ztl34245881-commits/dsh-task-planner
Task planning with experience muscle-memory for DeepSeek Harness: condition-reflex recall + LLM capability mat
Task planning with experience muscle-memory for DSH: give a task → the agent recalls past similar solutions (condition reflex via a 2–3-char sliding-window tokenizer so 'weekly report' hits a 'daily report' lesson), evaluates fit, and produces a dynamic plan matched against its capabilities — never hard-coded combos. Tools: plan_task (recall → LLM fit evaluation → decomposed steps with capability matching → risks → next actions), task_memory save/recall/list. Every plan auto-drafts a lesson (status draft); the agent marks it verified at loop close; a lesson reused 3× promotes to a formal skill, rejected 2× becomes obsolete. De-AI deliverable standard: textual outputs require a humanize-then-review pass.
Install
dsh plugin --profile web add /path/to/dsh-task-plannerLocal-bundle install per README: clone the repo and dsh plugin --profile web add /path/to/dsh-task-planner (the README's one-line example uses a github:<your-user> placeholder — substitute the real repo; npm package dsh-task-planner not published as of 2026-08-26). Optional config in cordis.patch.yml (lessonsDir default ~/.dsh/planner-lessons, capabilityFile optional). MIT.
Compatibility
DeepSeek Harness (requires llm, shell, tools services — all present in the standard harness). Model call uses agentDefaultModel with 8k internal maxTokens.
Details
- Repo: ztl34245881-commits/dsh-task-planner
- Category: Coding & Development
- Stars: 4
- Version: GitHub source (no npm package)
- Last push: 2026-08-14
- First seen: 2026-08-14
Recent updates
v0.1.0: experience library; condition-reflex planning; LLM-driven decomposition; auto-persist lessons; de-AI deliverable standard; zero keys, zero absolute paths.
FAQ
- How does recall work?
- task_memory recall uses a 2–3-char sliding-window tokenizer over plain-Markdown lessons with signature keywords, so 'weekly report' still hits a 'daily report' lesson.
- What happens to lessons?
- plan_task auto-drafts one (status draft); the agent updates it with the outcome at loop close (status verified); 3 successful reuses promote it to a formal skill, 2 rejections mark it obsolete.
- Where are lessons stored?
- ~/.dsh/planner-lessons by default (configurable) — plain Markdown, human-editable, greppable, portable. No keys and no absolute paths are required.
Alternatives
lhmd/dsh-director-toolkit · lhmd/dsh-promotion-toolkit · omdsh-dev/dsh-advisor