autoresearch-setup

name	autoresearch:setup
description	Scaffolds a new autonomous-research experiment directory (`.autoresearch/{YYMMDD}-{slug}/`) inside a deep-learning project so Claude can run a long train.py-mutation loop without blowing context. This skill should be used when the user asks to "start an autoresearch experiment", "set up autonomous research loop on this project", "create a new .autoresearch run", "scaffold autoresearch", "initialize autoresearch for this repo", "kick off an autonomous training loop", "set up Karpathy-style autoresearch here", or otherwise indicates they want Claude to begin autonomous iteration on their ML research code. The skill performs a venv preflight, analyzes the project's editable-install Python packages, surfaces primary-metric candidates from whichever tracker the host uses (wandb / tensorboard / plain stdout logs), introspects the host's training entrypoint (argparse-CLI script vs importable main() function vs hydra app), infers the distributed framework (accelerate / torchrun / FSDP / DDP / pytorch-lightning / none), detects checkpoint conventions (HF Trainer / Lightning / plain torch.save), runs a short interview, and then materializes the expr by calling `ar init` which renders the train.py / prepare.py / program.md templates.
argument-hint	["topic"]
allowed-tools	Read, Bash, Glob, Grep, AskUserQuestion, Write

autoresearch:setup

Scaffold a new .autoresearch/{YYMMDD}-{slug}/ experiment in the current project so the autoresearch loop (driven by autoresearch:run) can begin.

The skill performs four jobs, in order:

0. Preflight the venv + dependency state so ar init and ar run have a working environment (Phase 0, new in v0.2.0).
1. Analyze the project to propose intelligent defaults for the interview — including the new entry-point / distributed-framework / CLI-args probes (Phase 1).
2. Run an interview via AskUserQuestion — the only user-facing interaction in this skill (Phase 2).
3. Materialize the expr by calling ar init, bootstrapping the ar.py shim first if this is the project's first setup (Phases 3–5).

The interview is the only authorized user prompt. Do not interrupt mid-flow for confirmations, clarifications, or "are you sure?" prompts outside of it. Two exceptions: the Phase 0 remediation prompt (e.g. "symlink sibling .venv?") and the Phase 3 slug-collision prompt both use AskUserQuestion.

---

Phase 0 — venv preflight

Before any project analysis, verify the environment can actually run training. Each check is a one-line bash probe; stop and halt with actionable guidance on the first red signal.

Step 0.1 — pyproject.toml

test -f pyproject.toml && echo "ok" || echo "missing"

If missing, halt immediately:

autoresearch:setup must run from a Python project root — no pyproject.toml found in the current directory. Please cd into your project's root (the directory containing pyproject.toml) and re-invoke the skill.

Do NOT attempt to locate the project root automatically; this is almost always operator error and the correct response is to abort with a clear message.

Step 0.2 — venv + mandatory core dep

uv run --no-sync python -c "import torch" 2>&1

The --no-sync flag prevents uv from silently mutating the environment during the probe. torch is the only hard requirement — it's universal across every framework v0.3.0 supports (HF Trainer, Hydra, Lightning, plain-PyTorch, custom loops). Everything else is a soft signal.

• Exit 0, no output: proceed to step 0.3.
• Non-zero exit: diagnose. See "Diagnosis flowchart" below. When consent is needed, use AskUserQuestion.

Step 0.3 — optional-dep probe (soft signals, informational only)

Probe each in turn. Record each as present / absent for Phase 1 to consume — do NOT halt on any absence.

uv run --no-sync python -c "import accelerate"        2>/dev/null && echo "accelerate=present"        || echo "accelerate=absent"
uv run --no-sync python -c "import wandb"             2>/dev/null && echo "wandb=present"             || echo "wandb=absent"
uv run --no-sync python -c "import torch.distributed" 2>/dev/null && echo "torch_distributed=present" || echo "torch_distributed=absent"
uv run --no-sync python -c "import pytorch_lightning" 2>/dev/null && echo "lightning=present"         || echo "lightning=absent"
uv run --no-sync python -c "import lightning"         2>/dev/null && echo "lightning2=present"        || echo "lightning2=absent"
uv run --no-sync python -c "import tensorboard"       2>/dev/null && echo "tensorboard=present"       || echo "tensorboard=absent"

These signals feed Phase 1 detection; none of them block. If wandb=absent, the metric-backend interview question (Phase 2 item 6) will default to log. If accelerate=absent and lightning=absent and torch_distributed=absent, the distributed-framework question will default to none via the auto sentinel.

