v1.17.0

Added

• parent is a selectable Thinking level. Agent files, Subagent default thinking and the spawn wizard accept it, and the Subagent then runs at the parent session's current level. It sits in the position you set it in, so a file's parent beats pi's per-model level but loses to a wizard pick, and it resolves before the Subagent exists, so nothing downstream can show the word. Nothing changes unless you select it.
• Pre-spawn menus name what decided the Thinking level. Model settings and the spawn wizard read high (frontmatter) or minimal (parent) instead of a bare level, so an unexpected effort is traceable before you spawn. The widget, viewer and stats stay bare, because a running session stores only its number.
• Subagents are announced to Subagent-aware extensions. Creation, binding and disposal are broadcast on pi's event bus, so a permission gate can raise an ask from inside a Subagent on the parent session. An announcement can never fail a spawn or a teardown.
• New MessageAgent tool feeds input to an existing agent. A running agent takes it at its next turn boundary; a settled agent with a live session continues on it, context intact, with no second record.
• Short agent IDs work in StopAgent and MessageAgent. The 8-character ID the tools print is accepted when it matches one agent; an ambiguous prefix acts on none and names every candidate.

Changed

• The setting row is now Subagent default thinking. Label only, the key stays defaultThinking, and it no longer collides with pi's own defaultThinkingLevel. Unset reads pi default, and pi default in the list removes the key at the chosen layer.

Fixed

• An untrusted spawn target no longer hides your global per-model Thinking level. The gate discarded pi's whole modelThinkingLevels map rather than only the untrusted project's file, so the run ignored a level that Model settings printed.
• Both Thinking level annotations follow the same parent level. Model settings fell back to the current session context and the wizard did not, so the two rows disagreed about one run.
• A Thinking level picked in the spawn wizard is clamped to the model the Subagent gets. Choosing (inherits parent) was the one model change that skipped the clamp.
• The spawn wizard reads pi's per-model level for the model the Subagent actually runs on. It used the parent's model even where the agent file pins another, so it offered levels that model cannot run.