hkust-knowcomp/deeprefine-skill/dsh-skill

DeepRefine - DeepSeek Harness (dsh) Adapter

This file is the dsh-specific entrypoint. It keeps the platform rules small and loads longer DeepRefine procedure details only when needed:

• Full workflow, queue selection, refinement branch logic, and review rules:

references/deeprefine-workflow.md

• Verbatim judgement, abduction, and refinement prompts:

references/llm-prompts.md

• Checklist, command sequence, trace schema, paths, and CLI mode:

references/trace-and-commands.md

Do not reimplement or shorten the algorithm from memory. Load the relevant reference file before executing that part of the workflow.

dsh Invocation

Trigger this skill when the user:

• explicitly invokes /deeprefine;
• asks to refine, improve, diagnose, repair, inspect, or review a Graphify

knowledge graph;

• asks to apply a previously reviewed DeepRefine refinement.

Run from the knowledge-base project root, where graphify-out/graph.json exists. If the user is planning or asking how DeepRefine works, explain the workflow and do not mutate files.

If deeprefine is unavailable, tell the user to install it:

pip install deeprefine-cli

For source development:

pip install -e /path/to/DeepRefine-Skill

Hard Safety Policy

A normal /deeprefine invocation is dry-run only and MUST NEVER call deeprefine apply.

The default workflow must stop after:

1. deeprefine loop validate
2. deeprefine review
3. showing the proposed actions and HIGH/MEDIUM/LOW review report to the user

Then ask for explicit approval.

Only if the user's next message explicitly says to approve/apply/write the graph may you run:

deeprefine apply --trace-file ... --refinement-file ...
deeprefine loop finish --trace-file ... --refinement-file ...

Do not treat any of these as approval:

• generation of a <refinement> block;
• a valid loop_trace_<query_id>.json;
• a prior user message;
• a successful deeprefine review.

If the review contains any LOW-confidence action, deeprefine apply aborts and writes nothing (it does NOT apply HIGH/MEDIUM actions while skipping LOW ones). Use --allow-low-confidence only when the user's current approval message explicitly accepts that risk.

Mode Selection

Full workflow

Use for /deeprefine, or requests to refine/improve/fix the graph.

Follow the canonical reference in this order:

1. references/deeprefine-workflow.md
2. references/llm-prompts.md when producing tagged LLM outputs
3. references/trace-and-commands.md when writing traces or running commands

Do not copy only the latest query if pending history exists. Process all unrefined history queries first, preserving canonical dedupe/order rules.

Review only

Use when the user asks to review, audit, inspect, dry-run, check evidence, or show what would change.

Run validation and review only:

deeprefine loop validate --trace-file ... --refinement-file ...
deeprefine review --trace-file ... --refinement-file ...

Show the HIGH/MEDIUM/LOW evidence report. Do not modify graph.json.

Apply only

Use only when the user's current message explicitly approves a previously reviewed refinement.

Before applying, verify that the trace and refinement file match references/trace-and-commands.md and the review rules in references/deeprefine-workflow.md. Then run:

deeprefine loop validate --trace-file ... --refinement-file ...
deeprefine apply --trace-file ... --refinement-file ...
deeprefine loop finish --trace-file ... --refinement-file ...

Use the LOW-confidence override only with explicit risk acknowledgement in the same user message:

deeprefine apply --allow-low-confidence --trace-file ... --refinement-file ...

After review: closing options

After showing the HIGH/MEDIUM/LOW report, stop and present these options to the user instead of deciding on your own:

• Approve & apply — the user explicitly says approve/apply/write. Run

deeprefine apply (it aborts if any LOW action is present unless --allow-low-confidence is given) then deeprefine loop finish.

• Skip apply, mark done — the user declines the changes. Run

deeprefine loop finish --trace-file ... --refinement-file ... with no graph write.

• Leave as proposal — the user wants to stop. Do not finish; leave the

trace and review as-is for later.

Never run deeprefine apply before the user's current message explicitly approves it, regardless of which option they later pick.

Non-Negotiable Rules

These are restated here so the model always sees the hard stops before loading any reference.

Do not:

1. Run deeprefine refine unless the user explicitly asks for CLI/FAISS mode.
2. Call deeprefine apply without a valid loop_trace_<query_id>.json.
3. Call deeprefine apply before deeprefine review and explicit approval.
4. Ignore LOW-confidence review warnings without explicit risk acceptance.
5. Skip any hop's <judge>Yes</judge> / <judge>No</judge> judgement.
6. Skip error abduction when len(interaction_history) > 1.
7. Write <refinement> before abduction when refinement is required.
8. Hand-edit graphify-out/graph.json with Python or ad-hoc JSON patches.
9. Ignore pending history and refine only one latest query.
10. Invent a shorter pipeline such as "read file -> write refinement -> apply".

If validation fails, fix the trace or rerun the missing step. Do not bypass with --skip-trace-check in agent mode.

What to Load From References

Keep this adapter concise. Load the smallest reference needed:

• Full refinement pseudocode and safe review:

references/deeprefine-workflow.md

• Verbatim LLM prompts:

references/llm-prompts.md

• Required JSON shape, exact command sequence, and CLI/FAISS exception path:

references/trace-and-commands.md

Use the canonical commands and artifacts exactly as written there. This adapter only maps those rules onto dsh's /deeprefine invocation and approval behavior.