Getscipapers Requester

GetSciPapers Requester

Windows Runtime Commands

On native Windows, use the managed Windows runner and the native runtime command target. Set $runtime to the installed runtime root. Multi-agent installs usually use %LOCALAPPDATA%\ai-agents-skills\runtime. Then run:

$runtime = if ($env:AAS_RUNTIME_ROOT) { $env:AAS_RUNTIME_ROOT } else { "$env:LOCALAPPDATA\ai-agents-skills\runtime" }
& "$runtime\run_skill.ps1" "skills/getscipapers_requester/gsp_openclaw_helper.py" <args>

POSIX examples below use run_skill.sh and .sh command targets; use the Windows command target above on native Windows.

This is the external retrieval fallback. Do not use it before zotero for normal paper/library requests, and for review tasks that need a paper/book do not use it before both zotero and calibre have been checked. "Download" by itself does not bypass the library-first workflow; bypass only when the user explicitly says not to check/use the library or confirms outside retrieval after the local miss/ambiguity is shown.

When to use

• The paper is not in Zotero
• and, for review tasks, it is also not in Calibre
• The user explicitly says not to check/use the library, or confirms external retrieval after the library-first result is reported
• The task is DOI/ISBN/title resolution from external sources
• The user pasted many identifiers and wants batch retrieval

Base path

• $AAS_RUNTIME_WORKSPACE/skills/getscipapers_requester/

Use the managed runtime runner rather than invoking run_gsp_helper.sh directly. The runner sets OPENCLAW_WORKSPACE, PYTHONPATH, secrets, and workspace-local binaries.

Shared runner:

• bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh"

First-time setup

Before the core commands work, provision the dedicated venv once. This creates ~/.getscipapers_venv and installs the getscipapers fork from its default branch. The run scripts then export GETSCIPAPERS_BIN to the venv binary.

bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh" skills/getscipapers_requester/run_gsp_setup.sh setup

On native Windows, use the Windows runner and command target:

& "$runtime\run_skill.ps1" "skills/getscipapers_requester/run_gsp_setup.py" setup

Core commands

bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh" skills/getscipapers_requester/run_gsp_helper.sh run-getscipapers --timeout 180 -- getpapers --doi <DOI>

bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh" skills/getscipapers_requester/run_gsp_helper.sh resolve auto "<title>" --best

bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh" skills/getscipapers_requester/run_gsp_helper.sh make-manifest auto "<text-or-file>"

bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh" skills/getscipapers_requester/run_gsp_helper.sh doctor

Z-Library and DJVU fallback

Use the zlib backend only as an external fallback for books or papers after the library-first route has not satisfied the request, or when the user explicitly asks for zlib/external retrieval. For review tasks, check zotero and then calibre first unless the user explicitly opts out.

Before relying on the backend, check the local command help:

getscipapers zlib --help

Search before downloading:

getscipapers zlib --no-proxy --search "<title-or-isbn>" --search-limit 10

Download only after an exact match is established or the user selects a numbered candidate. Do not print zlib configuration values or credentials; when diagnosing configuration, report only file existence and key names. If the downloaded file is DJVU, do not attach it to Zotero automatically unless the user explicitly asks for that attachment. For "send it to me" requests, verify the file first and send the verified file directly through the requested channel.

For DJVU verification and reading, prefer DjVuLibre tools:

sudo apt-get install -y djvulibre-bin
djvused -e n book.djvu
djvudump book.djvu | sed -n '1,80p'
djvutxt book.djvu book.txt
djvutxt -page=<page> book.djvu page.txt
ddjvu -format=tiff -page=<page> book.djvu page.tiff

Use djvudump for structure, djvused -e n for page count, djvutxt or djvused -e 'print-pure-txt' for hidden text, and ddjvu for rendering or conversion. If Calibre is useful for conversion or metadata, use it as a fallback rather than the first verifier; on hosts with user-site Python package conflicts, run it with:

PYTHONNOUSERSITE=1 ebook-convert book.djvu book.txt

If the DJVU has no usable hidden text layer, render pages with ddjvu and then run OCR with tesseract or ocrmypdf when available. Avoid treating mutool, PIL, or generic ImageMagick probing as primary DJVU verification paths; they may fail or hang depending on local delegates. ImageMagick is still useful after a page has been rendered to TIFF/PNG.

Workflow

1. If DOI/ISBN is available, use it directly.
2. Otherwise resolve from title or text.
3. When title/text resolution returns multiple plausible matches, show the numbered candidates with title, authors, and year when available, then wait for the user's selected index before using --best, retrieval, attachment, send, or review steps. Exact DOI/ISBN requests do not need this disambiguation.
4. For many papers, create a manifest first.
5. For large batches, prefer dry-run style validation first.
6. If retrieval fails, report the failure precisely instead of hand-waving.

