v1.2.0

Highlights

• A new on-demand semantic-navigation Skill lets the model ask an installed local language server for definitions, references and single-file diagnostics, with before/after diagnostic comparison that keeps locations so a same-message replacement cannot hide behind an unchanged error count. All three bundled Skills are now declared in the package manifest. (#916)
• The Web session usage popover now shows the observed context composition of the latest call (system text, tool declarations, conversation), kept separate from Pi's context estimate and provider-reported tokens. (#914)
• Review diff excerpts and selected plan passages added to the composer now keep their source identity (repository, comparison, revision, or Session and result), including old paths of renamed files. (#915)
• The Web workbench adds a trajectory overview with context/model/tool tracks, search and an inspector for each call, plus a refreshed desktop shell, themes with preview cards, a compact model picker, more provider icons, and a narrower default sidebar (260px; saved widths are kept). (#868, #894, #904, #865)
• The Web composer recalls previously sent prompts with ↑/↓, using the same rules as Pi's TUI editor: ↑ only works when the box is empty, the cursor is at the start, or you are already browsing history. (#912)
• Custom models can declare their supported thinking levels in Web settings, and the conversation and Configuration pickers follow them immediately. Settings pages share one layout. (#864)
• Google Antigravity / Cloud Code Assist requests no longer fail with HTTP 400 when a tool schema uses a literal union (const is rewritten to a singleton enum). (#905)
• Long streamed answers, especially in Chinese, no longer make the Web client repeatedly disconnect and reload once a message exceeds the SSE event size; live messages are now trimmed by byte budget. (#854)
• The configured child-execution limit now bounds Workflow and Direct/BTW children together in one parent Session, and cancelling queued or starting work no longer leaves a slot held. (#828)
• Long sessions and active workflows do less CPU, allocation and disk work: unchanged checkpoint artifacts are not rewritten or re-serialized, and completed tool payloads are released after transcript eviction. (#895, #883)

Fixes

• Session worktrees keep executable bits on tracked scripts under a strict umask such as 077. (#911)
• Web recovery is more reliable: retry state, reading position, saved subagent history after a Host restart, chat switching while children run, archive Undo, preview limits and settings rollback. (#892, #893)
• A misnamed agent-type file now blocks fallback for both its filename and its declared name, so a narrowing override cannot fall back to a broader built-in role. (#833)
• Hidden terminal control strings in Web evidence are sanitized and oversized file identities are bounded. (#705)
• The TUI no longer stacks editor layers across session starts or /new. (#741)
• A pre-cancelled usage/quota wait now rejects instead of returning a stale result. (#734)
• The Cursor protobuf decoder rejects overflowing uint32 varints. (#749)

Compatibility and CI

• CI now also tests the minimum and latest supported Pi hosts on Node 22.19.0, and restores behavior on both: a Web config import missing on Pi 0.99.1, Pi 1.5.0's StringEnum schema and pi-schema package, and cross-model thinking replay on Pi 1.3+. Published support ranges are unchanged. (#719)
• The persisted sidecar compatibility boundary is documented with focused tests; no runtime change. (#902)

Full Changelog: v1.1.0...v1.2.0