Autoresearch: Autonomous Experiment Loop

An autonomous optimization loop where Claude edits code, runs a benchmark, measures a metric, and keeps improvements or reverts — repeating forever until stopped.

Core Concept

The loop is simple: edit → commit → run → measure → keep or discard → repeat.

• Primary metric is king. Lower (or higher, depending on direction) is better. Improved → keep the commit. Equal or worse → git revert.
• State survives context resets via autoresearch.jsonl (append-only log) and autoresearch.md (living session document).
• Domain-agnostic. Works for any measurable target: test speed, bundle size, LLM training loss, Lighthouse scores, build times, etc.
• Be careful not to overfit to the benchmarks and do not cheat on the benchmarks. Optimize the real workload, not the measurement harness.

Setup Phase

When the user triggers autoresearch, gather the following (ask if not provided). If $ARGUMENTS is provided, use it as the optimization goal:

1. Goal — what to optimize (e.g., "reduce unit test runtime") — use $ARGUMENTS if provided
2. Command — the benchmark to run (e.g., pnpm test, uv run train.py)
3. Primary metric — name, unit, and direction (lower or higher is better)
4. Secondary metrics — optional additional metrics to track for tradeoff monitoring (e.g., memory, compile time)
5. Files in scope — which files can be modified
6. Constraints — time budget, off-limits files, correctness requirements

Optionally check for .claude/autoresearch-ai-plugin.local.md in the project root for persistent configuration:

---
enabled: true
max_iterations: 50
working_dir: "/path/to/project"
benchmark_timeout: 600
checks_timeout: 300
---

# Autoresearch Configuration

Additional context or notes for this project's autoresearch setup.

• enabled — whether autoresearch is active (default: true)
• max_iterations — stop after N experiments (default: 0 = unlimited)
• working_dir — override directory for experiment files (default: current directory)
• benchmark_timeout — benchmark timeout in seconds (default: 600)
• checks_timeout — correctness checks timeout in seconds (default: 300)

If the file doesn't exist, use defaults. The file should be added to .gitignore (.claude/*.local.md).

Then execute these setup steps:

1. Create a branch: git checkout -b autoresearch/<goal>-<date>
2. Gitignore the living session files — the ones that change every run. This is critical for two reasons: git revert will fail if autoresearch.jsonl is tracked, and if autoresearch.md (which you update mid-loop) is tracked, a broad experiment commit could sweep it in and a later revert would erase your learnings:

   printf '%s\n' autoresearch.jsonl autoresearch.md autoresearch.ideas.md run.log >> .gitignore
   git add .gitignore && git commit -m "autoresearch: add session files to gitignore"

1. Read all files in scope thoroughly to understand the codebase
2. Write autoresearch.md — the session document (see examples/autoresearch.md). It is gitignored (living state), so it is never committed or reverted.
3. Write autoresearch.sh — the benchmark script (see examples/autoresearch.sh)
4. Optionally write autoresearch.checks.sh — correctness checks (tests, lint, types)
5. Commit only the immutable harness — autoresearch.sh and (if present) autoresearch.checks.sh. These never change after setup (the file-protection hook enforces it), so keeping them in history is safe and aids reproducibility. Do not commit autoresearch.md/.jsonl/.ideas.md (gitignored above).
6. Run baseline: bash autoresearch.sh
7. Parse metrics from output (lines matching METRIC name=value)
8. Record baseline in autoresearch.jsonl: write the "type":"config" header, then a {"type":"status","state":"running",...} marker, then the baseline result
9. Begin the experiment loop

The Experiment Loop

LOOP FOREVER. Never ask "should I continue?" — just keep going.

The user might be asleep, away from the computer, or expects you to work indefinitely. If each experiment takes ~5 minutes, you can run ~12/hour, ~100 overnight. The loop runs until the user interrupts you, period.

Each iteration:

0. Check stopping conditions: run `bash ${CLAUDE_SKILL_DIR}/scripts/session-status.sh --file autoresearch.jsonl [--max-iterations N]` and act on its VERDICT (concluded|max_iterations|wall|backstop|continue). Only continue if VERDICT=continue. (See "Stopping Conditions" for what the script computes and the spec it implements.)
1. Read current git state and autoresearch.md
2. Choose an experimental change (informed by past results and ASI notes)
3. Edit files in scope
4. git add <files-in-scope only> && git commit -m "experiment: <description>"
   (Commit ONLY the files in scope — never autoresearch.md/.jsonl/.ideas.md; they are gitignored so a revert can't erase them.)
5. Run: bash autoresearch.sh > run.log 2>&1
6. Parse METRIC lines from output
7. If autoresearch.checks.sh exists, run it (separate timeout, default 300s)
8. Decide: keep or discard
9. Log result to autoresearch.jsonl (include ASI annotations)
10. If discard/crash: git revert $(git rev-parse HEAD) --no-edit
11. Update autoresearch.md with learnings (every few experiments)
12. Repeat

