agent-context-isolation

Name: agent-context-isolation
Author: vibeeval

519stars44forks

How to install agent-context-isolation?

vibeeval/vibecosystem/agent-context-isolation

$npx -y skills add vibeeval/vibecosystem --skill agent-context-isolation

Installs into the current project.

Summary

--- name: agent-context-isolation description: Agent Context Isolation user-invocable: false ---

Files · 1

View on GitHub

SKILL.md

1	# Agent Context Isolation
2
3	Prevent agent output from polluting the main context window.
4
5	## Rules
6
7	### 1. Use Background Agents with File-Based Coordination
8	```
9	# RIGHT - background agent writes to file, main reads file
10	Task(subagent_type="...", run_in_background=true, prompt="... Output to: /path/to/file.md")
11
12	# WRONG - foreground agent dumps full transcript into main context
13	Task(subagent_type="...", run_in_background=false)
14	```
15
16	Background agents with `run_in_background=true` isolate their context. Have them write results to files in `.claude/cache/agents/<agent-type>/`.
17
18	### 2. Never Use TaskOutput to Retrieve Results
19	```
20	# WRONG - dumps entire transcript (70k+ tokens) into context
21	TaskOutput(task_id="<id>")
22	TaskOutput(task_id="<id>", block=true)
23
24	# RIGHT - check expected output files
25	Bash("ls -la .claude/cache/agents/<agent-type>/")
26	Bash("bun test") # verify with tests
27	```
28
29	TaskOutput returns the full agent transcript. Always use file-based coordination instead.
30
31	### 3. Monitor Agent Progress via System Reminders
32	```
33	# System reminders come automatically:
34	# "Agent a42a16e progress: 6 new tools used, 88914 new tokens"
35
36	# To detect completion:
37	# - Watch for progress reminders to stop arriving
38	# - Poll for expected output files: find .claude/cache/agents -name "*.md" -mmin -5
39	# - Check task output file size growth: wc -c /tmp/claude/.../tasks/<id>.output
40	```
41
42	Stuck agent detection:
43	1. Progress reminders stop arriving
44	2. Task output file size stops growing
45	3. Expected output file not created after reasonable time
46
47	### 4. Verify with Tests, Not Output
48	After agent work:
49	1. Run the test suite directly: `bun test`
50	2. Report pass/fail counts
51	3. Only investigate failures if tests fail
52
53	### 5. File-Based Agent Pipeline Pattern
54	```
55	Research agent → .claude/cache/agents/oracle/output.md
56	↓
57	Plan agent → .claude/cache/agents/plan-agent/output.md (reads research)
58	↓
59	Validate agent → .claude/cache/agents/validate-agent/output.md (reads plan)
60	↓
61	Implement agent → src/module.ts (reads validated plan)
62	```
63
64	Each agent reads the previous agent's file output, not TaskOutput.
65
66	## Why This Matters
67
68	Agent context isolation preserves the main conversation's context budget. Reading agent outputs via TaskOutput floods context, causing:
69	- Mid-conversation compaction
70	- Lost context about user's original request
71	- Repeated explanations needed
72
73	## Source
74	- Session where TaskOutput flooded 70k+ tokens into main context
75	- Session 2026-01-01: Successfully used background agents with file-based coordination for SDK Phase 3

People who liked this also viewed

6 picks

Type

lark-openapi-explorer飞书/Lark 原生 OpenAPI 探索：从官方文档库中挖掘未经 CLI 封装的原生 OpenAPI 接口。当用户的需求无法被现有 lark-* skill 或 lark-cli 已注册命令满足，需要查找并调用原生飞书 OpenAPI 时使用。SkillsJul 2026378k16k
lark-skill-maker创建 lark-cli 的自定义 Skill。当用户需要把飞书 API 操作封装成可复用的 Skill（包装原子 API 或编排多步流程）时使用。SkillsJul 2026377k16k
azure-cost-optimizationOfficial agent plugin providing skills and MCP server configurations for Azure scenarios.SkillsJul 2026209k1.3k
vercel-react-native-skillsVercel's official collection of agent skillsSkillsJul 2026172k29k
mcp-builderGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate externaSkillsJul 202693k163k
gws-sharedgws CLI: Shared patterns for authentication, global flags, and output formatting.SkillsJul 202653k30k

agent-context-isolation

519stars44forks

View on GitHub

How to install agent-context-isolation?

vibeeval/vibecosystem/agent-context-isolation

$npx -y skills add vibeeval/vibecosystem --skill agent-context-isolation

Installs into the current project.

Summary

--- name: agent-context-isolation description: Agent Context Isolation user-invocable: false ---

Files · 1

View on GitHub

SKILL.md

1	# Agent Context Isolation
2
3	Prevent agent output from polluting the main context window.
4
5	## Rules
6
7	### 1. Use Background Agents with File-Based Coordination
8	```
9	# RIGHT - background agent writes to file, main reads file
10	Task(subagent_type="...", run_in_background=true, prompt="... Output to: /path/to/file.md")
11
12	# WRONG - foreground agent dumps full transcript into main context
13	Task(subagent_type="...", run_in_background=false)
14	```
15
16	Background agents with `run_in_background=true` isolate their context. Have them write results to files in `.claude/cache/agents/<agent-type>/`.
17
18	### 2. Never Use TaskOutput to Retrieve Results
19	```
20	# WRONG - dumps entire transcript (70k+ tokens) into context
21	TaskOutput(task_id="<id>")
22	TaskOutput(task_id="<id>", block=true)
23
24	# RIGHT - check expected output files
25	Bash("ls -la .claude/cache/agents/<agent-type>/")
26	Bash("bun test") # verify with tests
27	```
28
29	TaskOutput returns the full agent transcript. Always use file-based coordination instead.
30
31	### 3. Monitor Agent Progress via System Reminders
32	```
33	# System reminders come automatically:
34	# "Agent a42a16e progress: 6 new tools used, 88914 new tokens"
35
36	# To detect completion:
37	# - Watch for progress reminders to stop arriving
38	# - Poll for expected output files: find .claude/cache/agents -name "*.md" -mmin -5
39	# - Check task output file size growth: wc -c /tmp/claude/.../tasks/<id>.output
40	```
41
42	Stuck agent detection:
43	1. Progress reminders stop arriving
44	2. Task output file size stops growing
45	3. Expected output file not created after reasonable time
46
47	### 4. Verify with Tests, Not Output
48	After agent work:
49	1. Run the test suite directly: `bun test`
50	2. Report pass/fail counts
51	3. Only investigate failures if tests fail
52
53	### 5. File-Based Agent Pipeline Pattern
54	```
55	Research agent → .claude/cache/agents/oracle/output.md
56	↓
57	Plan agent → .claude/cache/agents/plan-agent/output.md (reads research)
58	↓
59	Validate agent → .claude/cache/agents/validate-agent/output.md (reads plan)
60	↓
61	Implement agent → src/module.ts (reads validated plan)
62	```
63
64	Each agent reads the previous agent's file output, not TaskOutput.
65
66	## Why This Matters
67
68	Agent context isolation preserves the main conversation's context budget. Reading agent outputs via TaskOutput floods context, causing:
69	- Mid-conversation compaction
70	- Lost context about user's original request
71	- Repeated explanations needed
72
73	## Source
74	- Session where TaskOutput flooded 70k+ tokens into main context
75	- Session 2026-01-01: Successfully used background agents with file-based coordination for SDK Phase 3

People who liked this also viewed

6 picks

Type