autoresearch-nono

name	autoresearch-nono
description	Understands how to run autoresearch autonomous ML loops inside a nono sandbox with GPU access. Use when launching autoresearch, handling attestation, or diagnosing sandbox permission failures during ML training.

autoresearch inside nono

You are running an autonomous ML research loop (autoresearch) inside a nono security sandbox. nono enforces OS-level capability restrictions using Landlock (Linux). These restrictions cascade to all child processes, including GPU training subprocesses, and cannot be bypassed from within the session.

Launching

Use launch.sh from the repo root. It verifies attestation before starting:

./launch.sh

Or manually:

~/.local/bin/nono run \
  --profile claude-code-autoresearch \
  --allow-gpu \
  --allow-cwd \
  --workdir ~/autoresearch-nono \
  -- "$(dirname "$(command -v node)")/node" \
     "$(dirname "$(command -v node)")/../lib/node_modules/@anthropic-ai/claude-code/cli.js" \
     --dangerously-skip-permissions

The node binary must be called directly (not via the claude shim) to avoid picking up the old system node inside the sandbox's clean PATH.

Attestation

Before the first run, sign the program file for your corpus:

cd workload
nono trust sign ibd/program_ibd.md --keyref "file://$HOME/.config/nono/trust-key.pem"
nono trust verify ibd/program_ibd.md

launch.sh verifies this signature at startup. If verification fails, the session will not start.

GPU access

GPU access requires two things:

1. --allow-gpu CLI flag when launching nono
2. "allow_gpu": true at the top level of the active profile (not inside the security section)

The claude-code-autoresearch profile (installed by this pack) already includes allow_gpu: true.

When an operation is denied

If a file read, write, or bash command fails with EPERM, EACCES, or "Operation not permitted":

1. Do NOT retry the same operation.
2. Run nono why to diagnose:

   nono why --path /path/that/failed --op read 2>/dev/null
   

3. Present the user with options:
   • Quick fix: restart with --allow /path/to/needed added to the nono command
   • Profile edit: add the path to ~/.config/nono/profiles/claude-code-autoresearch.json under filesystem.allow

Common ML training permission issues

• CUDA libs (/usr/local/cuda/): already in the profile as read-only
• /proc/driver/nvidia: not in the allow list by default; add with --allow /proc/driver/nvidia
• /dev/shm: already allowed
• HuggingFace cache (~/.cache/huggingface): already allowed
• uv package cache (~/.cache/uv, ~/.local/share/uv): already allowed

Checking what is allowed

cat "$NONO_CAP_FILE" | jq .

Shows all allowed filesystem paths and whether network access is blocked.

Audit log

Every session is recorded. Query after a run:

nono audit list --recent 5
nono audit show <session-id> --json
nono audit show <session-id> --json | jq '.denials'

Experiment workflow

The agent operates autonomously:

1. Read program_*.md for the active corpus
2. Edit workload/train.py to try a change
3. Commit, run training, check val_bpb in run.log
4. Keep (leave committed) or discard (git reset) based on results
5. Record in results.tsv and loop

Never pause to ask the user during a run. If blocked by a sandbox denial, diagnose and present options — do not retry silently.