autoresearch（自动研究）

Autoresearch — Autonomous Iterative Optimization

You define the goal and how to measure it. The agent does the rest: hypothesize, edit, measure, keep or revert — running autonomously until you stop it or the budget runs out.

Inspired by Karpathy's autoresearch, adapted for Python, SQL, pytest, ty, and ML workflows.

[!IMPORTANT] Every experiment is committed before running, and reverted on failure. The branch only ever advances on real improvements.

Phase 1 — Setup (interactive)

Work through these questions with the user before touching any code. Do not skip or assume answers.

1.1 Goal

Ask:

What are you trying to improve? Examples: execution time, memory usage, pytest pass rate, ty error count, SQL query latency, model accuracy, training throughput, bundle size.

1.2 Metric command

Ask:

What command produces the metric, and how do I read the number from its output? 1. Command — the exact shell command to run 2. Extraction — regex, line number, JSON path, or description of what to parse 3. Direction — lower is better, or higher is better?

Refer to the domain quick-reference below for ready-made commands if the user is unsure.

Record:

• METRIC_COMMAND
• METRIC_EXTRACTION
• METRIC_DIRECTION (lower_is_better | higher_is_better)

1.3 Scope

Ask:

Which files or directories may I edit? Which are off-limits?

Record:

• IN_SCOPE — files/dirs the agent may modify
• OUT_OF_SCOPE — must not be touched

1.4 Constraints

Ask:

Any constraints I should respect? Examples: no new dependencies, must keep existing tests green, public API must stay stable, max 2 min per run, must stay type-clean (ty 0 errors), VRAM budget, complexity budget.

Record as CONSTRAINTS.

1.4b Web search

Ask:

May I search the web for optimization ideas, documentation, or techniques? Web search lets me look up library docs, algorithm papers, Stack Overflow answers, and benchmarking guides to generate better hypotheses — especially useful for ML tuning, SQL optimization, and unfamiliar libraries. Options: yes (search freely), ask (propose each query before running), no (stay offline, codebase only).

Record as WEB_SEARCH (yes | ask | no, default no).

1.5 Budget

Ask:

How many experiments, or keep going until interrupted?

Record as MAX_EXPERIMENTS (number or unlimited).

1.6 Simplicity policy

State the default and ask for adjustments:

Default: simpler beats marginally faster. Removing code while holding or improving the metric is a win. Complexity has a cost — weigh it honestly against the gain. OK to proceed with this policy, or do you want to adjust it?

Record any adjustment as SIMPLICITY_POLICY.

1.7 Confirm

Present a summary table and wait for explicit confirmation before continuing.

Parallel autoresearch

Run multiple experiment loops simultaneously, each on its own branch and working directory, using git worktree. Useful when you want to race two hypotheses, explore independent scopes concurrently, or keep your main editor on main while experiments run.

When to use parallel mode

• Two or more independent scopes (different files or modules) that won't conflict
• Long-running metric commands where waiting for one loop blocks progress
• Competing strategies you want to benchmark side-by-side

Setup

For each parallel run, create a dedicated worktree before starting its loop:

# In the repo root — repeat for each parallel run
RUN_ID_A="<goal-a>"   # e.g. etl-runtime
RUN_ID_B="<goal-b>"   # e.g. ty-errors

git worktree add ../<repo>-${RUN_ID_A} -b autoresearch/${RUN_ID_A}
git worktree add ../<repo>-${RUN_ID_B} -b autoresearch/${RUN_ID_B}

Each worktree is a sibling directory with its own HEAD and index — commits, resets, and file edits in one worktree are completely invisible to the others.

Guidelines

Cleanup

When a parallel run finishes, remove its worktree:

git worktree remove ../<repo>-${RUN_ID_A}   # removes directory + unregisters
# or, if the run produced no improvement:
git worktree remove --force ../<repo>-${RUN_ID_A}
git branch -D autoresearch/${RUN_ID_A}

List active worktrees at any time:

git worktree list

Phase 2 — Branch & baseline

1. Generate a run ID — a short goal slug, e.g. pytest-passrate. This ID is used for all artifacts of this run so multiple runs never collide. RUN_ID="<goal-slug>" # e.g. pytest-passrate RESULTS_FILE="autoresearch-${RUN_ID}.tsv" LOG_FILE="autoresearch-${RUN_ID}.log"
2. Create a branch — propose autoresearch/<run-id>, create it: git checkout -b autoresearch/${RUN_ID}
3. Read in-scope files — build full context before making any changes.
4. Initialize $RESULTS_FILE in the repo root: experiment commit metric status description Register both files in .git/info/exclude (append only — never modify tracked files): echo "${RESULTS_FILE}" >>.git/info/exclude echo "${LOG_FILE}" >>.git/info/exclude
5. Run the baseline — execute METRIC_COMMAND, extract the value, record as experiment 0 with status baseline.
6. Report to the user: Baseline: [metric] = [value]. Run ID: <run-id>. Starting experiment loop.

