Release v0.3.0 · spenceclark/Vessel

0.3.0 — 2026-09-19

Tools tab, Anthropic server tools, TypeSafe System One, MCP resources, and two new warnings.

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).

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).