dsh-batch-pipeline: batch task pipeline — progress, resume after restart, budget circuit-breaker, product checks

[Culeot]

What it is

dsh-batch-pipeline is a batch-task plugin for DeepSeek Harness. Batch work has one shared pain point: a single session cannot finish it — 50 essays, interrupted at #30, session restarted, progress gone. This plugin handles four things: progress, resume, budget, and product checks. It is a ledger, not a worker — the agent or a script does the work; the plugin records where you are, how much it cost, and whether the product passed.

• npm: https://www.npmjs.com/package/dsh-batch-pipeline
• source: https://github.com/Culeot/dsh-batch-pipeline (MIT)

Four capabilities

Capability	What it does
Progress	The item list lives in batch.json, persisted across sessions. You always know which item you are on.
Resume after restart	Session died? A new session calls batch_resume takeover=true and picks up from the last pointer. No restarting from scratch.
Budget circuit-breaker	Over budget, the batch auto-pauses so you don't burn tokens unnoticed.
Product checks	Each report verifies the product exists and is non-empty; PDF/zip are size-checked via stat, not read as text (which would garble them).

Why it works this way

• The agent reports; the plugin does not guess. Scanning a directory shows whether files exist, not whether the work is done. So after each item the agent calls batch_report with the item index; the plugin verifies it against the pointer and rejects mismatches. The ledger cannot be derailed by model hallucination.
• Inject on demand, not every turn. One short line per progress change (which item, budget left); identical states are not re-injected.
• Three stages, automatic. After the last item the batch enters the finishing stage — do the assembly (index, zip, merge), then batch_finish.
• costModel='none' for download/script-heavy batches: progress only, no budget accounting — don't charge fake LLM cost to non-LLM work.

How it differs from official todo/goal

Official todo/goal/plan are all session-scoped (their READMEs say same-session) — gone on restart. Batch work is by definition "work one session cannot finish", so this plugin is a cross-session ledger. todo is a sticky note; batch is the pipeline ledger.

Tools (6)

batch_create create · batch_status progress · batch_report report done · batch_pause / batch_resume pause / takeover · batch_finish archive

Quality

35 unit tests green; headless real-machine verification includes cross-session handoff (two independent sessions completing a 3-item task); six fixes driven by real usage (protocol wording, claim tolerance, binary assertions, costModel, lossless JSON, granularity guidance).

Install

cd ~/.dsh/profiles/<profile>
pnpm add dsh-batch-pipeline

# agent preset or cordis.patch.yml
- id: batch
  name: 'dsh-batch-pipeline'

Restart DSH; six batch_* tools appear. Prerequisite: the storage trio in the profile (web profile ships with it).