Step 0.4 — wandb login (soft, skipped if wandb absent)

Run only if step 0.3 reported wandb=present:

(test -f ~/.netrc && grep -q "api.wandb.ai" ~/.netrc) && echo "ok" || echo "missing"

If missing, warn once:

Note: wandb is installed but wandb login does not appear to be configured (~/.netrc lacks an api.wandb.ai entry). Training will still run, but wandb metrics will be offline-only. Run uv run wandb login any time to enable — does not block setup.

Proceed regardless — this is a soft signal. If wandb itself is absent, skip this step entirely.

Diagnosis flowchart (step 0.2 red path)

First, classify whether cwd is a git worktree:

if [ -f .git ]; then
  head -1 .git | grep -q "^gitdir: " && echo "worktree"
elif [ -d .git ]; then
  echo "main-repo"
else
  echo "not-a-git-repo"
fi

If "worktree": check siblings for a populated .venv (same project, different branch — likely has the same deps already downloaded). Candidate probe:

MAIN_WORKTREE=$(dirname "$(git rev-parse --git-common-dir)")
for cand in "$MAIN_WORKTREE" "$MAIN_WORKTREE"/../*; do
  if [ -d "$cand/.venv" ]; then
    uv run --project "$cand" --no-sync python -c "import torch" 2>/dev/null && \
      echo "candidate: $cand"
  fi
done

If a candidate is found, ask the user via AskUserQuestion with options:

• "Symlink: rm -rf .venv && ln -s <candidate>/.venv .venv" (fast — reuses downloaded torch/accelerate/flash-attn)
• "Run uv sync fresh (re-download deps)" (slow but hygienic)
• "Abort — I'll fix it myself"

Do NOT execute the symlink or sync without explicit consent.

If "main-repo" or "not-a-git-repo": don't offer the symlink path. Ask via AskUserQuestion:

• "Run uv sync to install/repair deps" (then re-run step 0.2)
• "Abort — I'll fix it myself"

After user confirms uv sync: run it, wait, re-probe. If still red after one remediation attempt, halt with:

Preflight probe still failing after uv sync. See references/venv-preflight.md for failure-mode table; fix manually and re-invoke /autoresearch:setup.

See references/venv-preflight.md for the complete failure-mode table (broken venv symlinks, CUDA mismatches, Python version drift, etc.).

Halt vs proceed

Proceed to Phase 1 only if steps 0.1 and 0.2 are both green. Steps 0.3 and 0.4 produce informational signals and soft warnings respectively — neither blocks.

---

Phase 1 — Project analysis

Goal: gather enough signal to propose strong defaults for the interview. Treat this phase as read-only reconnaissance. Target ≤ 14 reads total in this phase (up from 12 in v0.2 because the metric-backend + config-system + checkpoint-convention probes each add one read on average); prefer Glob/Grep over full-file reads. Do not read files already in session context.

Signals to collect

Read / glob these in priority order, and stop early once proposals are well-formed:

1. pyproject.toml — Read full. Extract project name, dependencies (torch / accelerate / wandb / deepspeed / pytorch-lightning / hydra-core / tensorboard / mlflow presence), package roots ([tool.setuptools.packages.find] / [project.scripts] / [tool.uv]).

2. CLAUDE.md (project-level, plus any under subpackages) — Read if present. Extract:

   • The project's own "how to run training" command block (usually under a "Commands" or "Training" heading). This is the strongest signal for both the runner and the CLI args — the exact command the user types to train locally.
   • Project description / research themes for goal candidates.

3. Metric-backend artifacts — do not read payloads; glob + stat only.

   • wandb/run-*/files/ exists? → wandb is a strong candidate. If up to 3 most recent runs exist, read wandb-summary.json (metric keys), config.yaml (the flattened training config — strongest signal for --cli-args-json — and contains the wandb project via _wandb.project), and wandb-metadata.json (program, args, surfaces the exact entry script and argv).
   • events.out.tfevents.* files anywhere under the project (common roots: runs/, tb_logs/, tensorboard/, lightning_logs/*/, outputs/*/tensorboard/) → tensorboard is a strong candidate. Record the glob that captures them; it feeds --tb-events-glob.
   • mlruns/ directory → mlflow — mark as custom for v0.3.0 and emit a TODO: "native mlflow backend planned for v0.4.0; a Python --metric-extract-code snippet is required until then".
   • None of the above → log (stdout-parsing fallback). Rank: wandb ≻ tensorboard ≻ custom (mlflow) ≻ log. Lowest-confidence detection (only one run dir, or only a stale events.out.tfevents.*) → mark detection confidence low so Phase 2 item 6 pushes harder for the user to confirm.

