docs

Name: docs
Author: stevesolun

560stars68forks

How to install docs?

stevesolun/ctx/docs

$npx -y skills add stevesolun/ctx --skill docs

Installs into the current project.

Summary

Repo-aware recommendation manager for ctx. Scans the active repository, identifies stack and workflow signals, recommends a capped set of skills, agents, and MCP servers, and unloads helpers that no longer match the current work after user confirmation. Harnesses are recommended

Files · 1

View on GitHub

SKILL.md

1	# Skill Router
2
3	Scan a repo. Know what the current work needs. Recommend only that. Keep the
4	wiki and graph as the durable catalog behind the decision.
5
6	## Scope
7
8	The router manages runtime recommendations for:
9
10	- skills
11	- agents
12	- MCP servers
13
14	Harnesses are separate catalog entities. They are recommended when a user wants
15	to run ctx with a non-Claude-Code host, local model, API model, or external loop
16	adapter. Once attached, the harness calls the same skills/agents/MCP
17	recommendation engine.
18
19	## Problem
20
21	Every loaded helper costs tokens, attention, and operational surface area. Most
22	sessions need a small top-scored bundle from the shipped graph, not every entity
23	ctx knows about. Loading too much:
24
25	- wastes context on irrelevant instructions,
26	- causes misfires when a helper matches the wrong task,
27	- slows the agent loop, and
28	- creates conflicting instructions.
29
30	## Architecture
31
32	```text
33	ctx/
34	\|-- src/scan_repo.py # Repo scanner -> stack profile
35	\|-- src/ctx/core/resolve/resolve_skills.py # Profile -> load/unload manifest
36	\|-- src/ctx/core/resolve/recommendations.py # Shared scoring/ranking engine
37	\|-- src/ctx/adapters/ # Host adapters + generic tools
38	\|-- src/harness_install.py # Custom-model harness install flow
39	`-- graph/wiki-graph-runtime.tar.gz # Shipped graph/wiki runtime
40	```
41
42	The router has three halves:
43
44	1. Scanner - analyzes a repo and produces stack/task evidence.
45	2. Resolver - scores graph/wiki entities and emits a capped manifest.
46	3. Wiki/graph - persistent catalog of entities, usage, quality, and links.
47
48	## Startup Flow
49
50	1. Read the shipped graph/wiki metadata and local user overrides.
51	2. Run `ctx-scan-repo --repo . --recommend` for the active repo.
52	3. Resolve a load/unload manifest with the shared recommendation engine.
53	4. Present changes with reasons and require confirmation unless the user enabled
54	automatic mode.
55	5. Record usage/quality changes after the user accepts or rejects suggestions.
56
57	## Scanner
58
59	The scanner reads repo structure and files to produce a stack profile. Detection
60	is evidence-based: every claim should map to a file, dependency, config value, or
61	import pattern.
62
63	### Detection Categories
64
65	- Languages - file extensions, shebangs, lock files.
66	- Frameworks and libraries - package manifests, imports, config files.
67	- Infrastructure and DevOps - Docker, CI/CD, IaC, cloud, Kubernetes.
68	- Data and storage - databases, migrations, queues, pipelines.
69	- Documentation and content - MkDocs, Docusaurus, Sphinx, API specs.
70	- Testing and quality - pytest, Jest, Playwright, Ruff, mypy, TypeScript.
71	- AI and agent tooling - MCP configs, LangGraph, CrewAI, prompt dirs, model
72	config names.
73	- Build and package - package managers, build tools, monorepos.
74
75	### Scanning Rules
76
77	1. Start with filenames and config files. Read source only when needed to
78	disambiguate.
79	2. Do not include speculative signals below the configured confidence floor.
80	3. Skip generated/vendor directories such as `.git`, `node_modules`,
81	`__pycache__`, `venv`, and `.venv`.
82	4. Keep initial scans bounded for large repos.
83	5. Redact secret values; only record secret/key names as evidence.
84
85	## Resolver
86
87	The resolver produces a manifest containing the exact helper set to load or
88	unload.
89
90	```json
91	{
92	"generated_at": "ISO-8601",
93	"repo_path": "/absolute/path",
94	"profile_hash": "sha256 of stack profile",
95	"load": [
96	{
97	"name": "fastapi",
98	"type": "skill",
99	"reason": "FastAPI detected in pyproject.toml dependencies",
100	"score": 0.94
101	}
102	],
103	"unload": [
104	{
105	"name": "react",
106	"type": "skill",
107	"reason": "No active frontend signal in the current repo window"
108	}
109	],
110	"suggestions": [
111	{
112	"name": "github",
113	"type": "mcp",
114	"reason": ".github/workflows exists and the repo uses GitHub Actions",
115	"install_command": "ctx-mcp-add ..."
116	}
117	],
118	"warnings": []
119	}
120	```
121
122	### Ranking Signals
123
124	Candidates can be scored by:
125
126	- tag/category/subcategory match
127	- semantic similarity edge weight
128	- direct wiki links
129	- source overlap
130	- type affinity
131	- usage score
132	- quality/security score
133	- user overrides (`always_load`, `never_load`)
134	- configured caps and minimum score gates
135
136	Recommendations are capped. ctx should not recommend at all costs; below-threshold
137	candidates should be withheld and the gap should be explained.
138
139	## Wiki And Graph Contract
140
141	The wiki follows a Karpathy-style durable memory pattern: schema, purpose, index,
142	log, lint, review ite

