Skip to content

Prepare initial RSI framework research release - #1

Merged
satsuki2486441738 merged 9 commits into
mainfrom
codex/open-source-readiness
Nov 9, 2026
Merged

satsuki2486441738 merged 9 commits into
mainfrom
codex/open-source-readiness

Conversation

@satsuki2486441738

@satsuki2486441738 satsuki2486441738 commented Nov 8, 2026 •

Copy link
Copy Markdown
Collaborator

Prepare RSISys as a 0.1 research framework around registered RSI controllers and Labs: controller → Lab execution → feedback → controller. External Python plugins and independent JSON configurations reuse the common runtime for authorization, routing, receipts, provenance and owned resource cleanup. Harnesses, memory, verifiers and training engines remain replaceable components inside those two modules.

The branch includes ReST^EM and RRSI implementations, SEAL reference and paper-profile training paths, TTRL reference GRPO and the pinned native ttrl_verl backend, plus optional FT-Agent and AIDE controllers. FT-Agent is a partial native-policy/local-fine-tuning adaptation; AIDE preserves the native Agent/search/review/Journal while adapting the task to fixed-model inference programs. docs/methods.md records fidelity, measured scope, deviations and result evidence. No full-scale paper performance reproduction is claimed.

Release preparation adds Apache-2.0 licensing and third-party notices, CPU quickstarts, public architecture and registration guides, compatibility/security/contribution guidance, CI and independent core/monitor package installation checks. The public architecture is separated from default meta-agent implementation details. Historical raw requests, logs and probes are archived outside the current tree; dated conclusions and curated experimental evidence remain. This does not remove those materials from Git history.

Runtime fixes cover independent SEAL/TTRL artifact ownership, SQLite initialization cleanup, isolated virtualenv entrypoints, exact terminal feedback, and GPU-worker descendants surviving an exited launcher. Linux regression fixtures also handle procfs disappearance during cleanup and allow runtime-file hashing within the medical evaluation test budget; production timeouts and scoring assertions remain unchanged.

Validation includes 2,096-test offline suites on local macOS (80 optional/environment skips) and b200 Linux (73 skips), 37 focused follow-up cases (one optional upstream skip), 147 frontend tests, 641 maintained documentation links, fatal static checks, core/monitor sdist-to-wheel builds, strict package metadata checks and isolated offline wheel installs outside the checkout. GitHub checks on the current head cover Linux/macOS × Python 3.12/3.13 plus source/wheel packaging; their actual status is shown on this PR.

Scope remains an early research release. Native search state and synchronous GPU recipes do not gain automatic recovery merely through registration; shared GPU leasing and complete long-job budget/cost integration remain core roadmap work. Repository visibility, merging, release tagging and PyPI publication are separate actions and are not performed by this PR.

@satsuki2486441738
satsuki2486441738 marked this pull request as ready for review November 9, 2026 05:53
@satsuki2486441738
satsuki2486441738 merged commit c7a8038 into main Nov 9, 2026
10 checks passed
@satsuki2486441738
satsuki2486441738 deleted the codex/open-source-readiness branch November 18, 2026 09:02
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant