Academic Research Skills for DeepSeek Harness

A complete DeepSeek Harness (DSH) port of academic-research-skills: contract-audited academic research, paper writing, simulated peer review, systematic-review screening, and an end-to-end research → write → review → revise → finalize pipeline.

The upstream repository content is shipped in payload/: 1,199 included files copied byte-for-byte from the pinned commit. Four full-page third-party web-page mirrors (IEEE, Springer Nature, ICMJE, and one CC BY article) are deliberately not distributed because their rights holders' terms are not covered by the upstream CC-BY-NC grant; the manifest that records their URLs, Wayback IDs, and SHA-256 is kept. Everything outside payload/ is the DSH adapter: skill/command registration, DeepSeek Harness tool and path adaptation, deterministic Python tooling, and the write-scope guard.

License and commercial use

This is an unofficial DeepSeek Harness port. It is not affiliated with, sponsored by, or endorsed by Cheng-I Wu or the upstream project.

The package is distributed under CC BY-NC 4.0 (SPDX CC-BY-NC-4.0), the same license as the upstream content. Non-commercial use only. This is source-public, non-commercial software; it is not OSI open source. Project policy (2026-10-28): the maintainer does not offer or pursue a commercial license. Commercial use requires permission from the upstream author (Cheng-I Wu); this project does not grant it. Do not contact the adapter author for a commercial license.

Attribution, as requested upstream:

Based on Academic Research Skills by Cheng-I Wu https://github.com/Imbad0202/academic-research-skills Licensed under CC BY-NC 4.0.

The runtime announcement and every adapted prelude repeat this attribution, so the credit travels with the text the model reads. See NOTICE.md, LICENSE, and ASSET-LICENSES.md.

Third-party content and API terms are audited separately in docs/third-party-audit.md. Two constraints survive the upstream licence: the default citation gate calls Semantic Scholar, whose API licence is non-commercial; and the payload's condensed COREQ/STROBE checklists require permission for commercial redistribution.

An optional technical helper remains for downstream users who independently obtain the necessary permissions; it is not an offered commercial path and does not grant a licence: tools/build-commercial.mjs produces a separate copy with Semantic Scholar skipped, COREQ/STROBE replaced by pointers, and IEEE quotations paraphrased; see docs/commercial-build.md. It does not grant a licence.

Runtime and search defaults

• Model: DeepSeek Flash through DeepSeek Harness. No Claude model, Claude CLI, or Claude API is required at runtime; upstream sonnet/opus labels are treated as quality tiers and rewritten to deepseek-flash.
• Context: DeepSeek Flash's 1,048,576-token context is used directly: the adapter tells the model to load complete references, keep one evidence ledger, and parallelize independent work with subagent/workflow instead of serializing it.
• Tools: deterministic ARS gates run through ars_script; the model is told to prefer scripts over model adjudication where a script can decide.
• Images: DeepSeek Flash accepts image input; figure/table/scan checks use read_image.
• Web search: the default source-discovery path is DeepSeek native web_search (provider deepseek-official), batched 1–4 queries per call. web_fetch handles known URLs. No Claude web tool is needed.

What the port exposes

Surface	Count	Names
Skills	5	deep-research, academic-paper, academic-paper-reviewer, academic-pipeline, sr-screener
Agent-role skills	4	ars-agent-synthesis, ars-agent-research-architect, ars-agent-report-compiler, ars-agent-screening-reviewer
Slash commands	16	/ars-full, /ars-plan, /ars-outline, /ars-revision, /ars-revision-coach, /ars-rebuttal-audit, /ars-abstract, /ars-lit-review, /ars-3w, /ars-reviewer, /ars-format-convert, /ars-citation-check, /ars-disclosure, /ars-mark-read, /ars-unmark-read, /ars-cache-invalidate
Tools	3	ars_status, ars_script, ars_env_setup
Hook layers	2	Session-start announce section; write-scope guard (infrastructure protection; the upstream phase fence is reported unavailable because DSH exposes no agent type)
Modes	35	Defined by the upstream skills; see payload/MODE_REGISTRY.md
Agent role files	47	43 unique roles: 4 plugin-exposed agent-role skills + 39 inline role prompts; the 4 root files are byte-identical mirrors

The port is faithful: the upstream instructions, mode protocols, templates, references, quality gates, Material Passport, citation verification scripts, and SR screening scripts are all included and reachable.

