LastTrainKing/deepseek-harness-batch-executor

A DSH skill that keeps the model in the loop only for judgment steps, runs deterministic work as one scripted pass, verifies results locally and stops unproductive retry loops.

Attacks the three cost centres of long agent sessions: repeated context cost, round-trip inflation and verbose feedback. A plan → batch execute → receipt workflow moves deterministic work into one programmatic pass (scripts/batch_exec.py), compressed-feedback helpers summarise search hits and test logs to stdout while full detail goes to a local log with stable ids (search_summary.py, test_summary.py), local verification reports counts, sha256 tree hashes and diffs as evidence (verify_state.py), and loop governance detects no-progress retries with evidence hashes and stops them with a reason (loop_guard.py). Destructive operations require explicit approval and a dry run first, and the README is explicit that when native Code Mode is available it should be preferred over the skill's manual path.

Skills ★ 0 updated 2026-10-17
View on GitHub ↗

Install

dsh plugin add github:LastTrainKing/deepseek-harness-batch-executor

Copied verbatim from the README's installation section. The package is not published to npm; it installs as a native DSH skill (SKILL.md plus helper scripts) from the repository.

Compatibility

The helper scripts are harness-agnostic: Python 3.8+, standard library only, no pip install and no network access. The README notes that the same shape is achieved natively by DSH's Code Mode / programmatic tool calling when it is enabled, and documents both paths plus the decision rules for choosing between them.

Details

Alternatives

tt-a1i/archify#integrations/deepseek-harness · Cavan-Ou/hermes-dsh-collab · zjsthmjialin/inspiration-deck-workshop

More plugins in Skills

Browse more in Skills

Guides for Skills plugins