Vessel: all versions and old releases

0.3.1 — 2026-10-02
System One replay between TypeSafe and Ollama, a tidier Compare view, IBM Plex Sans, and winget.
Added
- System One decisions replay between TypeSafe and Ollama 0.35, which serves the same
POST /v1/systemone API: a typesafe-systemone capture can replay to any ollama backend
or to an api.typesafe.ai backend. Credentials come only from the target, so a TypeSafe key
never reaches Ollama. OpenRouter's /api/alpha/decisions stays same-backend only (#137).
- Package-manager install on Windows with winget: winget install spenceclark.Vessel (#32).
Changed
- Compare hides metric rows that no column has a value for, in every format — for example
TTFT, Tok/s and Stop reason for System One (#137).
- For a decision pair, Compare's shared request shows only the parameter diff and state;
each answer column already carries its questions (#137).
- Decision values, confidence and the score scale are shown rounded to 3 places; Raw JSON
keeps full precision (#137).
- Choice bars follow the request's criteria order, so Compare columns line up. In the request
detail view the winning option is no longer always the top bar (#137).
- The UI font is now IBM Plex Sans, bundled with the app. The previous Inter font never
loaded because of a font-name mismatch, so the UI fell back to the system font (#127).
- Dependency bumps: @tanstack/react-query 5.104.0, @tanstack/react-virtual 3.14.13 (#128, #131).

0.3.0 — 2026-09-19
Tools tab, Anthropic server tools, TypeSafe System One, MCP resources, and two new warnings.
Added
- Tools tab: a request that declares tools gets a Tools (n) tab with one card per tool —
description, parameter table, and how many times it was called (#86).
- Anthropic server-tool blocks (server_tool_use, web search and web fetch results) render
as cards, with numbered citations and a sources list per message (#88).
- Tool call args and tool results that parse to an object render as a readable field list,
with multi-line strings unescaped; Raw JSON still shows the exact bytes (#91).
- TypeSafe System One (Jev) support: typesafe-systemone traffic — direct /v1/systemone or
OpenRouter's /api/alpha/decisions — is detected instead of falling to raw, with a
Decisions view (one card per question, probability bars) and same-backend replay (#122).
- Two new warnings: repetitive_output for a response stuck in a loop, and
slow_response for a non-streamed request slower than warnings.slowResponseMs
(default 120s, 0 disables; editable in the Config panel) (#90).
- MCP resources: vessel://requests/{id} and vessel://sessions/{id}, with
resources/list returning the most recent requests and sessions (#92).
- OpenRouter and TypeSafe in the "Add backend" catalog (#124).
Fixed
- A queued replay whose backend was repointed while it waited is refused with
409 replay_target_changed instead of sending the old credential to the new
destination (#106).
- MCP request reads decode bodies under capture.maxBodyMb rather than unbounded, and
report truncation with decodeTruncated (#107).
- While a listen-address change is pending restart, the UI/API stays reachable on the
address the process is actually bound to (#112).
- path_missing_v1 only fires for the missing-/v1 mistake: an OpenAI SDK path suffix, no
/v1/ segment anywhere, and a backend baseUrl with no path of its own (#123).
- Anthropic server-tool blocks are included in MCP include=text and full-text search;
encrypted_* fields never reach the flattened text (#89).
- openai-chat reasoning text under the reasoning key (e.g. Ollama) is no longer dropped
from the thinking block and search (#80).
- Anthropic rate-limit headers are parsed (#108).
- Parameter sweeps on OpenAI Responses replays use max_output_tokens (#111).
- Config validation rejects a backend with "type": null (#109).
- Multi-replay rejects a null variation with invalid_request and the offending index,
before dispatching anything (#110).
Changed
- Dependency bumps: react 19.3.0, tailwind-merge 3.7.0 (#117, #119).

Changelog
All notable changes are documented here. Vessel follows Semantic Versioning.
0.1.1 — 2026-08-30
Post-launch fixes and polish.
Fixed
- Copy-as-curl now emits the correct auth for Anthropic backends (x-api-key +
anthropic-version) rather than an OpenAI Bearer header — including backends left on
type: auto (#7).
- Copy-as-curl targets a reachable host instead of the bind address, so a command copied
while Vessel runs in a container actually connects (#8).
- Chrome DevTools' /.well-known/appspecific/… probe is no longer proxied and captured as
an errored request (#4).
Added
- Config validation rejects http:// for public/remote backends, so an API key is never
sent in plaintext; localhost and LAN backends are unaffected (#5).
- Keyboard navigation — ↑/↓ move through the request list (#6).
- "Add backend" is now a picker of known backends (OpenAI, Anthropic, Ollama, vLLM,
llama.cpp, LM Studio, …) that prefills URL, type, and auth env var (#9).
- First-run guidance when Ollama isn't reachable, pointing you to add a backend (#11).
Changed
- Clearer label and help text for the streamed-token-usage setting (#10).
0.1.0 — 2026-08-29
First public release of Vessel: a local-first, single-binary reverse proxy for LLM
traffic. It captures and searches OpenAI, Anthropic, Ollama, and raw traffic; provides
live metrics, replay and Compare; includes a read-only MCP endpoint; and ships native
Windows, Linux, and macOS archives plus a GHCR container image.