Autosci Daily Arxiv

/daily-arxiv

Run or manage the daily paper recommendation feed. Bare /daily-arxiv means "run today's recommendation pass now"; GitHub Actions is only the unattended scheduler for the same pipeline.

Load references only when needed:

• references/recommendation-and-ingest-policy.md — evidence, LLM decision schema, confidence gate, and auto-ingest guardrails
• references/automation-scaffold.md — GitHub Actions setup/status behavior, secrets, artifacts, and failure modes

Commands

• /daily-arxiv: run a one-off recommendation pass now. If config/daily-arxiv.yml is missing, infer defaults from the wiki and continue.
• /daily-arxiv setup: create or repair config/daily-arxiv.yml from config/daily-arxiv.yml.example; ensure .github/workflows/daily-arxiv.yml exists and that its daily-arxiv: job's env: block exposes both SEMANTIC_SCHOLAR_API_KEY and DEEPXIV_TOKEN — add the missing lines automatically rather than asking the user to hand-edit YAML (without these exposures the prepare step rate-limits out, identical symptom to never setting the secrets); and explain required secrets. See Setup Workflow below for the exact patch procedure.
• /daily-arxiv status: inspect config, workflow presence, schedule, mode, API/e-mail secret availability, and recent artifacts when available.
• /daily-arxiv disable: set schedule.enabled: false in config or tell the user what to change; manual /daily-arxiv must still work.

When running setup or status, treat S2/DeepXiv repo secrets as required (not optional) for any daily-cadence pipeline, and point the user at docs/daily-arxiv-deployment.md for the full setup checklist and symptom-keyed troubleshooting.

Inputs

• --mode inform|auto-ingest: default inform. Never infer auto-ingest from repo state.
• --hours N: pull papers from the last N hours; config/default is 24.
• --categories <cat...>: override configured arXiv categories.
• --max-recommendations N: maximum papers shown in the digest; config/default is 10.
• --max-auto-ingest N: cap for high-confidence auto-ingest; config/default is 1.
• --send-email true|false: workflow/setup preference for SMTP delivery.

Setup Workflow

Triggered by /daily-arxiv setup. Idempotent — re-running on a healthy repo is a no-op.

1. Config: if config/daily-arxiv.yml is missing, copy from config/daily-arxiv.yml.example. If present, leave it alone (the user's preferences are durable).

2. Workflow file: confirm .github/workflows/daily-arxiv.yml exists. If absent, point the user at docs/daily-arxiv-deployment.md and stop — re-creating the workflow from scratch is out of scope for setup.

3. Workflow env exposures (auto-patch): in .github/workflows/daily-arxiv.yml, locate the daily-arxiv: job's env: block. Use the Edit tool to ensure both lines are present as siblings of HAS_CLAUDE_CODE_AUTH:

   SEMANTIC_SCHOLAR_API_KEY: ${{ secrets.SEMANTIC_SCHOLAR_API_KEY }}
   DEEPXIV_TOKEN:            ${{ secrets.DEEPXIV_TOKEN }}
   

   • If both lines already exist, do nothing.
   • If only one is missing, append the missing line.
   • If the env: block doesn't exist at all (older workflow), insert it under the job with both lines plus the existing HAS_CLAUDE_CODE_AUTH / HAS_REVIEW_LLM flags. Do not touch any other step.
   • After any patch, tell the user what was changed and remind them to commit.

4. Secrets check: list which of ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN, SEMANTIC_SCHOLAR_API_KEY, DEEPXIV_TOKEN, and the optional SMTP secrets the user has configured. Use gh secret list when available; otherwise instruct the user to run it. Surface any missing-but-required secrets with the exact gh secret set command they need.

5. Summary: report what was created, patched, and what the user still needs to do (install the GitHub App, set missing secrets, verify with one gh workflow run daily-arxiv.yml).

Run Workflow

1. Resolve the Python interpreter and run the deterministic preparation:

   python3 tools/daily_arxiv.py prepare --wiki-root wiki --out .daily-arxiv/run/recommendation-context.json --out-feed .daily-arxiv/run/feed.json
   

2. Read .daily-arxiv/run/recommendation-context.json. Judge candidates with an LLM using the provided arXiv, wiki, Semantic Scholar, and DeepXiv evidence. Write .daily-arxiv/run/llm-decisions.json with decision, confidence, score, rationale, wiki_connections, and signals_used. In CI inform mode, an OpenAI-compatible review LLM may do this via:

   python3 tools/daily_arxiv.py recommend-llm --context .daily-arxiv/run/recommendation-context.json --out .daily-arxiv/run/llm-decisions.json
   

3. If mode is auto-ingest, use Claude Code runtime only: choose decision: ingest + confidence: high, obey max_auto_ingest, and invoke /ingest <arxiv-url> sequentially. Do not hand-write wiki or graph files. Third-party LLMs are recommendation-only and must not auto-ingest.

4. Finalize the digest:

   python3 tools/daily_arxiv.py finalize --context .daily-arxiv/run/recommendation-context.json --decisions .daily-arxiv/run/llm-decisions.json --out-md .daily-arxiv/run/digest.md --out-json .daily-arxiv/run/digest.json
   

5. Report strong recommendations, maybe-interesting papers, skipped duplicates, degraded signals, auto-ingest outcomes, and setup/status hints.

Wiki Interaction

Reads wiki/index.md, wiki/papers/, wiki/topics/, wiki/concepts/, wiki/methods/, wiki/ideas/, and wiki/log.md to build the interest profile and dedupe candidates.

Writes only scratch files under .daily-arxiv/ during inform runs. In auto-ingest, all durable wiki/raw mutations must come from /ingest.

Relationships

• /discover answers deliberate next-read requests from anchors, topics, or wiki state; it never ingests.
• /daily-arxiv watches the fresh arXiv stream and can notify daily or manually.
• /ingest is the only paper incorporation path. /daily-arxiv may call it only in explicit auto-ingest mode.