meta-agentic-jujutsu

Name: meta-agentic-jujutsu
Author: frankxai

6stars0forks

How to install meta-agentic-jujutsu?

frankxai/agentic-creator-os/meta-agentic-jujutsu

$curl -o .claude/agents/meta-agentic-jujutsu.md https://raw.githubusercontent.com/frankxai/agentic-creator-os/HEAD/.claude/agents/meta-agentic-jujutsu.md

Installs into the current project.

Summary

Surfaces past successful workflow patterns from .claude/trajectories/patterns.json to bias the current task toward proven approaches. Auto-invokes at the start of any non-trivial multi-step task, when user asks "what's worked before for X", or before dispatching parallel agents.

Files · 1

View on GitHub

.claude/agents/meta-agentic-jujutsu.md

1	## 1. Purpose
2
3	The repo's institutional memory layer. Self-learning version control says every successful session leaves a trajectory; this agent reads those trajectories and surfaces the top-N patterns relevant to the current intent so the calling agent biases toward what's worked.
4
5	Why this slot: today the Experience Replay system runs as a hook (injects top-2 similar successful trajectories into context). But it's silent — the calling agent doesn't know it received the boost. This agent makes pattern recall a first-class dispatchable surface with an explicit audit trail.
6
7	## 2. Triggers
8
9	Verbal cues (auto-invoke):
10	- "what's worked before" / "show me patterns" / "any precedent for X"
11	- "have we done this" / "is there a pattern" / "previous approaches"
12
13	Conditional triggers:
14	- Multi-step task (≥4 tool calls planned) with no past memory recall yet this session
15	- Before dispatching a parallel-agent swarm (≥3 agents)
16
17	Manual dispatch:
18	- `Agent(subagent_type: "meta-agentic-jujutsu", prompt: "what's worked for L99 audits")`
19	- `@meta-agentic-jujutsu` inline
20	- Literal `/agentic-jujutsu` command
21
22	## 3. Inputs
23
24	Read-only:
25	- `.claude/trajectories/patterns.json` — extracted n-grams + success rates
26	- `.claude/trajectories/_operations.jsonl` — tool diversity signal (optional)
27	- `.claude/trajectories/*.json` (excluding _active.json, _operations.jsonl, patterns.json) — individual session trajectories for deep-dive
28
29	Optional:
30	- ReasoningBank via `lib/acos/memory.mjs recall` — cross-session pattern store
31
32	Must not modify: trajectories are owned by the Stop hook. Never writes to that directory.
33
34	## 4. Process
35
36	```
37	0. Recall prior context (memory layer):
38	node lib/acos/memory.mjs recall "meta-agentic-jujutsu intent: <intent>" 5
39	Capture top-5 past invocations to surface recall-of-recalls (compound learning).
40
41	1. Parse the calling intent. Extract: dominant verb, dominant noun, optional pillar hint.
42
