[shinpr]

Summary

Required skill reads can lose instructions when either the shell output or the combined Code Mode response is truncated. Add a two-sentence instruction to the existing Required Skills sections in 18 recipes and 25 subagents: measure files with wc -c, budget both output layers, split oversized reads upfront, and recover truncated content before proceeding.

Bump the patch version from 1.8.4 to 1.8.5.

Validation

• npm test: 23 Node tests and 6 Python tests passed.
• All 25 agent TOML files parsed; git diff --check passed.
• Fresh startup runs with gpt-6.1-sol and high reasoning effort confirmed the actual model settings, size checks before reads, explicit inner and outer budgets, and received text identical to the complete source files:

Target	Required skill bytes	Inner / outer token budget	Result
recipe-plan	48,855	20,000 / 20,000	Full text received without truncation
technical-designer subagent	38,138	16,000 / 16,000	Full text received without truncation

The recipe test temporarily removed allow_implicit_invocation: false in a disposable copy, then restored it. The subagent test explicitly selected the changed agent definition. These runs cover startup and required-skill loading, not the full planning or design workflow.