feat: add OpenAI-optimized foundation bundle and context

[DavidKoleczek]

Summary

• Add context/shared/common-openai-base.md: OpenAI-specific system instructions with editing constraints, task management guidelines, frontend coding rules, pre-commit validation, and presentation formatting
• Add experiments/exp-foundation-openai.md: Experimental foundation bundle configuration optimized for OpenAI models (removes agents behavior, references the new OpenAI context file)

Details

This PR introduces an experimental bundle configuration tailored for OpenAI models. The key changes include:

1. OpenAI-optimized context file (common-openai-base.md):

   • Editing constraints (ASCII preference, code comments guidelines)
   • Git worktree handling instructions
   • Task management behaviors
   • Frontend/web development guidelines
   • Pre-commit validation rules
   • Response presentation formatting

2. Experimental bundle (exp-foundation-openai.md):

   • Based on the standard foundation bundle
   • Removes the agents behavior for a simpler configuration
   • References the new OpenAI-specific context file

Test plan

• Verify bundle loads correctly with amplifier chat --bundle experiments/exp-foundation-openai.md
• Confirm context file is properly included in the system prompt
• Test with OpenAI models to validate instructions are followed

🤖 Generated with Amplifier

[bkrabach]

Thanks for taking this on, @DavidKoleczek — experimenting with OpenAI model optimization is valuable work. A few pieces of feedback:

1. Fix apply_patch reference (functional bug)

The context file references apply_patch for single file edits — that's a Codex CLI tool, not available in Amplifier. The model will try to call a tool that doesn't exist. Should reference edit_file/write_file instead.

2. DRY to actually reduce context size

If one hypothesis driving this is that OpenAI models struggle with large instruction context (~20-40k tokens), duplicating content from common-system-base.md and common-agent-base.md works against that goal — the experiment carries redundant instructions that inflate the very context it's trying to shrink.

A tighter approach: @-mention only the minimal shared base needed, then add a focused OpenAI-specific delta on top. That way the experiment is a truer test of "less context = better performance" and won't drift as the base files evolve.

3. Move context file out of context/shared/

The context/shared/ convention is for universal, provider-agnostic content (all 4 existing files there are used by every bundle). Since this is experiment-specific, context/experiments/ or co-locating it alongside the experimental bundle would be a better fit.

4. Document the hypotheses in the bundle

The reasoning behind the design choices (delegation overuse with GPT-5.2, context size sensitivity, model calibration for sub-agents) would be great to capture as comments in the experimental bundle itself. That makes the experiment self-documenting and reviewable — when we evaluate results, we'll know exactly what was being tested and why.

5. Agents behavior removal and includes fork are fine

Disabling delegation to test whether OpenAI models perform better without it is a valid experiment. The forked includes list is acceptable for an experimental bundle — just worth noting it was forked from bundle.md v2.0.0 so future maintainers know when to sync.