Install

Requirements: DSH >= 0.2.1-rc.3 (Node ships with DSH), Python >= 3.11 on PATH, and network access for the first ars_env_setup (PyPI). Supported platforms: macOS and Linux; Windows is not supported yet because the interpreter wrapper is POSIX sh.

1. Open the Plugins page in the DSH sidebar.
2. Choose Add plugin and enter the absolute path to this checkout (a Git URL or a tarball also works).
3. Install, then click Enable now. Closing the dialog leaves the plugin installed but disabled.
4. In a session, call ars_env_setup once (or ask the model to), then confirm with ars_status that python.ready is true.

The bundle patch inserts one Host row (academic-research-skills) that registers everything. The row is enabled by default; disable it to switch the whole feature off. The row requires the profile's skills, commands, tools, and systemPrompt services: Cordis treats a missing injected service as a hard prerequisite, so a profile without one leaves the row inactive rather than partially registered.

The bundle is published as academic-research-skills-dsh on npm. From a terminal:

dsh plugin --profile <profile> add academic-research-skills-dsh

or enter that package name in Add plugin. The Desktop profile requires the Desktop-installed dsh CLI.

A scoped mirror is published to GitHub Packages as @guoqinchen/academic-research-skills-dsh by .github/workflows/publish-github-package.yml. Installing it requires an authenticated npm client:

@guoqinchen:registry=https://npm.pkg.github.com
//npm.pkg.github.com/:_authToken=${GITHUB_TOKEN}

The npmjs package remains the canonical install path.

Creator-mode agents can also run plugin_manager install_bundle target=<absolute path>; the default presets disable that tool, so the Plugins page is the normal path. The package has no npm dependencies and is a plain ESM host bundle.

Use

Slash commands

Type /ars-plan, /ars-full, /ars-citation-check, etc. in the composer. A command submits the adapted upstream command prompt as one user message, so the next model step runs the workflow. Command output stays in the command plane and does not enter model requests.

Skills

Ask naturally, or invoke a skill by name. The model loads the skill body with the skill tool. The skill's resourceBase is its payload directory, and its compatibility prelude tells the model which relative prefixes resolve to the skill directory versus the payload root.

Agent roles

The four upstream plugin-exposed agents are model-invocable agent-role skills. The model loads one and passes its instructions to the subagent tool when a separate agent is wanted. The other 43 role files are read inline by the skill workflows, exactly as upstream.

Tools

• ars_status — plugin/payload provenance, counts, Python readiness, guard config, and runtime verification of every registered /ars-* command.
• ars_script — run one upstream Python or shell script under the payload root with the managed interpreter, argument array (no shell), timeout, and bounded output. Use it for deterministic gates such as citation verification, pipeline integrity checks, revision patches, and SR screening builds.
• ars_env_setup — create/refresh the managed virtualenv and install requirements-runtime.txt. Requires network access to PyPI.

Python environment

The deterministic upstream scripts need Python packages that are not part of the base system. The plugin never installs them silently:

1. Call the ars_env_setup tool (preferred), or ask the model to do so. It creates a virtualenv at ~/.cache/ars-dsh/venv (configurable) and installs requirements-runtime.txt, preferring uv when available.
2. ars_status reports ready, the missing packages, and the interpreter in use.

The model is told to run ARS scripts through the plugin's bin/ars-python wrapper, which resolves ARS_DSH_VENV → this install's entry in the runtime-state table (so a custom envDir and pythonPath work from a model-issued shell too) → the default managed venv → python3. The table is keyed by plugin root, so two profiles with different settings cannot overwrite each other. ars_script uses that same wrapper, so tool and shell cannot disagree about the interpreter.