Watch state is a bounded, fail-closed local transaction. create-watch, list-watches, and update-watch share one cross-platform lock and use unique atomic replacements for watches.json. A corrupt, oversized, symlinked, or wrong-shaped watch store is reported without being reset or overwritten; repair or restore that state before retrying the handoff. Watch IDs include the full structured canonical watch key, require exact identity before reuse, and are made unique inside the locked transaction. Matching delimiter-era keys are migrated in place; collision-shaped nonmatching records are never reused. An exact-identity found record is already successful and is reused without reopening monitoring; failed or expired identities remain eligible for a new watch on retry. Stores with missing or duplicate IDs are rejected. Every stored record must also satisfy the complete baseline watch schema, including identity fields, service list, known status, timestamps, bounded counters, and sent-file hashes; optional deadline and note history fields are type- and range-checked too. Stored DOI, ISBN, or search identity must already be in its exact canonical grammar; watch_key must equal either the structured identity hash or its exact legacy predecessor. updated_at cannot precede created_at, and deadlines, last-check times, and note timestamps must remain within the record's declared lifetime. Store JSON rejects duplicate object members and non-standard constants such as NaN, and never serializes them, including in otherwise forward-compatible unknown fields. Creation, update, last-check, and note-event timestamps more than five minutes ahead of the current clock are rejected; accepted within-skew future records remain mutable by advancing event timestamps monotonically rather than moving them backward. The byte budget admits the bridge's full 3,000-item producer maximum even with bounded 500-code-point non-ASCII labels; individual label, identifier, initial/last/history note, and service fields remain bounded, and history records require valid timestamp/note shapes. Persisted string fields reject line breaks and Unicode control/format/surrogate characters, so helper-created state remains consumable by the bridge's matching schema.

Generated manifest JSON and DOI-list files use unique same-directory atomic stages. Pre-existing staging or destination symlinks are replaced as path entries and are never followed to an external target. Final manifest basenames use the full SHA-256 digest of kind plus source, so a short-prefix collision cannot overwrite another retained manifest generation.

make-manifest bounded-reads inline, stdin, or regular non-symlink input and admits at most 3,000 lines/items under a 2 MiB source ceiling, matching the bridge producer maximum when every DOI reaches the 500-character ASCII grammar limit. Unicode case-insensitive lookalikes are rejected rather than folded into another DOI. It caps metadata-resolution lines and output bytes separately, so bulk free text cannot fan out into unbounded requests or artifacts. Overlong inline text is validated as text rather than being mistaken for a filesystem path.

Only extract and make-manifest interpret their top-level source argument as stdin or a file. resolve queries and individual lines inside an already-read manifest source are always literal text; a title that resembles a readable local path is never reopened or sent as that file's contents.

Durability-bearing make-manifest and watch verbs parse an explicitly selected config strictly and require their configured manifest or state directory. They never switch those outputs to the temporary fallback root; an invalid config or unavailable configured directory is a nonzero handoff failure. The launcher exports a workspace-default config when the POSIX -f probe sees a regular file target; that probe may follow a live symlink, while a broken symlink is not exported. Every durability-bearing verb then performs strict directory-entry admission and rejects live/broken config symlinks and non-regular entries before durable state changes. Non-durable verbs retain the legacy permissive config reader and may use fallback settings; they do not establish manifest/watch durability. Relative paths inside an accepted strict config resolve against that config file's directory so producers and pollers do not depend on their current working directory. Strict config JSON rejects duplicate members and non-finite constants even in unknown fields; telegram_max_bytes must be an integer from 1 through 2 GiB. Invalid environment overrides use the bounded default only on non-durable configuration paths. Every required configured storage directory is then admitted by directory entry after create/existence, including the state_dir that lexically anchors a manifest directory. Configured and fallback live/broken symlinks, Windows reparse points such as junctions, and non-directory occupants are never followed for manifest or watch writes.

Crossref, Google Books, and OpenLibrary metadata requests stay on their exact documented HTTPS origins, refuse redirects, close every response, and stream under a fixed byte cap. Declared response framing is enforced, so a truncated body cannot be accepted merely because its prefix is valid JSON. Real requests run in isolated workers that the parent terminates and reaps under a whole-response wall-clock deadline. The deadline is independent of the socket idle timeout, so a slow-drip response cannot keep a resolver alive indefinitely; process creation itself may remain platform-dependent. Search queries and returned candidate counts are bounded, and malformed response shapes degrade without unbounded iteration. Metadata JSON rejects non-standard non-finite constants; candidate DOIs use an ASCII-only grammar, and ISBNs permit only ASCII digits/X separated by single hyphens or spaces before checksum validation. Remote score, year, and type scalars are bounded before ranking or manifest publication. Remote title, author, venue, publisher, and type text is HTML-unescaped, NFKC-normalized, tag-stripped, control/format/surrogate-cleaned, whitespace-collapsed, and bounded before candidate output. Selected and ranked summaries are sanitized again through their known-field schema at manifest publication, so a mocked or future resolver cannot persist extra or structurally unsafe metadata. Exact DOI fields, bare DOI lines, and DOI URLs retain every grammar-valid suffix character; free-text extraction alone applies surrounding-prose punctuation heuristics.