Decision Rules

• Metric improved → keep (commit stays, branch advances)
• Metric equal or worse → discard (run git revert $(git rev-parse HEAD) --no-edit)
• Crash or checks failed → discard (revert, note the failure in ASI)
• Simpler code for equal perf → keep (removing complexity is a win)
• Catastrophic secondary metric regression → consider discard even if primary improved (e.g., 1% speed gain but 10x memory usage)
• If stuck → think deeper, try a different approach. Consult autoresearch.ideas.md if it exists. Re-read source files for new angles. Try combining previous near-misses. Try more radical changes. Read any papers or docs referenced in the code.

Simplicity Criterion

All else being equal, simpler is better. Weigh complexity cost against improvement magnitude:

• A 0.001 improvement that adds 20 lines of hacky code? Probably not worth it.
• A 0.001 improvement from deleting code? Definitely keep.
• Equal performance with much simpler code? Keep.

Handling User Messages During Experiments

If the user sends a message while the loop is running:

1. Finish the current experiment cycle (don't abandon mid-run)
2. Address the user's feedback or question
3. Resume the loop immediately after — do not wait for permission

Benchmark Timeout

• Default benchmark timeout: 600 seconds (10 minutes)
• If a run exceeds the timeout, kill it and treat as a crash
• Checks timeout: 300 seconds (5 minutes), separate from benchmark

Don't Thrash

The consecutive-failure streak is the key signal, and it is defined so it can be recomputed from autoresearch.jsonl alone (never from memory — you may have been context-reset):

Starting from the last line of autoresearch.jsonl, count backward the run entries whose status is discard or crash. Stop counting (streak resets to 0) at the first keep, the first checks_failed, or the first run belonging to an earlier `segment`.

• Streak ≥ 3 → rethink (soft). Stop grinding on the current angle:

1. Think about why the last few failed.
2. Re-read the source files for new angles.
3. Try a fundamentally different approach.
4. Consult autoresearch.ideas.md for untried ideas.

• Streak > 8 → wall (terminal). The current segment is stuck. Stop the session and report it (an orchestrated run returns Status: WALL). Starting a genuinely different direction usually means a new segment, which resets the streak.

The > 8 boundary is deliberately high so a fruitful overnight loop with the occasional rough patch is never aborted — only a truly stuck segment trips it.

Stopping Conditions

Compute these with bash ${CLAUDE_SKILL_DIR}/scripts/session-status.sh --file autoresearch.jsonl [--max-iterations N] at the top of every iteration (loop step 0) — it emits SEGMENT, SEGMENT_RUNS, STREAK, LAST_STATUS, BACKSTOP_ACK, and a single VERDICT, all recomputed from `autoresearch.jsonl` and scoped to the current segment (delimited by config headers), so they survive context resets. The rules the script implements — and which you follow directly if the script is unavailable:

• Current-segment run count = the number of run entries (status in keep|discard|crash|checks_failed) whose segment equals the current segment. Count the matching entries directly — do not use the run number for this, since run is a session-wide sequential counter that keeps climbing across segments (segment 1 might start at run 51), so it would over-count a fresh segment.
• `max_iterations` reached — if max_iterations > 0 (from .claude/autoresearch-ai-plugin.local.md) and the current-segment run count ≥ it → stop (Status: DONE (max_iterations)). 0 or absent = unlimited.
• Unbounded-loop backstop (check-in, not a stop) — when max_iterations is 0/unlimited, pause for a human check-in every 200 current-segment runs (at counts 200, 400, 600, …). On reaching a boundary B (a multiple of 200) for which the log holds no acknowledgement, append a non-terminal marker recording it — {"type":"status","state":"running","backstop_ack":B,...} — and stop this dispatch to ask the user before continuing. Two properties make the continuation coherent: the status stays running (so a confirmed relaunch is not blocked by the resume guard), and the backstop_ack:B record is durable in the log — so the next dispatch sees B is already acknowledged, does not re-pause at it, and runs on to the next un-acknowledged boundary. The pause rule is precisely: pause at boundary `B` only if no status record already carries `backstop_ack` ≥ `B`. An explicit max_iterations replaces this backstop with a hard stop.
• Consecutive-failure wall — streak > 8 (see Don't Thrash) → stop (Status: WALL).
• Session already concluded — if the last {"type":"status",...} record in the log is not running (e.g. cancelled/done), the session is over; do not silently resume it.

