AIHawk review Job filtering and application state checks
Synthetic local tests of AIHawk search, deduplication, CV reading and attempt records, plus scripted post-submit outcomes and their limits.
AgentIndex · Published 2026-11-05 · Updated 2026-11-05
AIHawk is a terminal job-search assistant that builds a plan from a resume, filters postings and handles applications through a visible browser. It is relevant to people prepared to review fit explanations and form contents while keeping an application history. Finding a posting, judging its suitability and receiving a site's submission confirmation are separate stages.
We checked deterministic parts around that workflow. Synthetic postings could be filtered locally, rediscovery preserved an existing attempt state and an unknown scripted submission result remained unconfirmed. These findings do not establish real-site application success or the quality of model-generated resumes and judgments.
Version and eleven checks
On November 5, 2026, we pinned commit e74f532b83f0a376694309c86e1636fb6a6dd516, package version 1.0.0, under MIT licensing. We used Python 3.12.13; the project requires 3.11 or newer. Only dependencies needed for selected module checks were installed, and no browser engine was started. Package configuration
Inputs were five invented postings, a text CV and a minimal DOCX text package. A local fake supplied browser responses, and scripted values replaced model-gateway verdicts. All URLs used example.invalid. Python socket connections were blocked during the tests. The reproduction fixture records those boundaries.
| Check | Observed result |
|---|---|
| Local Parquet search for Python | Two role types; old and unrelated records excluded |
| Same company/title in different places | Merged into one result |
| Location restriction and intern exclusion | Corresponding engineer role remained |
| Synthetic TXT and minimal DOCX | Text read successfully |
| CV text below 50 characters | Rejected |
| Rediscover an existing job ID | No duplicate; previous attempt status retained |
| Scripted submitted / unknown / rejected | submitted / unconfirmed / refused |
| Fake click error plus verification failure | unconfirmed; no second fake click |
Together with CSV and control-classification checks, eleven checks passed. Submitted in this table describes a simulated state mapping, not a real application.
Search is separate from fit assessment
The local index is read from Parquet through DuckDB. Our dataset included different roles, a duplicate company/title and an old posting. Keyword search returned Python Engineer and Python Intern. Adding a location restriction and excluding intern left only the engineer result. Search implementation
This establishes the selected filter behavior. Title overlap does not establish suitability: responsibilities, seniority, work authorization and other conditions require the full posting. We did not generate a search plan or request model scores, so lexical matching should not be reported as measured recommendation accuracy.
Merging a company's same-title postings can reduce clutter, but they may represent different locations or requisitions. Inspect the source before selecting one. An index's publication date also does not guarantee that each included position remains open.
CV reading and inspectable history
The text CV and minimal DOCX were readable. Text shorter than 50 characters was rejected as insufficient readable content. That threshold is a parser rule, not an assessment of a person's experience or truthfulness. Scanned PDFs, complex formatting and generated CV quality were outside our tests. Resume reader
The SQLite store retained one row per job ID. Rediscovering our job preserved its unconfirmed state. An exported CSV kept the uncertainty label and wrote formula-like company, role and evidence cells as text. Attempt storage and export
A recorded score in this fixture was manually supplied test data, not a model assessment. Likewise, exporting a row does not prove that an employer received anything. Keep the distinction between stored claims and externally confirmed outcomes when reviewing a search history.
Preserve uncertainty after a click
The post-submit function clicks, waits for page changes and then reads the result. Scripted gateway responses for submitted, unknown and rejected produced different states. When the fake click returned an error and verification also failed, the result remained unconfirmed and there was no second click. Submission and verification path
That behavior avoids equating a failed result read with proof that nothing was sent. In actual use, inspect the page, messages or saved evidence before retrying an unconfirmed attempt. We verified how code handles verdicts; we did not verify a model's ability to understand every site's confirmation page.
Control wording remains consequential
The tested classifier identified Submit application as a sending control but treated a submit-type Continue button as a step. If a real site's final sending action uses that wording, a text-based refusal can be insufficient. The documentation also describes a fallback page-script path that can bypass some click restrictions. These mechanisms do not establish a strict authorization boundary.
Data handling is another adoption condition. The README states that CVs, answers and page content go to the model gateway and its providers, while applications go to employer sites. A visible browser helps observation but does not make this an entirely local data flow. Review submitted content, access scope and site requirements before using personal material. Data and runtime description
Adoption decision
Start with search and local records if the immediate need is to organize candidate jobs and inspect reasons and attempt states. Separately validate a chosen site's form flow. MIT licensing does not include model-service or browser operating costs, and the documented workflow depends on a model gateway.
This review supplies no application success rate, CV-quality measure or hiring conversion result. It covers local filtering, persistence and simulated post-submit states. For broader browser-control questions, see the Browser Harness review. For preserving work across sessions, continue with the Crew shared-log review.