Send Me Research

Configurable daily research digests powered by Codex.

This project automatically watches fresh papers from arXiv and OpenAlex every day, dedupes them, lets Codex decide what is actually worth reading, and sends polished HTML email digests to one or more recipients.

The hosted GitHub Actions workflow uses a dedicated Codex ChatGPT login. Do not upload your normal ~/.codex/auth.json: Codex refresh tokens are single-use, so sharing one token family between your laptop and GitHub-hosted runners will eventually invalidate one side.

A live sample digest is published at larstalian.github.io/send-me-research.

You can run it in two modes:

• one default digest from .env
• multiple custom digests with different recipients and different research priorities

Out of the box, the default audience is:

• LLMs
• agents
• robotics / embodied AI
• cyber / AI security

with extra weight on post-training, fine-tuning, code generation, tool use, and agentic-task improvement. You can keep that default, or define completely different stacks per recipient in digest_profiles.example.json or the Multiple People, Different Stacks section below.

Setup

1. Install uv and the Codex CLI.
2. Run codex login.
3. Copy .env.example to .env.
4. Fill in your SMTP settings.
5. Install deps:

uv sync

Use Gmail with an app password, not your normal password.

Default Mode

If you only use .env, the app sends one digest to EMAIL_TO using the built-in default audience:

• LLMs
• agents
• robotics
• cyber / security
• extra weight on post-training, fine-tuning, code generation, tool use, and agentic-task improvement

That is the fastest path if this is mainly for you.

Multiple People, Different Stacks

Create digest_profiles.json from digest_profiles.example.json.

Each profile can define:

• name
• recipients
• description
• priority_keywords
• top_n

Example:

{
  "profiles": [
    {
      "name": "Applied ML",
      "recipients": ["ml@example.com"],
      "description": "LLMs, agents, post-training, code generation, and evaluation.",
      "priority_keywords": ["fine-tuning", "post-training", "code generation"],
      "top_n": 15
    },
    {
      "name": "Security",
      "recipients": ["security@example.com"],
      "description": "Prompt injection, jailbreaks, agent security, and AI red teaming.",
      "priority_keywords": ["prompt injection", "jailbreak", "red teaming"],
      "top_n": 10
    }
  ]
}

If digest_profiles.json is present, the app sends one digest per profile. If it is absent, the app falls back to the default single profile from .env.

Local Commands

Pick a date:

TARGET_DATE="$(TZ=America/Los_Angeles date +%F)"

Load env:

set -a
source .env
set +a

Preview:

uv run send-me-research preview-digest --date "$TARGET_DATE"

Send:

uv run send-me-research run-digest --date "$TARGET_DATE" --send

Preview one profile only:

uv run send-me-research preview-digest --date "$TARGET_DATE" --profile Security

See resolved profiles:

uv run send-me-research list-profiles

Run tests:

uv run pytest -q

Required Env

Always required:

• EMAIL_FROM
• SMTP_HOST
• SMTP_PORT
• SMTP_USERNAME
• SMTP_PASSWORD

Only required when you are not using profile config:

• EMAIL_TO

Common optional settings:

• DIGEST_TIMEZONE
• CODEX_ENABLE_SEARCH
• CODEX_ENABLE_WILDCARD_DISCOVERY
• TOP_N
• CODEX_SHORTLIST_SIZE
• ARXIV_MAX_RESULTS
• ARXIV_CATEGORIES
• ROBOTICS_SPOTLIGHT_COUNT

GitHub Actions

Two workflows are included:

• .github/workflows/daily-digest-hosted.yml
• .github/workflows/daily-digest.yml

Hosted mode is the recommended setup if you want:

• a public repo
• no daily commits for state
• no always-on self-hosted runner

Hosted Setup

Set these GitHub secrets:

• CODEX_AUTH_JSON_BASE64
• CODEX_CONFIG_TOML_BASE64
• CODEX_SECRET_SYNC_TOKEN, required so hosted runs can persist refreshed Codex auth back to GitHub Secrets
• EMAIL_FROM
• SMTP_HOST
• SMTP_PORT
• SMTP_USERNAME
• SMTP_PASSWORD
• EMAIL_TO optional
• DIGEST_PROFILES_JSON optional
• OPENALEX_MAILTO optional

Best path:

./scripts/sync_github_hosted_secrets.sh

That script:

• starts a fresh hosted-only Codex device login
• uploads that temporary Codex auth
• deletes the local temporary auth copy
• uploads mail secrets from .env
• uploads digest_profiles.json as DIGEST_PROFILES_JSON if the file exists
• sets DIGEST_AUTOMATION_MODE=hosted
• refuses to sync if the hosted-only Codex login cannot pass a real codex exec probe

If DIGEST_PROFILES_JSON is set, the workflow uses that instead of a checked-in digest_profiles.json. That is the cleanest way to keep recipient emails and audience configs private in a public repo.

Create a repo-scoped GitHub token that can update Actions secrets and put it in .env as CODEX_SECRET_SYNC_TOKEN before running the sync script. The workflow refuses to consume Codex auth unless that token is present, and it writes the refreshed ~/.codex/auth.json back after every Codex attempt, including failed digest runs.

State And Outputs

The workflows do not commit state back to git. They restore the latest state artifact, use it for dedupe, prune it, and upload a fresh artifact after each run.

Local scratch paths:

• state/
• out/

Neither is tracked in git.