Summary

During live testing (CI-350, 2026-03-31), the rust-agent-handoff skill was selected at confidence 0.14 for a generic memory update prompt:

INFO zeph_core::agent::context::assembly: disambiguation selected skill skill=rust-agent-handoff confidence=0.14100000262260437

Prompt: "Now update: Alice was promoted and now works as a Senior Engineer at Acme Corp."

Observed Behavior

With disambiguation_threshold = 0.05 in testing.toml, rust-agent-handoff (0.141 > 0.05) passes the threshold and its SKILL.md context is injected. The skill's content references .local/handoff/ YAML files, causing the agent to:

1. Attempt edit on .local/handoff/Alice_work_update.yaml (file not found)
2. Call find_path to search for handoff YAML files
3. Eventually call memory_save as a fallback

Expected behavior: memory_save called directly without filesystem detour.

Impact

• Wasted tool calls (2 extra tool invocations before correct action)
• Longer response latency
• Potential user confusion from unexpected tool output
• disambiguation_threshold = 0.05 is too permissive for production use

Root Cause

disambiguation_threshold = 0.05 is intentionally low in testing.toml to exercise skill matching. However, skills like rust-agent-handoff contain strong filesystem/YAML operation patterns that redirect generic update tasks to file operations.

Suggestions

1. Raise disambiguation_threshold to at least 0.25 in testing.toml (and production defaults) — matching at 0.14 is noise
2. Consider adding a "scope guard" or intent classifier to reject skill injection when the user intent clearly maps to a core capability (memory, search, etc.)
3. Alternatively, add negative examples to rust-agent-handoff SKILL.md to prevent triggering on generic update/save prompts

Config

• testing.toml, disambiguation_threshold = 0.05
• cosine_weight = 0.7
• 27 skills loaded