Metric Output Format

Benchmark scripts output metrics as structured lines:

METRIC total_time=4.23
METRIC memory_mb=512
METRIC val_bpb=1.042

Parse these with the helper script at ${CLAUDE_SKILL_DIR}/scripts/parse-metrics.sh:

bash autoresearch.sh 2>&1 | bash ${CLAUDE_SKILL_DIR}/scripts/parse-metrics.sh

Secondary Metrics

Beyond the primary metric, output additional METRIC lines for tradeoff monitoring:

METRIC total_ms=4230        # primary
METRIC compile_ms=1200      # secondary — helps identify bottlenecks
METRIC memory_mb=512        # secondary — monitors resource usage
METRIC cache_hit_rate=0.85  # secondary — instrumentation data

Secondary metrics are tracked in the JSONL log and help guide future experiments, but they rarely affect keep/discard decisions (only discard if a catastrophic secondary regression accompanies a marginal primary improvement).

Output instrumentation data — phase timings, error counts, cache rates, domain-specific signals. This data guides the next iteration and helps identify where optimization effort should focus.

Actionable Side Information (ASI)

ASI is structured annotation per experiment that survives reverts. When code changes are discarded, only the description and ASI remain — making them the only structured memory of what happened.

Record ASI for every experiment:

{
  "hypothesis": "Reducing loop iterations by breaking early",
  "result": "Marginal speedup but code readability suffered",
  "next_action_hint": "Try vectorization instead of loop unrolling",
  "bottleneck": "Memory bandwidth on L2 cache misses"
}

ASI fields are free-form — use whatever keys are useful:

• hypothesis — what you expected
• result — what actually happened
• next_action_hint — guidance for the next experiment
• bottleneck — identified performance bottleneck
• error_details — crash/failure diagnostics
• Any other domain-specific observations

Logging to autoresearch.jsonl

Config Header (written once at setup)

{"type":"config","name":"Optimize unit test runtime","metricName":"total_ms","metricUnit":"ms","bestDirection":"lower"}

Status Marker (session lifecycle)

Immediately after the config header, append a status record, and append a new one whenever the session's lifecycle state changes:

{"type":"status","state":"running","timestamp":1700000000}

• state — one of: running (session active — this also covers a backstop check-in, which pauses for confirmation without concluding), done (concluded normally, e.g. max_iterations reached or the user ended an unbounded run), cancelled (user stopped it), wall (consecutive-failure wall), blocked (needs user input).
• backstop_ack (optional, integer) — on a running marker only: records the 200-run backstop boundary the loop paused at for a check-in, so a later dispatch doesn't re-pause at the same boundary. See the backstop rule under Stopping Conditions.
• Write running at setup; write done/wall/blocked when the loop stops; write cancelled in the Cancel procedure.
• Resume rule: a session is resumable only if the last status record is running (or there is no status record — legacy logs). If the last status is cancelled/done/wall, do not silently resume — a concluded session stays concluded until the user explicitly restarts it.
• This marker is a deliberate non-run record; it does not have a run number and is skipped by run-counting. It is agent-written, not runtime-enforced, so a hard crash could leave a stale running — acceptable, since the alternative (treating any log as active forever) is worse.

Experiment Results (appended after each run)

Each experiment appends one JSON line:

{"run":5,"commit":"abc1234","metric":4230,"metrics":{"compile_ms":1200,"memory_mb":512},"status":"keep","description":"parallelized test suites","timestamp":1700000000,"segment":0,"confidence":2.3,"asi":{"hypothesis":"parallel tests reduce wall time","next_action_hint":"try worker pool size tuning"}}

Fields:

• run — experiment number (1-indexed, sequential)
• commit — short git commit hash (7 chars)
• metric — primary metric value
• metrics — secondary metrics dict (optional)
• status — one of: keep, discard, crash, checks_failed
• description — brief description of what was tried
• timestamp — Unix timestamp (seconds)
• segment — session segment index (0-based, incremented when optimization target changes)
• confidence — MAD-based confidence score (null if < 3 experiments)
• asi — Actionable Side Information dict (optional, omit if empty)

Use ${CLAUDE_SKILL_DIR}/scripts/log-experiment.sh to append entries:

bash ${CLAUDE_SKILL_DIR}/scripts/log-experiment.sh \
  --run 5 \
  --commit "$(git rev-parse --short HEAD)" \
  --metric 4230 \
  --status keep \
  --description "parallelized test suites" \
  --metrics '{"compile_ms":1200,"memory_mb":512}' \
  --segment 0 \
  --confidence 2.3 \
  --asi '{"hypothesis":"parallel tests reduce wall time"}'

Valid statuses: keep, discard, crash, checks_failed

Logging Invariants

Keep the log analyzable and the run-counts/streaks trustworthy:

• Never log `keep` when correctness checks failed — the status is checks_failed regardless of how good the metric looks. (This also keeps the failure streak honest: checks_failed breaks a discard/crash streak, so mislabeling it as keep would hide a stuck segment.)
• Append-only — never rewrite or delete past run entries; the log is the reconstructible source of truth. (The {"type":"status",...} marker is the one intentional non-run append; it is expected, not a violation.)
• Secondary-metric consistency — once a secondary metric name has appeared in metrics, include it in every later run; if it is genuinely unavailable for a run, say so in ASI rather than silently dropping it.

