Acquire Research Papers
Deliver the official publisher PDF and the publisher's raw BibTeX export as a verified pair. Discovery metadata nominates candidates and supports comparison; it never becomes the final citation artifact.
Run the installed skill
Use this shared skill in Codex, Claude Code, Kiro IDE, or Kiro CLI. Resolve its actual installed directory, then run its portable launcher through the selected host's shell tool:
text
python <skill-directory>/scripts/arp.py --help
python <skill-directory>/scripts/arp.py doctor --host <actual-host> --skill-directory <skill-directory>
python <skill-directory>/scripts/arp.py config status
Python 3.11+ and
are required. The standard-library launcher installs the fixed release wheel from
only after verifying SHA-256, in a separate cache outside the skill. No developer checkout or project environment is needed. A release wheel that has not yet been published cannot bootstrap from its release URL; a matching local release wheel may be supplied through
for isolated validation.
Read host installation for host paths, Kiro custom-agent resources, and checked support limits. The host doctor performs local checks by default; host loading, remote availability, and institutional access remain separate and unverified unless checked explicitly.
Check
first. IEEE, MinerU, Elsevier, Semantic Scholar, and OpenAlex are optional and independent; configure only the capability needed for the current task. New installations use an editable local
outside the repository:
,
, and
. A user may fill it locally or use non-echoing
prompts. Never ask for secret values in chat or command arguments. Legacy DPAPI installations remain legacy until explicit
arp config migrate --from dpapi
on the original Windows account. See
credentials and cache.
Select the user's named institution profile for IEEE access. Stored configuration and historical checks do not prove current subscription or API entitlement. Public schema snapshots are included for reference;
arp schema <schema-name.json>
returns the authoritative schema bundled with the running release.
Route the task
| Request | Command and reference |
|---|
| Known DOI, publisher URL, exact title, or explicit paper list | ; read acquisition workflows |
| Venue/topic/year/count corpus | arp discover corpus --progress
|
| Nearest work, Related Work, claim citations, or gap evidence | → review → acquire → evidence; read research mode |
Start with
copyable recipes. Related Work and gap analysis share one research workflow; choose the question type in the brief. Discovery writes state only. Evidence review produces
; add
for tables. Never draft placeholder conclusions during discovery.
Advanced tasks:
author bibliographies,
library audit/update/import/organize and explicit chapters,
independent capabilities and run recovery,
institution profiles, and
explicit Markdown export. Read only the reference needed for the task. Author and library audit default to JSON plus
; CSV is opt-in.
Before planning venue coverage, use
arp sources --spec <corpus.yaml>
and read
sources and coverage. This offline diagnostic describes capability, not live coverage or universal completeness. Translate attached documents into the same explicit inputs. Keep corpus and research frozen hashes unchanged when running library or author tasks.
Review and completion
- Preserve original BibTeX bytes, BOM, encoding, and line endings. Decode only a parsing view. Never generate or repair citations with an LLM or a metadata API.
- Verify DOI, title, year, author scope, and venue. Report ambiguous identities and partial evidence explicitly.
- Corpus decisions inspect and record title-and-abstract evidence. Acquire only the frozen bound by ; keywords are optional and full text is not required for corpus relevance.
- The executing agent makes semantic decisions and records the actual host. New v2 decisions set and to that host; include model/version only when actually available. Preserve legacy Codex v1 review bytes and hashes. Never relabel another host as Codex or invent model metadata. Identity, relevance, and full-text evidence policies remain distinct.
- retains selections in ; never remove difficult papers from a frozen selection or lower relevance thresholds to fill a quota.
- Library audits distinguish the local PDF version, citation placement, and current metadata. A current citation does not prove the PDF is current. Review before explicit
arp library apply --plan <path>
- Author discovery reports source coverage, identity ambiguity, formal/early-access/preprint groups, and separate issue/online dates. Never claim an absolutely complete bibliography.
- Ordinary fetch, manual fetch, and corpus workflows produce no Markdown. Research uses an internal MinerU cache; export Markdown only after an explicit user destination request.
- A standalone needs no BibTeX but retains user-supplied trust. It does not satisfy the official-pair delivery contract. Capability formats structured records; it is separate from of full text.
- Capability runs record input/output hashes, item states, and next actions. Reuse the same and command inputs to resume; use
arp run status --run <directory>
Access boundaries
Read source policies before changing an acquisition contract and configuration before institution adaptation.
Never print, log, commit, or send credentials, tokens, cookies, browser storage, or decrypted values. Keep state and deliveries outside the public repository. IEEE uses an isolated visible profile for the selected institution; credentials are released only at its exact configured hostname. A successful identity-provider login is separate from return to IEEE, a verified download, and proof of institutional entitlement.
Never automate ScienceDirect article pages or ScienceDirect organization-login flows, attach to the user's normal Chrome, read its profile, or export Cookie/session data. Authorized users download PDF and raw BibTeX themselves; the CLI handles local stable files afterward. Do not bypass publisher/institution controls or substitute mirrors, discovery BibTeX, or generated citations.