Configuration

The plugin row accepts these fields in cordis.patch.yml (all optional):

Field	Default	Meaning
announce	true	Add the short ARS availability section to the system prompt.
commands	true	Register the 16 /ars-* commands.
tools	true	Register ars_status, ars_script, ars_env_setup.
guard	true	Register the write-scope guard.
guardInfra	true	Protect plugin enforcement files from writes.
pythonPath	null	Base interpreter used to create the virtualenv, and the interpreter scripts use when no managed venv exists. Recorded per plugin install, so ars_script, ars_status, and a model-issued bash bin/ars-python agree.
cacheRoot	~/.cache/ars-dsh	Parent directory for the virtualenv.
envDir	<cacheRoot>/venv	Managed virtualenv path.
scriptTimeoutMs	120000	Default timeout for ars_script; clamped to 1..2147483647.
maxScriptOutputBytes	200000	Per-stream output cap for ars_script; clamped to 4096..8388608.
modelProfile	deepseek-flash	Label reported by ars_status for the runtime profile. The prelude guidance itself always targets DeepSeek Flash.
webSearch	true	Include the DeepSeek native web_search default in the prelude/announce text.
commercialMode	false	Add the commercial boundary to the announce/prelude, export ARS_COMMERCIAL_MODE=1 to scripts, and let a commercial build skip Semantic Scholar.

DeepSeek Harness compatibility layer

The adapter rewrites, at registration time only (payload files stay untouched):

• ${CLAUDE_PLUGIN_ROOT} → the payload root;
• academic-research-skills:<skill> → the plain DSH skill name;
• Claude Code tool names → DSH tool names (Read→read, Write→write, Edit→edit, Bash→bash, Grep→grep, Glob→glob, WebSearch→web_search, WebFetch→web_fetch, TodoWrite→todo_write, AskUserQuestion→ask_user_question, Task→subagent);
• Python and shell script invocations → absolute payload paths through bin/ars-python;
• legacy model labels (sonnet/opus/haiku) → deepseek-flash; Claude Code → DeepSeek Harness; .claude/... runtime paths → the plugin layout;
• $ARGUMENTS in command bodies → the raw user input.

See docs/compat-report.md for the counted token inventory and the residual-pattern audit.

Write-scope guard

Upstream's PreToolUse hook is ported to a DSH ctx.tools.guard:

• Infrastructure protection applies to every caller and denies writes to the plugin's enforcement files (guard script, phase-scope manifest, agent definitions, hook/config files) anchored on the plugin root.
• Script output protection: ars_script accepts output flags (--out, --out-dir, --output, --dir, -o, ...). A declared output that resolves into the plugin's enforcement files is denied for every caller, and a single-phase agent may not run payload scripts at all. Positional arguments are not interpreted, so this is an infrastructure guard, not a sandbox: a script can still write wherever its own arguments point. Several payload scripts also make outbound HTTP requests (the arXiv, OpenAlex, Crossref, Semantic Scholar, and Chinese-literature clients), so ars_script is a network egress path as well — treat skill text as untrusted input when choosing which script to run.
• Phase fencing keeps upstream evaluate_decision's Bucket A logic for parity, but the registered guard never supplies an agent type: DSH tool executions have no agent_type field and the plugin keeps no role registry. ars_status reports guard.phaseFence: "unavailable" with the reason rather than implying a working fence. Normal sessions are unconstrained exactly like upstream's main session; infrastructure protection and schema-drift denial still apply to every caller.
• Schema-drift and workspace-escape denials match upstream behaviour.

Verification

node --test test/*.test.mjs          # unit, integration, and end-to-end suites
node tools/verify-install.mjs        # package, registration, guard parity, payload, schemas
node tools/check-payload.mjs         # payload integrity against the pinned digests

docs/verification.md is the dated record: it holds the exact pass/fail/skip counts, the host and toolchain each run used, and the known limitations. Those numbers are deliberately not repeated here, so this README cannot drift from a run. CI runs the same three commands plus the real @deepseek-ai/dsh-tools schema validator on every push.

What the suites establish:

• every upstream document parses, adapts, and lints clean, and the 5 skills, 4 agent-role skills, and 16 commands register with the live-host contract;
• the JS write-scope guard matches the upstream Python evaluate_decision on the captured fixtures and a generated decision matrix (decisions and advisory flags);
• the payload is byte- and mode-identical to the pinned snapshot (1,199 files), with PROVENANCE.json, REFERENCE-INDEX.json, and the requirements pins re-derived rather than trusted;
• ars_status, ars_script, and ars_env_setup conform to the DSH raw JSON Schema subset — enforced by a local subset test in CI and, additionally, by the real validator when the package is installed;
• payload immutability: PYTHONDONTWRITEBYTECODE=1 is set in both lib/env.js and bin/ars-python, so running the Python gates cannot create __pycache__ inside the read-only payload. If the payload is ever mutated, node tools/check-payload.mjs fails against the pinned digests and node tools/build-payload.mjs restores the committed snapshot.

