Automation of job application in hh.ru
The operating guide explains setup, the exact startup profiles,
questionnaire controls, background runs, full shutdown, answers/salary, CV maintenance, recovery
and every auxiliary feature. The CLI reference lists all 47 application
options and their actual defaults. .lenv.example lists every mapped environment key.
With dependencies installed and your cover letter reviewed:
# Silent visual alerts, manual CAPTCHA handling, browser review of uncertain answers.
bun run apply -- --configuration examples/operator.lenv
# Optional unattended stretch: defer questionnaires in RAM for this run.
bun run apply -- --configuration examples/operator.lenv --no-questionsThe checked-in operator profile contains no credentials. A normal run
allows questionnaires; --no-questions is optional. Test mode sends real applications too.
See full shutdown before leaving detached workers running.
2025-11-03.11-34-18.mp4
The cover letter is always your choice: pass --message-file <file> (MESSAGE_FILE) or
--message "<text>" (MESSAGE); the run does not start without one. The Russian version below
is in data/cover-letter.txt:
bun run apply -- --message-file data/cover-letter.txtRussian version:
Здравствуйте,
Мне понравилась ваша компания, я думаю моя кандидатура будет вам полезна и я смогу привнести ценность в работу компании.
В какой форме предлагается юридическое оформление удалённой работы?
Посмотреть мой код на GitHub можно тут:
github.com/konard
github.com/link-assistant
github.com/linksplatform
github.com/link-foundation
Для оперативной связи предлагаю использовать мой Telegram: @drakonard (+79582000567).
С уважением,
Константин Дьяченко
English version:
Hello,
I like your company, and I believe my candidacy would be useful to you and that I could bring value to the company’s work.
In what form is the legal arrangement for remote work offered?
You can view my code on GitHub here:
github.com/konard
github.com/link-assistant
github.com/linksplatform
github.com/link-foundation
For quick communication, I suggest using my Telegram: @drakonard (+7 958 200-05-67).
Kind regards,
Konstantin Diachenko
The application supports configuration via .lenv files using the lino-arguments library. This allows you to set default values without typing them every time.
Requirements: Bun 1.4.2 or newer (older Bun versions cannot attach Playwright
to the browser) and Chrome for Testing (npx playwright install chromium). The automation runs in
that browser, an app of its own: your Google Chrome is never used, so a click on Chrome in the Dock
or a new tab there is yours, the macOS keychain is never asked (--use-mock-keychain), and its
«Chrome for Testing is only for automated testing» bar is not shown (--test-type=gpu). A profile
Google Chrome used before is moved once, its logins kept (the cookies are read by Google Chrome
without a window and set in Chrome for Testing, in memory only). The browser is started the way a
person would start it (no automation infobars), and a resized window shows the page at its new size
(the recording's screenshots are taken by Chrome without a clip, browser-commander #180); the profile is kept in ~/.hh-automation/chrome-profile.
A restart keeps the hh.ru login: hh.ru keeps it in a session cookie, and the browser keeps session
cookies like «Continue where you left off» does (--restore-last-session, until browser-commander
keeps them for every persistent profile by default: browser-commander #187).
Run with defaults: bun run apply -- --message-file data/cover-letter.txt needs no other option
(the cover letter is the one choice the run asks of you): it logs in, defaults the matching search
to your Russian CV (the most recently updated Russian version when there are several), and
applies to the vacancies hh.ru suggests for it. If no Russian CV is available, it uses the latest CV.
Each application selects its own CV: dollar-denominated pay (USD, USDT or USDC) or English
questions select the English version; Russian questions select the Russian version. Vacancy
text decides when there are no questions, using the collected full description ahead of job titles
or localized application labels. The app opens the full application form, checks the
selected CV against local exports before sending. An already correct CV needs no chooser access;
otherwise the app scrolls the chooser into view, selects it once at the start and verifies the
resulting title or checked option without reopening the menu. Repeated form checks and worker restarts keep that choice and any
manual CV changes. A restart also preserves answers on an unfinished full application form.
The cover letter follows that language. A salary question naming only $ (for example,
compensation in $ or rubles) also selects English; shell variables and command substitutions
do not count as dollar pay. Questionnaire answers still follow their own questions: a Russian
questionnaire can have Russian answers alongside an English CV and cover letter. When a corrected
language decision replaces an earlier automatic choice, the app corrects an unchanged standard
letter too, preserving custom letter edits and all entered questionnaire answers. This prevents a one-click application
from sending HH.ru's default CV before the language check. Export both versions with
bun run cv -- export first. --no-auto-select-resume-language disables this behavior.
The vacancy is visited and collected before opening its application. Its full description, HTML,
skills and requirements stay in the private ~/.hh-automation/vacancies/<id>.lino app cache, outside the repository.
An unfinished form restores this context after a restart, for CV language, currency and saved-answer
matching. Recruiter chats and external form drafts also prefer the collected description over a new
network request. Disabling collection still retains context in RAM for the current run.
During the pause between applications, the app scrolls to the next unprocessed vacancy and logs its title. It stays on that card before opening the vacancy or application; previewing does not click Apply or shorten the pause. Processed vacancies and those waiting for a submission retry are excluded. If search cards reorder, the previewed vacancy is found again by ID.
- Alerts - on macOS, whenever the run waits for you (a question to answer or confirm, a
login, a captcha that still needs manual input after background work has finished, a form to fill by hand), it plays a sound and
shows a short notification («The job application automation needs your attention»; what is
needed is told in the terminal) with two buttons: «Show browser» brings the window
that waits for you to the front (un-minimized, on its tab), «Later» shows it again in 15 minutes
if the run still waits. The notification comes from a small app of the automation, «hh.ru
automation», built from
src/notifier/notifier.swiftinto~/.hh-automation/notifieron the first run (Xcode Command Line Tools); allow its notifications in System Settings → Notifications, and pick the «Alerts» style to keep them on screen. When it cannot be shown (a Focus or Do Not Disturb is on, the app is not allowed or not built) an alert window with the same buttons opens instead; «Show browser» is its default button, so a Return typed by mistake only brings the browser forward. No alert is shown while the browser window that waits for you is the app in front (checked again every 5 minutes). The sound is heard with a Focus on too; an alert not answered is shown again every 5 minutes, and it is taken away once the wait is over.NOTIFY=0turns them off. To mute sound while keeping visual alerts, use--no-notification-soundor setNOTIFY_SOUND: falsein.lenv(NOTIFY_SOUND=falsein the environment). Sound is enabled by default; set it back totrueor remove the setting to restore it. Detached runs also alert for unanswered questions, drafts needing review, and held validation errors. CV Save waits and external forms alert in their own browser slots. Repeated checks and overlapping waits share one alert per browser; completion or navigation removes it. - Form timing - ready full-page forms proceed after a short validation settle; there is no fixed 30-second review pause. Expanding a cover letter waits for the field to appear, and a disabled Send button is checked for up to five seconds. Questions and answers are checked again before sending. Forms requiring your input or review wait passively, keep saving answers, and show an explicit waiting message without repeatedly refilling or scrolling. The adaptive 1–4 minute pause between vacancies still applies.
- Login - if the automation profile is not logged in yet, the browsers on this machine are
checked for an hh.ru session or a VK, Mail.ru, OK, Google or Gosuslugi session (only cookie names
are counted). The hh.ru session of any browser you use (Google Chrome, Safari, Firefox, Edge,
Brave, Yandex, Arc, …) is tried first, most hh.ru cookies first: Chrome-family cookies are read by
the browser itself from a copy of its profile (no Keychain prompt), Safari's and Firefox's from
their cookie files (Safari needs Full Disk Access for the terminal). Else a social sign-in through
a Chromium-based profile is copied into the automation profile. Use
--login-from chrometo pick a browser, or--manual-loginto log in yourself. Every other browser slot (chats, forms, experience, LinkedIn) starts signed in the same way: with its own login, the login of another automation window, or else with a browser of yours. Log in once: a login made in any hh.ru window of the automation (main, chat, forms) is copied into the others that show signed out, checked every minute. Cookie values are never printed or stored outside the profiles. - Resume - suggested vacancies for the Russian CV are opened by default, unless
--urlis given. The update dates come from hh.ru's own state embedded in the resumes page (the page shows none), so choosing costs no extra request; the log names the resume and why it was chosen. A kept-open browser already on a search page keeps its filters and only switches the resume. - Pace - the pause between opened vacancies adapts: it starts at 2 minutes, each application
sent takes 10-20 s off it (down to 1 minute), and each hh.ru captcha adds 10-20 s (up to 4
minutes). A captcha holds speed-up for 5 minutes; repeated captchas cannot extend that hold
beyond 15 minutes from its first captcha. A successful application after that deadline can
speed up again. A random extra of up to a quarter is added, capped at 4 minutes for adaptive
pacing. Longer manual reviews can still extend a cycle. At its slowest, this schedule has
capacity for 360 cycles per 24 hours, leaving room for the goal of 200 sent applications; actual
throughput also depends on available vacancies, manual answers and CAPTCHAs. The pace is worked
out from
logs/stats.jsonl, so it carries over restarts;bun run statsshows it and the captchas per 100 applications at each pace.--job-application-interval <seconds>sets a fixed pace instead. The revised policy starts once at 2 minutes, retaining older statistics. - Daily application limit - a visible
negotiations-limit-exceedederror (200 responses per rolling 24 hours) pauses all application routes for two hours, including after a restart. Full forms, popups, the no-test button, manual submissions and standalone limit alerts use the same handling. Answers and the cover letter are backed up before the pause; the current form stays open and chats continue. Repeated checks reuse the original deadline. The application window's attention reminders are silent throughout the pause, including ones already displayed; actionable chat notifications keep working. After expiry the failed form is refreshed and its answers restored before retrying. This is an account-wide pause, never a successful application or a five-hour per-vacancy RAM skip. A fresh server refusal starts another pause; hidden banners and unrelated errors do not. - One tab - the automation browser keeps a single tab: when a run attaches, every other
tab is closed (the tab the automation used last is kept, else an hh.ru one), and tabs opened
during the run are closed as they appear. Turn it off with
--no-single-tab. - Browser lifecycle - Chrome is closed when the script exits. With
--keep-browser-openit keeps running, the next run reuses it and its open page, and it closes itself after--browser-idle-timeoutminutes (default 30) without use. Privatelogs/browser-lifecycle.jsonlrecords app-requested closes (with their reasons), page/target crashes, target loss, connection loss, and worker shutdowns separately. A page-close event alone does not establish a manual close. URL credentials, query strings and fragments are omitted from these diagnostics. - Debug traces - every run records, in
~/.hh-automation/traces/<time>-<name>-<id>/, a browser-commander trace (DOM snapshots, DOM mutations, interactions, console messages, page errors, failed requests) with its Links Notation exporttrace.lino, plus two Links Notation logs written alongside it:network.lino(document/XHR/fetch requests and responses with headers and bodies) anddom.lino(the DOM after every page load and every DOM mutation). Cookies, authorization and CSRF headers, password/token form fields and hidden input values are redacted (names kept), long bodies are cut with a truncation marker and binary bodies are skipped with their size. Traces stay in private machine-local app storage, outside the repository, and are never uploaded by the app or CI. They expire after three days by default. Cleanup runs at recording startup and every minute while recording, including waits on forms or account limits. Long recordings rotate daily, so an active session cannot keep old diagnostics indefinitely. Only timestamped directories with the recorder's own marker are pruned; unrelated files and symlink targets are preserved. Configure a positive--trace-retention-days/TRACE_RETENTION_DAYSvalue. Disable recording with--no-trace/TRACE=0. Old repository traces are not migrated into the new storage.
Quick setup:
-
Copy the example configuration:
cp .lenv.example .lenv
-
Edit
.lenvand uncomment/modify the options you want:# Enable verbose logging by default VERBOSE: true # 0 is the default: adaptive pacing within 1–4 minutes, including random variation. # A positive value selects a fixed interval in seconds instead. JOB_APPLICATION_INTERVAL: 0 # Keep the browser open between runs (closed after 30 idle minutes) KEEP_BROWSER_OPEN: true # Load a multi-line application message from a UTF-8 text file MESSAGE_FILE: ./message.txt -
Run the application (it will automatically load
.lenv):bun run apply
Configuration priority with the installed lino-arguments library:
- Explicit CLI arguments (highest priority), e.g.
--no-notification-sound. --configuration <file>/-c <file>, when supplied.- The local
.lenvfile (it overrides matching ambient environment variables). - Ambient environment variables not replaced by those files.
- Built-in defaults (lowest priority).
Use an explicit CLI flag when overriding an existing local setting for one run. Auxiliary commands with their own argument parser do not automatically accept the main application's flags.
Note: The .lenv file is gitignored to keep your personal settings private.
See .lenv.example for all available options with detailed comments.
Note: It's recommended to use --verbose flag for debugging to see detailed logs about which buttons are being clicked and which textareas are being detected.
The application now supports both Playwright and Puppeteer through a single unified command. Use the --engine flag to choose between them (default: playwright).
By default, the script will, on the short popup and on the full form alike:
- Auto-submit if the form has ONLY a cover letter (no test questions)
- Auto-submit a form with questions only when autofill answered every question with the saved
answer of the very same question (
--auto-send-exact-answers, see below) - Wait for you with any other form with questions: it is autofilled from
data/qa.linoby similarity and waits for youryon stdin, or, in an unattended run (--confirm none), for you to send or close it in the browser. Answers you type are saved todata/qa.lino; an autofilled answer of a similar question is saved under the form's wording only when the form is sent
To send forms with non-exact answers unattended anyway, use the --auto-submit-vacancy-response-form flag:
bun run apply -- --message-file data/cover-letter.txt --auto-submit-vacancy-response-form --verboseSafety Note: The default behavior (manual review) is recommended to ensure test answers are correct before submission.
Questions that the QA database cannot answer are never skipped by default: the form waits for
you (--on-missing-answers wait). Use --on-missing-answers skip for fully unattended runs.
Generic hh.ru submission errors («Произошла ошибка, попробуйте ещё раз» / “An error occurred,
please try again”, code unknown) automatically skip the vacancy for five hours using the existing
RAM retry queue. Before leaving the full form or closing its popup, the app saves every answer to
Q&A and writes a private recovery copy in logs/application-drafts/<vacancyId>.json, including the
cover letter, custom answers and selected choices. If reading or either save fails, the form stays
open. Passwords, hidden credentials and CAPTCHA fields are excluded from recovery copies.
Other applications continue while the failed vacancy is excluded from search. After five hours,
its answers can be restored for a retry under the usual review rules. The skip is held only in RAM:
normal process restarts clear the queue, and a recovery file never acts as a persistent generic-error
skip. Nothing is written to skipped-vacancies.lino for these errors.
Validation errors, limits and blocks still keep their form open for manual action; CAPTCHA handling stays manual. Dismissing such an error banner does not release its form for automatic sending. Confirmed submission releases the hold while retaining the local recovery copy.