4. Entry-point training scripts — Glob training/**/*.py, train.py, scripts/train*.py, tools/train*.py, src/**/train*.py. For each, check:

   • if __name__ == "__main__": block present? → candidate for argparse-cli pattern.
   • Contains argparse.ArgumentParser( or HfArgumentParser( or simple_parsing? → strongly argparse-cli. Record the dotted-module path (e.g. myproject.training.main).
   • Defines a top-level def main(...) with typed kwargs and is imported elsewhere? → candidate for function pattern.
   • Contains @hydra.main(...) decorator? → hydra pattern (v0.3.0-native entry pattern; see signal 5 below for config-system confirmation).
   • Neither clearly → custom pattern (agent will wire it manually in train.py).
   • Rank candidates by: (a) Hydra decorator wins over (b) presence of ArgumentParser + __main__ block wins over (c) main() function; ties broken by which file's path appears in CLAUDE.md's training command.

5. Config system detection (new in v0.3.0) — grep the entrypoint and its imports:

   • @hydra.main / from hydra / hydra.utils / from omegaconf import OmegaConf → hydra. Feeds --entry-pattern hydra.
   • LightningCLI( → lightning-cli. Not natively supported in v0.3.0 — record as custom with a TODO pointer to v0.4.0.
   • argparse.ArgumentParser( / HfArgumentParser( / simple_parsing → argparse-cli.
   • from fire import Fire / fire.Fire( → fire. Not natively supported in v0.3.0 — record as custom with a TODO pointer to v0.4.0.
   • None of the above but a top-level main() is importable → function.
   • Otherwise → custom.

6. Distributed framework — grep the training entrypoint (and nearby modules) for:

   • import pytorch_lightning / from pytorch_lightning / import lightning / from lightning.pytorch / pl.Trainer( / L.Trainer( → lightning (v0.3.0 recognizes this as a distinct framework; the launcher delegates to Lightning's own strategy resolver rather than wrapping with accelerate).
   • from accelerate / Accelerator( → accelerate
   • import deepspeed / deepspeed.initialize / deepspeed_config → deepspeed
   • FullyShardedDataParallel / FSDP( / torch.distributed.fsdp → fsdp
   • torch.distributed.init_process_group / DistributedDataParallel without the above → ddp
   • Nothing dist-related → none
   • Cross-check against Phase 0 step 0.3 signals: if grep says accelerate but accelerate=absent, downgrade to a lower-confidence candidate (the user may have inadvertently removed the dep; surface this tension in the interview).
   • If detection is ambiguous or multiple candidates tie, emit auto — the new v0.3.0 default that tells ar run to introspect the runner command at launch time.

7. Resume flag detection — grep the training entrypoint for argparse definitions matching add_argument\([\"']--?resume or similar. Common flag names in order of likelihood: resume_from_checkpoint, resume, resume_checkpoint, ckpt_path, pretrained_checkpoint. Record the detected flag without the -- prefix. If multiple candidates or none, mark as "unclear — ask user". For Hydra entrypoints, the equivalent is a config key (e.g. trainer.resume_from_checkpoint) rather than a flag — note that and let Phase 2 capture it as a Hydra override.

8. Checkpoint convention detection (expanded in v0.3.0) — grep the entrypoint and nearby modules for save/load patterns, then propose a --checkpoint-glob matching where files actually land:

   • trainer.save_model(, save_pretrained(, output_dir/checkpoint-*/ (HF Trainer convention) → suggest output_dir/checkpoint-*/ or the detected output_dir-relative equivalent.
   • *.ckpt references, ModelCheckpoint(, lightning_logs/ on disk → Lightning → suggest lightning_logs/*/checkpoints/*.ckpt (or outputs/*/checkpoints/best.pt if Hydra + Lightning).
   • accelerator.save_state( → accelerate → suggest checkpoints/*/ or <save_dir>/*/.
   • torch.save( with explicit path patterns → extract the path format from the literal and suggest checkpoints/*.pt or the project-specific equivalent.
   • Hydra projects with default outputs/ cwd + safetensors/torch saves → suggest outputs/*/model.safetensors or outputs/*/checkpoints/best.pt.
   • Multiple candidates → rank by filesystem evidence (a glob that matches existing files wins over a purely grep-derived guess). Offer the top 2-3 in the interview.

9. accelerate / deepspeed / Hydra configs — Glob configs/accelerate*.yaml, accelerate_config*.yaml, **/accelerate/*.yaml, conf/**/*.yaml, configs/**/*.yaml. If found, note exact filenames — used verbatim in the runner spec and, for Hydra, as the config search path.

10. Recent git commits — git log --oneline -n 20 if inside a git repo. Use commit messages as additional signal for goal candidates.