The /ars-* composer interaction itself is a host/UI path and was not typed into the Web composer during verification; the handler code and the live follow-up message delivery were verified separately.

Upgrading the upstream payload

The payload is pinned; upgrades are deliberate and read-only until you rebuild. Check a target checkout or ref first:

node tools/check-upstream.mjs --upstream ../repo
node tools/check-upstream.mjs --ref <branch-or-commit> --json

A scheduled Upstream check workflow runs weekly and files an upstream-drift issue when the pinned payload no longer matches upstream.

Exit codes are 0 clean, 1 drift, 2 error. --print-snapshot prints the PAYLOAD_SNAPSHOT / BY_TOP_LEVEL values needed before rebuilding. See UPGRADING.md for the full procedure (identity, pins, distribution exclusions, rebuild, gates, release).

Known limitations

• Command UI execution: /ars-* handlers are unit-tested with a fake invocation, and their registration is verified at runtime through ars_status; the plugin does not drive the Web composer itself.
• Guard phase fence: unavailable by design. DSH does not expose an agent_type on tool executions; ars_status reports the mode and reason instead of pretending the fence is active.
• Update checker: upstream's /plugin update reminder is not ported; the pinned payload is upgraded with tools/check-upstream.mjs and a rebuild (see UPGRADING.md), not in place.
• Claude Code hooks file: payload/hooks/hooks.json is kept for provenance but not executed; its two behaviors are replaced by the announce section and the tool guard.
• Non-commercial license: CC-BY-NC-4.0 applies to the whole package; see License and commercial use.

Layout

index.js                 plugin entry (wiring only)
lib/paths.js             payload path resolution
lib/config.js            row config normalization
lib/frontmatter.js       dependency-free frontmatter parser
lib/adapt.js             DSH compatibility transformations
lib/skills.js            5 skill registrations
lib/agents.js            4 agent-role skill registrations
lib/commands.js          16 /ars-* command registrations
lib/announce.js          system-prompt section
lib/env.js               managed Python virtualenv
lib/python-contract.js   interpreter/venv/wrapper contract
lib/payload-contract.js  one payload snapshot contract
lib/inventory.js         one ARS surface inventory (skills/commands/agents)
lib/intake.js            one document intake pipeline
lib/registration.js      registration report + effect ownership
lib/reference-rules.js   pure reference conventions
lib/reference-policy.js  reference policy bound to payload + inventory
lib/model-profile.js     DeepSeek Flash + web-search guidance
lib/tools.js             ars_status / ars_script / ars_env_setup
lib/status.js            status assembly
lib/guard.js             write-scope guard port
bin/ars-python           interpreter wrapper
payload/                 1,199 included upstream files, byte-for-byte
                         (4 third-party page mirrors withheld; see NOTICE.md)
tools/build-payload.mjs  deterministic payload builder
tools/check-payload.mjs  payload integrity checker
tools/build-python-wrapper.mjs  wrapper generator/checker
tools/check-upstream.mjs upstream payload diff checker
tools/build-commercial.mjs commercial-clearance build (see docs/commercial-build.md)
tools/jev-*.mjs          optional TypeSafe/Jev checks (see docs/jev-integrations.md)
tools/check-release-parity.mjs  npmjs/GitHub Packages version parity
test/                    node:test suites

UPGRADING.md             upgrade procedure
docs/jev-integrations.md optional Jev tools and data boundaries
docs/release-checklist.md release and parity checklist
ASSET-LICENSES.md        PDF/font asset inventory
CITATION.cff             citation metadata

Provenance and license

Upstream: https://github.com/Imbad0202/academic-research-skills (commit b34c63f828b3535f403079a4f42bc7955da2f64b, version 3.24.0), by Cheng-I Wu, CC-BY-NC-4.0. This is an unofficial port, not affiliated with or endorsed by the upstream author. The DSH adapter is copyright (c) 2026 Guoqin Chen. See NOTICE.md, LICENSE, and CITATION.cff, plus payload/PROVENANCE.json.

GitHub's license detector reports NOASSERTION for CC-BY-NC-4.0; the SPDX identifier is declared in package.json and NOTICE.md. PDF/font asset coverage is listed in ASSET-LICENSES.md.