dsh-retraction-watch

Has anything been published against the papers you cite?

Crossref work records + the verified filter=updates:<doi> query, a per-DOI status, an append-only timeline that shows changes, and REP evidence you can re-read. clear means no registered notice — never "this paper is valid".

English · 简体中文

---

What it does

• retraction_check(doi) → retracted / concern / corrected / notice / clear / unknown, with the notice DOIs and a REP evidence record holding both Crossref responses.
• retraction_watch(dois) → the same for a bibliography, with change detection: the timeline records every observed status, so "clear → retracted" is reported as a change.
• retraction_report(items) → Markdown table + SARIF 2.1.0 (retractions/concerns = errors, corrections = warnings).

Verified provider capability (not guessed)

Query	Result
GET /works/{doi}	the cited work; if it carries update-to, the DOI is a notice
GET /works?filter=updates:{doi}	works that update this DOI — verified live, HTTP 200 (2026-11-08)
GET /works?filter=update-to:{doi}	does not exist, HTTP 400
filter=relation.type:is-retracted-by	not a valid relation type, HTTP 400

The plugin only uses the two working queries. When the update query fails it reports unknown and a skipped check — never clear.

Status meanings

Status	Meaning
retracted	a retraction notice points at this DOI
concern	an expression of concern points at it
corrected	a correction notice points at it
notice	an update notice of another type points at it
clear	Crossref has no registered update notice
unknown	the DOI is not in Crossref, or the lookup failed

clear is absence of evidence, not evidence of validity: Crossref notices can lag publication, and Retraction Watch coverage is not queried in v0.1. unknown is never collapsed into clear.

What it does not do

• No Retraction Watch ingestion (licence/coverage to be checked before that step).
• No quality judgement: a correction does not make a paper wrong, and a retraction's reasons are not evaluated.
• No scraping beyond the two Crossref queries; no CNKI, no Sci-Hub.
• No scheduled runner: the timeline is updated whenever the tool is called; scheduling belongs to the host (dsh-jobs) or the user.

Compatibility & install

DeepSeek Harness	dsh-v0.2.1-rc.5 (verified 2026-11-08); peer >=0.2.1-0 <0.3.0 || >=0.2.2-0 <0.3.0
Node	^22.19.0 || >=24.9.0 (tested 24.22.1), ESM only
Peers	@perrylink/dsh-research-core ^0.1.0, @deepseek-ai/cordis ^4.0.5-alpha.1, @deepseek-ai/dsh-tools, @deepseek-ai/schemastery ^3.18.5-alpha.1
Runtime dependencies	zero

dsh plugin --profile <name> add @perrylink/dsh-research-core dsh-retraction-watch
dsh --profile <name> --dump-config | grep dsh-retraction-watch

Configuration

Key	Default	Meaning
enabled	true	false mounts nothing
root	.retraction-watch	store, cache, timeline and records root
userAgent	dsh-retraction-watch/0.1.0 (…)	sent to Crossref
mailto	—	Crossref polite-pool contact (optional)
timeoutMs	15000	per-request deadline
minIntervalMs	150	per-host interval
retries / maxRedirects	1 / 3	bounded retry and redirect budget
cacheTtlMs	604800000	response cache (7 days; fresh bypasses)
maxBatch / batchConcurrency	50 / 3	retraction_watch bounds

Security boundaries

• HTTPS only, exact-host allowlist (api.crossref.org), no credentials in URLs, no custom ports.
• Redirects are followed manually and each hop is re-validated against the allowlist.
• Responses are stored as immutable snapshots; the evidence records URLs, never secrets.
• Cached responses are best-effort: a cache failure is a miss, never a check result.

Evaluation

23 tests across three files: update-type mapping, notice filtering, aggregate status precedence, check construction (clear / retracted / unknown / network failure), runtime (retraction, clear with caveat, 404 → unknown, failed update query → unknown, correction, invalid DOI, re-check without an evidence conflict, no caching of a transient 5xx, polite-pool contact kept out of stored evidence), timeline change detection across three runs (the first observation is not a change), per-item batch isolation, batch cap, tool schemas/renderers, package contract.

Manual sample set (no benchmark is claimed for this plugin; its correctness is the Crossref contract, not a learned decision):

Scenario	Expected	Covered by
retraction notice points at the DOI	retracted, check fail	test
correction notice	corrected, check warn	test
expression of concern	concern, check fail	test
no notices	clear + absence caveat	test
DOI not in Crossref	unknown, not_found skipped	test
update query 5xx	unknown, network_error skipped	test
status changes between runs	changed: true, previousStatus	test
first observation of a DOI	changed: false, no previousStatus	test

Known limitations (v0.1)

• Retraction Watch data is not queried; Crossref-only coverage.
• A notice published without a Crossref update-to link is invisible here.
• The timeline is append-only JSONL with no compaction; prune manually for very large bibliographies.
• No email/webhook alerts; polling must be driven by the caller.

License

Apache-2.0. No third-party runtime code.