deeprefine_skill

Name: deeprefine_skill
Author: HKUST-KnowComp

85stars6forks

How to install deeprefine_skill?

hkust-knowcomp/deeprefine-skill/deeprefine_skill

$npx -y skills add HKUST-KnowComp/DeepRefine-Skill --skill deeprefine_skill

Installs into the current project.

Summary

Agent-native DeepRefine refinement loop — same control flow as DeepRefine.refine(), graphify search instead of FAISS, session LLM, dry-run review before approved graph writes.

Files · 1

View on GitHub

SKILL.md

1	# DeepRefine — Agent refinement loop (strict)
2
3	## Default safety policy: dry-run only
4
5	A normal `/deeprefine` invocation MUST NEVER call `deeprefine apply`.
6
7	The default `/deeprefine` workflow must stop after:
8
9	1. `deeprefine loop validate`
10	2. `deeprefine review`
11	3. showing the proposed actions and HIGH/MEDIUM/LOW review report to the user
12
13	Then ask the user for explicit approval.
14
15	Only if the user's next message explicitly says to approve/apply/write the graph may you run:
16
17	```bash
18	deeprefine apply --refresh-wiki --trace-file ... --refinement-file ...
19	deeprefine loop finish --trace-file ... --refinement-file ...
20	```
21
22	Do not treat generation of `<refinement>` actions as approval. Do not treat a valid trace as approval. Do not apply in the same `/deeprefine` turn.
23
24	---
25
26	You MUST implement the same control flow as `DeepRefine.refine()` in DeepRefine (`autorefiner/src/deeprefine.py`).
27
28	\| Component \| Agent mode \| CLI `deeprefine refine` \|
29	\|-----------\|------------\|-------------------------\|
30	\| Retrieval \| `graphify query` + k-hop from `graph.json` \| FAISS retriever \|
31	\| LLM \| Your session model \| External API / vLLM \|
32	\| Graph writes \| Dry-run proposal + `deeprefine review`; `deeprefine apply` only after user approval \| Dry-run by default; `--apply` persists \|
33
34	---
35
36	## FORBIDDEN (hard stop)
37
38	Do NOT:
39
40	1. Run `deeprefine refine` (unless the user explicitly asks for CLI/FAISS mode).
41	2. Call `deeprefine apply` without a valid `loop_trace_<query_id>.json` (CLI will reject).
42	3. Call `deeprefine apply` before running `deeprefine review` and receiving explicit user approval.
43	4. Ignore LOW-confidence review warnings unless the user explicitly requests `--allow-low-confidence`.
44	5. Skip any hop’s `<judge>Yes</judge>` / `<judge>No</judge>` judgement.
45	6. Skip error abduction when `len(interaction_history) > 1`.
46	7. Write `<refinement>` before abduction when refinement is required.
47	8. Hand-edit `graph.json` with Python or ad-hoc JSON patches.
48	9. Ignore pending history and refine only one latest query when unrefined queries already exist.
49	10. Invent a shorter pipeline (“read file → write refinement → apply”).
50
51	If validation fails, fix the trace and re-run the missing step — do not bypass with `--skip-trace-check`.
52
53	---
54
55	## Constants (match `refine_runner.py` / DeepRefine)
56
57	```text
58	MAX_HOPS = 4
59	INCREMENT_HOP = 1
60	BASE_TOP_K = 10
61	MAX_TRIPLE_NUM_BY_STEP = [5, 10, 15, 20] # cap triples per step
62	HISTORY_HORIZON = 4 # abduction uses last N steps
63	```
64
65	---
66
67	## Mandatory artifact
68
69	For each query, maintain:
70
71	`graphify-out/.deeprefine/loop_trace_<query_id>.json`
72
73	Create template:
74
75	```bash
76	deeprefine loop init --query "<exact question>"
77	```
78
79	Append each hop to `interaction_history` before starting the next hop.
80	Run `deeprefine loop validate --trace-file ...` after abduction and before `review` / approved `apply`.
81
82	---
83
84	## Query queue selection (default behavior of `/deeprefine`)
85
86	`/deeprefine` must process all unrefined history queries first, not just the latest one.
87
88	1. Sync graphify query memory into DeepRefine history:
89	- run: `deeprefine history sync-memory`
90	- source dir: `graphify-out/memory/query_*.md`
91	- target file: `graphify-out/.deeprefine/history.jsonl`
92	2. Read pending queue from `graphify-out/.deeprefine/history.jsonl`:
93	- include rows where `refined != true`
94	- dedupe by `id` (first occurrence)
95	- preserve file order
96	3. If pending queue is non-empty: set `target_queries = pending_queue`.
97	4. If pending queue is empty: set `target_queries = [current session question]`.
98	5. Run the full refinement loop for each query in `target_queries`, one by one.
99	6. For early-exit queries, finish immediately. For refinement-path queries, generate review output and wait for user approval before `apply` + `loop finish`.
100
101	---
102
103	## Control flow (must match `DeepRefine.refine()`)
104
105	Pseudocode — follow exactly:
106
107	```text
108	target_queries = pending_history_queries() # refined != true, dedupe by id, keep order
109	run "deeprefine history sync-memory" before loading pending history
110	if target_queries is empty:
111	target_queries = [current session question]
112
113	for question in target_queries:
114	interaction_history = []
115	for step in 1..MAX_HOPS:
116	print "[Step: {step}]" # show in chat
117
118	if step == 1:
119	# Vector-retrieval equivalent: graphify query on full question
120	RUN: graphify query "<question>"
121	triples = parse NODE/EDGE → [{subject, relation, object}, ...]
122	cap = MAX_TRIPLE_NUM_BY_STEP[0] # 5
123	record retrieval.method = "graphify_query"
124	else:
125	# k-hop expansion from entities in previous hop (NOT a new random search)
126	entities = uni

