ops

Reply-watch mode uses a deterministic keyword-based state machine in reply-matcher.mjs to classify employer replies into categories like Interview, Offer, Rejected, or Noise, then suggests tracker updates that require human approval before applying.

The reply-watch mode in santifer/career-ops transforms raw mailbox data into actionable tracker updates through a three-phase pipeline. This article breaks down exactly how the classification logic works, what signals it prioritizes, and how the human-in-the-loop approval mechanism prevents accidental data corruption.

Architecture Overview

The reply-watch system processes employer emails through three distinct phases orchestrated by reply-watch.mjs with core logic residing in reply-matcher.mjs.

Phase 1: Load and Match

The pipeline begins with candidate loading and fuzzy matching against existing tracker entries.

• Input: data/reply-candidates.json (or user-supplied path)
• Tracker sources: data/applications.md and data/follow-ups.md
• Matching function: matchCandidates builds text context from from + subject + body_snippet

The matcher scores candidates against tracker rows using three weighted signals:

Signal Combination	Confidence Level
Company name + Role title	High
Sender domain + Role title	High
Single signal only	Medium/Low

Domain safety measures exclude SHARED_DOMAINS (LinkedIn, Greenhouse, Gmail, etc.) to prevent false matches. Short company names (≤3 characters) require word-boundary matching to avoid substring collisions like "HP" inside "PHP".

Phase 2: Classification via Deterministic State Machine

For matched candidates, classifyReply executes a strict precedence-ordered keyword check in reply-matcher.mjs:

1. Noise — Marketing-alert keywords match first; no tracker action suggested
2. Rejected — Rejection keywords or signal: 'rejection' override all; suggests Rejected status
3. Offer — Offer phrases ("offer letter", "congratulations on the offer"); suggests Offer status
4. Auto-confirmation — Generic "application received" strings; no tracker action
5. Need Action — Scheduling or task requests; suggests Interview if scheduling detected, else Responded
6. Interview — Interview invite keywords including Chinese "邀您面试" or signal: 'interview_invite'; suggests Interview status
7. Responded — Follow-up style phrases; suggests Responded status
8. Unknown — Recruiting-related terms present but unclear; flags "Needs Review"
9. Fallback — Everything else defaults to "Needs Review"

Critical precedence rule: Rejection beats offer, noise beats everything. No LLM inference occurs—all decisions are rule-based.

Phase 3: Human-in-the-Loop Update

Classification outputs aggregate into a digest presented to the user. The updateTrackerStatuses function in reply-watch.mjs only executes after explicit "y"/"yes" confirmation:

• Acquires file lock on data/applications.md
• Rewrites only rows where current status ≠ recommended status
• Runs node tracker.mjs sync to refresh SQLite index

No mutations occur without user approval regardless of classification confidence.

Internationalization Support

The classifier handles non-English content through normalizeChinese in reply-matcher.mjs:

• Strips corporate suffixes: "有限公司", "公司"
• Enables substring matching on Chinese characters (space-agnostic)
• Recognizes interview invitation patterns like "邀您面试"

Running Reply-Watch Mode

Basic invocation with default candidates file:

node reply-watch.mjs

Custom candidates file path:

node reply-watch.mjs /path/to/my-replies.json

Example Candidate Input and Output

Sample reply-candidates.json structure:

[
  {
    "message_id": "msg1",
    "from": "hr@awesome-tech.com",
    "subject": "Interview invitation – Senior Engineer",
    "body_snippet": "We'd like to schedule a video interview next week.",
    "signal": "interview_invite"
  },
  {
    "message_id": "msg2",
    "from": "noreply@job-alerts.com",
    "subject": "New opportunities for you",
    "body_snippet": "Check out these openings…",
    "signal": null
  }
]

Generated digest:

Today: 2 application updates need review

1. Awesome Tech — Senior Engineer
   Type: Interview
   Signal: interview_invite
   Evidence: interview invitation; schedule an interview
   Suggested tracker update: Interview

2. job-alerts.com
   Type: Noise
   Evidence: new opportunities for you
   Suggested tracker update: none

Confirming "y" rewrites the Awesome Tech entry from Applied to Interview in the tracker.

Key Source Files

File	Purpose
reply-watch.mjs	Orchestrates loading, matching, classification, and conditional tracker updates
reply-matcher.mjs	Implements matchCandidates, domain/company matching, and classifyReply state machine
modes/reply-watch.md	User documentation for inputs, invocation, and workflow

Summary

• Deterministic classification via precedence-ordered keyword matching in reply-matcher.mjs—no LLM dependency
• Multi-signal matching using company name, sender domain, and role title with safety guards for shared domains and short names
• Strict precedence hierarchy: Noise → Rejected → Offer → Auto-confirmation → Need Action → Interview → Responded → Unknown
• Mandatory human approval before any data/applications.md mutation via updateTrackerStatuses
• Chinese language support through normalized substring matching

Frequently Asked Questions

Does reply-watch use AI or machine learning to classify emails?

No. The classification relies entirely on deterministic keyword matching and explicit signal flags. The classifyReply function in reply-matcher.mjs uses static keyword lists with hardcoded precedence rules, making behavior predictable and debuggable without external API dependencies.

What prevents reply-watch from matching the wrong application entry?

Three safeguards operate in matchCandidates: excluded SHARED_DOMAINS prevent generic email providers from skewing results; company names of three characters or fewer require word-boundary matching; and scoring requires multiple signal alignment for high-confidence matches. Low-confidence matches surface for user review rather than auto-association.

Can reply-watch automatically update the tracker without asking me?

No. Even with perfect signal alignment (e.g., explicit interview_invite flag), updateTrackerStatuses in reply-watch.mjs requires explicit "y" or "yes" input. The file-lock mechanism only activates after confirmation, and the sync command runs post-rewrite to maintain index consistency.

How does the system handle non-English employer replies?

The normalizeChinese function in reply-matcher.mjs preprocesses Chinese text by removing common corporate suffixes and enabling character-level substring matching. Specific interview invitation patterns like "邀您面试" are directly encoded in the classification rules alongside English equivalents.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}