gpt-5-6

Imported from rami-maalouf/fable-5-vs-gpt-5-6 (swift-rewrite/gpt-5-6/AGENTS.md). Install upstream with npx skills add rami-maalouf/fable-5-vs-gpt-5-6 --skill gpt-5-6. Copyright stays with the author.

Expo HAS CHANGED

Read the exact versioned docs at https://docs.expo.dev/versions/v57.0.0/ before writing any code.

Token Budget & Agent Cost Rules

To keep context small and costs low, observe these rules during UI builds:

1. Prefer Text over Pixels: Use describe or debugger-component-tree to verify state. Only use screenshot when a check is strictly visual (spacing, color, clipping).
2. Minimize Screenshot Weight: The system is configured with ARGENT_SCREENSHOT_SCALE=0.2. Use includeImageInContext: false when capturing baselines just for diffing.
3. Run Sequences: Use run-sequence for multi-step actions instead of tapping and screenshotting repeatedly. Use await-ui-element to wait for state to settle.
4. One Session Per Checkpoint: Use checkpoints.md as the handoff artifact. Start fresh sessions per checkpoint to avoid dragging the whole build history along.
5. Delegate Exploration: Use subagents to explore and dump files, returning only the conclusion to the main context.
6. Use the Cheapest Capable Subagent: When the active client supports subagents, delegate bounded exploration, file inspection, and log analysis to the least expensive available model that can reliably complete the task. Use only model identifiers supported by the active client.
7. Keep Heavy Device Work Isolated: When the active client supports subagents, delegate repeated Argent inspection loops to one bounded subagent and have it return a concise conclusion. Use the active client's own context-management features instead of assuming commands or tool parameters from another client.