lunheng-article-pipeline

dsh plugin --profile web add lunheng-article-pipeline

Lunheng (lunheng-article-pipeline) — a DeepSeek Harness bundle for multi-agent long-form writing

🌐 English (this file) ｜ [中文](README-zh.md) ｜ [Español](README-es.md) ｜ [Português](README-pt.md) ｜ [हिन्दी](README-hi.md)

版本：v19.27.16（DSH bundle：package.json + cordis.patch.yml + lib/index.js）

A DeepSeek Harness (DSH) bundle that registers two on-demand agent skills:

- lunheng-article-pipeline — the main 9-role pipeline (T1–T9, see below).

- lunheng-commands — a thin wrapper exposing 13 /lunheng-* slash commands (draft / resume / cite / audit / journal / ppt / history / rollback / status / stats / gate / abort / help).

No new role, no new M-gate item; see skills/lunheng-commands/SKILL.md.

The main skill turns long-form production — academic papers, industry analysis, business commentary, and long-form articles — into a 9-role pipeline with a human in the loop.

What this is

Lunheng is a writing pipeline, not a text generator. It splits a long-form deliverable into 9 independent roles (T1–T9) across 6 phases, orchestrated with DSH subagent calls, and produces output with an evidence base, counter-argument review, independent audit, and human checkpoints.

The nine roles are independent and interchangeable with nothing else: T1 literature scout, T2 data scout, T3 case scout, T4 analyst, T5 writer, T6 critical companion, T7 auditor, T8 finalizer (executed by the coordinator itself), T9 peer reviewer.

When to use it

• You need a long-form piece (over 2000 characters) that has to hold up under scrutiny, and you can wait several hours (the pipeline deliberately gives no hour range — see skills/lunheng-article-pipeline/SKILL.md §边界).
• The topic involves facts, figures, or multiple viewpoints, so it needs an evidence base rather than opinion only.
• You want human checkpoints: confirm the outline before drafting, and review the final draft.

When not to use it

Lunheng actively collects published evidence and integrates evidence you supply. It cannot produce the following on its own; supply the material first, or use another tool:

• First-hand data collection — experiments, surveys, interviews, field work.
• Statistical analysis — it can cite results but does not run SPSS/R/Python.
• Raw chart data collection — it renders data visualizations; scraping, OCR, and speech-to-text need dedicated tools.
• Original images or video — DSH has no built-in text-to-image. Covers fall back to local SVG or supplied files.
• Code execution — the pipeline runs only whitelisted scripts; anything else needs your explicit approval.

Rule of thumb: ask whether the evidence is already published. If yes, Lunheng collects it. If not, supply it first.

What you get

Item · Content

Literature cards [Lxx] · Published sources with A/B/C confidence grading and a pioneer list for originality checks

Data cards [Dxx] · Figures with source, year, freshness grading, trust level, and conflicting figures shown side by side

Case cards [Cxx] · Event structure (who/when/what/each side's account), or an explicit [C-空] empty marker

Analysis outline · Argument thread, claim-to-evidence mapping, counter-argument plan, load-bearing evidence list

Drafts · Successive versions with AI-trace cleanup, each independent writer run

Review reports · Critical report (C1–C7), audit report (G0–G14), peer-review report (6 dimensions + journal matching), AI-trace report

Final deliverables · final/定稿.md, figures, evidence bundle, delivery notes, M-gate report

DSH integration (bundle install) · Four read-only tools — lunheng_m_gate (M-gate mechanical pre-check), lunheng_char_count (pure Chinese-character count), lunheng_handoff_check (handoff-report shape / version / pairs / agents-log verification at the role-handoff boundary), and lunheng_ethics_sanitize (ethics de-identification: masks IDs / phones / e-mails / bank cards / person names / place names in interview transcripts and field notes; returns the masked text so the caller decides where to write it; v18.60.1); if your session does not expose them, call the same scripts with pwsh as before (same source of truth). Human commands /lunheng-status (reads run/<project>/status.md; produces no model message) and /lunheng-stats (cross-project telemetry dashboard; spawns the host-process scripts/lunheng-stats.mjs with --json whitelisted only; output truncated to 8000 chars; produces no model message). Mechanism-file write protection: a global guard rejects write/edit-style tool calls that target the skill package, so a session cannot quietly rewrite the pipeline's own rules. Boundary, stated plainly: the guard only sees tool calls — pwsh and any subprocess are not behind this gate; the owner's escape hatch is LUNHENG_ALLOW_MECH_EDIT=1 (or config: { allowMechanismEdit: true }). Plugin Config (deployment switch, this plugin's row in your profile) covers quiet, allowMechanismEdit, scriptTimeoutMs, scriptMaxOutputBytes, handoffLevel, hookRewriteContent, hookMaxBlockChars, hookMaxTotalChars, and ethicsHook, and panel (v18.90.0; on by default since v18.93.0 — read-only run panel registers a read-only path prefix on the host's existing web carrier so the Web GUI can show one project's phase / roles / gates / artifacts live; to disable use LUNHENG_PANEL=0 or panel: false, see SECURITY.md) — handoffLevel toggles the handoff-check strictness (basic = A1/A2/B1 only; strict = + structure/version/pairs/agents-log); hookRewriteContent defaults to false (H2 listener only injects metadata without altering result.content — recommended); hookMaxBlockChars defaults to 262144 (256 KB) and caps single-text-block sanitization to keep the listener off the hot path for huge payloads (use the lunheng_ethics_sanitize tool for paging); the first two and panel are mirrored by the `LUNH

…(truncated)