Mercury CLI
Documentation

Releases

1.0.5

published 4 November 2026

Mercury CLI 1.0.5 rewrites the tool kit so the model reads each fact once: every tool description states what the tool does and returns, parameter meaning lives on the fields, and the eager tools at session start take 15.5 KB on the wire instead of 25.2 KB. The bundled Craft extensions move to FilmCraft 0.17.0, PhotoCraft 0.20.1 and LightCraft 0.18.0, and the in-house skills are smaller and current. Stale statements in the Bash, PowerShell, Read, Eval, ScheduleWakeup, NotebookEdit, Skill, Crewmate, Workflow and Journey descriptions are corrected.

Changed

  • Changed every tool's model-facing text: each description states what the tool does and returns, once; units, bounds and defaults live on the parameter fields; only the six tools with a separate reference (Browser, Crewmate, Debug, Document, Eval, Workflow) send a first-use manual.
  • Changed the deferred tool roster to list each tool with its full one-line purpose, with search keywords kept in an off-wire index; in ToolSearch a +word term matches a tool's name, description or search words.
  • Changed the bundled FilmCraft extension to 0.17.0: the control port requires a session token that Mercury creates per launch, keeps in a private file and passes through the environment; attaching to a window you started by hand takes tokenFile; craft_connect describes all seven of its fields.
  • Changed the bundled PhotoCraft extension to 0.20.1 (one archive per platform; on macOS the command-line tool bundled in the app) and LightCraft to 0.18.0.
  • Changed the bundled skills to be smaller and current: provider-apis points at the live model classifier and per-dialect references instead of copied inventories; image-generation, mcp-smithy, skill-forge and research follow the connected servers and the requested deliverable instead of fixed procedures; the FilmCraft, PhotoCraft and LightCraft skills keep each app's safety boundaries and current commands.

Fixed

  • Fixed the model-facing input schemas of AskUserQuestion, Workflow and Computer: AskUserQuestion exposes only its questions, Workflow accepts only the five parameters it reads, and Computer's step items no longer admit a nested sequence action.
  • Fixed the Bash and PowerShell descriptions: the transcript shows a collapsed preview of command output (not nothing); exit 1 is a result only when the last command is a search or test command, with a per-command label; the working directory persists between calls only in the main conversation.
  • Fixed the Read, NotebookEdit and Skill descriptions: Read selects PDF pages as pictures only on models with image input and as text otherwise; NotebookEdit resolves a relative path against the working directory; a skill declared to fork runs in a crewmate.
  • Fixed the Eval and ScheduleWakeup descriptions: Eval raises in the cell whenever a nested tool call returns an error, with tool.attempt returning the error as a value; ScheduleWakeup's reason heads the woken prompt and shows on the wake's row.
  • Fixed the Crewmate, Workflow and Journey descriptions: an omitted effort or model takes the agent definition's setting, then the configured crewmate default, then yours; Workflow's agents wait for a free crewmate slot only while a model call is in flight; a Journey's kept services end with the session.