Segments (Multi-Phase Sessions)

When the optimization target changes mid-session (different benchmark, metric, or workload):

1. Write a new config header to autoresearch.jsonl with the updated target
2. Increment the segment counter
3. Old results stay in the JSONL but are filtered as previous phase
4. Establish a new baseline for the new segment

This allows a single session to evolve — e.g., first optimize compilation speed, then switch to runtime performance.

Resuming After Context Reset

If autoresearch.jsonl exists in the working directory (the authoritative session state):

1. Read autoresearch.jsonl to see all past experiments, current best, ASI annotations, and the last `{"type":"status",...}` record. If that last status is cancelled/done/wall, the session is concluded — do not silently resume; confirm with the user first. Only resume when the last status is running (or there is no status record — a legacy log).
2. Read autoresearch.md for full context (goal, metrics, files, constraints, learnings). If it is missing (it is gitignored, so a fresh clone won't have it), reconstruct it from the JSONL config header and git log.
3. Check autoresearch.ideas.md if it exists — prune stale entries, experiment with remaining ideas
4. Check git log to verify current branch state matches expected state
5. Re-check the Stopping Conditions (max_iterations, backstop, failure streak) from the log before continuing
6. Resume the loop from where it left off — no re-setup needed
7. Resume immediately if the session is active — do not ask "should I continue?"

Confidence Scoring

After 3+ experiments, assess whether improvements are real or noise:

• Compute the Median Absolute Deviation (MAD) of all metric values in the current segment as a noise floor
• Confidence = |best improvement| / MAD
• ≥2.0× → likely real improvement (green)
• 1.0–2.0× → marginal, could be noise (yellow)
• <1.0× → within noise floor (red) — consider re-running to confirm

Record confidence on each experiment result in the JSONL log. When confidence is low, consider:

• Running the benchmark multiple times inside autoresearch.sh and reporting the median
• Pinning CPU frequency or reducing system noise
• Making larger changes that produce clearer signal

See references/confidence-scoring.md for detailed methodology.

Session Files

Cancel and Status

Cancelling an Autoresearch Session

When the user asks to cancel or stop autoresearch:

1. Finish the current experiment cycle if one is running
2. Read autoresearch.jsonl to count total experiments and results
3. Report a summary: goal, total runs, kept improvements, best metric
4. Append a {"type":"status","state":"cancelled","timestamp":...} marker to autoresearch.jsonl — this is what makes the session stay stopped instead of being silently resumed later.
5. Do NOT delete autoresearch.jsonl or autoresearch.md — they contain valuable history
6. Do NOT delete .claude/autoresearch-ai-plugin.local.md — it is the user's persistent per-project config; the cancelled marker (not config removal) is what ends the session
7. Do NOT revert any kept commits — the improvements are real
8. Inform the user they can resume later with /run-autoresearch (which will see the cancelled marker and ask before restarting)

Checking Session Status

When the user asks about autoresearch status or progress:

1. Check if autoresearch.jsonl exists — if not, report "No active session"
2. Read autoresearch.md for the goal and primary metric
3. Parse autoresearch.jsonl to compute: total runs, kept/discarded/crashed counts, baseline vs best, improvement percentage, confidence score, and the lifecycle state — the state of the last {"type":"status",...} record (running/done/cancelled/wall/blocked)
4. Display a formatted summary, including whether the session is active (last status running) or concluded

Additional Resources

Reference Files

• `references/confidence-scoring.md` — Detailed MAD-based confidence methodology
• `references/best-practices.md` — Tips for writing good benchmarks, choosing experiments, ASI patterns, and avoiding pitfalls

Example Files

• `examples/autoresearch.md` — Example session document template
• `examples/autoresearch.sh` — Example benchmark script with METRIC output
• `examples/autoresearch.checks.sh` — Example correctness checks script

Utility Scripts

• `scripts/parse-metrics.sh` — Extract METRIC lines from benchmark output
• `scripts/log-experiment.sh` — Append an experiment result to autoresearch.jsonl
• `scripts/session-status.sh` — Compute the stopping-condition state (SEGMENT_RUNS, STREAK, LAST_STATUS, BACKSTOP_ACK, VERDICT) from autoresearch.jsonl