Job Application Agent

The material sets narrow browser boundaries, forbids reading cookies or session files, and requires pauses for passwords, MFA, CAPTCHA, legal attestations, government identifiers, and uncertain facts. Two autonomy modes, revocable grants, and confirmed-submission-only recording provide strong user-confirmation and external-effect controls. The README gives unusually specific disclosure of OS keyring storage, optional private cloud state, default telemetry, identity sharing, community sharing, retention behavior, and separate opt-out commands. Deductions apply because these safeguards are predominantly documentation claims without the implementation files needed for line-level verification, while analytics and community sharing are enabled by default rather than using strict default data minimization. Dependencies are few and pinned, deployment actions use commit hashes, and workflow contents permission is read-only, but no lockfile, vulnerability report, or dependency-update policy is supplied. Autonomy can be revoked, updates are staged, and private state is kept outside the replaceable installation, but submitted applications cannot realistically be rolled back and no comprehensive transactional recovery process is shown. Repository, npm, issue, and private security-reporting paths are identifiable, although the license copyright line names no responsible party and publisher identity remains unknown.

README, package metadata, security policy, and workflows agree on the Node requirement, test entry points, safety boundaries, and product purpose, with no material internal contradiction. Node 20, browser capability, Linux secret-tool, keyring-daemon requirements, and headless setup are disclosed adequately. The deduction is that browser hosts, OS credential stores, optional Cloudflare services, and external applicant-tracking systems create substantial environmental dependencies whose fallback behavior is not fully documented. The README says blockers and empty results are reported and provides an attention queue, but the supplied evidence contains no representative error messages, error taxonomy, or complete failure-handling implementation, limiting the failure-messages score.

The intended personal job-search audience and discovery, qualification, application, tracking, and outcome-review scenarios are concrete, with examples for individual URLs, thread batches, rounds, attention queues, and outcome recording. Capability boundaries are particularly strong: use is restricted to the candidate's own search, access controls may not be bypassed, unverifiable claims are refused, and sensitive or judgment-heavy questions return to the user; outreach is disabled by default and does not automate social-platform access or sending. Natural-language examples and the review-each and routine-auto modes provide precise activation semantics. Environment-fit deductions reflect the continued need for Node 20, a browser-capable host, and platform-specific credential facilities; headless Linux requires extra daemon setup, and the Grok native sharing path is explicitly still pending.

The README has strong navigation and separates stages, privacy, safety, workflow, installation, development, and responsible-use material with useful tables and a flow diagram. Installation, updates, platform templates, Linux credential storage, and development commands are documented thoroughly. Naming is mostly consistent across the package, CLI, and skill directory, but the product also spans several installation paths, sites, and hosted Workers without an explicit compatibility or naming-stability guarantee. Numerous usage examples are present, though there is no consolidated FAQ or troubleshooting section. Known limitations and prohibited uses are explicit. MIT metadata matches the complete license text. Version 3.7.1 is declared, but no changelog, release history, or semantic-versioning policy is supplied, so versioning and changelog coverage is thin. Issue and private vulnerability-reporting routes are clear and a seven-day acknowledgement target is stated, but no named maintainer, support lifetime, or succession policy establishes complete maintenance responsibility.

The described outputs are operationally useful: eligibility decisions, duplicate prevention, attention queues, confirmed-submission ledgers, outcome records, discovery-source coverage, and delivery reconciliation all map to concrete job-search work. The product adds meaningful value beyond generic form filling through verified-fact constraints, confirmation-based accounting, multi-source discovery, outcome learning, and privacy-aware state management. The cost-benefit deduction reflects the complexity introduced by browser-agent integration, OS keyrings, optional cloud state, telemetry, a community registry, leases, and several ledgers. Default identity analytics and anonymous community sharing also impose an opt-out and comprehension burden on users.

Many claims are tied to named documents, commands, script categories, test scripts, and CI checks, while package.json, SECURITY.md, and deployment workflows partially corroborate version, testing, security-reporting, dependency, and permission statements. However, the supplied material omits SKILL.md, analytics and cloud-state specifications, test bodies, and core implementation, so the major privacy and submission-control claims cannot be traced statically to their enforcing code. Cross-file corroboration is meaningful but incomplete because many detailed behavioral claims appear only in the README. The source carefully distinguishes confirmed submissions from filled forms, derived data from candidate facts, unknown receipts from verified delivery, and unverifiable claims from acceptable facts, fully satisfying fact-versus-inference separation.