job-application-agent — Deep Dive Report

What It Does (TL;DR)

An Agent Skill + CLI that lets a coding agent discover, evaluate, fill out, submit, and track job applications on a candidate’s behalf. It packages:

• A deterministic scoring engine in pure Node.js (no LLM calls required for the math)
• Local state management with append-only ledgers, attention queues, and friction logs
• OS-backed secure storage (macOS Keychain / Windows Credential Manager) for the candidate profile
• Privacy-preserving telemetry that ships through a Cloudflare Worker → PostHog pipeline with strict identity stripping
• An installer that drops the skill into ~/.agents/skills/ and registers an hourly auto-update background job

The agent is told to use it via natural language (“search jobs”, “apply https://…”, “show attention queue”) and the CLI handles all the deterministic bookkeeping. The agent itself only does browser interaction + discovery.

Repository Stats

File Layout

job-application-agent/
├── bin/
│   └── job-application-agent.mjs      ← npm bin entry
├── installer/
│   └── src/
│       ├── cli.mjs                     ← install/update/status commands
│       ├── installer.mjs               ← staging, rollback, vendor copy
│       ├── runner.mjs                  ← auto-update shell script gen
│       └── scheduler.mjs               ← launchd / systemd / Task Scheduler
├── job-application-agent/
│   ├── SKILL.md                        ← agent-facing spec
│   ├── references/
│   │   ├── ANALYTICS.md
│   │   ├── APPLICATION_GUIDANCE.md
│   │   ├── AUTONOMY.md
│   │   ├── BROWSER_UPLOADS.md
│   │   ├── RUNS.md
│   │   └── SCHEMAS.md
│   ├── scripts/
│   │   ├── job-application.mjs         ← main CLI (command router)
│   │   ├── secret-store.mjs            ← keychain/cred manager wrapper
│   │   ├── telemetry-client.mjs        ← outbound telemetry
│   │   ├── telemetry-schema.mjs        ← event validation
│   │   └── windows-profile-store.ps1   ← DPAPI wrapper for Win
│   └── tests/                          ← 7 test files
├── LICENSE
├── package.json
├── README.md
├── SECURITY.md
└── scripts/
    ├── ci/
    ├── smoke-package.mjs
    └── ...

Architecture & Components

The system is composed of several tightly-scoped modules, each with clear responsibilities:

Architecture Diagram

Workflow: How a Job Gets Applied To

The system uses a round-based execution model for batch operations. When the agent starts a batch, it initiates a “round” with a target count:

{
  roundId: "round-2026-08-20-001",
  requestedCount: 30,
  startedAt: "2026-08-20T10:00:00Z",
  status: "active"
}

Each roundId gets attached to ledger entries throughout the round. The round-complete logic rejects if you didn’t hit the target count, enabling resumable workflows across agent sessions and crash recoveries.

Workflow Steps

1. Discovery — Agent searches job boards or follows provided URLs
2. Scoring — scoreJob(input, target) evaluates each position against the candidate profile
3. Decision — Gate engine returns auto-submit, review, or reject
4. Application — Agent fills out the form via browser automation
5. Ledger — Result is appended to the application ledger with round ID
6. Follow-up — Attention queue surfaces applications needing responses

Scoring Engine Deep Dive

The core scoreJob(input, target) function is a deterministic gate-decision engine. No LLM calls — just pure math.

Decision Gates (in order)

Score Components

Auto-Submit Guard

The auto-submit threshold is intentionally strict — all conditions must be met:

decision === "review"
seniority === "senior" || seniority === "staff"
no experience mismatch
score >= 80
must-have coverage >= 70%
all gates passed

Local Storage & Privacy Boundaries

Secret Store

The secret store abstracts OS-native credential management:

Ledger System

All state changes are written to append-only ledgers with:

• Idempotency keys — prevents duplicate applications
• Duplicate detection — same job URL + company flagged
• Round accounting — tracks counts per round for batch completion checks
• Friction logs — records UX issues encountered during application for later review

Telemetry Pipeline

The telemetry system is designed to be privacy-preserving while still useful for the developer.

Event Architecture

• 13 documented event types with strict enum validation
• Max 4 KB payload per event
• Events validated against schema before sending
• Ships to Cloudflare Worker → PostHog

Identity Stripping

The containsDirectIdentity() function uses regex to reject payloads containing:

• Email addresses
• Phone numbers
• URLs (may contain identifying paths)
• LinkedIn/GitHub profile references

Installer & Auto-Updates

Installation Flow

1. Validates packaged skill integrity
2. Stages to a temp directory
3. Atomic replace — swaps the old version with the new in one operation
4. Rollback — if anything fails mid-install, the previous version is restored

Vendor Sync

The installer copies the skill to all known agent vendor directories:

~/.codex/skills/
~/.claude/skills/
~/.cursor/skills/
~/.copilot/skills/
~/.gemini/skills/

Auto-Update Scheduler

Auto-update can be disabled via CLI. The installer also supports manual update and status commands.

Safety Guardrails

Hard Stops — Agent Will NOT Do These

Browser Session Rules

• Never inspect cookies or local storage
• Reuses the existing browser session (doesn’t create new logins)
• Privileged path-based upload — uses known upload paths, not page scraping

Submission Rules

• No submission without visible confirmation (agent must see a success state)
• Duplicate detection prevents applying to the same job twice
• Default mode: review-each — human reviews each application before submit

Final Verdict

What It Does Well ✅

• Architecture is thoughtful — clean separation of deterministic vs. non-deterministic logic
• Privacy claims backed by code — keychain integration is real, telemetry stripping is implemented
• Credential handling is real — not just “store in a JSON file”, actual OS keychain
• Append-only ledgers — crash-safe, auditable, resumable
• Auto-update opt-out — respects user autonomy
• Honest about scope — doesn’t overpromise

Caveats ⚠️

• 9 versions in 5 days — very early, API may still be shifting
• Linux unsupported for profile store (major gap for developer audience)
• Telemetry on by default — should be opt-in for a privacy-focused tool
• PostHog retention caveat — data lives on third-party infra
• Single-author project — bus factor of 1
• Browser-uploads still requires agent for form-filling — not fully autonomous

Recommendation

Bottom line: Pin v3.1.1, disable telemetry first, and give it a shot on macOS. The architecture is sound — the main risks are project maturity and single-author dependency.