v0.4.0

0.4.0 (2026-06-12)

Features

• add timeout multiplier flags for slow hosts (7be54a0)
• agent setup hook and host-credential passthrough (374b68f)
• collect task-declared artifacts from the trial container (5cfa485)
• curated clean-room verifier, per-step separate verifier and setup.sh (df75edb)
• emit harbor-compatible per-key reward stats in the job result (286fd81)
• environment and per-step healthcheck (d5fd189)
• honor task-configured agent/verifier user (46c5aa2)
• make --strict-resources mirror harbor cpu/memory limits (76a08c7)
• model verifier rewards as a named map (6e508a6)
• run multi-step tasks (86277f3)
• run the verifier in a separate environment when declared (f8519ed)
• support reward.json and fractional rewards (bdc9ad6)
• whole-trial retries with a non-retryable default set (cadba66)

Bug Fixes

• give image pulls at least 30 minutes (19ca62e)
• honor image WORKDIR and mount task solution at /solution (36e3659)
• let the verifier reward decide the trial, not the script exit code (1a9a105)
• request explicit platform for prebuilt foreign-arch images (a9837b9)
• sweep orphaned trial containers and quiet cleanup races (f01f4f7)

Performance Improvements

• give each trial a fair share of host CPUs; lower default concurrency (d9c40d4)
• move /app to docker volumes, boost task resource caps, drop preflight barrier (079c957)

v0.3.0

0.3.0 (2026-06-09)

Features

• add upgrade command (238a9ab)

v0.2.0

0.2.0 (2026-06-09)

Features

• add deterministic evaluation core (e5f1b47)
• add direct nop agent execution (e158cfe)
• add sandboxed docker execution backend (530513a)
• add seaport CLI and project naming (e5bc6c9)
• align task environment parsing with harbor (658a188)
• group resolution progress under run header (bf94c57)
• improve run logging (80444fe)
• make run output terminal friendly (8b2ecaf)
• pass phase environment variables (f272892)
• resolve git-backed registry tasks (46b5995)
• resolve harbor local registry datasets (a8f31c9)
• resolve remote package datasets (65d0c16)
• run direct registry and git tasks (76904ae)
• run harbor-compatible local datasets (6d8b7de)
• run local oracle tasks (8a6fe5d)
• run sandboxed external agent commands (48e018b)
• run task attempts concurrently (94a5e40)
• show run startup progress (d7c4949)
• show task durations in run output (f9aa331)
• stream task execution logs (b07261f)

Bug Fixes

• align docker runtime with benchmark toolchains (904ed2f)
• allow sandboxed tmp binaries to execute (1f8a034)
• build task dockerfiles on native platform (1d09989)
• expose cobol copybooks at workspace root (a86f426)
• gate docker api response parsing to unix (411c805)
• honor task docker resource limits (e84f8a0)
• infer amd64 platform for legacy Java images (a4af2ce)
• infer amd64 platform for x86 assembly tasks (8402b98)
• match benchmark container layout (62e5385)
• prefer native docker platform when available (903dabf)
• preserve docker image workspace (7b0867d)
• record failed trials without stopping runs (3333674)
• report each task's share of the execution timeline (0cbeb07)
• retry package archive downloads (ed9efbd)
• retry transient docker build failures (52f24b9)
• satisfy formatting and clippy checks (e9925c5)
• set docker platform for task containers (1235408)
• stage task files in execution workspaces (6c0f46e)
• stream docker build progress (c8604f9)
• use buildkit-compatible network mode (ac36a85)

Performance Improvements

• adapt default run concurrency (f209a1b)
• cache docker task environments (acec736)
• deduplicate docker image pulls (10bfeaf)
• preflight docker environments (fdf0f00)
• raise adaptive concurrency cap (9631112)
• schedule long-running trials first (e1196c1)
• schedule work by run phase (db643a1)
• speed up workspace snapshot restores (f3195fb)
• use a managed buildkit builder (358f7cd)
• use docker api for control-plane checks (8b5cfea)