daily-paper-digest

name

daily-paper-digest

description

Discover arXiv papers for a configurable recurring time window or a specified date, retrieve conference papers through Crossref, OpenReview, or a normalized JSON proceedings export, filter them by a user's research preferences, and create budget-controlled six-part paper reviews with optional scheduled email delivery. Use when Codex is asked for a daily paper feed, conference-paper shortlist, research-interest alert, cheap-API paper digest, GitHub Actions literature scan, email paper alert, or one-to-many review run with either an OpenAI-compatible provider or Codex.

Daily Paper Digest

Run the project pipeline without changing the six-part review semantics. Separate discovery/ranking from full-text synthesis so broad candidate scans remain inexpensive.

Workflow

1. Locate the project root containing pyproject.toml and the single committed config.toml.
2. Edit only non-secret preferences, provider metadata, limits, output, and email settings in config.toml. Keep API keys and SMTP credentials in the environment variables named by the config.
3. Read references/configuration.md before changing a source, deploying GitHub Actions, or enabling SMTP delivery.
4. Run a zero-model preview first for a new preference profile:

python .agents/skills/daily-paper-digest/scripts/run_digest.py run --config config.toml --dry-run

5. Inspect selection-decisions.json and selected.json. Adjust interests, include/exclude keywords, categories, the decision policy, or safety limits when selection is visibly off.
6. Generate with an inexpensive OpenAI-compatible provider or the current Codex login:

python .agents/skills/daily-paper-digest/scripts/run_digest.py run --config config.toml --backend openai_compatible
python .agents/skills/daily-paper-digest/scripts/run_digest.py run --config config.toml --backend codex

7. Report the manifest status, selected/completed counts, reserved budget, failures, and output paths. Do not claim completion when the manifest says partial.

Source selection

• Use arxiv with discovery.window for a recurring local-time half-open range [start, end). The pipeline converts its boundaries to GMT for arXiv. An explicit --date disables the window for that run and accepts today, yesterday, or YYYY-MM-DD.
• Use crossref for unattended proceedings discovery by conference name and publication-date range.
• Use openreview for a venue ID such as ICLR.cc/2026/Conference. Prefer accepted; use all only when under-review submissions are intended.
• Use json for non-OpenReview proceedings or a curated list. Read references/configuration.md before mapping unfamiliar metadata.

Cost and evidence rules

• Use deterministic ranking for zero-cost previews. Use selection.ranker = "llm" when semantic relevance is the primary selection signal and the configured budget covers the candidate set.
• For large LLM scans, keep discovery.max_candidates, selection.max_selected_papers, and review.max_papers distinct. Retaining 500 papers must not silently trigger 500 deep reviews.
• Apply hard negative-keyword exclusions before LLM selection. Give the model an explicit natural-language decision policy and require one include or exclude decision with a reason for every remaining paper. Do not request or derive numeric relevance scores. Reject incomplete batch responses instead of silently dropping papers.
• When using DeepSeek V4 Flash for filtering, use model deepseek-v4-flash, enable JSON Output, and disable thinking for the selection calls.
• Apply the configured token and estimated-dollar limits as hard gates. Never raise them silently.
• Download and read full main text only for selected papers. Stop before appendices/references by default.
• Preserve the fixed narrative sections: Background, Motivation, Core Idea, implementation-level Method, Experiments and what they demonstrate, and Conclusion.
• Explain Idea conceptually and Method as an input-to-output implementation trace with essential interpreted equations. Do not substitute translation or abstract paraphrase for synthesis.
• Label abstract-only fallbacks honestly and do not invent missing experimental or method details.
• Preserve clickable arXiv, OpenReview, proceedings, or DOI links in outputs.

Resume and scheduling

Reuse existing per-paper summary checkpoints unless the user explicitly requests regeneration with --force. For GitHub Actions, keep non-secrets in config.toml, put model and SMTP credentials in repository Secrets, and use the included workflow. Treat a partial manifest as a failed automation that needs inspection. Send success, partial, and failure notifications without exposing credentials in logs. Never commit .env, API keys, SMTP authorization codes, or generated private-paper content.