Repository navigation

feat: token-level capture for RL training (Switchyard ↔ NeMo Gym) - #63

Merged

linj-glitch merged 15 commits into

NVIDIA-NeMo:rl-token-capture-betafrom

drifold:drifold/feature/token-capture

Aug 5, 2026

Merged

feat: token-level capture for RL training (Switchyard ↔ NeMo Gym)#63
linj-glitch merged 15 commits into
NVIDIA-NeMo:rl-token-capture-betafrom
drifold:drifold/feature/token-capture

Commits on Jul 14, 2026
Commits on Jul 15, 2026
- fix(observability): harden token-capture session dirs, config, and provenance
  drifold
  committed
  ed4ca12
Commits on Jul 20, 2026
- feat(observability): read opencode's native X-Session-Id as capture session id
  drifold
  committed
  7b33c2c
- feat(observability): capture parent session id and expose GET /v1/sessions
  drifold
  committed
  be0b2c5
Commits on Jul 21, 2026
- feat(backends): add token-continuity injection for vllm capture routes
  drifold
  committed
  7f20b2a
- fix(backends): stop non-extending calls from pinning injection fallback and clamp injected max_tokens
  drifold
  committed
  940b767
Commits on Jul 22, 2026
- fix(backends): route calls across per-session conversation chains and clamp budgets to model context
  drifold
  committed
  8307fd9
Commits on Jul 23, 2026
- feat(observability): expose upstream max_model_len on /v1/models for capture routes
  drifold
  committed
  021cb8f
- docs: cover token capture and injection in lib-core skill and rl-logging help
  drifold
  committed
  c448253
Commits on Aug 5, 2026
- fix(backends): drop vllm version gates that break on fork version strings
  
  Dafna Rifold
  authored and
  drifold
  committed
  9364d61
- feat(backends): add prefix_chat injection transport and /tokenize fallback for max_model_len
  
  Dafna Rifold
  authored and
  drifold
  committed
  7098508