Before opening, please confirm:
Operating System
darwin
Kiro Version
0.8.140
Bug Description
Kiro Context Overflow Bug Report
Issue Summary
New sessions immediately hit 85%+ context usage and enter summarization loop, even with minimal steering docs and simple prompts. This worked fine for 6+ months prior.
Environment
OS: macOS (darwin)
Kiro Version: [Check Help > About in Kiro]
Workspace: homeassistant-kiro
MCP Servers: home-assistant, influxdb, grafana (configured globally)
Reproduction Steps
Start fresh Kiro session
Send prompt: "It's been a while, are there any insights from homeassistant?"
AI responds with clarification request (correct behavior)
Context hits 85%+ threshold
"Conversation above has been summarized" appears
"Conversation summarization failed" appears
Session becomes unusable
Expected Behavior
Simple question + response should use <20% context
Worked correctly for 6+ months with same setup
Should not trigger summarization on first exchange
Current Context Load
Always-Loaded Steering Files (3 files, ~33 lines total)
8 lines - .kiro/steering/git.md
15 lines - .kiro/steering/homeassistant-mcp.md
10 lines - .kiro/steering/sensors.md
Manual Steering Files (5 files, only loaded when referenced)
187 lines - deployment.md
131 lines - devices.md
17 lines - product.md
87 lines - structure.md
48 lines - tech.md
No Active Specs
Deleted completed specs (1,186 lines removed)
.kiro/specs/ directory is empty
What We've Tried
✅ Reduced always-loaded steering from ~400 to ~33 lines
✅ Moved large guides to docs/ directory
✅ Deleted completed spec files
✅ Set most steering to inclusion: manual
✅ Made steering docs ultra-concise
❌ Added directories to .kiroignore - DID NOT WORK
✅ Physically moved files out of workspace - THIS WORKED
Root Cause: .kiroignore Not Working
The Problem
.kiroignore file contained:
ha-live-config/
timelapse_photos/
timelapse_videos/
Expected behavior: These directories should be excluded from file tree
Actual behavior: All 10,080 files were still loaded into context (100,000+ tokens)
Evidence
Added entries to .kiroignore
Restarted Kiro
Context still overflowed (85%+ usage)
Physically moved files out of workspace
Restarted Kiro
Context usage dropped to normal (<20%)
Conclusion
.kiroignore is not being respected by Kiro's file tree loader.
This is either:
A bug in .kiroignore implementation
File tree caching issue (doesn't refresh on .kiroignore changes)
Syntax issue (trailing slashes not working as expected)
Requires workspace reload, not just Kiro restart
Impact
Users cannot rely on .kiroignore to exclude large directories. The only reliable workaround is to physically move files outside the workspace directory.
This is a significant usability issue for workspaces with:
Photo/video collections
Build artifacts
Cache directories
Large datasets
Synced/mirrored directories
Minimal Test Case
Prompt: "What's the temperature?"
Expected:
AI queries specific sensor
Returns temperature value
Uses <10% context
Actual:
[Test this in new session and document result]
Hypothesis
One of these changed in recent Kiro update:
Context limit decreased
Summarization threshold became more aggressive (85% → lower?)
MCP tool definitions now counted in context
Base system prompt grew larger
Context calculation method changed
Additional Context
MCP servers configured globally at ~/.kiro/settings/mcp.json
Home Assistant MCP returns large responses if not filtered
Steering explicitly warns to use filters
AI correctly asks for clarification (not calling MCP without filters)
Failure happens AFTER AI response, not during
Questions for Kiro Team
Is .kiroignore supposed to exclude files from the file tree context?
Does .kiroignore require workspace reload (not just Kiro restart)?
What's the correct syntax for directory exclusion?
Is there file tree caching that prevents .kiroignore from taking effect?
Was there a recent change to context limits or summarization?
Are MCP tool definitions counted toward context usage?
What's the actual context limit in tokens?
Can we see context usage breakdown (system prompt vs steering vs file tree vs conversation)?
Is there a way to disable auto-summarization for debugging?
Recommended Fix
Kiro should either:
Fix .kiroignore to actually work - most important
Show warning when file tree exceeds threshold (e.g., "10,000+ files detected, consider using .kiroignore")
Provide context usage breakdown in UI so users can diagnose issues
Document .kiroignore behavior clearly in official docs
Workaround
User must be extremely specific in prompts:
❌ "Any insights?"
✅ "Show me Pico W 01 temperature reading"
This defeats the purpose of conversational AI.
Date: January 18, 2026 Reporter: User via AI assistant
Steps to Reproduce
Start fresh Kiro session
Send prompt: "It's been a while, are there any insights from homeassistant?"
AI responds with clarification request (correct behavior)
Context hits 85%+ threshold
"Conversation above has been summarized" appears
"Conversation summarization failed" appears
Session becomes unusable
Expected Behavior
AI queries specific sensor
Returns temperature value
Uses <10% context
Conversation ID
142a08bb-b283-4390-a6c8-4ab652df4711
8730600f-07d3-4559-84c1-c96c74943f71
Additional Context
Kiro captured & logged content in the bug. Retried rebooting OS, different kiro workspaces (no issue), and existing sessions were fine, only creating new sessions bug was evident. Luckily with a few open sessions in this workspace kiro was able to help me narrow down the fact the kiroignore wasn't working as it previously did in previous versions of kiro (been using this workspace for a while, and haven't even pulled in new files to the workspace).
