Repository navigation
Repo cleanup: remove run logs and forwarding stubs, single source for pins, merge weight and lag bounds - #239
Open
michaellchung wants to merge 11 commits into
Open
michaellchung wants to merge 11 commits into
michaellchung wants to merge 11 commits into
Conversation
Untrack 236 run-log files (run-logs/, *.log, *.jsonl, *.out, *.err) under openspec/changes/*. Copies live in infra-drafts/evidence/<change>/; each old directory keeps an EVIDENCE.md pointer. Drop log re-include lines from .gitignore and add trailing log rules. Add tests/test_no_tracked_logs.py; test_fn_learning_signal reads the try27 log from the out-of-repo evidence dir. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Remove yeto/rl/engine/miles_adapter/ (28 stubs) and 17 yeto/rl/ stubs; rewrite every import, mock target, -m command and path string in yeto/, tests/, scripts/, tools/, docker/ and docs/ to the real module. yeto/rl/learner.py is now a command-line entry only: python -m yeto.rl.learner runs island_entry, importing it raises ImportError. test_import_boundaries forbids the old paths everywhere. Removing the miles_overlay stub exposed a pre-existing launcher -> adapter import the check could not see; it is now an allowlist entry (cap 10 -> 11). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
pins.py gains MILES_FORK_LINEAGE, CRITIC_FORK_LINEAGE and miles_fork_lineage_from(); loss_variants.FORK_COMMITS, critic_fork and the Miles declaration pins in entry.py derive from them instead of hand-copied SHA sets (values unchanged). The torch_memory_saver preload pins move from ssh_harness.py into pins.py. test_rl_engine_pins asserts all sets agree. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
syncer --island-lease-s defaulted to 30 s while the launcher always passes 300 s and the CLI help says 300. Change the syncer default to 300.0 and add a test that the syncer, CLI help and launcher defaults agree. cargo test: 155 passed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Remove ElasticRlBridge and make_island_bridge from yeto/rl/bridge.py (no production caller; the elastic island path is engine.bridges.ElasticAvgSync), TwoRoleStrictAvg with its helpers from critic_state.py (critic sync is DualStrictAvgSync), and the unused PPO role lane in local_learner.py (role lanes are SAO-only in production). Delete or narrow the tests that only covered this code. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both strict paths push byte-identical delta payloads and round metadata for the same base and local result, and the reference token^2/steps merge of those payloads is elementwise equal (three seeds, two islands). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
elastic.rs merge_weight now calls merge::learner_weight and keeps its zero-tokens/zero-steps branch. learner_weight takes c_steps as u64 so the elastic wire value needs no narrowing; strict callers widen u32 losslessly. A new Rust test checks the result is bit-identical to the old formula. cargo test: 156 passed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
test_rl_inter_island_rust_golden writes merge_weight_golden.txt from Python island_ledger.merge_weight; a new merge.rs test replays it bit for bit. Lag bounds are read from elastic.rs (MAX_SAMPLE_OUTER_LAG, DEFAULT_MAX_CARRY_LAG) and compared with island_ledger, printing both values on a mismatch. The golden tape must keep its lag=3 carry_lag_exceeded rejection. cargo test: 157 passed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The engine/bridges.py module docstring now holds the merge weight, staleness and finish rules of the three sync modes; test_rl_sync_mode_table checks the header, the three rows, and that the change spec carries the same table. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…bsent The run logs this test reads were moved out of the repository, so in a clean checkout the arm discovery raises FileNotFoundError instead of skipping. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
First batch of the repository cleanup (OpenSpec change
rl-code-quality-cleanup, stages 0 and 2). No training numerics or scheduling behaviour change.Stage 0 - repo hygiene and one import path
EVIDENCE.mdpointer..gitignorecoversrun-logs/,*.stderr,*.stdout,*metrics*.jsonand no longer whitelists logs. Newtests/test_no_tracked_logs.pyfails if a log is tracked again.yeto/rl/, 28 inyeto/rl/engine/miles_adapter/); every import uses the real module path.python -m yeto.rl.learnerstill works as a CLI entry point.pins.py;loss_variants,critic_forkand the Miles entry derive from it.--island-lease-sdefault is now 300 s, matching the launcher and CLI help.Stage 2 - one bridge, one weight and staleness definition
ElasticRlBridge,make_island_bridge,TwoRoleStrictAvg, and the unused PPO role-lane.StrictRlBridgevsStrictAvgSyncparity test (byte-identical).merge.rs;elastic.rscalls it. The step argument is widened tou64losslessly, so values are unchanged.Known deviation
tests/import_boundary_allowlist.txtcap goes from 10 to 11. Removing themiles_overlaystub exposed a pre-existing launcher -> adapter overlay import that the stub had hidden. This is not a new import, but it does relax the "allowlist only shrinks" rule. Follow-up: remove that import.Test plan
pytest tests/, excluding the Ray-spawningmismatch_tape/upstream_parse_argstests): 17 failed / 33 errors before the evidence-skip fix, and the failure set matched the pre-change baseline except for the one test fixed in the last commit. That test now passes or skips.cargo testinsyncer/: 157 passed, 0 failed.Stacked on #238 (base
s22-agentic-sft-modal). Retarget tomainafter #238 merges.🤖 Generated with Claude Code