By default, before applying, the browser visits the full vacancy page, saves its description,
HTML, skills, employer and other visible job details in ~/.hh-automation/vacancies/<id>.lino,
scrolls to the description and shows it for three seconds, then opens that vacancy's application
directly. --vacancy-preview-seconds 0 removes the preview pause; --no-apply-from-vacancy-page
keeps the previous route through the search page before opening the form. The settings also work
as VACANCY_PREVIEW_SECONDS and APPLY_FROM_VACANCY_PAGE in .lenv.
Employer descriptions are kept outside the repository and never published. Existing ignored
data/vacancies/ records are copied automatically, retaining their full content and the originals;
existing app records are not overwritten. --vacancy-cache-dir changes the private location.
Each selected vacancy is collected, including ones the filters later exclude.
Updates use the same vacancy ID, so repeated reads do not inflate demand statistics. Successful
applications are marked in the cache. If the full description cannot be read, that application
is left unsent on this pass.
A separate terminal report shows the technology coverage of the résumé, matched technologies, names without résumé evidence, and alternative spellings. This is a keyword comparison, not a prediction of hiring outcomes, and a low score does not block an application. Explicit skills and requirement sections take priority over general prose; preferred requirements are kept separately. The complete original text remains available for later analysis.
Known equivalent names are listed in data/technology-aliases.lino. For example, Golang can match
Go and Postgres can match PostgreSQL. Related technologies and versions are kept distinct: knowing
Angular does not establish AngularJS experience, and .NET does not establish .NET Core experience.
Verified equivalent spellings are maintained in both CVs' About inventories and every relevant
position's technology list, using the canonical brand with observed alternatives in parentheses,
such as Go (Golang). Each field is checked independently; an alias in another job or language
does not suppress it here. The private vacancy cache supplies the evidence; duplicate variants,
known typos and aliases already in the same label are omitted. Legacy name-variant sections are removed.
Each technical category is sorted alphabetically. k8s stays lowercase; npm (JS/TS) and
Nginx Proxy Manager (NPM) remain distinct. Bare uppercase NPM in vacancies is reported as
ambiguous rather than silently assigned to either product.
One reviewed update at a time is prepared in its own browser slot (9352); the owner presses final
Save. The app then exports both saved CVs and replans before preparing the next field, retaining
owner edits. Original prose is preserved; no missing technology is added as experience. The private
plan and reviewer status are under logs/experience/alias-sync-9352.*; sanitized .lino snapshots
are refreshed after verified saves. Existing drafts are observed without overwriting them.
The reviewer watches the vacancy cache for new spellings and remains separate from applications
and their RAM questionnaire queue. Only one reviewer owns the slot across app restarts.
bun run vacancy-stats # ranked technologies, spellings, gaps and requirements
bun run vacancy-stats -- --top 50 --json # machine-readable statistics from the private cache
bun run apply -- --message-file data/cover-letter.txt --no-vacancy-insights
bun run apply -- --message-file data/cover-letter.txt --no-update-resume-aliasesIndividual switches are --no-collect-vacancies, --no-compare-resume, and
--no-update-resume-aliases; --no-vacancy-insights disables the analysis bundle. The vacancy-page
entry route is controlled independently by --apply-from-vacancy-page.
The comparisons and statistics run locally and use no AI tokens.
Optional Oxion credentials belong in ignored .lenv: OXION_BASE_URL, OXION_API_KEY, and
OXION_MODEL. bun run ai-status checks the authenticated catalog;
bun run ai-status -- --probe also makes a four-token text request to confirm model access.
An authenticated key does not mean every model is available. Unsupported models are reported
explicitly, and a different model is never substituted silently. Credentials are not logged.
To check what will be sent, apply to a single vacancy step by step:
bun run apply -- --message-file data/cover-letter.txt --test-mode --max-applications 1By default the run asks before sending a form with questions (--confirm send) and waits
for you when a question has no saved answer. A form is sent without asking only when autofill
alone answered it exactly: every question is a saved question word for word and its field holds
the saved answer (the same text, the same checked options). Any other form is autofilled as far as
possible and waits for you; turn the exception off with --no-auto-send-exact-answers. --test-mode is a preset: --max-applications 1
and also confirm answers. Each part can be set
on its own:
| Option | Values | Effect |
|---|---|---|
--confirm |
answers, cover-letter, send (default), popup (comma-separated), or none |
Steps that wait for y on stdin: typed answers and radio/checkbox choices on the full form, its cover letter, the click that sends a form with questions (full form or popup; forms with only a cover letter are sent right away), everything in the short popup form |
--questions / --no-questions |
enabled by default | Optional no-questions mode: skip every questionnaire in RAM until questionnaires are enabled again; saved skip/deferred lists are unchanged |
--on-missing-answers |
wait (default), skip |
Questions without a saved answer: wait for you to answer them in the browser, or skip the vacancy (kept in data/deferred-questions.lino) |
--auto-send-exact-answers |
true (default), false |
Send forms whose every question autofill answered with the saved answer of the very same question, without asking |
--auto-reuse-answers |
false (default), true |
When a recruiter bot in a chat offers the answers you gave it before («Похоже, вы отвечали на часть вопросов… Используем эти ответы?»), each is compared with your saved answer and the comparison is printed (✓ the same, ✗ differs, ? not saved); by default you choose, with this «Да» is clicked when all are the same and «Нет» when one differs (the bot then asks again and the saved answers are given) |
--skip-question |
text, repeatable | Skip vacancies that ask a matching question for now (see below) |
--max-applications |
number, 0 = no limit |
Stop after sending this many applications |
Other buttons are clicked right away. Before a confirmed step the field is scrolled into view
(nothing is drawn on the page, so screenshots and recordings stay clean), and the terminal shows
the field and the exact value. Type y to go on or q to stop. When confirmations are on and a
question has no saved answer, the run lists it and waits until you have answered it in the
browser; your answers are saved to data/qa.lino.
A continuous supervised run, which only asks before sending forms with questions and when an answer is missing:
bun run apply -- --message-file data/cover-letter.txt --keep-browser-openAlongside the application loop, unread rejections («Отказ») in hh.ru chats are opened every 2
hours in the separate chat browser slot (port 9340), so they leave the unread chats. Closed chats
are only read. When replies are still allowed, the default is to send
«Здравствуйте, могу я узнать причину?» once. The full conversation is checked first: questions,
existing drafts, prior replies after a rejection, and active or unverified window state prevent
automatic sending. Private receipts in logs/rejection-replies/ prevent repeated requests across
workers/restarts, including when a send was not confirmed. --no-ask-rejection-reason (or
ASK_REJECTION_REASON: false in .lenv) keeps read-only cleanup; --rejection-reason-reply
(REJECTION_REASON_REPLY in .lenv) changes the message. Between those checks, the unread list is checked every 10 seconds while the chat
window is inactive. The pass covers the whole unread list, including rejections below the visible
rows, and returns to the selected chat with its draft, selection, unread filter and scroll position
restored. A draft does not block this background cleanup. The pass pauses when you return to the
browser; choosing another chat keeps your new selection. A conversation ending with a clear
review-and-follow-up acknowledgement is finished: without a draft, it is not reopened after cleanup.
On macOS the actual foreground browser
process determines whether the window is active. Turn cleanup off with --no-clear-rejections,
or use --no-live-rejections with answer-chats to keep only the scheduled passes.
Rejections are recognized by either the status badge or explicit rejection wording in the full
message preview, even when it has the normal text color. Messages asking a question or requesting
a reply remain unread. Cleanup leaves the window alone while it is active, even after an idle pause.
