Skip to content

CHANGELOG.md

Latest commit

 

History

History
121 lines (101 loc) · 6.96 KB

CHANGELOG.md

File metadata and controls

121 lines (101 loc) · 6.96 KB
Raw

Changelog

All notable changes to this project are documented here.

[Unreleased]

[0.3.0a1] - 2026-09-10

  • Mark the workspace-capable feature milestone: provider-neutral persistent workspaces, verified Sprite lifecycle recovery, supervised retention, and offline destination cost comparison. Includes all 0.2.0a2 capabilities below.

  • Publish as an explicitly approved alpha prerelease with independent-user onboarding and release-matched public lifecycle evidence still outstanding. Publication does not certify beta maturity; see docs/RELEASE_0.3.md.

  • Add offline, provider-neutral elsewhere compare for caller-supplied utilization, total job cost, eligibility, deadline and budget estimates. Recommendations remain advisory and do not authorize execution.

[0.2.0a2] - 2026-09-10

  • Add provider-neutral persistent workspace execution with Sprites as its first adapter: fresh work, approved projects, and isolated repository snapshots.

  • Bind workspace execution to exact source content, identity, access policies, resources, runtime, estimated cost, and supervised retention expiry.

  • Keep disconnected tasks alive only after runner readiness, recover verified changes, reconcile interrupted sessions, and protect shared projects from task deletion.

  • Add workspace-reap --execute to recover results and apply overdue retention when the developer reconnects. Native checkpoint forks remain unsupported.

  • Stop treating retained macOS swap as current memory pressure: idle machines may run one declared build or test when it fits, while live paging, memory stalls, unknown swap at low headroom, and the protected OS floor still block.

  • Automatically consume the installed macOS host sampler so sandboxed clients receive swap and paging truth without a private environment-variable override.

  • Report that sampler as available in elsewhere doctor, matching the live capacity decision instead of showing a contradictory installation warning.

  • Add a Herdr workload pane with explicit execution, project-scoped job receipts, and cleanup gated on verified result recovery.

  • Assess remote placement before portable verification and when local capacity denies work; expose queued placement opportunities with stable event identities.

  • Follow job changes through a bounded MCP wait and return completion receipts with transported-source fingerprints, results, estimates, and cleanup status.

  • Try approved fallback regions when Fly rejects new resources in a retired region.

  • Add guided elsewhere init and read-only elsewhere doctor commands so a new user can configure Fly/Tigris or Azure and see the exact remaining setup step.

  • Add Fly-native Tigris source and result transport with endpoint pinning, credential-safe presigned URLs, authenticated downloads, and verified cleanup.

  • Keep explicit project trust denial and revocation from inheriting unrelated global approval while preserving compatibility for older project configurations.

  • Add Linux and macOS CI coverage for Python 3.11-3.13, public-content guardrails, and an executable release gate for current multi-user lifecycle evidence.

  • Sense local capacity natively on Linux via /proc/meminfo and the /proc/pressure/memory PSI signal, so the run-here-or-elsewhere decision works without a sampler daemon.

  • Report local sensing as explicitly unavailable on unsupported platforms instead of silently claiming the machine has zero RAM; remote execution stays available.

  • Retry the remote result upload on transient network or 5xx/429 failures so a momentary blip no longer discards a completed job's result.

  • Match Fly Machines strictly by dispatched Machine ID and read the ID from structured JSON first, so a stale same-named machine from a prior run can no longer be mistaken for the current job.

[0.2.0a1] - 2026-07-18

  • Prevent quiet retained swap from starving builds, return actionable privacy-safe denial details, and add cleanup --stale for orphaned reservations and dead local jobs.
  • Persist normalized remote state and stable provider identity across processes.
  • Scope Fly logs to the exact Machine and retain completed Machines until verified cleanup.
  • Put Fly and Azure lifecycle behavior behind a shared provider contract.
  • Verify compute and Azure Blob absence before reporting cleanup as complete.
  • Protect unavailable results during cleanup, with an explicit --discard-results escape hatch that still verifies all billable resources are gone.
  • Use the package version in MCP metadata and keep repository version labels aligned.
  • Return a checksum-verified result bundle containing stdout, stderr, the exact exit code, and explicitly requested files after remote compute has stopped.
  • Validate MCP requests and tool arguments before any planning or execution action.
  • Redact signed URLs and credential-shaped values from persisted and displayed state.
  • Verify provider compute, source artifacts, and result artifacts are absent before a job reaches cleaned.

Added

  • Add a sandbox-safe macOS host sampler for memory pressure, retained swap, and swap/page-out activity.
  • Add a service workload for preview servers and other small persistent work.
  • Persist denied local work and resume it automatically when shared capacity returns.
  • Add elsewhere queue and a loopback control room for waiting/running work, recent history, standalone reservations, cancellation, and deliberate release.
  • Add durable trust receipts covering provider accounts, regions, artifact storage, source roots, private/uncommitted source, resources, runtime, and estimated cost.
  • Ship a typed Codex plugin for trust inspection, planning, dispatch, queue visibility, and job control.

Changed

  • Replace the below-50%-means-one-slot cliff with adaptive memory budgets, graded pressure bands, swap-activity brakes, and lease-age accounting.
  • Keep managed local-job leases alive for the lifetime of their process while preserving expiry for standalone reservations.
  • Rename the product to Elsewhere around the promise that work can continue beyond the device where it started.
  • Add elsewhere as the primary command while retaining agent-capacity as a compatibility alias.
  • Reframe the README around “Your work doesn't need your laptop” and “Close the lid.”
  • Document the public product promise, CTFL acceptance test, and copy guardrails.

[0.1.0] - 2026-07-11

Added

  • Coordinate local memory reservations across Codex, Claude, browser automation, builds, and tests.
  • Route OCI workloads through a shared provider contract, with Fly and Azure adapters.
  • Retry Fly workloads across regions and fall back between configured providers.
  • Package local source safely, excluding common secrets and heavyweight generated folders.
  • Transport source through short-lived Azure Blob URLs independently of the compute provider.
  • Inspect, log, cancel, and clean up dispatched workloads.
  • Verify Azure Blob to Fly source transport with regional retry, URL redaction, secret exclusion, and zero-resource cleanup.