fix: prevent empty responses after load_skill

Merge #6226

# fix(skills): prevent empty responses after `load_skill`

Fixes #6225

## Problem

When an agent uses the skill tools, ADK appends a skill system instruction to the
prompt (`_build_skill_system_instruction` in
`src/google/adk/tools/skill_toolset.py`). The instruction does not tell the model
what to do *after* a `load_skill` call returns. With Gemini models we observed a
significant rate of **empty model responses** right after a `load_skill` call — the
model loads the skill and ends its turn with no visible output.

We saw this as a spike in empty responses in a Verizon production environment running
ADK + Gemini, most acutely for tool-heavy skills whose next correct action after
loading is to call more tools rather than reply.

## Change

`_build_skill_system_instruction` only — covers both `DEFAULT_SKILL_SYSTEM_INSTRUCTION`
and the per-toolset prefixed variant.

**Add rule 7** (continue-after-load guard), appended after the existing rule 6:

> 7. Loading a skill only retrieves its instructions; it does NOT complete your turn.
> After a `load_skill` call returns, continue in the SAME turn: call whatever tools
> the skill's steps require (search, data retrieval, render), then write your reply.
> Never end your turn with an empty response right after loading a skill.

No other rule is modified — rules 1–6 are left exactly as-is. The new rule uses the
`{prefix}` substitution like the other rules (`{p}load_skill`), so the prefixed system
instruction carries it too.

## Result

This is the exact change we validated in production: appending rule 7 (rules 1–6
unchanged) dropped the empty-response rate for Gemini after skill loading
substantially.

## Tests

Adds focused tests in `tests/unittests/tools/test_skill_toolset.py`:
- `test_system_instruction_marks_load_skill_as_non_terminal` — rule 7 present in the
  default instruction.
- `test_prefixed_system_instruction_includes_continue_after_load_rule` — prefixed
  variant carries rule 7 with the prefix.

Existing tests reference `DEFAULT_SKILL_SYSTEM_INSTRUCTION` by constant (not hardcoded
text), so they are unaffected.

## Backwards compatibility

Prompt-text-only, strictly additive change. No public API, signature, or existing-rule
change.

Co-authored-by: Yifan Wang <wanyif@google.com>
COPYBARA_INTEGRATE_REVIEW=#6226 from lwangverizon:fix/skill-load-empty-response 663b28b
PiperOrigin-RevId: 940550604

Author: 2 people

src/google/adk/tools/skill_toolset.py

@@ -97,6 +97,11 @@ def _build_skill_system_instruction(prefix: str | None = None) -> str:
       f"6. If `{p}run_skill_script` returns an error (for example "
       f"`SCRIPT_NOT_FOUND`), do not retry the same script or guess a "
       "different script path. Report the error to the user and stop.\n"
+      f"7. Loading a skill only retrieves its instructions; it does NOT "
+      f"complete your turn. After a `{p}load_skill` call returns, continue "
+      "in the SAME turn: call whatever tools the skill's steps require "
+      "(search, data retrieval, render), then write your reply. Never end "
+      "your turn with an empty response right after loading a skill.\n"
   )
 
 

tests/unittests/tools/test_skill_toolset.py

@@ -1758,6 +1758,24 @@ def test_system_instruction_references_run_skill_script():
   )
 
 
+def test_system_instruction_marks_load_skill_as_non_terminal():
+  """Rule 7 must tell the model load_skill does not complete the turn.
+
+  Without it, some models (notably Gemini) treat the load_skill tool call as
+  the entire turn and stop with no visible output, producing empty responses.
+  """
+  instruction = skill_toolset.DEFAULT_SKILL_SYSTEM_INSTRUCTION
+  assert "does NOT complete your turn" in instruction
+  assert "empty response" in instruction
+
+
+def test_prefixed_system_instruction_includes_continue_after_load_rule():
+  """The prefixed builder variant must also carry rule 7 (with the prefix)."""
+  instruction = skill_toolset._build_skill_system_instruction(prefix="my")
+  assert "does NOT complete your turn" in instruction
+  assert "my_load_skill" in instruction
+
+
 # ── Finding 2: empty files are mounted (not silently dropped) ──