The same pass handles unread template messages: a chat whose last message, as the chat list shows it,
is a saved template of data/chat-templates.lino word for word (greeting and signature aside, «Спасибо
за отклик! Мы рассмотрим Ваше резюме и позже сообщим о своем решении.») is opened and answered with
its reply (one of those you gave to it, with --auto-send-exact-answers, on by default), or only read
when it needs none. A chat with anything else waiting is left for you, and other chats stay unread.
General review notices also qualify without an exact saved wording, including initial screening with a no-calls notice and requests for more time followed by a promise to return with feedback. They get the saved acknowledgement (for example, «Здравствуйте, благодарю, ожидаю.»). Questions, tasks, contact requests, invitations and rejections remain separate. Automatic template replies preserve existing drafts and stop if the window becomes active or the conversation changes.
To also answer unread hh.ru chats automatically, add --process-chats (the rejections are always
cleared first). The chat processor runs in the separate chat browser slot (port 9340, never
the application browser) and checks for unread messages every 2 hours (change with
--chats-interval-minutes). Chats with a known reply or no reply needed are handled quietly;
when a chat cannot be handled it is left open in the chat slot with any available draft, and
apply prints 💬 Chat needs your reply: <title> <url> — applications keep running meanwhile:
bun run apply -- --message-file data/cover-letter.txt --process-chats
bun run apply -- --message-file data/cover-letter.txt --process-chats --chats-interval-minutes 60You can also send a form yourself in the browser: the run notices the sent application (it
counts towards --max-applications), withdraws its pending question, keeps the interval
pause and returns to the vacancy list.
Some questions take time to answer well (see the case study). The run does not have to wait for them:
- At the open-questions prompt type
s: the vacancy is skipped for now, and its ID is kept under each open question indata/deferred-questions.lino. - A later form whose open question matches a deferred one is skipped right away and added there.
- A question added to the file while its form waits at the prompt skips that form too: only that
question is kept for it, not the form's other open questions.
For an unattended stretch, use
bun run apply -- --message-file data/cover-letter.txt --no-questions. This optional mode is off by default. It automatically leaves full forms and popups with questions, even when their answers are already known. Cover-letter-only applications continue. Answers and recovery drafts are preserved, but questionnaire questions and vacancy IDs are not added to either saved skip list. The exclusion and revisit queue exist only in the worker's RAM; a normal restart clears them and includes questionnaires again unless--no-questionsis passed.
On macOS/Linux, kill -USR1 <application-worker-pid> enables the mode while running, including
when a questionnaire is already waiting for review. kill -USR2 <application-worker-pid> enables
questionnaires again and revisits the RAM queue using the normal vacancy collection, pacing,
resume-selection and review rules. Separately saved exclusions remain in force. Recruiter chat
panels opened after sending are left for later while this mode is active. CAPTCHA handling and
account application-limit pauses retain their configured behavior.
-
--skip-question "<text>"skips forms that ask a matching question (a part of it is enough), answered or not. Repeat it for several questions, or pass several values after it:bun run apply -- --message-file data/cover-letter.txt --skip-question "портфель автоматизаций" bun run apply -- --message-file data/cover-letter.txt \ --skip-question "портфель автоматизаций" --skip-question "руководить командой"
In
.lenvor the environment:SKIP_QUESTIONS: портфель автоматизаций|руководить командой.
The file has the format of qa.lino, with vacancy IDs under each question:
Какой портфель автоматизаций и AI-продуктов вам удалось реализовать? Как считали эффекты для бизнеса?
137956393
Its vacancies are not opened again while the question has no answer in data/qa.lino. Once you
add the answer, they are opened like any other vacancy and the answer is filled in; a vacancy you
have applied to is removed from the file.
Vacancies that are not programming jobs (electrical installation, circuit design), and manual or
field work that needs physical presence, are skipped without asking, by the rules in
data/vacancy-filters.lino. A part of the text is enough; case and ё/е do not matter:
vacancy
схемотехник
question
дифавтомат
on-site
выезды на объект
работа на складе
programming
разработчик
remote
можно удаленно
vacancy: the vacancy card in the search list (title, company, labels), checked before the vacancy is opened, so it costs no request; and the vacancy name on its response formquestion: a question of the response form (full form or popup)page: any text of the response formon-site: physical presence (installation, field trips, warehouse, production line, shifts), looked for in the card, the vacancy description, the questions and the form. It filters only a vacancy that is clearly not programming: when its card, name or description has aprogrammingorremotefragment, on-site rules do not apply. So a Moscow office Go developer is never filtered, even when its description mentions the warehouses it automatesprogramming,remote: those guards
With vacancy insights disabled, the vacancy description is read only when it is needed: the card has no programming or remote sign and no rule matched it already (a «Сервисный инженер» card, not a «Go-разработчик» one). Then the run requests the vacancy page once, from the vacancy list (a same-origin request for the HTML only, no navigation and no images or scripts), reads its title, work format and description, and keeps the usual pause between vacancies after it. The search card and the response forms do not show the description, so this is the cheapest way to it.
Every filtered vacancy is logged with the rule (🚫 Vacancy … filtered out by vacancy-filters.lino (question "кв.мм": …)) and kept in data/filtered-vacancies.lino with its title and the time, so
it is not opened again. Add a line to the rules when you see a vacancy that should have been skipped.
hh.ru can show its interface in Russian or English (the account's setting). The tool reads it by
its markup, the same in every language: the «Откликнуться» (Apply) buttons, the «Только
непрочитанные» (Unread only) filter, a rejection in the chat list (its status in red), the vacancy
of a chat, the chat link of a sent application, the hidden resume notice and the direct
application popup. Where only the text tells (a message titled «Отказ» / "Rejection", the
application «Отклик на вакансию» / "Vacancy application", a chat the employer turned off, the
captcha, «Вам недоступна эта вакансия»), the words are in src/hh-words.mjs, in both languages.
Drafted answers are written in the language of the question, and a question or a message in English
is never answered in Russian: a saved answer in Russian is not reused for it (a similar question, a
pattern:, a chat template), a salary question gets an English answer («From $7 500 gross per month
($6 000 net).»), and an acknowledgement gets an English thanks («Hello, thank you.», or the one you
saved). A question in Russian may take an answer in English («Docker, Kubernetes»).
hh.ru sometimes sends the application at the click on «Откликнуться», with no popup: the card says «Вы откликнулись» / «Ваш отклик отправлен работодателю». This is taken as a sent application at once (no waiting for a popup), and the cover letter is attached right away:
- by the card's «Приложить письмо» on the list (no page is opened);
- else by «Приложить сопроводительное письмо» on the vacancy page, then back to the list;
- else the vacancy is kept in
data/skipped-vacancies.linoasletter_not_attachedwith its link, tried again at the start of the next runs (a few a run) until it is attached or hh.ru no longer offers it.
The letter counts as delivered once its popup has closed and hh.ru no longer offers to attach it. The same check runs after every application from the list and after a full form: a letter that did not go with the application is attached the same way.
No vacancy is skipped silently. Every other skip is logged and kept in
data/skipped-vacancies.lino with the reason, the vacancy title, the employer's link when there is
one, the time and the number of attempts:
138295841
"reason: external_site"
"title: Go-разработчик | Ромашка"
"url: https://career.example.com/jobs/1"
"time: 2026-10-10T12:00:00.000Z"
"attempts: 1"
Whether a skipped vacancy is opened again on the next run depends on the reason:
| Reason | Opened again |
|---|---|
external_site (applied on the employer's site), resume_not_visible, skipped_by_user (s, or a popup you closed) |
No, until you clear it |
questionnaire_ignored |
No, while --ignore-vacancies-with-questionnaire is on |
questions_deferred |
Once its questions are answered (data/deferred-questions.lino) |
modal_timeout, apply_button_disabled, apply_click_failed, button_not_found, button_disabled, click_failed, timeout, unanswered_questions |
Once more on a later run; skipped a second time, no |
A vacancy you apply to is removed from the file. List the skipped vacancies, with the
bun run prefill-form -- <url> command for each one applied on the employer's site, or clear a
reason or a vacancy so it is opened again:
bun run skipped
bun run skipped -- --clear resume_not_visible
bun run skipped -- --clear 138295841Keep these runtime records and data/deferred-questions.lino local, alongside your private answers.
