Describe the bug

Starting in v1.25.0, an agent running a recipe via goose run --recipe systematically deviates from the workflow defined in recipe.instructions. The same recipe ran correctly in v1.24.0 with no changes to the recipe file.

The deviation is consistent and reproducible: the agent ignores prescribed phases, skips required steps, and reverts to default helpful-assistant behavior. Drift is most pronounced after context compaction but also occurs in fresh sessions.

To Reproduce

1. Use the goose-coder.yaml recipe posted in the comments below
2. Run with v1.24.0: goose run --recipe goose-coder.yaml — agent follows the workflow exactly
3. Run with v1.25.0: same command, same recipe — agent deviates systematically

Expected behavior

Agent follows the multi-step workflow defined in recipe.instructions as it did in v1.24.0.

Please provide the following information

• OS & Arch: Linux x86_64
• Interface: CLI
• Version: v1.25.0 (last working: v1.24.0)
• Extensions enabled: developer, context7, brave_search, aptu
• Provider & Model: Google – Claude Sonnet (GCP Vertex AI)

Additional context

The recipe.instructions injection path appears unchanged between v1.24.0 and v1.25.0 (extract_from_cli.rs:55 → additional_system_prompt → extend_system_prompt("additional", ...)). We have not isolated the root cause.

Commit 12a346fec (#6991, v1.25.0) is a candidate — it shortened system.md and removed desktop_recipe_instruction.md. We suspect the reduction in base system prompt scaffolding weakens the model's tendency to follow structured multi-step workflows, but we have not confirmed this is the cause.

Workaround: Prepend the following to the recipe instructions field:

IMPORTANT: Follow these instructions exactly and validate each action against the workflow before proceeding.

Recipe: see first comment below.