Repository navigation
Audit fixes: broken features, browser agent, security, performance, cleanup - #4
Merged
Merged
Conversation
…d switching logic
- GET /persona/export downloads the persona as JSON, Markdown or a CV-style PDF - POST /jobs/from-url scrapes an arbitrary job posting (ATS API, JSON-LD, OpenGraph) into a job for the normal analysis flow - Profile "Download Twin" menu and a paste-a-link form on the job feed
- Pass the query as a list to the ATS level of live search; it iterated the string's characters, so the 60+ ATS job boards never returned anything - Only mark an application Applied when the browser session reports a confirmed submission (skip, cancel, timeout and errors used to count) - Map persona certification fields (issuing_organization / issue_date) into generated CVs - Require JWT_SECRET instead of falling back to a hard-coded default; docker compose fails fast without it - Only refresh a shared job description from the stored job's own URL, never from a client-supplied id/URL - Pass BRAVE_SEARCH_API_KEY through docker compose
- Use browser-use's own Chat* LLM wrappers (0.11 rejects LangChain models); drop the langchain packages and pin browser-use==0.11.13 - Reimplement the live screencast and user takeover on the CDP API (0.11 removed Playwright, so both were silently dead) - The agent fills the form and stops before the final submit; the user reviews, then submit / done / skip - Always kill() the browser, including on client disconnect (0.11 has no close(), so every application leaked a Chromium process)
…rompts - Encrypt the job-site password with Fernet (CREDENTIALS_KEY) in its own column, migrating any plaintext; GET /persona only returns a mask - Send LLMs an allowlisted persona: no credentials, date of birth, contact details, diversity data or salary - Guard server-side fetches of user-influenced URLs: public IPs only, every redirect re-checked; validate the /apply job URL - Browser agent: password via a host-scoped sensitive_data placeholder, judge call disabled (it saw unredacted step history), raw IPs and internal hosts blocked, disable_security removed, job page opened explicitly - Strip credential fields at any depth in tool inputs
- Live discovery runs in a background worker process, one run per query/location with a cooldown; /jobs reports `discovering` and the feed re-polls page 1 - Embed only new or vector-less jobs, in batched Ollama /api/embed calls, without holding the DB write lock; short embedding timeout for requests - Query, tags and category now filter in SQL; whole-word location matching; hidden jobs excluded before pagination; correct has_more - Share one TLS context across httpx clients (each new client loaded the CA bundle synchronously, stalling the event loop) - Parse Greenhouse boards off the event loop; never persist fixture jobs; route Adzuna by whole-word country names; bound in-memory caches
- Remove jobs-mcp and civic-guardrails (nothing called them), the /stream agent pipeline, /guardrails/audit, unused models and frontend components - Scope application status updates to the owning user; mark previously saved jobs Applied on submit; close the connection after a duplicate insert (it left applications.db locked) - Keep each user's uploaded CV and use it when applying without a tailored CV - Drop the unused JOB_EMAIL / JOB_PASSWORD env vars; rewrite the README - Add a pytest suite (fast + browser-marked) and GitHub Actions CI
- Add pydantic-settings: browser-use 0.11.13 imports it at package import but does not declare it, so a fresh install (CI, Docker) failed - Put api/ on pytest's path so a bare `pytest` can import `app`
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes from a full audit of the codebase, one commit per priority. Each commit builds and imports on its own.
Commits
3336a91feature/agent-model, which was never pushed)ac58fb3c633f09JWT_SECREThad a hard-coded fallback, shared job rows could be rewritten from a client-supplied URL8756de2159f791CREDENTIALS_KEY); LLMs only see an allowlisted persona; SSRF guard on server-side fetches; password passed to the agent as a placeholder so the LLM never sees it (the judge call is disabled because it leaked it); internal hosts blocked6adfc64ed6df96/streampipeline and the unused jobs-mcp / civic-guardrails services; status updates are scoped to the owning user (fixes an IDOR); uploaded CVs are kept; pytest suite and CI added; README rewrittenBreaking changes / deploy notes
JWT_SECRETandCREDENTIALS_KEYare now required. Docker compose won't start without them. Generate a key with:python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"JOB_EMAIL/JOB_PASSWORDenv vars and thejobs-mcp/civic-guardrailsservices have been removed.browser-useis pinned to0.11.13, and thelangchain-*packages are gone.Testing
cd api && pytest: 94 passed (fast suite, browser tests deselected)pytest -m browser: 8 passed against a local headless Chromium (fill, review, takeover, submit / done / skip, disconnect cleanup, internal-host blocking, password placeholder never sent to the model).github/workflows/ci.yml) runs the api, browser and frontend jobs.