They can be modified in the working tree because they are historically tracked; stage source and
sanitized CV snapshots explicitly rather than publishing the runtime stores with git add ..
A vacancy with hh.ru's notice «поменяйте видимость резюме на «Видно всем работодателям…»» is skipped
without asking (resume_not_visible). Only a rendered notice counts: every response popup carries it
in a collapsed hidden-resume-warning block that opens only for a hidden resume.
Questionnaires that employers send outside hh.ru (Google Forms, Yandex Forms, a company's own job form) are prefilled for review, each in its own browser slot:
bun run prefill-form -- https://forms.gle/... https://practicum.yandex.ru/job/vacancy-364- Slot
nis a separate Chrome (profile~/.hh-automation/form-slot-<n>, port 9330+n), apart from the hh.ru automation browser; it stays open for review.--first-slot <n>picks the first slot. - Contacts come from the exported resume (
bun run resume);data/profile.lino(not committed) overrides or adds values. Other answers come fromdata/qa.linoby the same matching as hh.ru forms; the rest are drafted by local Claude Code from the resume and saved answers, with «[уточнить: …]» where only you know the fact (--no-draftturns drafts off). - A site captcha is left to you; the slot fills the form after it. Nothing is ever submitted.
- What you finally send is learned: a small detached watcher per slot (
src/form-watch.mjs) attaches to the slot browser, reads the answers on every page (a multi-page Google Form's «Далее» pages too) and, once the form is really sent (its confirmation such as «Ваш ответ записан» or «Спасибо», or the form gone after a click on its send button), saves them todata/qa.linoand adds the saved questions to the report. Contacts are stored as{{placeholders}}; empty answers, answers with «[уточнить: …]» left in, contacts the prefill took from the profile and you did not change, test assignment links, and answersqa.linoalready gives are not saved. The watcher exits after the send, when the slot browser closes, or when the slot is prefilled again; its log islogs/forms/watch-slot-<n>.log.--no-learnturns it off (also inanswer-chats). - Contacts that are public anyway (Telegram, phone, LinkedIn, GitHub) are kept in
data/contacts.lino. Saved answers and the cover letter can use them as placeholders ({{telegram}},{{phone}},{{linkedin}},{{github}}), so a changed contact is changed in one place; contact values in answers you save are stored as placeholders. - Answers that depend on the vacancy are set in
data/answer-contexts.lino: each context (in a group) lists the words that show it and the placeholders it gives other values. A salary saved as{{salary_net}}is typed in dollars when the question, the vacancy card, its description or the form's other questions speak of dollars («$», «USD», «в валюте»), and in rubles otherwise, whichever currency it was saved in. The question's own words decide first, then the vacancy's (the card with its pay first, then the description and the form's other questions), then the group's(default)context. This works everywhere answers are typed: hh.ru forms and popups, the chat panel after an application, chats (answer-chats: the employer's messages, else the vacancy page, read only when the reply depends on it) and forms prefilled in slots. Sums you type («$6000», «6000 USDT», «450 000 ₽») are saved as placeholders, so they follow the vacancy too. Euros work the same way («€», «EUR», «евро»): the minimum issalary_net_eurindata/contacts.lino, and «до вычета НДФЛ» becomes «до вычета налогов» in dollars and euros. A gross in dollars or euros is what a contract pays your own company so that you get the net: the company's taxes are set indata/company-taxes.lino(personal, not committed): the profit tax by the company's revenue a year (15% up to 3000000000 VND, the band's sum in any currency the Central Bank has), thenpayout(the tax on the profit paid to you); it is rounded up to 100. Without that file the ruble tax's brackets are used at the Central Bank's rate. A blockchain vacancy («blockchain», «блокчейн», «криптовалют», «web3», «USDT»…) is paid in USDT: the dollar sums are typed as «6 000 USDT», unless the question names its own currency. - A salary question not saved word for word is answered by pattern (
src/salary-questions.mjs), learned fromqa.lino: it is a salary question by its words (built in, and those found in saved salary questions only), it asks for the gross by «gross», «брутто», «до вычета», and the answer is one of your own saved wordings of that kind (net; gross with net), the one picked by the question, so different questions get a bit different wordings with the same sums. Questions about a past job's pay, a given range («устраивает?»), or a year are left to the saved answers. Hourly expectations («почасовая ставка», «руб/час», “hourly rate”) are calculated from the monthly targets in the requested currency and tax basis.salary_hours_per_monthindata/contacts.linosets the calculation basis (default 160). For example, ₽450,000 net/month gives ₽3,000 net/hour. Hourly minimums round upward to steps of ₽500 or $5/€5/5 USDT; each currency uses its own target, and an exact step stays unchanged. Hourly placeholders ({{salary_net_hourly}},{{salary_gross_hourly}}and their currency suffixes) are kept separate from monthly amounts in forms, chat, choices and saved answers. When the currency is told (by the question, the vacancy's pay or its description), such an answer counts as saved, so a form whose other answers are saved too is sent without asking. The answer is saved under the new question once the form is sent. - The report (every question, its source and answer) is printed and saved to
logs/forms/. bun run formalize-qachecks how Formal AI (its JavaScript version, nothing to compile; cloned into~/.hh-automation/formal-ai) formalizes every saved question and answer: the kinds of mistakes are counted with examples, every result goes tologs/formalize-qa.json. Answers are not matched through it until it can formalize them (#1222, #1223).
See the requirements for what is done and what is planned, and the full list of requirements for everything else.
bun run test-assignment -- ai-business-management-system --dry-run # show the issue only
bun run test-assignment -- ai-business-management-system # create the repository and issueWhen a form contains a test assignment, it gets its own GitHub repository before the questions are
answered. The tool takes the assignment from the form open in a form slot (or --from <file>, - for
stdin), has local Claude Code restate it in English in its own words, and checks the result before
publishing: no employer name (the form title and «quoted» names, in any case ending), English only,
no six words in a row from the original.
The vacancy sets the stack: --vacancy <id> (or, for a form sent in an hh.ru chat, the chat's vacancy,
remembered by answer-chats) is read once in the chat slot. The language it names (or, when it names
none, the most fitting one, chosen by local Claude Code from the vacancy and the assignment;
--language overrides) picks the CI/CD template recommended by
hive-mind
(link-foundation/<language>-ai-driven-development-pipeline-template), and the issue ends with a
«Stack» section that asks for the most fitting stack and keeping the template's checks green.