People who liked this also viewed

6 picks

Type

deeprefine_skill

85stars6forks

View on GitHub

How to install deeprefine_skill?

hkust-knowcomp/deeprefine-skill/deeprefine_skill

$npx -y skills add HKUST-KnowComp/DeepRefine-Skill --skill deeprefine_skill

Installs into the current project.

Summary

Agent-native DeepRefine refinement loop — same control flow as DeepRefine.refine(), graphify search instead of FAISS, session LLM, dry-run review before approved graph writes.

Files · 1

View on GitHub

SKILL.md

1	# DeepRefine — Agent refinement loop (strict)
2
3	## Default safety policy: dry-run only
4
5	A normal `/deeprefine` invocation MUST NEVER call `deeprefine apply`.
6
7	The default `/deeprefine` workflow must stop after:
8
9	1. `deeprefine loop validate`
10	2. `deeprefine review`
11	3. showing the proposed actions and HIGH/MEDIUM/LOW review report to the user
12
13	Then ask the user for explicit approval.
14
15	Only if the user's next message explicitly says to approve/apply/write the graph may you run:
16
17	```bash
18	deeprefine apply --refresh-wiki --trace-file ... --refinement-file ...
19	deeprefine loop finish --trace-file ... --refinement-file ...
20	```
21
22	Do not treat generation of `<refinement>` actions as approval. Do not treat a valid trace as approval. Do not apply in the same `/deeprefine` turn.
23
24	---
25
26	You MUST implement the same control flow as `DeepRefine.refine()` in DeepRefine (`autorefiner/src/deeprefine.py`).
27
28	\| Component \| Agent mode \| CLI `deeprefine refine` \|
29	\|-----------\|------------\|-------------------------\|
30	\| Retrieval \| `graphify query` + k-hop from `graph.json` \| FAISS retriever \|
31	\| LLM \| Your session model \| External API / vLLM \|
32	\| Graph writes \| Dry-run proposal + `deeprefine review`; `deeprefine apply` only after user approval \| Dry-run by default; `--apply` persists \|
33
34	---
35
36	## FORBIDDEN (hard stop)
37
38	Do NOT:
39
40	1. Run `deeprefine refine` (unless the user explicitly asks for CLI/FAISS mode).
41	2. Call `deeprefine apply` without a valid `loop_trace_<query_id>.json` (CLI will reject).
42	3. Call `deeprefine apply` before running `deeprefine review` and receiving explicit user approval.
43	4. Ignore LOW-confidence review warnings unless the user explicitly requests `--allow-low-confidence`.
44	5. Skip any hop’s `<judge>Yes</judge>` / `<judge>No</judge>` judgement.
45	6. Skip error abduction when `len(interaction_history) > 1`.
46	7. Write `<refinement>` before abduction when refinement is required.
47	8. Hand-edit `graph.json` with Python or ad-hoc JSON patches.
48	9. Ignore pending history and refine only one latest query when unrefined queries already exist.
49	10. Invent a shorter pipeline (“read file → write refinement → apply”).
50
51	If validation fails, fix the trace and re-run the missing step — do not bypass with `--skip-trace-check`.
52
53	---
54
55	## Constants (match `refine_runner.py` / DeepRefine)
56
57	```text
58	MAX_HOPS = 4
59	INCREMENT_HOP = 1
60	BASE_TOP_K = 10
61	MAX_TRIPLE_NUM_BY_STEP = [5, 10, 15, 20] # cap triples per step
62	HISTORY_HORIZON = 4 # abduction uses last N steps
63	```
64
65	---
66
67	## Mandatory artifact
68
69	For each query, maintain:
70
71	`graphify-out/.deeprefine/loop_trace_<query_id>.json`
72
73	Create template:
74
75	```bash
76	deeprefine loop init --query "<exact question>"
77	```
78
79	Append each hop to `interaction_history` before starting the next hop.
80	Run `deeprefine loop validate --trace-file ...` after abduction and before `review` / approved `apply`.
81
82	---
83
84	## Query queue selection (default behavior of `/deeprefine`)
85
86	`/deeprefine` must process all unrefined history queries first, not just the latest one.
87
88	1. Sync graphify query memory into DeepRefine history:
89	- run: `deeprefine history sync-memory`
90	- source dir: `graphify-out/memory/query_*.md`
91	- target file: `graphify-out/.deeprefine/history.jsonl`
92	2. Read pending queue from `graphify-out/.deeprefine/history.jsonl`:
93	- include rows where `refined != true`
94	- dedupe by `id` (first occurrence)
95	- preserve file order
96	3. If pending queue is non-empty: set `target_queries = pending_queue`.
97	4. If pending queue is empty: set `target_queries = [current session question]`.
98	5. Run the full refinement loop for each query in `target_queries`, one by one.
99	6. For early-exit queries, finish immediately. For refinement-path queries, generate review output and wait for user approval before `apply` + `loop finish`.
100
101	---
102
103	## Control flow (must match `DeepRefine.refine()`)
104
105	Pseudocode — follow exactly:
106
107	```text
108	target_queries = pending_history_queries() # refined != true, dedupe by id, keep order
109	run "deeprefine history sync-memory" before loading pending history
110	if target_queries is empty:
111	target_queries = [current session question]
112
113	for question in target_queries:
114	interaction_history = []
115	for step in 1..MAX_HOPS:
116	print "[Step: {step}]" # show in chat
117
118	if step == 1:
119	# Vector-retrieval equivalent: graphify query on full question
120	RUN: graphify query "<question>"
121	triples = parse NODE/EDGE → [{subject, relation, object}, ...]
122	cap = MAX_TRIPLE_NUM_BY_STEP[0] # 5
123	record retrieval.method = "graphify_query"
124	else:
125	# k-hop expansion from entities in previous hop (NOT a new random search)
126	entities = uni

People who liked this also viewed

6 picks

Type