With a database that publishes deltas (manifest schema 2), the index on
this computer is downloaded whole once, kept as data/index/index.parquet
with the build it is at, and from then on brought up to date by applying
the deltas it has not seen, in one rewrite where a key's newest action
wins, as in Delta Lake's log: a posting added, changed and then gone ends
gone. Only a gap that no delta covers sends it back to the snapshot.
Descriptions are found by the posting's key in the current build's
files, whose key ranges the manifest gives, with the keys as constants so
the reader skips every row group that cannot hold them. That reaches jobs
found in any earlier build, whose old file-and-row refs point into
releases since deleted. A database from before schema 2, a local one
included, is read exactly as before.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The browser gets the environment without the model key: the one the settings name, and the default.
rich is declared, as tui.py imports it.
Dead code removed: Store.to_apply, the threshold setting, and tests of removed names.
A new resume replaces the plan and keeps the person's filters, such as exclude.
An unreadable description database is reported once per search, and the search goes on.
_places has no empty part, so a place without a city no longer reads ", Italy".
The apply protocol is named Browser, not Session.
Confirm, Finish and Complete are held back from any element, as sends do.
Judge comments say why, with no numbers a reader cannot check.
A key that OrcaRouter refuses is forgotten when it came from the key file, and asked for again; a key from the
environment is named by its variable and not forgotten. Ctrl+C leaves at every prompt, and leaving while a prompt
is open no longer raises NoMatches from the heading. Log lines wrap to the whole width, and Quit stays on the
footer at 80 columns. Database and model failures say what happened in plain words, with the detail after them.
Files saved with a byte-order mark are read. The README gives the install commands checked in new venvs, the
index size from the manifest, and how to change the key. CI runs the unit job on Python 3.11 and 3.12 on Ubuntu,
Windows and macOS; gate still needs all of it. The classifiers name the three systems.
The default form model moves to deepseek/deepseek-v4.1-flash: about 3 minutes and 1.5 US cents a form on 11 real sites,
85 of 100 on the hard local forms, level with the best within noise, where glm-5.3-flash took twice as long at three
times the cost.
Enter is refused everywhere except a bare Enter in a text input or textarea whose own placeholder or aria-label asks
for it (Paylocity's skills, Teamtailor's tags). Control+Enter and other combinations stay refused.
One question before a batch is sent (Enter sends, Escape cancels); no per-job confirmation and no --auto-submit.
A form's questions are asked after the others are sent, then the form is tried again.
After the click the verifying model gets the plugin's text as its rules and the page as untrusted text.
Each clicked job keeps its time, final address, verdict and evidence, page text and a screenshot under data/proofs.
Key s lists every job tried with its plain status; applied.csv is rewritten after each attempt, cells that start like a formula are made text.
Place words match whole parts of the location, so Wales is not New South Wales.
A location answered no scores 1, out of the top.
Progress counts while descriptions and jevos read; jevos is said to download or start only when it does.
Looking for shows the first wordings of the plan's nodes, not their ids.
A failed jevos verdict is counted once per batch, not taken for no opinion.
A scan's log lines reach the screen from the loop, not from its thread.
A judged posting with no description is marked no text in its status.
jevos weighs the next 100 postings by title with six closed questions; the best 25 by the title's fit and that evidence go to the judge model for a score, and the rest are kept as passed over. The title screen, its status and its tests are gone.
jevos is pinned to jevos-v5 and downloaded at first use, checked by SHA-256 before unpacking. Where it cannot run, the cloud model answers the same questions; where there is no key either, the title decides.
The search plan is a graph of nodes and weighted forms with conditional arcs, after the owner's harrier engine, kept in
aihawk.yaml so it can be tuned. Search scores titles in Python from it, by rarity among the postings of the place, and a
town's name no longer matches where the country column says otherwise.
Judging asks closed questions and computes the score from the answers. Plugins now describe how each site's pages are
built; SmartRecruiters, Teamtailor, JazzHR, BambooHR and Personio are new, from rendered pages.
The browser window is visible. The model gets the server's own instructions, its rungs of coordinates and screenshots,
and any tool it marks read-only; a click at a point goes through the same check as a click on a selector. The form rules
now send certifications and unstated facts back as questions, after a real test ticked them. A daily run tries the
newest invisible-playwright-mcp.
`aihawk resume.pdf` is all a person types. The model reads the resume and works out what to search and the contact
details it states; the list of jobs, best first, fills in on a terminal screen as they are judged; space selects, a
applies, and what a form asks that the files do not say is asked there and remembered. Nothing is sent without y.
Searching the published files over the network took two minutes each time, so the database now publishes a light index
that is downloaded once a day and searched locally. A description is read from the database only for a job being judged.
Filters match whole words, so java is not javascript and roma is not Romania.
A plugin is a Markdown file: the sites it is for, how a posting's link becomes the form's, and what the model must know
about the form. AIHawk applies where a plugin exists and lists the rest, so adding a site takes no code, and a person
can add their own beside aihawk.yaml. The first seven were written from real forms read in the browser, with nothing
filled in. What protects the person stays in apply.py, ahead of any plugin's words.
Reading those forms showed the refusal was wrong for Ashby: its Yes and No answers are untyped buttons inside the form,
which count as submit buttons by HTML's rule, so the model could never have answered them.
This repository is AIHawk again. Dots moved to feder-cr/invisible_dots with its history.
Every day a job merges the public job databases into one (pipeline/), one row per posting, and publishes it as a
release. `aihawk scan` queries it in place, so a user scrapes nothing and downloads only what the search reads. `score`
judges jobs against the CV through OrcaRouter, `apply` tailors the CV and fills the form through invisible-playwright-mcp.
The model never sends an application: it fills the form and names the submit control, and AIHawk clicks it after the
person's yes, or by itself with --auto-submit. While the model works, a click on what would send the form and the Enter
key are refused. What the CV and the answers file do not say comes back as a question, never as a guess.
