v1.0.3

Changes since v1.0.2

• Prepare 1.0.3 release (c08a00a)
• Use home state dir and record daemon intervals (#16) (2a87f75)

Full Changelog: v1.0.2...v1.0.3

v1.0.2

Changes since v1.0.1

• Prepare 1.0.2 release (#15) (82d9625)
• Harden daemon lifecycle and clarify reports (#14) (d402ef4)

Full Changelog: v1.0.1...v1.0.2

v1.0.1

Changes since v1.0.0

• Prepare 1.0.1 release (4e63bce)
• Make gua daemon run in background (#13) (91a2057)

Full Changelog: v1.0.0...v1.0.1

v1.0.0

Changes since v0.4.1

• Prepare bare-metal 1.0 release (#12) (59e06c7)
• Prune post-1.0 runtime planning (#11) (944583d)
• Document bare metal 1.0 status (3fd8745)
• PR B: Packaging And Install UX (#10) (06b6599)
• Bare Metal Scope Reset (#9) (dd1d41d)
• Document bare metal 1.0 scope (856e0e0)
• Default daemon and report DB to tmp path (#8) (f15afdc)
• RuntimePlan and doctor v1 (#7) (239a333)
• Document uv tool update and uninstall (#6) (d2cee3b)

Full Changelog: v0.4.1...v1.0.0

v0.4.1

Changes since v0.4.0

• Enable PyPI trusted publishing (#5) (140500d)
• Document uv tool install download flow (#4) (bb1c44d)

Full Changelog: v0.4.0...v0.4.1

v0.4.0

Changes since v0.3.0

• Release hardening and download path (#3) (18a371f)
• PR 2: Command Surface Skeleton (#2) (3708b11)
• Merge pull request #1 from AI-Ocean/proposal/auto-runtime-architecture (e2880b8)
• ci(release): replace --generate-notes with git-log-based notes (0315928)

Full Changelog: v0.3.0...v0.4.0

v0.3.0

Changes since v0.2.0

• docs+bump: 0.3.0 — separate demo/daemon, document --since semantics (3c8ade9)
• feat(cli): add 'demo' subcommand for self-contained fake-tier demos (804121c)
• docs(readme): trim out-of-scope; add Usage section with daemon/report semantics (0e90d3c)
• docs(readme): revive Quick demo, simplify install, add NVML section (ffa4556)
• chore: drop CHANGELOG.md; release workflow uses --generate-notes (6040b22)

Full Changelog: v0.2.0...v0.3.0

v0.2.0

First Python stable. Real NVIDIA NVML telemetry support added — daemon
now runs on real GPU hosts via --tier nvml.

Added

• NVMLTier (gpu_usage_audit.nvml): real telemetry via pynvml
  (nvidia-ml-py) — compute-running processes, per-card UUID +
  utilization, bytes→MB memory conversion. Late-bound import so the
  package works without the [nvml] extra; --tier nvml raises a
  friendly install hint if the extra is missing.
• daemon --tier {fake,nvml} flag (default fake). The fake source
  remains usable on any host for the funnel demo; the NVML source is
  for real GPU hosts.
• [nvml] optional dependency: pip install gpu-usage-audit[nvml] or
  uvx --with nvidia-ml-py gpu-usage-audit ....

Fixed

• FakeTier now pins synthetic loginuid_user values (alice / bob /
  None) so the daemon's system_user_lookup no longer accidentally
  resolves a real local user when a synthetic PID happens to exist on
  the host.

v0.2.0a1

First Python alpha. The 5-section report and daemon loop are ported
from the Go v0.1.0 design; no real NVML yet.

Added

• daemon subcommand — FakeTier sampling into SQLite with anti-drift
  scheduling and signal-driven shutdown (threading.Event cancels
  stop.wait(delay)).
• report subcommand — five-section retrospective report:
  • §1 Headline three-bar (active / idle-held / truly-idle).
  • §2 Waste (idle GPU-hours, equivalently-unused GPUs).
  • §3 Per-GPU breakdown.
  • §4 Top identities (loginuid-resolved or unknown).
  • §5 Day-of-week × hour activity heatmap.
• FakeTier — deterministic 5-tick GPU-0 cycle
  (active → idle-held → truly-idle → repeat), invariant GPU-1/2.
• Classify / Summarize / detect_env_kind (bare/docker/k8s)
  ports of the Go v0.1.0 decisions.
• SQLite layer: journal_mode=WAL, busy_timeout=5000, indexes on
  (gpu_uuid, ts), transactional write_snapshot.
• version / help subcommands alongside --version / --help.
• _duration argparse type: "30s" / "1h" / "200ms" parsing
  (Go time.ParseDuration subset).
• Test suite (85 tests, standard testing-style with pytest): unit
  tests for every domain module plus CLI smoke and integration fixtures.
• GitHub Actions CI: ruff + mypy --strict + pytest on every push
  / PR via uv sync --all-groups --locked.
• Release workflow: tag push (v*) → uv build → GitHub Release with
  sdist + wheel attached, release notes extracted from this CHANGELOG.

Notes

• This is an alpha — the binary's --help works, daemon/report run
  end-to-end on fake telemetry, but real NVML is not wired yet.
• PyPI distribution requires trusted-publishing setup; until then the
  wheel/sdist are downloadable from the GitHub Release page.
• Go v0.1.0 remains downloadable at the v0.1.0 tag / go-archive
  branch.

v0.1.0

First public release.

Added

• daemon subcommand — periodic GPU/process sampling into SQLite with
  anti-drift scheduling, signal-driven shutdown, and single-transaction
  per tick.
• report subcommand — five-section retrospective report from any
  accumulated database file:
  • §1 Headline: active / idle-held / truly-idle proportions with a
    glyph-differentiated three-bar.
  • §2 Waste: idle GPU-hours and equivalent unused GPU count.
  • §3 Per-GPU: idle-held breakdown by card.
  • §4 Top identities: by-user GPU-hours and idle-held share.
  • §5 Heatmap: day-of-week × hour activity grid.
• FakeTier — deterministic time-varying fake telemetry source so the
  daemon is exercisable on any host (no NVIDIA driver required).
• Identity resolution via /proc/<pid>/loginuid with UserLookupFunc
  abstraction; pluggable table-based lookup for tests.
• Host environment auto-detection (bare / docker / k8s) from
  /proc/1/cgroup.
• Three-table schema (host, gpu_sample, proc_sample) — minimal
  surface aligned to the idle-held question.
• SQLite journal_mode=WAL + busy_timeout=5000, so the daemon and
  report can share the same database file without SQLITE_BUSY.
• Indexes idx_gpu_sample_uuid_ts and idx_proc_sample_uuid_ts on
  (gpu_uuid, ts) for card-keyed time-window queries.
• help / version subcommands (alongside --help / --version).
• Unit and DB-layer test coverage (standard testing only, no
  third-party deps): Classify, DetectEnvKind, Summarize,
  FakeTier phase cycle, and all Load* report queries against
  a real on-disk SQLite fixture.
• GitHub Actions CI: vet + race-enabled test + build on every
  push and pull request.
• Apache 2.0 license, Makefile (build / run / test / clean),
  --version injected at link time.

Notes

• v0.1.0 ships fake telemetry only — the daemon is exercisable on any
  host. Real NVML support is targeted for v0.2.0.
• The legacy gpu-usage-audit (v0.1.x) project is archived in favour
  of this rewrite.