v0.17.0

[0.17.0] - 2026-07-13

New Features

• Better guidance on smaller models — and fewer dead ends on every model. A "not found" from any code-graph tool now names the closest real symbols and ends with a ready-to-send follow-up call instead of leaving the agent to guess. On Haiku sessions, SpecShip goes further: the per-prompt nudge becomes a precise do-this template (and steers away from costly subagent fan-out), call-path answers render as numbered step-by-step hops with each connection explained inline, and the tool menu slims to the three core tools so tool choice stays reliable — all without removing any evidence, and all off on frontier models. The benchmark harness also gained an EVAL_MODEL switch so these behaviors are validated per model tier.