Then gh creates the repository from the template (public unless --private; the owner is the gh
user unless --owner; an existing repository without a pipeline gets the template merged in) and
the issue (a rerun updates the issue it created). The next prefill of that form puts the repository
link into the question asking for the completed assignment.
bun run answer-chats -- https://hh.ru/chat/5698827730 # one chat
bun run answer-chats -- --watch # every chat waiting for an answer, one at a timeChats open in their own browser slot (port 9340), logged in with the hh.ru session of the automation
browser. Without --auto-send, replies are typed into the message field for your review:
- a template message (the same text from many companies, greeting and signature aside) gets its saved
reply from
data/chat-templates.lino, or one of the replies you have given to it before, picked at random (those you used more often more likely; a reply to someone by name or a long one of its own is not repeated elsewhere); a template you only thank for gets one of all the thanks you have given to templates or saved for them («Здравствуйте, благодарю.», «Здравствуйте, благодарю, ожидаю.», …); a rejection gets the saved question about its reason (Отказ), a closed chat is only read; - a question gets the saved answer of a close question in
data/qa.lino, otherwise local Claude Code drafts the reply from the chat, the resume and saved answers; - questionnaire links (and hh.ru's employer rating poll) are prefilled in form slots;
- what you have answered in chats is learned into
qa.linoandchat-templates.lino, and what you send in those questionnaires intoqa.lino(--no-learnturns the questionnaire part off); Template messages you answered the same way are generalized into patterns, e.g.pattern: ваше резюме … опыт* … позиции … свяже* с вами(shared words in order,…for any words,*for any ending), which match other companies' wordings of the same message. A pattern match counts as an exact one (sent with--auto-send) when the message says nothing more than the pattern (up to 40 words, or 3 times the pattern's, a company's and a vacancy's names in the gaps) and asks nothing: no invitation, test, form, link or call. An acknowledgement of the application in any company's words («Спасибо за интерес к нашей вакансии! Мы рассмотрим ваше резюме… мы свяжемся с вами», «ознакомимся с резюме и дадим обратную связь», «продолжаем рассматривать кандидатов и анализировать результаты собеседований… свяжемся с вами») is thanked for exactly, as the saved ones, unless it asks something, invites or rejects. A status such as «Ваше резюме находится на этапе рассмотрения» / “Your application is under review” is handled the same way, including from the chat list, without requiring a promise to contact you. An optional sentence offering other vacancies on an hh.ru employer page is allowed; surveys, tests and other links still need review. The full list preview is checked before opening a chat, and its actual message checked again before replying. Thanks replies count as the same reply when templates are generalized. A template inside a longer message is an exact one when what the company adds asks nothing (no more question marks, no «расскажите», «уточните», invitation, form or link). The closing and the signature («С уважением, Егор / HR команды Нонтон») are left out, and «позиции», «вакансии» and «должности» are the same word in templates. Templates are answered as they show in the chat window, as rejections are read.
--watch waits until you send the reply (or type s to skip) before the next chat. With --auto-send,
buttons are selected for exact saved answers, unique accepted answers from a completed bot summary,
and unambiguous salary ranges, including in the live
chat window and the application panel. A contradictory net/gross question is answered automatically
only when both configured amounts fit the same range. Ambiguous ranges and unknown answers wait
for review. Strict bot options (including Yes/No without a “choose” instruction) never get a
free-text draft; HH's generic conversation suggestions are excluded. A bot repeating the
question after an answer does not get the same text suggested again. Buttons arriving after their
question are detected too; clearing an unsent draft retries the choice. Saved answers and frequency counts are retained; plain salary
suggestions are filtered to the requested net or gross basis.
Salary radio/select prefills in HH.ru and external forms use numeric ranges rather than fuzzy
text matching. They understand тыс., k, ranges and open upper bands such as «Свыше 241 тыс. руб».
The configured currency and net/gross target choose the containing band, or the unique nearest
band for a reviewed form prefill. Ties and currency mismatches remain open. Chat auto-send still
requires a uniquely containing band; a previously saved lower band does not override the current target.
Salary follow-ups such as «цифры в долларах и до вычета налогов» inherit the recent salary
topic even without repeating «зарплата». The reply's language, currency and net/gross basis are
chosen independently: Russian messages can request USD gross, and a new explicit currency
overrides an earlier one. Unrelated questions stop that inheritance. A missing amount in the
requested currency is left for review, never replaced by an amount in rubles. Inferred follow-ups
are prefilled for review under the same rules as other salary patterns.
Explicitly approved availability-answer replacements live in data/answer-preferences.lino:
the preferred answer is followed by its indented equivalent wordings. They apply to start-date
questions on save/load and when reading chat history, combining the equivalents' usage counts.
Original chat messages remain unchanged. Unrelated historical replies, such as a performance
metric accidentally sent to a start-date question, are excluded from its suggestions.
Answer alternatives differing only by a single final period are combined into the version with
the period by default. Use --answer-final-period without or ANSWER_FINAL_PERIOD: without in
.lenv to prefer the version without it. This preference applies on save, load and frequency
sorting only when both versions exist. The first save keeps the answer exactly as entered;
an individual answer never gains or loses a period just because of the setting. Usage counts
are combined; original chat transcripts, other punctuation and different answers remain intact.
The “You've viewed the vacancy but haven't applied. Why?” survey automatically gets “Maybe
I'll apply later”, even if it appears while the application loop is waiting. Use
--no-auto-answer-viewed-vacancy or AUTO_ANSWER_VIEWED_VACANCY: false in .lenv to disable
this. A different reason already selected by the user is left alone.
While the chat window is open (--watch, and the rejection checks of apply), each sent message
is captured as the page changes, including delivery-status and bubble-style updates, without
requests to hh.ru. Switching chats immediately after sending does not lose the reply. The
application browser's chat frames are recorded too. Navigation waits for the conversation pane
to match its URL, so the previous chat's messages are not recorded under the next chat's ID. A new
question in the chat window is prefilled from saved answers, leaving text you already typed alone.
Full-history learning scrolls toward older messages using the negative scroll position of hh.ru's
reversed message column; it keeps reading older pages until the original application is reached.
Recruiter questions containing private access links stay in the local transcript and are excluded
from the shared Q&A file.
Prefills run one at a time, including at startup, to prevent duplicate text.
Questions seen while automatic chat processing owns the window are queued and retried when it
finishes, even if the conversation does not change again. A newer chat replaces the pending one;
existing drafts stay untouched. Restarting a signed-in chat slot keeps its current page open.
A bot's question is saved without its acknowledgement of your last answer («Понял, спасибо. А
есть ли у вас опыт с NestJS?» is saved as «Есть ли у вас опыт с NestJS?»), so it is answered word for
word next time, whatever the bot says first; hh.ru's AI assistant's summary («В разговоре с
кандидатом я узнал: …») is learned too and needs no reply. Salary inference uses the question after
the acknowledgement, so thanking you for your salary information before asking for GitHub does
not produce another salary answer. Inferred answers stay subject to review and never become
exact template matches. A short confirmation to «Можем продолжить общение?» is learned for that
reminder; a substantive answer after it still belongs to the earlier technical question. A bare
«Да» does not prefill a request for examples or an explanation. A check waits while you are typing or
moving around in that window. bun run answer-chats -- --learn-all reads the chats you answered
before, slowly, in a window of its own (port 9341). It opens unread chats too (marking them read),
scrolls through older message pages, and records each page before it scrolls out of view. A chat
is marked complete only when its original application is reached; an incomplete load is logged
and remains eligible for another pass. Repeated last-message previews do not skip conversations.
Every chat read (in the chat window, the chat panel after an application, or --learn-all) is kept
as it is in logs/chat-history.json (personal, not committed), with writes locked across processes.
Changed messages are updated and earlier versions of your sent text kept. Replies sent in several
bubbles are learned together. Existing answers in qa.lino are kept as you wrote them; every
distinct answer from the archive is also available to prefill, separated by | when there are
several, sorted by how often you sent each wording (most frequent first; ties keep their existing
order). Each chat stores answerCounts alongside its transcript. A sent reply counts once by its
message IDs, even when read repeatedly; a bot repeating it in a summary adds no use. An answer
known only from a summary or the saved Q&A file has count 0 until a sent occurrence is found.
Such alternatives wait for review. Checkbox selections keep their existing meaning.
bun run answer-chats -- --relearn learns the answers again from the archive without opening hh.ru.