Phase 3 — Experiment loop

Run continuously. Never pause to ask "should I continue?". Stop only when:

• MAX_EXPERIMENTS is reached, or
• the user interrupts

Each iteration

QUESTION  (Questioner) Profile the current state before hypothesizing.
          Ask structured questions about the code/data to surface bottlenecks:
          1. What changed since last iteration? (diff awareness)
          2. Where is time/resource actually spent? (profile or re-read metrics)
          3. What patterns in prior results suggest a direction?
             - 2+ consecutive keeps in same area → probe deeper
             - 3+ discards → pivot strategy
             - crash → structural issue, not parameter issue
          4. Are there external signals to incorporate?
             (error messages, profiler output, log patterns)
          Record answers as QUESTIONER_NOTES for this iteration.

THINK     Synthesize QUESTIONER_NOTES into a hypothesis.
          Use domain reasoning vocabulary (see below) to sharpen the hypothesis.
          Form: "X should improve Y because Z."
          If WEB_SEARCH is `yes` or `ask` and you are stuck or entering a new
          domain (e.g. unfamiliar library, ML algorithm, SQL planner behavior),
          search the web for relevant techniques, docs, or benchmarks before
          forming the hypothesis. For `ask`, state the proposed query and wait
          for confirmation. Log the source URL in the description column of the
          TSV when a web result directly inspired the experiment.
          Follow experiment strategy priority below.

SCORE     Rate the hypothesis before acting (1–10 each):
          - Impact: how much metric improvement expected?
          - Feasibility: how likely to work without breaking things?
          - Novelty: how different from what was already tried? (check $RESULTS_FILE)
          Average ≥ 5 → proceed. Below 5 → generate a better hypothesis.
          Skip SCORE on experiment #1 (no prior data to compare against).

REFLECT   Self-check before editing:
          - What assumption am I making that could be wrong?
          - Has something similar already been tried and failed? (scan $RESULTS_FILE)
          - Am I stuck in a local optimum? (3+ keeps in same area → try different axis)
          - Could this change make things worse in a way I won't measure?
          If reflection reveals a flaw → revise hypothesis and re-SCORE.

EDIT      Make one focused change to in-scope files.
          Keep it minimal — one idea per experiment.

COMMIT    Stage only in-scope files, then commit:
          git add <IN_SCOPE files> && git commit
          Message format: "experiment: <short description>"

RUN       Execute METRIC_COMMAND, redirect all output:
          <command> > $LOG_FILE 2>&1

MEASURE   Extract the metric from $LOG_FILE.
          On failure: read the last 50 lines of run.log for the error.

INSPECT   (Inspector) Validate the experiment beyond just the metric.
          Checklist — all must pass for a "keep" decision:
          ☐ Metric improved (or held, if simplification pass)
          ☐ Change is minimal and focused (one idea, not a kitchen sink)
          ☐ No unrelated regressions introduced (test suite, type checker)
          ☐ Code complexity did not increase disproportionately to the gain
          ☐ The change is understandable — would a reviewer accept it?
          ☐ No hardcoded magic values that only work for the benchmark
          ☐ Description matches actual change (no hallucinated improvements)
          ☐ Metric improvement is real, not positive spin on noise
          If any check fails, downgrade to "discard" even if metric improved.
          Record inspector verdict as INSPECTOR_NOTES.

DECIDE    Compare to current best (incorporating INSPECTOR_NOTES):
          ✅ IMPROVED  → keep commit, update best, log status = "keep"
          ❌ SAME/WORSE → revert only in-scope files:
                         git reset HEAD~1               # soft-reset: undo commit, keep working tree
                         git restore <IN_SCOPE files>   # discard changes to in-scope files only
                         log status = "discard"
          ⚠️ METRIC UP BUT INSPECTOR FAIL → revert, log status = "inspector-reject"
                         Record why: "metric +12% but added 40 LOC of unmaintainable code"
          💥 CRASH      → attempt up to 2 quick fixes (typo, import, simple error),
                         amend commit, re-run. If still broken, soft-reset and
                         restore only in-scope files; log status = "crash".

