Empty model responses after load_skill with Gemini (skill system instruction)

[lwangverizon]

Empty model responses after load_skill with Gemini (skill system instruction)

Summary

When an agent uses the skill tools (SkillToolset), ADK appends a skill system
instruction to the prompt. With Gemini models we observed a significant rate of
empty model responses immediately after a load_skill tool call: the model
loads the skill, then ends its turn with no visible output to the user.

We observed this as a spike in empty responses in a Verizon production environment
running ADK + Gemini. The empty turns were most frequent for tool-heavy skills,
where the correct next action after loading is to call more tools rather than to
reply.

Affected component

src/google/adk/tools/skill_toolset.py — _build_skill_system_instruction(),
which produces DEFAULT_SKILL_SYSTEM_INSTRUCTION.

Root cause

The skill system instruction never states what the model should do after a
load_skill call returns. There is no rule telling the model that load_skill is a
non-terminal step, so some models (notably Gemini) treat the load_skill tool call
as the entire turn — they load the skill and stop, emitting no content.

(The existing rule 2 says to follow loaded instructions "before replying to the
user", which reinforces this reading, but the core gap is the missing
"continue after load" guidance.)

Impact

• User-visible empty responses (the model appears to "do nothing").
• For deployments with an empty-response retry backstop, each occurrence costs an
  extra LLM pass.

Proposed fix

Add a rule 7 to the skill system instruction that explicitly states load_skill
only retrieves instructions and does not complete the turn — the model should
continue in the same turn (calling whatever tools the skill's steps require) and
never end with an empty response right after loading a skill. Rules 1–6 are left
unchanged.

We deployed exactly this guidance (appended rule 7, rules 1–6 untouched) in
production and saw the empty-response rate for Gemini after skill loading drop
substantially.

A PR implementing this is attached.

Environment

• ADK: current main
• Model: Gemini (observed with Gemini 3.5 Flash)