| name | amp-skill |
| description | Interruption pattern detection and retrieval from Amp thread history. Use for analyzing tool rejection patterns and improving agent behavior. |
| version | 1.0.0 |
Amp-Skill: Interruption Pattern Detection and Retrieval
GF(3) Trit: 0 (ERGODIC - coordination layer)
Foundation: DuckDB ACSet from Amp file-changes
Overview
Amp-Skill distills usage patterns from Amp thread history, specifically focusing on interruption patterns where tool suggestions were rejected in favor of:
- Two-lock processes - user requested confirmation before proceeding
- Complete discontinuation - user abandoned the thread entirely
- Reverted operations - user explicitly undid an AI action
Retrieval Benchmark Metrics
| Metric |
Value |
| Total Threads |
616 |
| Total Tool Calls |
2,535 |
| Threads with Interruptions |
282 (45.8%) |
| Reverted Tool Calls |
84 (3.3%) |
| Two-Lock Cascades |
47 |
| Complete Discontinuations |
22 |
Interruption Pattern Distribution
| Pattern Type |
Count |
Percentage |
| abandoned |
269 |
62.4% |
| reverted |
84 |
19.5% |
| two_lock |
47 |
10.9% |
| discontinuation |
22 |
5.1% |
| high_rejection |
9 |
2.1% |
GF(3) Distribution
| Role |
Threads |
Tool Calls |
Reverted |
| MINUS |
218 |
769 |
27 |
| ERGODIC |
209 |
976 |
26 |
| PLUS |
189 |
790 |
31 |
File Types Most Frequently Reverted
.org - 28 reverts (Emacs org-mode files)
.py - 20 reverts (Python scripts)
.sh - 18 reverts (Shell scripts)
.md - 15 reverts (Markdown documentation)
SQL Access
-- Query interruption patterns
SELECT * FROM amp_interruptions ORDER BY ts DESC;
-- Find high-rejection threads
SELECT thread_id, tool_call_count, reverted_count,
ROUND(100.0 * reverted_count / tool_call_count, 1) as revert_pct
FROM amp_threads
WHERE reverted_count > 0
ORDER BY revert_pct DESC;
-- Detect two-lock cascades
SELECT thread_id, COUNT(*) as consecutive_reverts
FROM amp_interruptions
WHERE pattern_type = 'two_lock'
GROUP BY thread_id
ORDER BY consecutive_reverts DESC;
Key Insights
1. Abandoned Threads (62.4%)
- Single tool call threads indicate quick task completion OR user abandonment
- Requires context analysis to distinguish
2. Two-Lock Cascades (10.9%)
- 47 instances of consecutive reverts within 60 seconds
- Indicates AI attempting same action repeatedly despite rejection
- Action: Implement backoff after 2 consecutive reverts
3. High-Rejection Threads (2.1%)
- 9 threads with >50% revert rate
- Top offender: 100% revert rate on
capability-signer-prototype.sh
- Pattern: Security-sensitive code rejected multiple times
4. File Type Sensitivity
.org files most frequently rejected (28)
- Personal/organizational files have higher rejection rate
- Action: Add confirmation prompt for org-mode edits
Retrieval Benchmark Success Criteria
Usage
# Load/refresh Amp threads
bb scripts/amp_thread_loader.bb
# Query via DuckDB
duckdb trit_stream.duckdb "SELECT * FROM amp_interruptions"
Integration with Triadic Protocol
When Amp-Skill detects a high-rejection pattern:
- MINUS (-1): Validate the rejection pattern
- ERGODIC (0): Coordinate alternative approaches
- PLUS (+1): Generate new solution avoiding rejected pattern
Data Sources
- Local:
~/.amp/file-changes/T-* (2,535 files)
- API: None currently (file-changes only contain diffs)
- Potential: Amp GraphQL API for full conversation history