People who liked this also viewed

6 picks

Type

docs

560stars68forks

View on GitHub

How to install docs?

stevesolun/ctx/docs

$npx -y skills add stevesolun/ctx --skill docs

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Skill Router
2
3	Scan a repo. Know what the current work needs. Recommend only that. Keep the
4	wiki and graph as the durable catalog behind the decision.
5
6	## Scope
7
8	The router manages runtime recommendations for:
9
10	- skills
11	- agents
12	- MCP servers
13
14	Harnesses are separate catalog entities. They are recommended when a user wants
15	to run ctx with a non-Claude-Code host, local model, API model, or external loop
16	adapter. Once attached, the harness calls the same skills/agents/MCP
17	recommendation engine.
18
19	## Problem
20
21	Every loaded helper costs tokens, attention, and operational surface area. Most
22	sessions need a small top-scored bundle from the shipped graph, not every entity
23	ctx knows about. Loading too much:
24
25	- wastes context on irrelevant instructions,
26	- causes misfires when a helper matches the wrong task,
27	- slows the agent loop, and
28	- creates conflicting instructions.
29
30	## Architecture
31
32	```text
33	ctx/
34	\|-- src/scan_repo.py # Repo scanner -> stack profile
35	\|-- src/ctx/core/resolve/resolve_skills.py # Profile -> load/unload manifest
36	\|-- src/ctx/core/resolve/recommendations.py # Shared scoring/ranking engine
37	\|-- src/ctx/adapters/ # Host adapters + generic tools
38	\|-- src/harness_install.py # Custom-model harness install flow
39	`-- graph/wiki-graph-runtime.tar.gz # Shipped graph/wiki runtime
40	```
41
42	The router has three halves:
43
44	1. Scanner - analyzes a repo and produces stack/task evidence.
45	2. Resolver - scores graph/wiki entities and emits a capped manifest.
46	3. Wiki/graph - persistent catalog of entities, usage, quality, and links.
47
48	## Startup Flow
49
50	1. Read the shipped graph/wiki metadata and local user overrides.
51	2. Run `ctx-scan-repo --repo . --recommend` for the active repo.
52	3. Resolve a load/unload manifest with the shared recommendation engine.
53	4. Present changes with reasons and require confirmation unless the user enabled
54	automatic mode.
55	5. Record usage/quality changes after the user accepts or rejects suggestions.
56
57	## Scanner
58
59	The scanner reads repo structure and files to produce a stack profile. Detection
60	is evidence-based: every claim should map to a file, dependency, config value, or
61	import pattern.
62
63	### Detection Categories
64
65	- Languages - file extensions, shebangs, lock files.
66	- Frameworks and libraries - package manifests, imports, config files.
67	- Infrastructure and DevOps - Docker, CI/CD, IaC, cloud, Kubernetes.
68	- Data and storage - databases, migrations, queues, pipelines.
69	- Documentation and content - MkDocs, Docusaurus, Sphinx, API specs.
70	- Testing and quality - pytest, Jest, Playwright, Ruff, mypy, TypeScript.
71	- AI and agent tooling - MCP configs, LangGraph, CrewAI, prompt dirs, model
72	config names.
73	- Build and package - package managers, build tools, monorepos.
74
75	### Scanning Rules
76
77	1. Start with filenames and config files. Read source only when needed to
78	disambiguate.
79	2. Do not include speculative signals below the configured confidence floor.
80	3. Skip generated/vendor directories such as `.git`, `node_modules`,
81	`__pycache__`, `venv`, and `.venv`.
82	4. Keep initial scans bounded for large repos.
83	5. Redact secret values; only record secret/key names as evidence.
84
85	## Resolver
86
87	The resolver produces a manifest containing the exact helper set to load or
88	unload.
89
90	```json
91	{
92	"generated_at": "ISO-8601",
93	"repo_path": "/absolute/path",
94	"profile_hash": "sha256 of stack profile",
95	"load": [
96	{
97	"name": "fastapi",
98	"type": "skill",
99	"reason": "FastAPI detected in pyproject.toml dependencies",
100	"score": 0.94
101	}
102	],
103	"unload": [
104	{
105	"name": "react",
106	"type": "skill",
107	"reason": "No active frontend signal in the current repo window"
108	}
109	],
110	"suggestions": [
111	{
112	"name": "github",
113	"type": "mcp",
114	"reason": ".github/workflows exists and the repo uses GitHub Actions",
115	"install_command": "ctx-mcp-add ..."
116	}
117	],
118	"warnings": []
119	}
120	```
121
122	### Ranking Signals
123
124	Candidates can be scored by:
125
126	- tag/category/subcategory match
127	- semantic similarity edge weight
128	- direct wiki links
129	- source overlap
130	- type affinity
131	- usage score
132	- quality/security score
133	- user overrides (`always_load`, `never_load`)
134	- configured caps and minimum score gates
135
136	Recommendations are capped. ctx should not recommend at all costs; below-threshold
137	candidates should be withheld and the gap should be explained.
138
139	## Wiki And Graph Contract
140
141	The wiki follows a Karpathy-style durable memory pattern: schema, purpose, index,
142	log, lint, review ite

People who liked this also viewed

6 picks

Type