Before opening, please confirm:
Operating System
darwin
Kiro Version
0.8.140
Bug Description
Kiro Context Overflow Bug Report
Issue Summary
New sessions immediately hit 85%+ context usage and enter summarization loop, even with minimal steering docs and simple prompts. This worked fine for 6+ months prior.
Environment
OS: macOS (darwin)
Kiro Version: [Check Help > About in Kiro]
Workspace: homeassistant-kiro
MCP Servers: home-assistant, influxdb, grafana (configured globally)
Reproduction Steps
Start fresh Kiro session
Send prompt: "It's been a while, are there any insights from homeassistant?"
AI responds with clarification request (correct behavior)
Context hits 85%+ threshold
"Conversation above has been summarized" appears
"Conversation summarization failed" appears
Session becomes unusable
Expected Behavior
Simple question + response should use <20% context
Worked correctly for 6+ months with same setup
Should not trigger summarization on first exchange
Current Context Load
Always-Loaded Steering Files (3 files, ~33 lines total)
8 lines - .kiro/steering/git.md
15 lines - .kiro/steering/homeassistant-mcp.md
10 lines - .kiro/steering/sensors.md
Manual Steering Files (5 files, only loaded when referenced)
187 lines - deployment.md
131 lines - devices.md
17 lines - product.md
87 lines - structure.md
48 lines - tech.md
No Active Specs
Deleted completed specs (1,186 lines removed)
.kiro/specs/ directory is empty
What We've Tried
✅ Reduced always-loaded steering from ~400 to ~33 lines
✅ Moved large guides to docs/ directory
✅ Deleted completed spec files
✅ Set most steering to inclusion: manual
✅ Made steering docs ultra-concise
❌ Added directories to .kiroignore - DID NOT WORK
✅ Physically moved files out of workspace - THIS WORKED
Root Cause: .kiroignore Not Working
The Problem
.kiroignore file contained:
ha-live-config/
timelapse_photos/
timelapse_videos/
Expected behavior: These directories should be excluded from file tree
Actual behavior: All 10,080 files were still loaded into context (100,000+ tokens)
Evidence
Added entries to .kiroignore
Restarted Kiro
Context still overflowed (85%+ usage)
Physically moved files out of workspace
Restarted Kiro
Context usage dropped to normal (<20%)
Conclusion
.kiroignore is not being respected by Kiro's file tree loader.
This is either:
A bug in .kiroignore implementation
File tree caching issue (doesn't refresh on .kiroignore changes)
Syntax issue (trailing slashes not working as expected)
Requires workspace reload, not just Kiro restart
Impact
Users cannot rely on .kiroignore to exclude large directories. The only reliable workaround is to physically move files outside the workspace directory.
This is a significant usability issue for workspaces with:
Photo/video collections
Build artifacts
Cache directories
Large datasets
Synced/mirrored directories
Minimal Test Case
Prompt: "What's the temperature?"
Expected:
AI queries specific sensor
Returns temperature value
Uses <10% context
Actual:
[Test this in new session and document result]
Hypothesis
One of these changed in recent Kiro update:
Context limit decreased
Summarization threshold became more aggressive (85% → lower?)
MCP tool definitions now counted in context
Base system prompt grew larger
Context calculation method changed
Additional Context
MCP servers configured globally at ~/.kiro/settings/mcp.json
Home Assistant MCP returns large responses if not filtered
Steering explicitly warns to use filters
AI correctly asks for clarification (not calling MCP without filters)
Failure happens AFTER AI response, not during
Questions for Kiro Team
Is .kiroignore supposed to exclude files from the file tree context?
Does .kiroignore require workspace reload (not just Kiro restart)?
What's the correct syntax for directory exclusion?
Is there file tree caching that prevents .kiroignore from taking effect?
Was there a recent change to context limits or summarization?
Are MCP tool definitions counted toward context usage?
What's the actual context limit in tokens?
Can we see context usage breakdown (system prompt vs steering vs file tree vs conversation)?
Is there a way to disable auto-summarization for debugging?
Recommended Fix
Kiro should either:
Fix .kiroignore to actually work - most important
Show warning when file tree exceeds threshold (e.g., "10,000+ files detected, consider using .kiroignore")
Provide context usage breakdown in UI so users can diagnose issues
Document .kiroignore behavior clearly in official docs
Workaround
User must be extremely specific in prompts:
❌ "Any insights?"
✅ "Show me Pico W 01 temperature reading"
This defeats the purpose of conversational AI.
Date: January 18, 2026 Reporter: User via AI assistant
Steps to Reproduce
Start fresh Kiro session
Send prompt: "It's been a while, are there any insights from homeassistant?"
AI responds with clarification request (correct behavior)
Context hits 85%+ threshold
"Conversation above has been summarized" appears
"Conversation summarization failed" appears
Session becomes unusable
Expected Behavior
AI queries specific sensor
Returns temperature value
Uses <10% context
Conversation ID
142a08bb-b283-4390-a6c8-4ab652df4711
8730600f-07d3-4559-84c1-c96c74943f71
Additional Context
Kiro captured & logged content in the bug. Retried rebooting OS, different kiro workspaces (no issue), and existing sessions were fine, only creating new sessions bug was evident. Luckily with a few open sessions in this workspace kiro was able to help me narrow down the fact the kiroignore wasn't working as it previously did in previous versions of kiro (been using this workspace for a while, and haven't even pulled in new files to the workspace).