LLM-enhanced blueprint fails with from_json error when the model wraps output in a ```json code fence

[yosef-chai]

Summary

The LLM-enhanced local assist blueprint parses the model response with a bare
from_json:

llm_result: '{{ result.response.speech.plain.speech | from_json }}'

The prompt instructs the model to reply with raw JSON ("So also no code tags.
Only the JSON!"), but several models — most notably Google Generative AI
(Gemini) — frequently ignore this and wrap the payload in a Markdown code
fence:

```json
{"action_data": {"media_id": "...", "media_type": "track"}, "media_description": "..."}
```

from_json then fails on the leading ```json and the automation aborts:

ValueError: Template error: from_json got invalid input '```json
...
'

The automation never reaches the play_media step, so nothing plays. This is
intermittent because whether the model adds the fence varies run-to-run for the
same prompt.

Environment

Component	Version
Blueprint	llm-enhanced-local-assist-blueprint (all language variants: _en, _fr, _hu, _nl, community translations)
Home Assistant Core	2026.6.0
Conversation agent	Google Generative AI (Gemini)

Steps to reproduce

1. Use the blueprint with a Google Generative AI conversation agent.
2. Issue a normal music request (e.g. "play ").
3. Intermittently the agent returns the JSON inside a ```json fence and
   the automation errors at the from_json step; the request silently does
   nothing.

Root cause

The blueprint trusts the prompt's "no code tags" instruction and does no
defensive sanitization before from_json. Prompt instructions are not a
guarantee, and Gemini in particular tends to Markdown-format JSON.

Proposed fix

Strip an optional leading/trailing Markdown code fence before parsing. This is
backward compatible (raw JSON is unaffected) and tolerant of the most common
deviation:

llm_result: >-
  {% set r = result.response.speech.plain.speech | string %}
  {% if r.startswith("```") %}
    {% set r = r.split("\n", 1)[1] %}
    {% if r.rstrip().endswith("```") %}{% set r = r.rstrip()[:-3] %}{% endif %}
  {% endif %}
  {{ r | from_json }}

(A regex-based variant — e.g. r | regex_replace('^```(json)?\\s*|\\s*```$', '')
— works too; the explicit form above avoids edge cases with fences appearing
mid-string.)

Verified locally: with this change, both raw-JSON and ```json-wrapped
responses from Gemini parse correctly and playback proceeds.

Impact

• Intermittent, hard-to-diagnose failures for users on Google Generative AI
  (one of the most common HA conversation agents).
• The failure is silent from the user's perspective (assistant says nothing /
  generic response, no music), which makes it look like the voice pipeline is
  broken rather than a parsing issue.