sublera-cli 0.0.0-stage → 1.3.0

package/CHANGELOG.md

@@ -0,0 +1,172 @@
+# Changelog
+
+All notable changes to `sublera-cli` (formerly `sublet-cli`) are listed here. The project follows
+[Semantic Versioning](https://semver.org/). The stable surface is: command
+names, flags, `SUBLET_*` variables, `agent.json` keys, exit codes, the
+`--json` output of `doctor` and `config` (versioned by `schemaVersion`), and
+the files under `~/.sublet`. A change to any of these needs a major version.
+
+## [Unreleased]
+
+## [1.3.0] - 2026-11-14
+
+First release as `sublera-cli` (formerly `sublet-cli`).
+
+### Changed
+- The default server is now `https://sublera.com` (was `https://sublet.duckdns.org`).
+  `--server`, `SUBLERA_SERVER` and saved configs still choose another one.
+
+### Security
+- The device-login link from the server is opened only when it is on the
+  server's own origin. On Windows it is opened through `rundll32.exe` by full
+  path instead of `cmd /c start`, so a link can no longer run commands.
+- Plain-http servers are refused unless they are on this machine, because the
+  host key travels in every request. `SUBLERA_ALLOW_HTTP=1` allows a test server
+  on a trusted network.
+- Server text in the login prompt is printed without control characters.
+
+### Fixed
+- `start <code>` and `jobs` no longer exit when a different booking appears. They
+  exit when their own booking has ended and no other booking uses the GPU.
+- A booking that ended is no longer added back (and its earnings counted twice)
+  by a status answer from before the end. The server's clock decides whether a
+  listed booking is over.
+- A load for another request is refused with "Host is busy" while a loaded request
+  is still in its model check or a benchmark runs, instead of restarting
+  llama-server or replacing that request.
+- `start <code>` and `jobs` also keep running while another request on this host
+  is claimed, loading or in its model check.
+- A booking the agent learns about from the status poll now keeps the machine awake.
+- A booking frame no longer clears the busy flag of a model load in progress.
+- llama-server is restarted when the slot count changes, but only when no job or
+  other booking uses it; otherwise the agent keeps it and logs a warning.
+- A load for the file that is already running with the same settings reuses
+  llama-server and skips the VRAM check, which counted the loaded model twice.
+  A load that would replace the model under other bookings or public buyers is refused.
+- The saved booking keeps the slot count of its load instead of 1.
+- Per-token earnings are no longer overwritten on the dashboard when a booking starts.
+- `--help` named `~/.sublera` for logs and `agent.json`; the files are in `~/.sublet`.
+- Update the CI packed-package smoke test to install `sublera-cli` and exercise
+  both the new `sublera` executable and the legacy `sublet` alias.
+- Refuse linked model roots before pruning; concurrent instance-lock release
+  shares one operation and cannot unlink a later owner's lock.
+- Correct download budgets when an obsolete partial must be restarted.
+- Serialize engine installation per home, protect `doctor --fix` with instance
+  ownership, use unique staging directories and atomic manifest writes, and
+  attempt rollback when replacement installation fails.
+- Keep instance ownership until pending message handlers settle; clean up jobs
+  after wake-lock acquisition failures and suppress work/output after stop or
+  cancellation. Reject contradictory or malformed readiness snapshots.
+- Attempt all shutdown cleanup steps even if one fails, retaining ownership
+  through pending handlers and final state/health writes before reporting errors.
+
+### Added
+- Sublera branding, local package name `sublera-cli`, and `sublera` executable
+  aliases. Legacy executables and `SUBLET_*` environment names remain supported;
+  `SUBLERA_*` names take precedence. The default `.sublet` data directory and
+  repository identity are unchanged.
+- Several bookings at once on one loaded model: speed seats (a reserved tok/s)
+  side by side, a full-GPU rental, and per-token buyers in public mode. Each
+  booking has its own end timer, and a `reject` ends only its own request.
+  Bookings on the loaded model are accepted without a `load` frame.
+- `--price-hour <usd>` (`SUBLERA_PRICE_HOUR`, legacy `SUBLET_PRICE_HOUR`,
+  `agent.json` key `priceHour`, 0 to 50): the price per hour for the full GPU in
+  `sublera host`. It is sent as `priceHourUsd` in `hello` (default 0.50; `start`
+  and `jobs` send 0). The server sets seat prices from it.
+- The dashboard shows every rental and its reserved speed, the one that ends
+  first, and the hourly price in public mode. Earnings are per-token earnings
+  plus the time share of every booking. `sublera status` lists every booking.
+- `state.json` saves every booking in a `bookings` list, and a restart restores
+  all of them with their slot count. Files with the old single `booking` field
+  are still read.
+- The agent reads `activeRequests` from `/api/agent/me` and falls back to
+  `activeRequest` on older servers. Bookings the server no longer lists are dropped.