LOG       Append to results.tsv:
          <N>	<commit>	<value>	<status>	<description>

Domain reasoning vocabulary

Insert these domain-specific keywords into your THINK step to sharpen hypothesis quality. Using precise terminology activates better reasoning patterns.

Experiment strategy

Follow this priority order:

1. Low-hanging fruit — obvious inefficiencies, trivial parameter changes
2. Follow promising directions — if something worked, probe further
3. Diversify after plateaus — 3–5 consecutive failures → switch strategy
4. Combine winners — if A and B each improved independently, try A+B
5. Simplification passes — periodically try removing code; hold the metric
6. Bigger changes — algorithmic or architectural changes after incremental ideas run dry

Constraint enforcement

• Time budget: if a run exceeds 2× baseline duration, kill and treat as crash
• Test integrity: if constraints require green tests, run them after each experiment; revert if they break, even if the primary metric improved
• ty/type safety: if type-cleanliness is a constraint, run uv run ty check after each change and revert if new errors appear

Phase 4 — Report, cleanup & next steps

When the loop ends (budget reached or interrupted), work through all four sub-phases.

4.1 Results report

1. Print $RESULTS_FILE as a formatted table.
2. Summarize:

- Total experiments / kept / discarded / crashed - Baseline → final metric, improvement % - Top 3 most impactful changes (by metric delta)

1. Show the git log of kept experiments: git log --oneline <start-commit>..HEAD

4.1b Run-level review

After reporting results, perform a structured review of the entire optimization run. Rate the run on these axes (1–10):

Present scores and a 2–3 sentence summary. Flag:

• Experiments where metric improved but the change may be fragile
• Promising directions that were not explored (ideas scored but never tried)
• Any sign of overfitting to the specific benchmark input
• Inspector-rejected experiments that might be worth revisiting with a different approach

If the run-level review reveals concerns, note them in the report before cleanup.

4.2 Cleanup

Remove run artifacts that are no longer needed. Ask the user once before deleting:

Run complete. Clean up $RESULTS_FILE and $LOG_FILE from the working directory? (They stay in git history if you need them later.)

If confirmed:

rm -f "${RESULTS_FILE}" "${LOG_FILE}"

Also remove the excludes entries added in Phase 2 so the file is left tidy:

# removes the two lines added during setup (grep -v is safe here — no tracked files touched)
grep -v "^${RESULTS_FILE}$\|^${LOG_FILE}$" .git/info/exclude > /tmp/_exclude_tmp \
  && mv /tmp/_exclude_tmp .git/info/exclude

4.3 Proposed next steps

Present these as a checklist. Mark which apply based on what the run actually changed.

Code quality

• Run /deslop on changed files — automated optimization often leaves mechanical patterns, naming inconsistencies, or removed comments that need a pass
• Run ty across the full project to confirm no new type errors leaked in
• Run the full test suite one final time on the current branch tip

Commit hygiene

• Squash experiment commits into one or a few logical commits before merging: bash git rebase -i <start-commit> Replace all experiment: commits with meaningful messages describing *what changed and why it helped*.
• Update any affected docstrings or inline comments that describe the old behavior

Integration

• Open a PR from autoresearch/<run-id> into your base branch
• Add a note in the PR description linking to the $RESULTS_FILE (or paste the summary table) so reviewers understand the methodology

Further experimentation

• Things not tried (ideas the loop skipped as too risky or complex): — algorithmic rewrites that would change public interfaces — dependency upgrades — schema or data-structure changes — parallelism / concurrency changes Decide which are worth pursuing manually.

4.4 Branch lifecycle

If the run produced no net improvement, offer to delete the branch cleanly:

git checkout <base-branch>
git branch -D autoresearch/${RUN_ID}

If improvements were made, leave the branch for PR review.

Domain metric quick-reference

Use these when the user isn't sure how to phrase their metric command.

