DeepSeek Harness Long-Horizon Task Suite

DSH’s native goal / todo / subagent modules are suitable for short tasks. When handling long-term research or projects, they can exhibit rigid task planning, stop-on-failure behavior, reliance on manual wake-up, ambiguous success criteria, excessive sub-agent consumption, loss of summary information, primitive collaboration, model laziness and degradation, experience loss, inflated self-evaluation, and unrecoverable reminders.

This suite addresses the above issues through Mission state files, the Claim Pool/Lease claiming mechanism, Blackboard artifact communication, memoryless blind review, LLM Wiki memory, and forced replanning on failure, without hard-coding domain-specific workflows.

Core Features¶

• Long-term autonomous task manager
• Long-Run Captain preset
• LLM verifier
• Autonomous scheduled wake-up
• Task state machine
• Claim Pool/Lease claiming mechanism
• Blackboard artifact communication
• Memoryless blind review
• LLM Wiki memory
• Forced replanning on failure

Component Description¶

This suite includes three independent plugins and one preset directory:

• dsh-mission-control: task state machine, mission tools, Claim Pool/Lease, Blackboard artifacts, blind review and LLM Wiki tools.
• Long-Run Captain: a complete system-prompt preset, applicable to any model; the first-round context is heavier.
• Long-Run Captain Router: a minimal first-round system-prompt preset, tuned for the DeepSeek V4 Flash series, while preserving the collective planning style.
• dsh-plugin-llm-verifier: a strict-review LLM verifier with correction capabilities, providing verify_rollout / verify_select / verify_compare / verify_track functions.
• dsh-timer-scheduler-ui: provides the schedule_reminder autonomous scheduled wake-up function and a timed reminder entry in the top session header.

State Machine and Mechanisms¶

Task states are divided into five stages: open → active → needs_review → accepted / rejected.

• Rejections require follow-up: A rejected task must include a replaces attribute pointing to a follow-up task; otherwise it cannot be completed.
• Completion requires evidence: By default termination_policy=success; it must map to success criteria, and outcome=success is required.
• Meta-verifier: mission_check only checks evidence completeness (missing evidence, missing review, or missing final audit all count as failure).
• Claim Pool / Lease: Tasks can declare capabilities, and workers claim tasks based on capabilities. Long-running tasks use mission_heartbeat to renew the lease, mission_release to release it, and expired leases are automatically reclaimed.
• Blackboard communication: Workers exchange typed artifacts through mission_publish_artifact and mission_consume_artifacts.
• LLM Wiki: Cross-task experience is maintained in the .memory/ directory and supports wiki_write, wiki_search, and wiki_lint.
• Memoryless blind review: Before substantive deliverable tasks are completed, they must undergo mission_blind_review, generating blind_review.md and calibration_gap.

Typical Usage¶

In a session using the Long-Run Captain or Long-Run Captain Router preset, you can directly issue instructions:

启动一个 mission：开发一个命令行工具，递归扫描指定目录下的 Markdown 文件，按标题生成带层级、文件路径和更新时间的索引 index.md。
termination_policy: success
budget: { maxRounds: 6, maxHours: 4 }
成功标准：
- CLI 能递归扫描目录并生成 index.md
- 索引按标题层级组织，包含文件相对路径和更新时间
- 提供 3 个测试用例并全部通过
- 输出 README 说明安装和使用方式
- 由 reviewer 独立验证索引内容正确

Compatibility and Installation¶

This plugin is compatible with DSH versions >=0.1.0-rc.8 <0.2.0.

No explicit official installation command has been found yet. The repository is a monorepo for three independent plugins; when publishing to DSH Store, explicit subpaths must be submitted separately.

Notes¶

• When using DeepSeek V4 Flash series models, it is recommended to choose the Long-Run Captain Router preset.
• For other models or when a heavier system prompt is needed, choose the Long-Run Captain preset.
• Both presets depend on dsh-mission-control, and the task workflow, state machine, and review rules are identical.