+- Exclusive `SUBLET_HOME` ownership for serving agents and storage maintenance.
+  A lock left by a process that is no longer running is taken over, so an agent
+  restarted after a crash restores its booking; a live or unreadable owner's lock
+  is refused with recovery instructions.
+- Local `health [live|ready]` snapshots, bounded SIGTERM inference drain, and
+  `storage status|prune` with dry-run-by-default pruning of old partial files.
+- Operational regression tests for cross-process ownership, cancellation,
+  reconnect readiness and safe maintenance. Async server message errors are
+  caught; busy hosts refuse conflicting model loads/new inference.
+- Storage guards: model-cache downloads default to a 20 GiB budget and preserve
+  5 GiB disk headroom, configurable through `SUBLET_MODEL_CACHE_GB` and
+  `SUBLET_MIN_FREE_DISK_GB`. Partial files count; existing data is not evicted.
+- Engine archive/cache limits, bounded tar extraction/link materialization,
+  and staging-space preflight checks. Model downloads serialize within one
+  process to avoid spending the same budget concurrently.
+
+
+### Added
+- **Corporate proxy support.** `HTTPS_PROXY`, `HTTP_PROXY`, and `NO_PROXY`
+  (upper or lower case) now apply to every outbound connection: API calls,
+  model and engine downloads, and the WebSocket tunnel (HTTP CONNECT, with
+  `user:password@` proxy login). `127.0.0.1`/`localhost` (llama-server) is
+  never proxied. `NODE_EXTRA_CA_CERTS` adds a corporate root CA. Proxy
+  passwords are never printed.
+- **Configuration without flags.** Supported settings can come from a `SUBLET_*`
+  environment variable or `~/.sublet/agent.json` (override the path with
+  `SUBLET_CONFIG`). Precedence: flag > environment > file > default. Unknown
+  keys and bad values are errors, not silently ignored.
+- **`sublet config`** shows the settings in use and where each one came from;
+  `sublet config path` prints the settings file path. `--json` supported.
+- **Agent log file.** `start`, `host`, and `jobs` write JSON lines to
+  `~/.sublet/logs/agent.log`, rotated at 10 MB with 5 old files kept.
+  `--log-file <path|off>`, `--log-level <level>`.
+- **Documented exit codes:** 0 ok, 1 runtime error, 2 usage or settings error,
+  130 Ctrl+C, 143 SIGTERM.
+- `doctor` reports the network path (direct or proxy) and the extra CA file;
+  `doctor --json` now includes `schemaVersion` and `agentVersion`.
+- Release gate (`npm run check`, `scripts/check-package.mjs`), a tag-driven
+  release workflow with npm provenance, and Dependabot.
+- **Engine self-check.** After every model load the agent asks the engine one
+  fixed question twice at temperature 0 and compares the token probabilities;
+  after a crash restart it compares against the load-time answer. A difference
+  is logged as a diagnostic warning, not a hardware diagnosis or attestation.
+  Incomplete or invalid evidence is inconclusive. It never stops a booking,
+  and a buyer's job cancels it.
+
+### Changed
+- A model already on disk is hashed again before it is served (once per agent
+  run). Before, a booking trusted the sidecar file and reported "hash-checked"
+  without hashing. A measured expected-hash mismatch triggers a replacement
+  download without first deleting the old model; IO errors preserve the file.
+- Model `verified` means a valid reference hash matched, not just that hashing
+  completed. Local checks do not authenticate a hostile host.
+- Streaming requires an explicit `[DONE]` marker, rejects malformed JSON and
+  token counts, bounds buffers/output, and releases the response reader.
+- Added unattended `--non-interactive` operation, server-scoped saved keys,
+  nested secret redaction, and serialized atomic state/config writes.
+- llama.cpp's environment settings (`LLAMA_ARG_*`, `LLAMA_LOG_*`,
+  `LLAMA_API_KEY`) no longer reach `llama-server`.
+- `start` and `doctor` say so when the engine is the host's own build
+  (`LLAMA_SERVER` or `PATH`) instead of the one Sublet installs.
+- Host keys, API keys, bearer tokens, and `token=` parameters are redacted from
+  every log line, on screen and on disk.
+- `status` shows HELD only when the server marks the host as held (with the
+  server's reason), never from the reliability number alone.
+- The tunnel refuses server frames over 16 MB.
+- Minimum Node.js is now 20.18.1 (needed by `undici`, used for proxies).
+- The npm package no longer ships source maps.
+
+## [1.2.2] - 2026-11-09
+
+### Added
+- Shared bookings: llama-server starts with the slot count the server sends.
+
+### Fixed
+- A benchmark load is shown as a benchmark, not as a buyer request.
+- `--gpu`/`--vram` apply to `jobs` and `doctor`; an unknown VRAM size is shown as unknown.
+
+## [1.2.1] - 2026-11-09
+
+### Changed
+- HELD is shown only when the server holds the host.
+
+## [1.2.0] - 2026-11-09
+
+- First public beta on npm.
+
+[Unreleased]: https://github.com/somemusic69-sudo/Sublet-cli/compare/v1.3.0...HEAD
+[1.3.0]: https://github.com/somemusic69-sudo/Sublet-cli/releases/tag/v1.3.0

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Sublet
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.