What to derive

From the above signals, derive the proposal lists that feed the interview:

Proposal	Source	Output
Goal candidates (3–4)	CLAUDE.md + recent commits + top-level module names	Short phrases.
Primary-metric candidates	wandb-summary.json keys, tensorboard scalar tags, or grepped stdout regex anchors	Ordered list with inferred direction.
Mutation-scope candidates	Training entrypoint imports + obvious module roots	Dotted paths.
Runner inference	CLAUDE.md + accelerate configs + entrypoint	accelerate launch --config_file ..., torchrun --nproc-per-node N, python, custom.
Entry pattern	Signals 4 + 5	argparse-cli / function / hydra / custom.
Entry main module	Signal 4	Dotted path, e.g. myproject.training.main. Only required if pattern is argparse-cli or hydra.
CLI args baseline	wandb config.yaml + CLAUDE.md command block	Flat JSON-like dict of scalar flag values (e.g. {"config":"base","learning_rate":1.5e-4}) for argparse-cli, or a list of Hydra override strings (e.g. ["optimizer.lr=3e-4","model.hidden_dim=512"]) for hydra. Pick the top 5–10 most-distinctive keys; skip obviously-cosmetic ones (log_level, wandb_mode).
Metric backend (new in v0.3.0)	Signal 3	wandb / tensorboard / log / custom / auto.
TensorBoard events glob (new in v0.3.0)	Signal 3	Glob string, e.g. runs/*/events.out.tfevents.*. Only required if metric backend is tensorboard.
Custom metric snippet (new in v0.3.0)	Phase 2 follow-up AskUserQuestion	Short Python body that reads run_dir / run_log_text and returns a dict[str, float]. Required if metric backend is custom.
wandb project	wandb config.yaml (_wandb.project)	Project name string. Only relevant if metric backend is wandb.
Distributed framework	Signal 6	accelerate / deepspeed / fsdp / ddp / lightning / none / auto.
Resume flag name	Signal 7	String without -- prefix, or null. For Hydra, capture as an override key instead (e.g. trainer.resume_from_checkpoint).
Checkpoint glob (new in v0.3.0)	Signal 8	Glob string relative to project cwd (e.g. output_dir/checkpoint-*/, lightning_logs/*/checkpoints/*.ckpt, outputs/*/model.safetensors) or null (host doesn't save to a conventional location).

prepare.py mode selection (thin-wrapper vs full-prep) stays unchanged — ar init handles it based on presence of build_train_loader / data-module imports.

Read budget

Soft cap: 14 reads across this phase. If probes 4–6 produce enough signal on the first training script read, skip additional entry-point candidates. If signal 3 finds no tracker artifacts at all, fall back to log + stdout-regex scan (the user confirms the metric name in the interview).

Do not read the same file twice. Do not read .pt / .safetensors / .ckpt / wandb/run-*/logs/ / events.out.tfevents.* binary payloads — large and worthless (globs + stat are enough).

See references/project-analysis.md for extended grep patterns if any of {goal, metric, scope, runner, entry-pattern, metric-backend, checkpoint-glob} produced fewer than 2 candidates.

---

Phase 2 — Interview

One AskUserQuestion call with the items below, in this exact order. Use Phase 1 detections to pre-populate options; always include a free-text escape hatch. Present detected values with a preview so the user can confirm or override.

Items

1. Entry pattern (entry_pattern)

• question: "What shape is your training entrypoint?" — include a preview line showing the detected main module, e.g. Detected: myproject.training.main (argparse.ArgumentParser present + __main__ block).
• options:
  • "argparse-cli — runpy-invoke an if name == 'main': script" (pre-selected if Phase 1 detected ArgumentParser + __main__)
  • "hydra — invoke a @hydra.main-decorated entrypoint with key=value overrides" (pre-selected if Phase 1 detected @hydra.main / hydra deps)
  • "function — call an importable main(**kwargs)"
  • "custom — I'll wire it manually in train.py"
• Detected main module feeds --entry-main-module for argparse-cli / hydra / function. If detection was ambiguous, ask as a follow-up free-text.

2. Baseline CLI args / Hydra overrides (cli_args)

• Skip entirely if user picked function or custom in item 1.
• question: "These are the baseline overrides the agent will start from (edit CLI_OVERRIDES in train.py per run). Keep as-is, or edit?" — show the detected dict/list as preview. For argparse-cli:

  {"config": "base", "learning_rate": 1.5e-4, "num_train_epochs": 1}
  

  For hydra:

  ["optimizer.lr=3e-4", "model.hidden_dim=512", "trainer.max_epochs=1"]
  

• options:
  • "Keep as-is" (pre-selected)
  • "Edit — I'll paste a JSON value" (free-text; accept either a dict for argparse-cli or a list of strings for hydra)
  • "Empty — start with {} / []" (useful if agent should rebuild from scratch)
• Record as a JSON string; becomes --cli-args-json.