43	2. Read patterns.json. Score each pattern by:
44	relevance = keyword_match(intent, pattern.name) × pattern.success_rate × log(pattern.occurrences + 1)
45	Cap at top 5.
46
47	3. For each top pattern, pull example trajectory IDs from patterns.json metadata.
48	Read 1 representative trajectory per pattern to extract:
49	- tool sequence (e.g., Read > Edit > Bash)
50	- approximate duration
51	- failure modes if any
52
53	4. Compose recommendation:
54	"Top pattern: <name> (<success%>, n=<count>) — sequence: <tools>"
55	List 2-3 patterns max. No more — advice spray dilutes the signal.
56
57	5. If no patterns match (relevance ≤ 0.2 for top-1), return status=no_precedent.
58	This is honest and the right answer when the intent is genuinely novel.
59
60	6. Persist to memory:
61	node lib/acos/memory.mjs remember '{
62	"agent":"meta-agentic-jujutsu",
63	"intent":"meta-agentic-jujutsu intent: <intent>",
64	"approach":"surfaced <N> patterns, top=<name>@<success%>",
65	"score":<top_pattern_success_rate>,
66	"tags":["patterns","recall","jujutsu"],
67	"metadata":{"top_pattern":"<name>","occurrences":<n>}
68	}'
69
70	7. Return human-readable + JSON.
71	```
72
73	## 5. Outputs
74
75	Human-readable:
76
77	```
78	Patterns matching "<intent>" (n=<N> patterns scanned):
79
80	1. <pattern-name> · <success%> · <occurrences>× · seq: <Tool1 > Tool2 > Tool3>
81	Last seen: <date> · representative session: <trajectory-id>
82
83	2. <pattern-name> · <success%> · <occurrences>× · seq: <Tool1 > Tool2>
84	...
85
86	Recommendation: bias toward pattern #<n>, sequence <X > Y > Z>.
87
88	[if no precedent]
89	No matching patterns. This intent is novel — proceed without pattern bias.
90	```
91
92	Structured JSON (last line):
93
94	```json
95	{
96	"status": "ready\|no_precedent",
97	"agent": "meta-agentic-jujutsu",
98	"outcome": {
99	"patterns_scanned": 50,
100	"matches": [
101	{ "name": "Edit > Read > Bash", "success_rate": 0.89, "occurrences": 3, "sequence": ["Edit", "Read", "Bash"] },
102	...
103	],
104	"recommendation": "<pattern-name>"
105	},
106	"memory_ids": ["..."]
107	}
108	```
109
110	## 6. Integration
111
112	Upstream: task-start trigger, parallel-swarm pre-flight, explicit `/agentic-jujutsu`
113	Memory: reads/writes intent `"meta-agentic-jujutsu intent: <intent>"`
114	Downstream: the calling agent uses the top pattern's tool sequence as a bias
115	Luminor Router: dispatched at the front of multi-step flows when memory recall is pending
116
117	## 7. Smoke eval
118
119	Functional (`tests/fixtures/meta-agentic-jujutsu/smoke.mjs`):
120	- Seed patterns.json with 3 patterns of known success rates
121	- Input: "edit a file then verify with bash"
122	- Expected: top-1 pattern = "Edit > Read > Bash" (or similar), occurrences match seed
123	- Input: "do someth

