DreamShift

DreamShift is an autonomous job application system built for quality, not volume. It finds current roles, tailors truthful materials, and prepares applications while enforcing checks that prevent duplicates, fabricated claims, and low-quality submissions.

It automates the repeatable work and keeps a human in the loop for the decisions where judgment or personal data matters.

What happens automatically

1. Refresh vetted official employer feeds.
2. Remove expired, duplicate, previously submitted, and rejected postings.
3. Score fit with deterministic rules before using a model.
4. Use a model only for strong finalists that need semantic judgment.
5. Tailor a visible resume using verified candidate facts.
6. Validate the resume and prepare a small named approval batch.
7. Track every state, stop reason, receipt, and model-resource total.

DreamShift can safely work unattended through the prepared-batch stage. Final submission remains candidate-controlled. It stops on CAPTCHA, login walls, changed questions, unsupported claims, required data outside the approved scope, or weak receipt evidence.

Two operating lanes

Review lane is the default: prepare a small batch, inspect it, and decide what to submit. It handles new questions, custom answers, and applications where judgment matters.

Certified fast lane is opt-in and deliberately narrow: standard Ashby forms containing only name, email, resume, and an explicitly authorized required phone field. It never uses a model, never touches optional fields, never falls back to a less-safe browser path, and stops on CAPTCHAs, login walls, custom questions, or any schema change. A candidate can authorize a bounded batch in this lane without reviewing every routine form one by one.

Start here

Requires Python 3.11+, Git, and Chrome or Chromium.

git clone https://github.com/Iamevanjordan/DreamShift.git
cd DreamShift
python3 scripts/bootstrap.py

The setup wizard creates an isolated candidate workspace under ~/.applypilot/. Personal data, resumes, credentials, answers, and browser state never belong in the repository.

The legacy applypilot command remains available for upstream compatibility. Candidate-facing examples use dreamshift:

source .venv/bin/activate
dreamshift official-discover
dreamshift official-triage
dreamshift run score tailor cover pdf --validation strict
dreamshift control-status
dreamshift review-queue
dreamshift secure-profile
dreamshift fast-lane --enable --limit 2

Read START_HERE.md for candidate onboarding and APPLICATION_CONTROL.md for the approval boundary.

The human handoff

Every batch identifies:

• employer, role, compensation, and current official URL;
• fit evidence and eligibility risks;
• exact resume and cover-letter artifacts;
• every nontrivial answer;
• the personal data that would be transmitted;
• the one action the candidate must take.

The browser may stage an approved form. The candidate reviews the visible result, handles any human-verification step, and makes the final submission decision. A model claiming success never counts as a confirmed application; DreamShift requires independent receipt evidence.

Resource discipline

Discovery, deduplication, state transitions, validation, and receipt checks are deterministic. Model calls are reserved for matching, truthful rewriting, and ambiguous questions. The default Codex path uses a low-cost model at low reasoning effort and escalates only when an evaluation identifies a specific failure.

Run the zero-model controls at any time:

dreamshift control-eval
pytest -q

Portability

The portable system is the repository, versioned schema, deterministic evaluations, and candidate-profile contract. Each user creates their own local private workspace. A clean installation must pass the control suite before it can prepare a live application.

For a second person or a separate test candidate, choose a separate workspace:

python3 scripts/bootstrap.py --data-dir ~/.dreamshift/my-candidate

Public-release gate

Never publish a private working history that once contained candidate data. Export the sanitized tracked tree into a new repository and run:

python scripts/public_release_check.py --history
pytest -q
dreamshift control-eval

See docs/RELEASE_READINESS.md for the verified clean-export process and the remaining publication decisions.

Upstream and license

DreamShift began with Pickle-Pixel/ApplyPilot, first published in 2026. ApplyPilot provided the pipeline foundation that made this system possible; DreamShift builds on it with candidate isolation, deterministic controls, official-feed discovery, auditable approval batches, receipt verification, resource accounting, and safer portability.

The project remains licensed under AGPL-3.0. Preserve the upstream attribution and license in every fork or hosted distribution.