feat: per-env reset routing in the bridge (fleet mode PR-A, BRG-5) - #172
Merged
Conversation
The batched-env probe measured the surprise: n_envs=4 already steps at ~the single-env wall cost (3.6 ms vs 3.5 ms, 3.8x env-step throughput), and per-env command routing (BRG-5), per-env guard state, and per-env state publishing all pre-exist. The one true gap was RESET: - teleport_reset(seed, env_id): env-sliced state injection (boxes and arm home via envs_idx) with per-env command drain and dropped-count clearing; env_id None keeps the whole-scene semantics bit for bit. - ResetQuarantine is per-env (arm/held/tick): a reset of env k no longer freezes the other envs' episodes; the legacy hold() contract is preserved and re-tested. - CommandQueue.drain(env_id): env-scoped drain for per-env resets. - The reset handler validates metadata env_id per BRG-5 (int, in range), teleports the slice, and replies reset_done with that env_id. The shared publish grid is NOT re-anchored on per-env resets (fleet cadence anchors to the first whole-scene reset — a documented fleet-mode difference from ADR-25's per-episode anchoring). Cameras remain env-0-only (ADR-7): the fleet MVP targets the L0 rung, where neither policy nor the oracle verifier needs rendering. Frozen-set edit (dora_genesis.py): env_hash re-pinned. Gates: ruff clean; 1051 unit tests (per-env quarantine/drain suites); live T0 episode green (single-env regression). Co-Authored-By: Claude Fable 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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
First fleet-mode increment (§8.4.3). The probe measured that batched physics, per-env command routing, guard state, and state publishing already exist — the true gap was per-env RESET:
teleport_reset(seed, env_id): env-sliced injection viaenvs_idx; whole-scene path bit-identical.ResetQuarantine+CommandQueue.drain(env_id): one agent's reset no longer freezes or flushes its neighbours.env_idon reset requests;reset_donereplies carry it; the shared publish grid anchors to the first whole-scene reset (documented fleet-mode deviation from ADR-25).Measured: n_envs=4 steps at ~single-env wall cost (3.8× env-step throughput). PR-B (the multiplexing scheduler + fleet CLI) follows.
Gates: ruff clean · 1051 unit tests · live T0 episode green (single-env regression) · env_hash re-pinned.
🤖 Generated with Claude Code