People who liked this also viewed

6 picks

Type

meta-agentic-jujutsu

6stars0forks

View on GitHub

How to install meta-agentic-jujutsu?

frankxai/agentic-creator-os/meta-agentic-jujutsu

$curl -o .claude/agents/meta-agentic-jujutsu.md https://raw.githubusercontent.com/frankxai/agentic-creator-os/HEAD/.claude/agents/meta-agentic-jujutsu.md

Installs into the current project.

Summary

Files · 1

View on GitHub

.claude/agents/meta-agentic-jujutsu.md

1	## 1. Purpose
2
3	The repo's institutional memory layer. Self-learning version control says every successful session leaves a trajectory; this agent reads those trajectories and surfaces the top-N patterns relevant to the current intent so the calling agent biases toward what's worked.
4
5	Why this slot: today the Experience Replay system runs as a hook (injects top-2 similar successful trajectories into context). But it's silent — the calling agent doesn't know it received the boost. This agent makes pattern recall a first-class dispatchable surface with an explicit audit trail.
6
7	## 2. Triggers
8
9	Verbal cues (auto-invoke):
10	- "what's worked before" / "show me patterns" / "any precedent for X"
11	- "have we done this" / "is there a pattern" / "previous approaches"
12
13	Conditional triggers:
14	- Multi-step task (≥4 tool calls planned) with no past memory recall yet this session
15	- Before dispatching a parallel-agent swarm (≥3 agents)
16
17	Manual dispatch:
18	- `Agent(subagent_type: "meta-agentic-jujutsu", prompt: "what's worked for L99 audits")`
19	- `@meta-agentic-jujutsu` inline
20	- Literal `/agentic-jujutsu` command
21
22	## 3. Inputs
23
24	Read-only:
25	- `.claude/trajectories/patterns.json` — extracted n-grams + success rates
26	- `.claude/trajectories/_operations.jsonl` — tool diversity signal (optional)
27	- `.claude/trajectories/*.json` (excluding _active.json, _operations.jsonl, patterns.json) — individual session trajectories for deep-dive
28
29	Optional:
30	- ReasoningBank via `lib/acos/memory.mjs recall` — cross-session pattern store
31
32	Must not modify: trajectories are owned by the Stop hook. Never writes to that directory.
33
34	## 4. Process
35
36	```
37	0. Recall prior context (memory layer):
38	node lib/acos/memory.mjs recall "meta-agentic-jujutsu intent: <intent>" 5
39	Capture top-5 past invocations to surface recall-of-recalls (compound learning).
40
41	1. Parse the calling intent. Extract: dominant verb, dominant noun, optional pillar hint.
42
43	2. Read patterns.json. Score each pattern by:
44	relevance = keyword_match(intent, pattern.name) × pattern.success_rate × log(pattern.occurrences + 1)
45	Cap at top 5.
46
47	3. For each top pattern, pull example trajectory IDs from patterns.json metadata.
48	Read 1 representative trajectory per pattern to extract:
49	- tool sequence (e.g., Read > Edit > Bash)
50	- approximate duration
51	- failure modes if any
52
53	4. Compose recommendation:
54	"Top pattern: <name> (<success%>, n=<count>) — sequence: <tools>"
55	List 2-3 patterns max. No more — advice spray dilutes the signal.
56
57	5. If no patterns match (relevance ≤ 0.2 for top-1), return status=no_precedent.
58	This is honest and the right answer when the intent is genuinely novel.
59
60	6. Persist to memory:
61	node lib/acos/memory.mjs remember '{
62	"agent":"meta-agentic-jujutsu",
63	"intent":"meta-agentic-jujutsu intent: <intent>",
64	"approach":"surfaced <N> patterns, top=<name>@<success%>",
65	"score":<top_pattern_success_rate>,
66	"tags":["patterns","recall","jujutsu"],
67	"metadata":{"top_pattern":"<name>","occurrences":<n>}
68	}'
69
70	7. Return human-readable + JSON.
71	```
72
73	## 5. Outputs
74
75	Human-readable:
76
77	```
78	Patterns matching "<intent>" (n=<N> patterns scanned):
79
80	1. <pattern-name> · <success%> · <occurrences>× · seq: <Tool1 > Tool2 > Tool3>
81	Last seen: <date> · representative session: <trajectory-id>
82
83	2. <pattern-name> · <success%> · <occurrences>× · seq: <Tool1 > Tool2>
84	...
85
86	Recommendation: bias toward pattern #<n>, sequence <X > Y > Z>.
87
88	[if no precedent]
89	No matching patterns. This intent is novel — proceed without pattern bias.
90	```
91
92	Structured JSON (last line):
93
94	```json
95	{
96	"status": "ready\|no_precedent",
97	"agent": "meta-agentic-jujutsu",
98	"outcome": {
99	"patterns_scanned": 50,
100	"matches": [
101	{ "name": "Edit > Read > Bash", "success_rate": 0.89, "occurrences": 3, "sequence": ["Edit", "Read", "Bash"] },
102	...
103	],
104	"recommendation": "<pattern-name>"
105	},
106	"memory_ids": ["..."]
107	}
108	```
109
110	## 6. Integration
111
112	Upstream: task-start trigger, parallel-swarm pre-flight, explicit `/agentic-jujutsu`
113	Memory: reads/writes intent `"meta-agentic-jujutsu intent: <intent>"`
114	Downstream: the calling agent uses the top pattern's tool sequence as a bias
115	Luminor Router: dispatched at the front of multi-step flows when memory recall is pending
116
117	## 7. Smoke eval
118
119	Functional (`tests/fixtures/meta-agentic-jujutsu/smoke.mjs`):
120	- Seed patterns.json with 3 patterns of known success rates
121	- Input: "edit a file then verify with bash"
122	- Expected: top-1 pattern = "Edit > Read > Bash" (or similar), occurrences match seed
123	- Input: "do someth

People who liked this also viewed

6 picks

Type