A little pause. A clearer perspective.
Phission is an interactive learning workspace for looking more carefully at email requests and link destinations. This release is a synthetic-only demo, not a security product or a live phishing detector.
Choose one of five authored emails, select a link, and explicitly run demo analysis. Every result is simulated. No email URL is opened, fetched, or submitted to a service. There is no mailbox connection, credential setup, arbitrary URL input, upload feature, analytics, database, or hidden live-mode switch.
Local production-build screenshot, captured in a 1440 × 1000 Chromium desktop viewport from 29e4418 on 2026-09-26. Messages and reputation results are synthetic; this is not live detection.
Requirements: uv, Python 3.13 or 3.14, and the platform prerequisites in the Reflex installation guide (including unzip on Linux). Python dependencies are locked in uv.lock; Reflex is pinned to 0.9.12. The generated reflex.lock/bun.lock and reflex.lock/package.json pin frontend dependencies. .python-version selects Python 3.13 by default. This setup is validated on Linux/WSL; native Windows and macOS are not yet verified.
make setup # locked Python dependencies + initialize project-local Bun
make runOpen http://127.0.0.1:3000. make run builds and serves the frontend and Python backend together in production mode, bound to loopback. Stop it with Ctrl+C. Changes require a restart/rebuild. The Makefile invokes uv with the committed lockfile and keeps the uv cache in .cache/uv.
Dependencies and the first frontend build require network downloads. make setup initializes Reflex's project-local Bun 1.4.0; app builds do not require a separate Node/npm installation. The Makefile selects local Bun and refuses to build if it is missing, preventing silent npm fallback and lock replacement. rxconfig.py enables frozen frontend installs. Reflex initialization and Bun use different JSON whitespace styles, so scripts/frontend.py preserves the committed manifest formatting only after verifying identical decoded metadata and identical Bun lock bytes. Any dependency change fails and remains on disk for diagnosis. CI checks for drift after setup as well as after builds. The Makefile keeps downloaded tooling and caches under the project’s ignored .cache/, with the virtual environment in .venv/. Framework telemetry is disabled in rxconfig.py and the Makefile. Runtime example analysis itself does not use external services. System fonts and local assets are used; there are no remote fonts, CDNs, or embedded media.
To select Python 3.14 explicitly, use UV_PYTHON=3.14 make setup and the same UV_PYTHON=3.14 prefix on subsequent commands. Do not add API keys or real email to this repository.
After installing dependencies:
make check # Ruff lint + formatting check, then pytest
make test # behavioral tests only
make format # explicitly apply Ruff formatting
make build # optimized production frontend in .web/build/client/Check/test targets do not install dependencies, write Git hooks, or modify source. CI runs checks and a production build on Python 3.13 and 3.14. Tests cover MIME/header/Unicode handling, nested HTML destinations, duplicates and prefix-overlapping links, incomplete coverage, malformed responses, score boundaries including zero, target isolation, loading/empty/error states, and credential-free application import/runtime with outbound connections blocked.
With make run active in one terminal, use a second terminal with Node 22.23.2+:
npm ci --ignore-scripts
npx playwright install chromium
npm run test:browserChromium also needs its platform libraries. These Node dependencies are test tooling only, separate from the app's Bun frontend lock. Set PHISSION_URL=http://127.0.0.1:3157 if using another local port. The checks cover keyboard navigation, score/unknown/no-links flows, session isolation, responsive widths, four axe accessibility scans, no-JS messaging, nested 404s and attempted external browser requests. Screenshots and JSON evidence go to ignored .scratch/browser-validation/; CI retains them as artifacts for 30 days. Automated accessibility checks are not a screen-reader or physical-device audit.
For deliberate dependency updates, regenerate and review both frontend lock files with the pinned Bun/Reflex toolchain, then verify a clean build leaves them unchanged. Do not delete a lock or switch package managers merely to make CI pass.
