v0.2.0

0.2.0 — 2026-09-07

Named sessions, multi-replay with scoring, reports, export, and package-manager installs.

Added

• Named sessions: assign requests to a run with X-Vessel-Session (created on first use),
  with a session picker, per-session stats, and session deletion (#29, #41).
• Multi-replay: fan one captured request across several models, or one model across
  several parameter values, and see every response side by side with per-column metric
  deltas (#48).
• Scoring: rate any response 1–5 from the Compare view — original included — and a
  leaderboard in Reports ranks models and parameter sets by mean score and win-rate (#49).
• Reports view: requests, tokens, tok/s, duration, cache efficiency, and warnings by
  model/tag/backend, plus a context-growth chart per session or tag (#25, #26).
• Export the current filtered request list to CSV or JSONL (#24).
• Replay dialect fix-ups: max_tokens ↔ max_completion_tokens is renamed automatically
  for OpenAI Chat replays and shown as (auto) in Compare's parameter diff (#28).
• Package-manager installs: Homebrew (macOS and Linux) and Scoop (#31, #32).
• Tool calls a model emits as plain text instead of structured output are detected and
  flagged (#40).
• A hint on 404s from OpenAI-compatible backends when the client's base_url is missing
  /v1 (#57).
• README: tagging agents and naming runs from LangChain / LangGraph.

Changed

• Compare is now a grid rather than a fixed pair (a pair is the two-column case); metrics
  render as a table across columns (#48).
• MCP search_requests and get_request include each request's score and replay linkage
  (#48, #49).
• Config panel: backend cards compacted to two rows with shared explainers (#42).
• macOS unblock steps updated for Sequoia (#39).

The database schema migrates automatically on first start; the additions are additive, so
an older Vessel can still open the file.