{"content":"Dev journal — DSH Phase 1 tool-behavior fix (feat/dsh-phase1): live session logs showed the tools were correctly delivered in every model request (write/edit/read/todo_write plus the bridged search tools) yet the model made zero tool calls — pure chat. Root cause: the Claude path inherits tool-use discipline from the claude_code system-prompt preset, while the DSH base prompt carries none, so with only the persona prompt the model describes tasks instead of acting. Added an idbots:tool-use guidance section (order 150) to every DSH session: mandatory tool use for file/command/search work, write-then-read-edit discipline, todo tracking, and never claiming work the tools did not confirm. Also fixed along the way: zod raw-shape schemas ({key: ZodType}) now convert to JSON schema at the host bridge (the top-level-only check missed bot_browser_screenshot's shape and leaked zod internals into provider requests).","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}