3. Resume flag / override key (resume_flag_name)

• Skip entirely if user picked function or custom in item 1.
• question: "Which flag (argparse) or override key (hydra) resumes from a checkpoint?" — with detected preview if Phase 1 found one.
• options (argparse-cli): detected flag first (pre-selected, e.g. resume_from_checkpoint (detected)), then common alternatives resume, ckpt_path, pretrained_checkpoint, "None — no resume flag (use AR_RESUME_CKPT env var instead)", "Other (type)".
• options (hydra): detected override first (e.g. trainer.resume_from_checkpoint), common alternatives trainer.ckpt_path, resume_from, "None — use AR_RESUME_CKPT env var", "Other (type)".
• Record name without -- prefix. Becomes --resume-flag-name (argparse) or --resume-override-key (hydra — same underlying template field, surfaced as an override).

4. Distributed framework (distributed_framework)

• question: "Which distributed-training framework does the host use?"
• options: Phase 1 detection first (pre-selected), then accelerate, deepspeed, fsdp, ddp, lightning, none, auto (let ar run introspect the runner command at launch time — new v0.3.0 default when Phase 1 detection is ambiguous).
• Becomes --distributed-framework.

5. Research goal (research_goal)

• question: "What should this experiment optimize for?"
• options: 3–4 candidate themes from Phase 1, plus "Other (describe)". Free text accepted.
• Becomes --goal. Kebab form seeds the slug if $ARGUMENTS was not provided.

6. Metric backend (metric_backend) — NEW in v0.3.0

• question: "How does the host training run report metrics the agent should read?"
• options: detected primary first (pre-selected), then the alternatives:
  • wandb — read wandb/run-*/files/wandb-summary.json`` (pre-selected if Phase 1 found a wandb/ dir)
  • tensorboard — read events.out.tfevents.*files viatensorboard.backend.event_processing`` (pre-selected if Phase 1 found event files)
  • log — parse training's stdout/stderr (captured by arintorun.log) with a user-supplied regex
  • custom — paste a Python snippet that reads run_dirand returns adict[str, float]`` (forces follow-up AskUserQuestion below)
  • auto — let ar run try wandb → tensorboard → log in that order on each run
• Pushy branch: if Phase 1 detection confidence is low (e.g. a single stale tfevents file and no wandb runs), phrase the question more assertively: include a preview line like Low-confidence detection — please confirm; defaulting to log may be safer than auto.
• Becomes --metric-backend.
• Follow-up AskUserQuestion fires only if custom was selected. Question: "Paste a short Python body (indent-free). It receives run_dir: Path and run_log_text: str; it must return dict[str, float]. Example: return {\"val_loss\": float(re.search(r'val_loss=([0-9.]+)', run_log_text).group(1))}". Free-text only. Becomes --metric-extract-code (passed as a base64-encoded argument; ar init writes it into program.md and train.py).
• If tensorboard was selected, the follow-up is a free-text AskUserQuestion for the events glob ("Which glob matches the event files you want to read?"), pre-populated with Phase 1's suggestion. Becomes --tb-events-glob.

7. Checkpoint glob (checkpoint_glob) — NEW in v0.3.0

• question: "Where does training save checkpoints? The AR-SAVE block uses this glob to discover new files after each run and promote the best one."
• options: Phase 1-detected candidates first (pre-selected), plus common alternatives:
  • output_dir/checkpoint-*/ (HF Trainer)
  • lightning_logs/*/checkpoints/*.ckpt (Lightning defaults)
  • outputs/*/checkpoints/best.pt (Hydra + Lightning composed)
  • checkpoints/*.pt (plain torch.save)
  • accelerate_state/*/ (accelerate.save_state)
  • "skip — host doesn't save to a conventional location" (AR-SAVE falls back to its priority-1 torch-state snapshot path; agent may need to monkey-patch a save call from train.py)
  • "Other (glob)"
• Becomes --checkpoint-glob. Drives the wrapper-mode AR-SAVE priority-0 discovery path: when set, AR-SAVE expands the glob after each run, picks the newest matching file/dir by mtime, and copies into runs/{id}/state.pt (or state/ if a directory). When skip, AR-SAVE uses the legacy in-process torch-state capture.

8. Primary metric (primary_metric)

• question: "Which metric is the single source of truth for 'did this run improve'?"
• options: top metric candidates from Phase 1 with inferred direction, e.g. val/loss (min), eval/top1 (max). Include "Other (type metric=direction)".
• Becomes --primary-metric + --primary-direction.

