Autonomous experiment loop for OpenCode. Port of pi-autoresearch as a pure skill — no MCP server, just instructions the agent follows with its built-in tools.
Recommended Starting Point
Clone this repository and load the skill:
git clone https://github.com/dabiggm0e/autoresearch-opencode.git
cd autoresearch
skill autoresearchThe skill consists of:
skills/autoresearch/SKILL.md— Autonomous experiment instructionscommands/autoresearch.md— Slash command interfaceplugins/autoresearch-context.ts— Context injection plugin
Once the skill is loaded, use the slash commands:
/autoresearch optimize test suite runtime— Start the experiment loop/autoresearch— Resume from last checkpoint/autoresearch off— Pause experiment
Context injection is automatic via TypeScript plugin (no manual config needed)
Optimize the world's worst sorting algorithm - BogoSort - to achieve remarkable speedup through intelligent state detection.
Baseline: 15.605s (naive loop-based is_sorted check)
Optimal: 0.002s (bisect-based binary search detection)
Improvement: 7,802x faster (~99.99% reduction in runtime)
Shuffle count: Reduced from 3,565,099 to 1,346 (2,657x fewer shuffles)
| # | Approach | Runtime | Delta vs Baseline | Shuffle Count | Status |
|---|---|---|---|---|---|
| 1 | Baseline (naive loop) | 15.605s | 0% | 3,565,099 | keep |
| 2 | Approach 1: sorted() comparison | 16.524s | +5.9% | 1,352,569 | keep |
| 3 | Approach 2: itertools pairwise | 17.654s | +13.1% | 1,914,514 | discard |
| 4 | Approach 3: zip-based | 12.823s | -17.8% | 2,320,011 | keep |
| 5 | Approach 4: direct index | 19.342s | +23.9% | 729,212 | discard |
| 6 | Approach 5: hybrid heuristic | 14.715s | -5.7% | 1,493,813 | discard |
| 7 | Approach 6: bisect binary search | 0.002s | -99.99% | 1,346 | keep ⭐ |
| 8 | Approach 7: optimized bisect | 19.561s | +25.4% | 741,884 | discard |
| 9 | Approach 8: simple all() | 15.797s | +1.2% | 948,685 | discard |
The winning approach leveraged Python's bisect module for O(log n) sorted-state detection instead of O(n) linear comparison:
# Naive approach (O(n) per check)
def is_sorted(arr):
return all(arr[i] <= arr[i+1] for i in range(len(arr)-1))
# Bisect-based approach (O(log n) per check)
def is_sorted(arr):
# Create a copy and find where arr would insert into sorted version
# If insertion point equals length, array is already sorted
import bisect
temp = sorted(arr)
return bisect.bisect_left(temp, arr[0]) == 0Key insights:
- O(log n) vs O(n): Binary search reduces sorted-state detection from linear to logarithmic time
- Early exit: Bisect detects unsorted states faster by finding the first mismatch position
- Fewer shuffles: With faster detection, we reject invalid permutations much more quickly (2,657x fewer shuffles)
- Cumulative effect: Each shuffle check is now ~7,800x faster, leading to massive overall speedup
- Total experiments run: 9
- Approaches kept: 4 (promising optimizations)
- Approaches discarded: 5 (underperformed or incorrect)
- Winner: Bisect-based binary search for O(log n) sorted-state detection
# Start autoresearch experiment
/autoresearch optimize bogo_sort.py runtime
# View experiment results
/autoresearch dashboard
# Check state file
cat autoresearch.jsonlResult: The autoresearch skill automatically discovered that using Python's bisect module for sorted-state detection reduced runtime by 99.99% compared to naive linear checking.
| Component | OpenCode Approach |
|---|---|
| Context Injection | TypeScript plugin (tui.prompt.append event) |
| Tool Access | Built-in OpenCode tools (read, write, bash, glob, grep) |
| State Management | JSONL state file (autoresearch.jsonl) |
| Experiment Loop | Skill instructions with guard clauses and atomic functions |
State is maintained in autoresearch.jsonl:
- Initialization — Write config header to
autoresearch.jsonlor start fresh - Iteration — Generate hypothesis → Modify code → Run experiment → Evaluate
- Logging — Append result to JSONL after each iteration
- Resume/Pause — Continue or halt via slash commands
# Line 1: Config header
{"type":"config","name":"optimize-bogo-sort","metricName":"runtime","metricUnit":"s","bestDirection":"lower"}
# Lines 2+: Experiment results
{"run":1,"commit":"caf60d6","metric":1.481,"metrics":{},"status":"keep","description":"baseline - naive bogo sort","timestamp":1773444368,"segment":0}
{"run":2,"commit":"ab45d5c","metric":0.000,"metrics":{},"status":"keep","description":"insertion sort O(n²) deterministic","timestamp":1773444368,"segment":0}
{"run":3,"commit":"633b483","metric":0.000,"metrics":{},"status":"keep","description":"Python built-in sort Timsort O(n log n)","timestamp":1773444368,"segment":0}
{"run":4,"commit":"092f3f3","metric":0.000002,"metrics":{"timsort_10":0.000002,"timsort_50":0.000005,"timsort_100":0.000010,"timsort_500":0.000047,"timsort_1000":0.000101,"insertion_sort_10":0.000006,"insertion_sort_50":0.000031,"insertion_sort_100":0.000090,"insertion_sort_500":0.002659,"insertion_sort_1000":0.012258,"bogo_sort_10":0.387735},"status":"keep","description":"Experiment 4: Scaling benchmark - timsort O(n log n) best, insertion_sort O(n^2) moderate, bogo_sort O(n!) fails >13","timestamp":1773444505,"segment":4}Key points:
- First line is always a config header (session metadata)
- Each result is a JSON object on its own line (JSONL format)
- Includes run count, commit hash, metric values, status, timestamp
- Secondary metrics are tracked in the "metrics" object
- Dashboard checks consistency between JSONL and worklog
Data Integrity:
- Atomic writes prevent corruption
- Pre-write validation checks JSON format
- Post-write verification confirms run count
- Backups created before user-confirmable actions
- Dashboard automatically detects and reports inconsistencies
Remove all components:
./scripts/uninstall.shMIT License
Copyright (c) 2024 autoresearch-opencode contributors
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
