v0.2.0

Added

• Self-hosted plugin marketplace — the repo now doubles as a Claude Code marketplace:
  /plugin marketplace add jhesham/cross-llm-delivery then /plugin install cross-llm-<provider>@cross-llm-delivery. Three per-provider plugins; every push is a new
  installable version.
• "Using it" guide in the README — the conversational flow (plan with Claude → "use
  cross-llm- to run this plan" → Claude delivers + judges).
• Git preflight — a missing git or a non-git target now aborts with a clear message
  (install hint / git init) instead of a raw traceback on the first slice.

Changed

• README depth pass: badges, real --status/gate output samples, a "Safety rails" section, a
  providers/models table with validation status, an honest "Is this for you?" filter, and a
  neutral design-stance section (no comparative claims about other projects).

Fixed

• --status elapsed time was always 0s from the CLI (now defaults to wall-clock).
• --executor <provider>:<model> with an explicit model is honored verbatim instead of being
  silently routed to a catalogued workhorse; stale kimi-k2.7 catalog id → kimi-k2.7-code.
• First CodeQL scan findings (read-only CI workflow token); clean-machine CI failures.

Security

• Enabled repo secret scanning + push protection, private vulnerability reporting, CodeQL,
  Dependabot; neutral commit identity enforced.

v0.1.0 — initial public release

Changelog

0.1.0 — 2026-07-03 (initial public release)

• Provider-blind engine (engine/cld/): orchestrator with DAG layering + parallel dispatch,
  per-slice git-worktree isolation, resumable JSON ledger, integration gate.
• Deterministic judge: committed failing acceptance tests are the dispatch contract; pass/fail
  is the real pytest exit code; an allowed-files diff rule rejects out-of-scope edits; failures
  retry with structured judge feedback.
• Three executor providers behind a drop-in registry: antigravity (agy), opencode,
  cursor (incl. cursor:composer-2.5), each with a model catalog, headless-dispatch quirks
  handled (shim bypass, stdin detachment, console-encoding safety), and validate-before-trust
  probing for untested models.
• Complexity routing + escalation ladder: untagged slices route to the cheapest viable model
  and climb on failure; explicit --executor models are honored verbatim; per-slice executor:
  tags pin a slice.
• Telemetry: every lifecycle moment emits to .cld/events.jsonl; --status digest (layer
  position, in-flight slices + elapsed, tokens, cost-by-model, gate); --watch live view;
  opt-in OpenTelemetry export (GenAI semantic spans) to any OTLP backend, with a two-env-var
  Langfuse convenience.
• Generator: produces self-contained per-provider Claude Code skills (vendored engine, no
  pip install) into dist/, with an optional mirror-publishing helper.
• Safety rails: caller-merge preflight (warns when a pending layer depends on accepted but
  unmerged slices), executor-CLI preflight (friendly message instead of a traceback when no CLI
  is installed), non-destructive diff preservation for rejected slices.