Messages and form fields with several questions retain the complete question and reply. When
numbered replies align with numbered, bulleted or separate questions, each individual pair is
also saved to qa.lino and counted separately in chat history. Conditional follow-ups stay with
their parent question; code fences and instructions are not treated as extra questions. Ambiguous
replies remain intact rather than being guessed into separate answers.
Chats, hh.ru applications and external forms can reuse individual answers from a questionnaire, or assemble a complete draft from saved individual answers. Each component uses its own salary and contact context. A newly assembled block requires review; it is not an exact saved reply for automatic submission. Missing components leave the block open. Answer variants retain their frequency order. Existing saved questionnaires are indexed when the app or form tools start; to backfill both saved Q&A and archived chat counts without opening a browser, run:
bun run backfill-questionnairesbun run experience -- export # hh.ru and LinkedIn work experience → data/resume/experience.json
bun run experience -- diff # differences per job, report in logs/experience/
bun run experience -- sync --to linkedin # prefill LinkedIn from hh.ru (or --to hh)
bun run experience -- sync --to hh --auto # export again, diff and sync in one go
bun run experience -- edit --company Evirma --stack "PHP 8, Symfony 7, Go, ClickHouse"
bun run experience -- edit --company Capico --add "OpenAI Codex CLI" # appended to its list-
Edit one job's technologies:
editopens that job's hh.ru edit form (company in Cyrillic or Latin, «ООО "ЕВИРМА"» is found byEvirma) in the hh.ru slot (port 9351) and sets its «Используемые технологии:» list under the description, the rest of the text kept as hh.ru has it (--addappends to the list the job has, skipping what it lists already). It is saved only when you typey, or when you click «Сохранить» in the slot window yourself. The technologies missing from the resume's «Навыки» (one list for the whole resume) are listed. -
Export: hh.ru jobs come from
bun run resume(data/resume/resume.json, exported first when missing,--refresh-hhexports again). LinkedIn's experience page (https://www.linkedin.com/in/konard/details/experience/,--profile) is read in its own browser slot (profile~/.hh-automation/linkedin-slot, port 9350). When the slot has no LinkedIn session (noli_atcookie, or a login wall), the run says so and waits: log in in that window yourself; passwords are never read or typed. Both sides are saved in one shape (company, title, location, dates, description, skills) todata/resume/experience.json(gitignored). -
Diff: jobs are matched by company (Cyrillic or Latin spelling, legal forms ignored) and time; titles and descriptions are compared through translations; skills, dates and locations are compared directly. Jobs only on one side are listed.
-
Translation: every text is translated by three translators, shown side by side in the report: Claude Haiku (
claude -p --model claude-haiku-5-5), Codex with the latest Luna model (as the captcha prefill picks it), and Formal AI (cargo install formal-ai, run withFORMAL_AI_LIVE_API=1). The chosen translation is Haiku's, else Luna's, else Formal AI's; an answer that is no translation is passed over. Results are cached indata/resume/translations.json.--translators haiku,luna,formal-aipicks them. -
Formal AI failures (an error, an empty or placeholder answer, «I could not translate», a web search instead of a translation, text not in the target language) are reported to
link-assistant/formal-ai: an open issue on the same failure gets one comment per Formal AI version, otherwise a new issue is filed with the version, the input, the command, the output and what the other translators gave.--dry-run-issuesonly prints them,--no-report-formal-aiturns it off. -
Sync: the changes that bring the target in line with the source (titles, dates, descriptions, missing skills, jobs only on the source) are translated and prefilled one by one in the target's form in a browser slot (LinkedIn: port 9350; hh.ru:
hh-experience-slot, port 9351, logged in by copying the automation browser's hh.ru cookies). Each is saved only after you typey;sdiscards it,qstops and leaves the current form prefilled for you. Nothing is ever deleted; hh.ru keeps skills per resume, so missing skills are only listed.
See the requirements.
bun run resume # the most recently updated resume
bun run resume -- --resume YOUR_RESUME_HASH # a specific one, from https://hh.ru/resume/YOUR_RESUME_HASH
bun run resume -- --formats pdf,doc,rtf,txt --out ~/cvThe export runs in its own headless Chrome. It copies the hh.ru login, in memory only, from
the automation browser on --browser-port (9322), so an application run in progress, with its
half-filled answers, is left untouched. It writes to data/resume/, which is gitignored
because it holds personal data:
| File | Content |
|---|---|
resume.pdf, resume.rtf (resume.doc) |
hh.ru's own exports; hh.ru's "doc" is the same RTF file, so it is only downloaded on request |
resume.html |
hh.ru's HTML ("txt") export |
resume.md, resume.txt |
Markdown and plain text converted from it |
resume.json |
Name, position, update date, key skills, jobs and the collected stack |
stack.md |
Every technology once, key skills, "Технические навыки" categories, and per job the listed technologies ("Используемые технологии", "Использовавшиеся навыки", …) plus those only mentioned in the description |
bun run resume -- --all exports every resume of the account, each to <out>/YOUR_RESUME_HASH/.
bun run cv -- export # every resume to examples/cv.<language>.lino
bun run cv -- import --lang en # bring hh.ru in line with examples/cv.en.lino
bun run cv -- import ~/my-cv.ru.lino --resume YOUR_RESUME_HASH
bun run cv -- import ~/my-cv.en.lino --resume YOUR_RESUME_HASH --one --detach-extra
bun run cv -- edit logs/experience/my-change.json --resume YOUR_RESUME_HASH --snapshots-out examples
bun run cv -- watch --hh-port 9351 --out examples # observe, never edit or navigate
bun run cv -- batch ~/cv-changes.json --resume YOUR_RESUME_HASH --snapshots-out examples
bun run cv -- technologies # private bilingual technology audit and drafts
bun run cv -- technologies --review --snapshots-out examplesThe export source is normally the authenticated main browser on port 9322. Editing/review/watch
commands attach to an existing CV slot (9351 by default); the
operating guide shows how to
open it without starting applications. --hh-browser-port selects the export/login source,
whereas --hh-port selects the editor.
Each resume becomes one indented file: title, «Навыки», «Обо мне» and the jobs, a text line
per line (' ' is an empty line), dates as 2026-05 or present. The Russian and English
files have the same layout, so they compare line by line:
examples/cv.ru.lino, examples/cv.en.lino.
Contacts, salary and birth date are never written: the export stops when a text holds one.
Each workplace also includes area, website as displayed by HH, and industries with the
stable HH id, localized name and parent. Archived company websites retain their full paths.
Import compares industry IDs rather than translated captions, and can update employer names,
cities, websites and activities. Older files without these fields preserve the saved metadata.
If HH hides metadata for a linked employer, the app exports a fresh baseline before exposing
and restoring the native controls, so an employer rename cannot clear the other fields.
HH shares the personal name between all CVs; an imported name difference is reported for
separate profile review, never silently applied to both languages.
City labels and activity captions are resolved through HH's localized dictionaries and stable IDs,
so an English editor can import and verify Russian CV metadata without false mismatches.
import reads the resume again and goes to the one with the file's title (else the only one in
its language, or --resume). Each changed or new job and the About text is prefilled in the hh.ru slot (port 9351)
and the owner saves it in the browser. The app captures the fields at the final Save click,
including changes made after prefilling, and verifies the result against fresh exports. --one
stops after one position. --confirm-save opts into the older terminal y confirmation.
Save attempts are recorded before export verification. Transient verification failures retry
without another click. A resumed About queue can recover an earlier Save only when HH.ru's
matching Save-return URL, changed saved export and displayed About text agree; closing or
cancelling an unchanged draft never counts as a Save. Export navigation waits for the response
commit so unrelated slow page resources do not stall the workflow.