Python runtime

# hyperfine (install: brew install hyperfine / pip install hyperfine)
hyperfine --warmup 3 'python my_script.py'
# → parse: "mean" field (lower is better)

# built-in timing
python -m timeit -n 100 -r 5 "import my_module; my_module.run()"

# cProfile summary
python -m cProfile -s cumtime my_script.py 2>&1 | head -20

Python memory

# tracemalloc (add to script, or use wrapper)
python -c "
import tracemalloc, my_module
tracemalloc.start()
my_module.run()
current, peak = tracemalloc.get_traced_memory()
print(f'peak_kb={peak/1024:.1f}')
"
# → parse: peak_kb= line (lower is better)

# memory_profiler (pip install memory_profiler)
python -m memory_profiler my_script.py

pytest

# pass rate
pytest --tb=no -q 2>&1 | tail -1
# → parse: "X passed" (higher is better)

# coverage
pytest --cov=src --cov-report=term-missing --tb=no -q 2>&1 | grep "TOTAL"
# → parse: last percentage (higher is better)

# duration
pytest --tb=no -q 2>&1 | grep "passed"
# → parse duration from summary line (lower is better)

ty

# count type errors (lower is better)
uv run ty check 2>/dev/null | grep -c "^error\["

# exit code only: 0 = clean, 1 = errors
uv run ty check 2>/dev/null; echo "exit=$?"

SQL (PostgreSQL)

# query duration via psql (wrap your query in EXPLAIN ANALYZE)
psql $DATABASE_URL -c "EXPLAIN (ANALYZE, FORMAT JSON) <your query>" \
  | python -c "import sys,json; d=json.load(sys.stdin); print(d[0]['Execution Time'])"
# → parse: Execution Time (lower is better)

# if using pgbench
pgbench -c 5 -T 30 $DATABASE_URL 2>&1 | grep "tps ="
# → parse: tps value (higher is better)

Machine learning

# training run — capture final metric from stdout
python train.py --epochs 10 2>&1 | grep "val_loss" | tail -1
# → parse: val_loss= value (lower is better)

# sklearn cross-validation
python -c "
from sklearn.model_selection import cross_val_score
import numpy as np, my_model, my_data
X, y = my_data.load()
scores = cross_val_score(my_model.build(), X, y, cv=5, scoring='f1_macro')
print(f'f1={np.mean(scores):.4f}')
"
# → parse: f1= (higher is better)

Spark / Databricks (via Spark Connect)

Run PySpark scripts locally against a remote Databricks cluster using databricks-connect. The cluster does the heavy lifting; the local process is your experiment coordinator. See the spark-connect skill for setup details.

# Time a PySpark script end-to-end (local process + cluster execution)
hyperfine --warmup 1 'uv run python notebooks/my_pipeline.py'
# → parse: "mean" field (lower is better)

# Capture a row count or computed metric from Spark SQL
uv run python -c "
from utils.databricks import get_spark_session
spark = get_spark_session()
result = spark.sql('''
    SELECT COUNT(*) as n FROM my_catalog.my_schema.my_table
    WHERE condition = true
''').collect()[0]['n']
print(f'metric={result}')
"
# → parse: metric= (direction depends on goal)

# Capture MLflow metric from a training run
uv run python -c "
import mlflow
client = mlflow.tracking.MlflowClient()
run = client.search_runs(['<experiment_id>'], order_by=['metrics.val_f1 DESC'], max_results=1)[0]
print(f'val_f1={run.data.metrics[\"val_f1\"]:.4f}')
"
# → parse: val_f1= (higher is better)

Autoresearch loop model with Spark Connect:

Local machine (autoresearch loop)
  └─ edits notebook / script
  └─ commits the change
  └─ runs: uv run python notebooks/my_script.py
            └─ get_spark_session()  →  Spark Connect  →  Databricks Cluster
            └─ reads metric from stdout / MLflow
  └─ keep or revert based on metric
  └─ next iteration

Constraints to consider for Spark runs:

• Each run spins up or warms up the Databricks cluster — cold starts add ~1–2 min
• Use hyperfine --warmup 1 (not 3+) to avoid excessive cluster cost
• Set a generous time budget cap: 2× baseline is fine for inter-cluster runs
• Prefer spark.sql(...) over PySpark DataFrame API for readability and planability
• If DATABRICKS_CLUSTER_ID or DATABRICKS_PROFILE are not set, fail fast with a clear message rather than hanging

Results TSV format

Filename: autoresearch-<run-id>.tsv (e.g. autoresearch-pytest-passrate.tsv)

Tab-separated, 5 columns:

experiment	commit	metric	status	description
0	a1b2c3d	142.3	baseline	unmodified code
1	b2c3d4e	138.1	keep	replace list comprehension with generator
2	c3d4e5f	145.0	discard	switch to numpy vectorization (slower on small data)
3	d4e5f6g	0.0	crash	add numba jit (import error, unfixable)
4	e5f6g7h	131.4	keep	cache repeated db lookups with lru_cache
5	f6g7h8i	128.9	inspector-reject	inline dict — metric +2% but +40 LOC unmaintainable

Key principles