Bootstrap actor-based autoresearch loop #126

Closed
lost-rob0t wants to merge 6 commits from feature/autoresearch-bootstrap into master
lost-rob0t commented 2026-08-21 22:50:29 +00:00 (Migrated from github.com)

Summary

  • add a generic metric-driven autonomous research loop using Zara's existing bounded actor model
  • run each LLM experiment in a disposable repository copy and only promote allowlisted file changes when the metric improves
  • keep evaluator stdout/stderr out of the LLM context; persist structured numeric decisions in an append-only JSONL ledger
  • expose zara-autoresearch, graceful STOP-file handling, tests, and wiki documentation

Safety / boundaries

  • evaluator commands run without a shell and with a fixed timeout
  • agent changes outside the explicit research file allowlist are rejected before evaluation
  • this is process/workspace isolation, not a hostile-code sandbox; evaluator code remains trusted local development code

Verification

  • CI must run the repository's full test gate (nix develop -c bash scripts/test-all.sh / nix flake check)
  • local full-suite execution was not available through the GitHub connector, so this PR remains draft until CI is green

Follow-ups intentionally not in this bootstrap

  • contract hash + robust resume semantics
  • noisy-metric replication/statistical admission gate
  • parallel hypothesis workers / successive halving
  • UI/status integration beyond the CLI and actor status message
## Summary - add a generic metric-driven autonomous research loop using Zara's existing bounded actor model - run each LLM experiment in a disposable repository copy and only promote allowlisted file changes when the metric improves - keep evaluator stdout/stderr out of the LLM context; persist structured numeric decisions in an append-only JSONL ledger - expose `zara-autoresearch`, graceful STOP-file handling, tests, and wiki documentation ## Safety / boundaries - evaluator commands run without a shell and with a fixed timeout - agent changes outside the explicit research file allowlist are rejected before evaluation - this is process/workspace isolation, not a hostile-code sandbox; evaluator code remains trusted local development code ## Verification - CI must run the repository's full test gate (`nix develop -c bash scripts/test-all.sh` / `nix flake check`) - local full-suite execution was not available through the GitHub connector, so this PR remains draft until CI is green ## Follow-ups intentionally not in this bootstrap - contract hash + robust resume semantics - noisy-metric replication/statistical admission gate - parallel hypothesis workers / successive halving - UI/status integration beyond the CLI and actor status message
nsaspy changed title from WIP: Bootstrap actor-based autoresearch loop to Bootstrap actor-based autoresearch loop 2026-08-28 03:31:13 +00:00
Owner

Closing as parked: the actor-based autoresearch loop is off-roadmap — no roadmap issue owns it, and RAGE queue discipline forbids entertaining unqueued work while P0 gates (#134) are open. The module is self-contained (zara/autoresearch.py + tests + wiki) and the branch is preserved; if this becomes roadmap work, open a dedicated issue and re-derive it against current master.

Closing as parked: the actor-based autoresearch loop is off-roadmap — no roadmap issue owns it, and RAGE queue discipline forbids entertaining unqueued work while P0 gates (#134) are open. The module is self-contained (zara/autoresearch.py + tests + wiki) and the branch is preserved; if this becomes roadmap work, open a dedicated issue and re-derive it against current master.
nsaspy closed this pull request 2026-09-04 14:40:03 +00:00

Pull request closed

Sign in to join this conversation.
No description provided.