This is not a standalone static site. JavaScript and a running Reflex Python backend/WebSocket connection are required for inbox selection and analysis. An exported frontend alone does not make the demo interactive. Production Reflex 0.9.12 serves both parts on a single port; see the self-hosting documentation before any separate deployment work. This repository does not provision or deploy a service. Do not use a development server as a public deployment.
| Real application logic | Authored simulation |
|---|---|
| Standard-library RFC822/MIME parsing | All five emails, names and destinations |
| Plain text and HTML anchor extraction | Local IPQS-shaped result objects |
| Validation of provider success and integer score range | Scores, concern levels and unavailable result |
| Per-session target selection and stale-result prevention | No live reputation lookup or independent phishing classification |
The scenarios include an ordinary workshop notice, an urgent credential lure, a friendly HTML label pointing to a different destination, an unavailable result with an uninspected attachment, and a note without links. Destinations use reserved .example/.test domains; the recipient uses example.org.
A score of 0 is valid, not an error. Values 0–74 represent a lower signal, 75–84 caution, and 85–100 high concern in this demonstration. These illustrative thresholds are not validated detection performance. A score is not a probability or a guarantee of safety. Invalid responses are errors; unavailable results are unknown, not safe. The normalizer accepts already-decoded Python objects and validates them; there is deliberately no live HTTP/JSON transport adapter, timeout/retry layer, or provider authentication.
One Python Reflex application; no separate API service, database, queues, or new machine learning.
phission/models.py: typed messages, extracted links, coverage and scan outcomes.phission/parsing.py: bounded stdlib MIME parser (256 KB / 100 parts), decoded headers/text, HTML-to-text conversion and HTTP(S) link extraction.phission/fixtures/andphission/demo.py: fixed synthetic manifest and deterministic, message-and-URL-bound responses. Files load on the inbox event, not on module import.phission/scanning.py: success/unknown/error normalization, separate score and cautious guidance.phission/state.py: in-memory per-session selection/results. Nothing is persisted to disk; a server restart loses sessions. No deep-link or browser-history routing for message selection.phission/phission.pyandassets/styles.css: one fixed page, semantic controls and responsive reading workspace. No runtime page generation or compilation on inbox actions.
Original email HTML is never rendered. The reading pane shows plain text; extracted destinations and HTML labels are shown separately, as text inside selection buttons. Text is not destructively rewritten to replace URLs. Attachments, embedded messages, image-only content, relative/non-HTTP links, redirects, obfuscated destinations and some malformed content are outside parsing coverage. Partial, unsupported and malformed messages are distinguished from “no HTTP(S) links found.” This bounded parser is not a production email sanitizer or a comprehensive threat detector.
The interface uses semantic buttons/headings/landmarks, visible keyboard focus, selected-state text, a polite live status region, reduced-motion styling and a mobile back-to-messages link. Message selection moves focus to the reading heading; tab to destinations and use Enter/Space to select and analyze. Warnings use text, not color alone. Original images and HTML are not embedded.
Read-aloud is not included in this slice: there is no server speech engine or browser speech service. Browser/OS reading tools and assistive technology may be used independently. JavaScript-disabled users can read the static introduction and limitation text but cannot interact with the demo. Browser, mobile viewport and assistive-technology checks complement the unit tests; those tests alone do not establish accessibility conformance or physical-device support.
The original Phission prototype was built in 2023 for HCI and affective-computing coursework, including usability testing. The original findings are no longer available. This refresh does not make outcome claims or reinterpret historical survey scores.
The original Gmail/IPQS integrations, server-local speech, Pynecone configuration, database and raw study artifacts are retired from the maintained tree. They are not archived as a second application. The historical source remains available at exact commit 0ced1f54baccc2ed6abd307d43971e82315c3961. No history has been rewritten. Reintroducing real mail or live scanning would require a separate security/privacy scope, not an environment-variable toggle.