HH.ru keeps workplaces for the whole profile. Changing a shared entry automatically prepares
a separate copy for the target CV, retaining the city, website, industries, dates and unchanged
description. After its reviewed save, a checkpoint in logs/experience/cv-edits/ gives the next
cv edit ...next.json command: deselect the old entry from this CV in another reviewed save.
It does not delete the workplace or change other CVs. An existing open form blocks new editing,
so unsaved user changes are never navigated away from. --detach-extra offers the same reviewed
deselection for extra entries absent from the desired file. Exact saved copies are matched first.
The opened record is checked against its company, title and full description before fields or
checkboxes change. Saved checkpoint follow-ups also use the exact HH.ru workplace record ID,
so language copies with the same company and title cannot be confused.
Existing jobs are opened through the full workplace list, including older jobs hidden on the
résumé overview. The app locates one matching Edit control, waits for its handler, then uses
a native click before verifying the opened record's identity.
edit accepts JSON with kind (add, update, copy, detach, about), target (company, title and
description to distinguish language copies), and fields. Fields can include company, title,
description, start/end (YYYY-MM), current, area, website, and industries (whole labels,
stable IDs such as [{"id":"7.541"},{"id":"49.404"}], or full parent/leaf paths; commas within
an industry name are preserved). bun run industries -- --locale EN --json exports the complete
official industry tree, including choices outside the virtualized selector's visible branches.
Localized dictionaries are cached privately under ~/.hh-automation/dictionaries.
This is the same reusable workflow
used by .lino import. An about change uses fields.about and optionally target.about to reject
stale plans. The owner saves the About dialog too; edits made before Save are retained and verified.
Title and «Навыки» differences remain listed for separate editing.
The HH.ru branches of experience sync --to hh and experience edit --stack/--add use this
same reviewed workflow, including shared-entry protection and capture of the owner's final edits.
batch accepts an array of these changes and advances after each verified browser Save. Each
item can specify its own resumeHash for a mixed-language queue; otherwise --resume is used,
automatically inserting a copy's separate deselection step. --after-checkpoint <file> waits
for an already-running review to record saved_by_user before starting the queue. An idle or
closed browser is insufficient; the owner does not need to send a separate “Saved” chat message.
watch attaches to the existing slot without opening, clicking, saving or navigating any form.
It observes both workplace forms and About dialogs.
It exports actual saved CVs after each confirmed Save, retains all owner edits, and writes status
to logs/experience/cv-watch.json. Cancel and failed validation are not recorded as successful
saves. --once stops after one confirmed save. --snapshots-out examples does the same after
import/edit; use a separate desired file so snapshots do not overwrite your pending plan.
Publication is opt-in: --publish commits and pushes only the sanitized .lino snapshot files.
Full private exports remain gitignored; existing PDFs and backups are kept during refreshes.
technologies reads all saved CVs and audits raw position lists before deduplication. It consolidates
the existing About inventories and technologies evidenced in jobs into matching bilingual inventories,
retaining the biography and personal text. Equivalent names are grouped, distinct products and versions
stay separate, and vacancy-observed valid spellings go inline in each relevant field's parentheses.
No separate alias block is generated. Job-specific
lists stay under their jobs and receive the same verified aliases, canonical branding, alphabetical
ordering and deduplication as About. Differences from a uniquely matched counterpart are also
proposed for review. Ambiguous employer matches block the review instead of silently skipping a job.
Drafts and changes.json default to gitignored data/resume/technology-plan/; --out overrides this.
--review prepares each About/position change in turn, waits for the owner's verified Save, and runs a
final coverage check against fresh exports. Technology coverage does not assert that every translated
sentence is equivalent. Snapshots and publication always reflect actual saved CVs, never pending drafts.
Browser workflow checks use an isolated fixture, never a real account:
bun tests/browser/hh-cv-editor.mjs (requires local Chrome).
- CAPTCHA: the operator profile disables prefill and automatic sending; the run waits for you to finish the challenge in the indicated browser. A confirmed daily application limit takes priority: the unusable dialog is dismissed, answers are preserved and the account waits.
- Optional assistance settings: the implementation also exposes
--captcha-prefilland--captcha-auto-send, both enabled in the built-in defaults, unlike the operator profile. They depend on configured local model tooling; a configured model name does not establish availability. Use--no-captcha-prefill --no-captcha-auto-sendto select manual handling explicitly. - Unconfirmed application: a click alone is not success. If HH does not confirm the application, the worker waits for review or stops instead of counting it and opening another vacancy.
- Pauses: adaptive pacing defaults to a two-minute starting interval and moves within 1–4 minutes,
including random variation. Work already spent on a vacancy counts toward its interval; manual
reviews, login and CAPTCHA waits can extend the actual cycle. A positive
--job-application-intervalselects a fixed interval instead. Server account limits have their own pause. - Stats:
bun run stats(-- --days <n>, 14 by default) shows each day: applications, per hour, the time of work on one (typical and longest), how many took over 7 minutes with the pause, captchas (how many AI solved, and per 100 applications), skipped vacancies, rejections read (how many as they came) and letters attached later; then the pace and the captchas per 100 applications at each pace. A captcha is written when it shows, in any window (the main one, the chat window, a form slot), with the site, page and window, and again once it is gone. A run prints today's line when it starts. The events are kept inlogs/stats.jsonl(not in git). - Chat panel: hh.ru opens the employer chat after an application; the run closes it. When the employer turns the conversation off («Работодатель отключил переписку по данной вакансии») while the run waits for your answer there, the panel is closed and the run goes on.
Matching searches start at the configured salary_net_rub target when no salary filter is set.
After all pages at that level are exhausted, the salary ladder lowers the floor by 50,000 RUB
(or the equivalent in the selected filter currency) and starts from the first page. Existing
salary filters are respected. Vacancies without salary remain eligible at every level, and
processed vacancies are not applied to again. An advertised upper bound below half the configured
target in that currency is excluded. A low minimum without an upper bound is deferred to its
salary tier, rather than excluded. These deferrals are session-only, separate from questionnaire
skips. --no-salary-ladder restores the previous search behavior. Amounts are compared as
advertised; the ladder does not guess tax treatment or convert unknown project/shift fees to monthly pay.
If you want to skip vacancies that require any additional questionnaire fields beyond the cover letter, use:
bun run apply -- --message-file data/cover-letter.txt --ignore-vacancies-with-questionnaire --verboseThis works for both modal response forms and full vacancy_response pages. The vacancy will be skipped as soon as the script detects extra questionnaire fields, and kept in data/skipped-vacancies.lino (reason questionnaire_ignored; the former data/ignored-vacancy-ids.txt is moved there on the next run).
bun run playwright -- --configuration examples/operator.lenvOr use the main entry point explicitly:
bun run apply -- --engine playwright --configuration examples/operator.lenvFor your own search, copy its URL from hh.ru and quote it:
bun run apply -- --configuration examples/operator.lenv \
--url 'https://hh.ru/search/vacancy?resume=YOUR_RESUME_HASH&from=resumelist'The operator profile uses adaptive 1–4 minute pacing. --job-application-interval SECONDS
selects a fixed interval instead. --verbose adds debugging output. The
operating guide shows
how to capture logs and a worker PID without losing live questionnaire controls.
bun run puppeteer -- --configuration examples/operator.lenvOr explicitly:
bun run apply -- --engine puppeteer --configuration examples/operator.lenvBoth engines use the same configured profile by default. Run one worker per profile; switch engines
between runs. For an isolated profile, use the CLI option --user-data-dir /path/to/profile.