9. Hard constraints (hard_constraints)

• question: "Any hard constraints? (Violations mark the run invalid regardless of primary.)"
• options: "None" (default), plus 2–3 common candidates if auxiliary metrics look like resource bounds (peak_vram_mb, tokens_per_sec). Multi-select + free-text.
• Parse each as {name, op, threshold}; operators <=, <, >=, >.

10. Mutation scope (mutation_scope)

• question: "Which modules/classes are in-scope for train.py to monkey-patch?"
• options: dotted-path candidates from Phase 1; multi-select + free-text. Include "(advisory — will skip)".

11. Default run duration (seconds)

• question: "Default per-run time budget?"
• options: 300, 600, 900, 1800, "Custom".

12. Runner (runner)

• question: "How should ar launch training?"
• options: Phase 1's inferred runner first (pre-selected), then alternatives, then "Custom". Examples:
  • accelerate launch --config_file configs/accelerate_8gpu.yaml
  • torchrun --nproc-per-node 8
  • python -m <your_project>.training.main (Hydra apps often do this)
  • python
  • custom
• Stored verbatim.

13. wandb project (wandb_project)

• Skip entirely if metric backend (item 6) is not wandb or auto.
• question: "Which wandb project should runs log to?" — with detected preview if Phase 1 found one.
• options: detected value (pre-selected), then "Other (type)". Optional — accept empty.
• Becomes --wandb-project.

Structuring the AskUserQuestion call

The 13 items go into one AskUserQuestion invocation (the tool supports multiple items per call). Items 2, 3, and 13 are conditionally relevant based on items 1 and 6. The skill may split into two rounds — first item 1 alone (to branch on entry pattern), then items 2–13 with the CLI-args / resume-flag / wandb-project items included or omitted accordingly. Both approaches are acceptable; the one-call variant is preferred when Phase 1 detected a high-confidence entry pattern + metric backend.

The custom metric-backend path and the tensorboard events-glob capture are follow-up AskUserQuestion calls by design — each needs free-text input that doesn't fit the main interview's option/multi-select schema. These follow-ups are exempt from the "three permitted prompts" rule.

Why these items

Items 1–3 + 12 exist because a v0.1-style flow would generate a TODO-stub train.py that crashes on first ar run. Capturing the entry pattern, baseline overrides, resume anchor, and runner lets ar init render a working wrapper-mode train.py that actually executes on iteration 1.

Items 4 + 6 + 7 are the v0.3.0 framework-agnostic additions: they decouple the rendered templates from the HF-Trainer-plus-wandb assumption baked into earlier versions, letting any mainstream pipeline (HF, Hydra, Lightning, plain PyTorch, DeepSpeed, FSDP) drop in with the right metric probe and checkpoint discovery wired up.

Items 5 + 8–11 + 13 are the minimum program-spec items needed to stamp program.md with enough context for the agent's loop.

See references/interview-questions.md for verbatim option copy and direction-inference heuristics.

---

Phase 3 — Slug and collision handling

Slug generation

Format: {YYMMDD}-{kebab}, lowercase, ASCII only.

• YYMMDD = today's date in local time.
• kebab = kebab-case of either:
  • the [topic] positional argument ($ARGUMENTS), if provided; or
  • the chosen goal phrase from interview item 5, truncated to 4–5 meaningful words.

Strip articles, stopwords, punctuation. Collapse whitespace to -. Max kebab length: 40 chars.

Collision handling

If .autoresearch/{slug}/ already exists, AskUserQuestion with: "Reuse existing / New slug with suffix / Abort". This is the third and final permitted mid-flow prompt (after Phase 0 remediation and Phase 2 interview).

---

Phase 4 — Bootstrap ar.py shim (first-time-only)

Before calling ar init, check whether .autoresearch/ar.py exists. If it does, skip this phase. Otherwise:

mkdir -p .autoresearch

Write the shim (unchanged from v0.1) to .autoresearch/ar.py via the Write tool:

#!/usr/bin/env python
"""Launcher shim for the autoresearch plugin.

Resolves the plugin's lib/ directory, injects it on sys.path, and delegates
to ar.cli.main(). Runs under the target project's uv-managed venv so torch,
and any framework the host uses (accelerate / lightning / hydra / etc.), are
all resolvable.

