Step 0 of every turn could only call rectification-read-case with the
default projection, a result the server can compute. The runner now runs the
agent's own read-case tool after the Skill is bound (same tool, receipts and
projection), passes the call and result to the model as the turn's opening,
and gives step 0 the full tool set. Synthetic chunks go through the same loop
(case.loaded, activity and Skill receipt once each) but not into step timings.
The synthetic call carries a short reasoning line: thinking-mode DeepSeek
rejects a prior tool call without reasoning_content (checked live). read-case
stays callable for resume / full_diagnostics. Agents without listTools keep
the step-0 read. The Skill package has no "read first" text; prompts updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