Deps `jinja2` and `psutil` are not part of typical ML project dependencies, so
if they're missing in the active interpreter, this shim re-exec's itself via
`uv run --with jinja2 --with psutil python <self> <args>`. The re-exec is
self-suppressing (AR_SHIM_BOOTSTRAPPED sentinel) to avoid infinite loops.
"""
from __future__ import annotations

import os
import sys
from pathlib import Path


def _bootstrap_deps_if_needed() -> None:
    if os.environ.get("AR_SHIM_BOOTSTRAPPED") == "1":
        return
    missing: list[str] = []
    try:
        import jinja2  # noqa: F401
    except ImportError:
        missing.append("jinja2")
    try:
        import psutil  # noqa: F401
    except ImportError:
        missing.append("psutil")
    if not missing:
        return
    with_flags: list[str] = []
    for pkg in missing:
        with_flags.extend(["--with", pkg])
    new_env = dict(os.environ)
    new_env["AR_SHIM_BOOTSTRAPPED"] = "1"
    cmd = ["uv", "run", *with_flags, "python", __file__, *sys.argv[1:]]
    os.execvpe("uv", cmd, new_env)


def _resolve_plugin_lib() -> Path:
    override = os.environ.get("AR_PLUGIN_ROOT")
    if override:
        lib = Path(override).expanduser().resolve() / "lib"
        if (lib / "ar" / "cli.py").exists():
            return lib
        raise SystemExit(f"AR_PLUGIN_ROOT set but {lib}/ar/cli.py not found")

    home = Path.home() / ".claude" / "plugins"
    candidates = sorted(
        home.rglob("autoresearch/plugin/lib/ar/cli.py"),
        key=lambda p: p.stat().st_mtime,
        reverse=True,
    )
    if candidates:
        return candidates[0].parent.parent
    raise SystemExit(
        "autoresearch plugin not found; set AR_PLUGIN_ROOT or install the plugin"
    )


_bootstrap_deps_if_needed()
sys.path.insert(0, str(_resolve_plugin_lib()))
from ar.cli import main  # noqa: E402

if __name__ == "__main__":
    main()

Why this design

• AR_PLUGIN_ROOT is the local-dev escape hatch: export it when iterating on the plugin source itself.
• The rglob fallback tolerates the nested ~/.claude/plugins/<cache>/<marketplace>/<plugin-name>/ layout without hardcoding paths.

---

Phase 5 — Call ar init

With the shim in place, materialize the expr. Build the command with the flags derived from the interview:

uv run python .autoresearch/ar.py init \
  --expr {slug} \
  --goal "<from interview 5>" \
  --primary-metric "<from interview 8 name>" \
  --primary-direction <min|max> \
  --runner "<from interview 12 verbatim>" \
  --seconds <from interview 11> \
  [--constraints "<name> <op> <threshold>" ...] \
  [--mutation-scope "<comma,separated,dotted.paths>"] \
  [--entry-pattern <argparse-cli|function|hydra|custom>] \
  [--entry-main-module <dotted.path>] \
  [--cli-args-json '<JSON value>'] \
  [--distributed-framework <accelerate|deepspeed|fsdp|ddp|lightning|none|auto>] \
  [--resume-flag-name <flag_or_override_key_without_leading_dashes>] \
  [--metric-backend <wandb|tensorboard|log|custom|auto>] \
  [--tb-events-glob '<glob>'] \
  [--metric-extract-code '<base64-encoded python body>'] \
  [--checkpoint-glob '<glob>'] \
  [--wandb-project <name>]

Flag semantics (must match ar init exactly)

Core flags:

• --expr — the generated slug. Required.
• --goal — free-form goal string. Required. Quote it.
• --primary-metric — metric key verbatim (e.g. val/loss, val_loss, eval/top1). Required.
• --primary-direction — literal min or max. Required.
• --runner — full invocation string. Required.
• --seconds — integer seconds. Required.
• --constraints — repeatable; each value is "name op value". Optional.
• --mutation-scope — single flag, comma-separated dotted paths. Optional.

Entry-point flags:

• --entry-pattern — argparse-cli / function / hydra / custom. Default is equivalent to custom. Set to argparse-cli or hydra to render the wrapper-mode train_wrapper.py.jinja; Hydra mode additionally formats CLI_OVERRIDES as a list of override strings.
• --entry-main-module — dotted path for argparse-cli / hydra / function modes, e.g. myproject.training.main. Required when --entry-pattern is argparse-cli or hydra; ar init rejects the call otherwise.
• --cli-args-json — single JSON value string. Dict for argparse-cli (becomes CLI_OVERRIDES = {...}), list-of-strings for hydra (becomes CLI_OVERRIDES = [...]). Default {}.
• --resume-flag-name — string without -- prefix (e.g. resume_from_checkpoint) for argparse-cli, or a Hydra override key (e.g. trainer.resume_from_checkpoint) for hydra. If unset, the wrapper exports AR_RESUME_CKPT env var but doesn't inject a flag/override.
• --distributed-framework — one of accelerate / deepspeed / fsdp / ddp / lightning / none / auto. Default auto (new v0.3.0 default — ar run introspects the --runner command at launch time).

Metric + checkpoint flags (v0.3.0):

• --metric-backend — wandb / tensorboard / log / custom / auto. Default auto. Drives how ar run extracts the primary metric after each run:
  • wandb → read wandb/run-*/files/wandb-summary.json via pointer file (v0.2.0 behavior).
  • tensorboard → expand --tb-events-glob and read the last scalar event for --primary-metric.
  • log → regex-scan run.log for a user-supplied pattern (default: {primary_metric}=<float>).
  • custom → eval the body from --metric-extract-code with run_dir and run_log_text in scope.
  • auto → try wandb → tensorboard → log in order; first non-empty result wins.
• --tb-events-glob — glob relative to project cwd, e.g. runs/*/events.out.tfevents.* or lightning_logs/*/events.out.tfevents.*. Required when --metric-backend=tensorboard.
• --metric-extract-code — base64-encoded Python body. Required when --metric-backend=custom; ignored otherwise. ar init decodes and embeds into program.md + train.py so both are self-contained.
• --checkpoint-glob — glob relative to project cwd, e.g. output_dir/checkpoint-*/, lightning_logs/*/checkpoints/*.ckpt, outputs/*/model.safetensors. Drives AR-SAVE's priority-0 path: the wrapper expands this glob after the training subprocess exits, picks the newest match, and copies into runs/{id}/state.pt (or state/ for a directory). If unset, AR-SAVE uses the priority-1 in-process torch-state capture.
• --wandb-project — project name string. Only relevant when metric backend is wandb or auto. Becomes the wandb.init(project=...) value in wrapper mode.

Do NOT pass --parent-ckpt — that flag is reserved for chain-init.

What ar init does

Source of truth for:

• Creating .autoresearch/{slug}/ with program.md, prepare.py, train.py (or train_wrapper.py body if wrapper mode), empty runs/, results.tsv with header, and .ar-session.json.
• Selecting train.py template by --entry-pattern:
  • argparse-cli → train_wrapper.py.jinja in argparse mode (uses runpy.run_module with synthesized sys.argv).
  • hydra → train_wrapper.py.jinja in hydra mode (invokes the entry module with a list of key=value overrides).
  • function / custom → legacy train.py.jinja (function-mode or pure-agent-wiring entry).
• Stamping program.md with the "Entry point" section (pattern, main module, metric backend, checkpoint glob, distributed framework, resume flag, baseline overrides).
• Picking prepare.py mode (thin-wrapper vs full-prep).

The skill does NOT render any templates itself. All file materialization is ar init's job.

Error handling

• Non-zero exit from ar init: print the tail of its stderr (10–20 lines) and stop. Common causes: missing pyproject.toml, broken venv, --entry-pattern=argparse-cli or hydra without --entry-main-module, --metric-backend=tensorboard without --tb-events-glob, --metric-backend=custom without --metric-extract-code. Do not retry automatically.
• If uv is not installed: print install uv: https://docs.astral.sh/uv/ and stop.

---

Phase 6 — Completion

On successful ar init, print exactly one next-step line:

Next: /autoresearch:run {slug}

Nothing else. Do not summarize interview answers or list created files — the user knows what they asked for. Users can inspect .autoresearch/{slug}/program.md (and the new "Entry point" section there) for the full record.

---

Rules of engagement

• Three permitted prompts only. Phase 0 remediation (if probe red), Phase 2 interview, Phase 3 collision. No other AskUserQuestion calls — with two exempted micro-followups: (a) when Phase 2 item 6 chooses tensorboard, one free-text prompt for the events glob; (b) when item 6 chooses custom, one free-text prompt for the Python extraction snippet. Both are mechanical captures of free-text input that the main interview's option schema cannot carry.
• Read budget applies to Phase 1 only. Once the interview starts, no further project reads unless ar init fails.
• Write budget: exactly one file via the Write tool. That file is .autoresearch/ar.py, only if absent. Everything else is created by ar init via Bash.
• No git operations. Setup does not commit, branch, or stash.
• Honor $ARGUMENTS. If invoked with a topic argument, use it as the kebab seed and as additional context for goal candidates — but still run the full interview.
• Phase 0 cannot be skipped. Even if the user has run setup before in the same session, re-run the preflight — .venv state can change between invocations.

---

Resources

• references/venv-preflight.md — Complete failure-mode table + detection recipes for Phase 0. Read when step 0.2 red path needs diagnosis beyond the inline flowchart.
• references/project-analysis.md — Extended heuristics for Phase 1 on unusual project layouts. Read if any of {goal, metric, scope, runner, entry-pattern} produced fewer than 2 strong candidates.
• references/interview-questions.md — Verbatim AskUserQuestion question/option copy + direction-inference table. Read before constructing the interview call if uncertain about wording.