Talk To Experts

Talk to Experts

Launches 5 subagents IN PARALLEL - each consults one expert and returns the response. Your main context stays clean.

How It Works

1. You prepare the question + context
2. Launch 5 Task subagents in ONE message (all parallel, each with run_in_background: true)
3. Each subagent runs its Bash command to consult one expert
4. Collect results from all 5 subagents
5. Synthesize the expert opinions

Expert Council

Workflow

1. Prepare Context

CRITICAL: Experts need sufficient context.

Gather:

• Language/Framework versions (PHP 8.4, Laravel 12, Livewire 3)
• Database type and version
• Relevant file paths (absolute!)
• Problem description

2. Launch 5 Subagents in ONE Message

Launch ALL 5 Task calls in a SINGLE message. Each subagent runs ONE Bash command.

Task 1 - Grok:

subagent_type: "expert-consultant"
run_in_background: true
prompt: |
  Run this Bash command and return the output:

  opencode run "@expert [QUESTION_WITH_CONTEXT]" -m xai/grok-4-1-fast -f [FILES] --format json 2>&1 | jq -r 'select(.type == "text") | "response: \(.part.text)\nsessionid: \(.sessionID)"'

Task 2 - Kimi:

subagent_type: "expert-consultant"
run_in_background: true
prompt: |
  Run this Bash command and return the output:

  opencode run "@expert [QUESTION_WITH_CONTEXT]" -m opencode/kimi-k2-thinking -f [FILES] --format json 2>&1 | jq -r 'select(.type == "text") | "response: \(.part.text)\nsessionid: \(.sessionID)"'

Task 3 - Gemini:

subagent_type: "expert-consultant"
run_in_background: true
prompt: |
  Run this Bash command and return the output:

  opencode run "@expert [QUESTION_WITH_CONTEXT]" -m google/gemini-3-pro-preview -f [FILES] --format json 2>&1 | jq -r 'select(.type == "text") | "response: \(.part.text)\nsessionid: \(.sessionID)"'

Task 4 - MiniMax:

subagent_type: "expert-consultant"
run_in_background: true
prompt: |
  Run this Bash command and return the output:

  opencode run "@expert [QUESTION_WITH_CONTEXT]" -m openrouter/minimax/minimax-m2.1 -f [FILES] --format json 2>&1 | jq -r 'select(.type == "text") | "response: \(.part.text)\nsessionid: \(.sessionID)"'

Task 5 - GPT:

subagent_type: "expert-consultant"
run_in_background: true
prompt: |
  Run this Bash command and return the response + session ID:

  codex exec --profile expert --sandbox read-only "[QUESTION_WITH_CONTEXT]

  Files to analyze:
  [FILE_PATHS]" -m gpt-5.2 -c model_reasoning_effort=\"high\" 2>&1

IMPORTANT: codex MUST use --sandbox read-only to prevent any file modifications!

3. Collect Results

Use TaskOutput to collect all 5 responses. Each subagent returns:

• The expert's response
• The session ID for follow-ups

4. Synthesize Results

Present to user:

## Expert Council Results

### Individual Responses
**Grok:** [summary]
**Kimi:** [summary]
**Gemini:** [summary]
**MiniMax:** [summary]
**GPT:** [summary]

### Consensus
- [Points where experts agree]

### Divergent Views
| Expert | Position |
|--------|----------|

### Recommendation
[Best solution combining insights]

### Session IDs
- Grok: [id]
- Kimi: [id]
- Gemini: [id]
- MiniMax: [id]
- GPT: [id]

Follow-Up Conversations

opencode:

opencode run "Follow-up" -s SESSION_ID -m [MODEL] 2>&1

codex:

codex exec resume SESSION_ID "Follow-up" 2>&1

File Handling

opencode JSON Output Format

When using --format json, opencode outputs one JSON object per line (NDJSON). Example:

{"type":"step_start","timestamp":1765763743412,"sessionID":"ses_xxx",...}
{"type":"text","timestamp":1765763744943,"sessionID":"ses_xxx","part":{"type":"text","text":"The actual response",...}}
{"type":"step_finish","timestamp":1765763744961,"sessionID":"ses_xxx",...}

Parsing with jq (recommended):

Extract both response and session ID in one clean command:

opencode run "@expert [QUESTION]" -m [MODEL] --format json 2>&1 | jq -r 'select(.type == "text") | "response: \(.part.text)\nsessionid: \(.sessionID)"'

Output format:

response: The expert's answer here
sessionid: ses_xxxxxxxxxxxxx

This filters for the type="text" line and extracts both the response text and session ID in a clean, parseable format.

Rules

• Launch ALL 5 Tasks in ONE message - critical for parallelism
• Use run_in_background: true for each Task
• Use subagent_type: "expert-consultant" - custom agent with Bash(opencode *), Bash(codex *) permissions
• Each subagent runs ONE Bash command to consult ONE expert
• Always use absolute paths for files
• Always use --format json for opencode commands to get session IDs
• codex MUST use --sandbox read-only - prevents any file modifications!
• opencode expert agent has write/edit/bash disabled - already read-only
• Bash commands run INSIDE subagents, not in main thread
• Main context only receives final synthesized results

Required: Custom Agent

This skill requires the expert-consultant agent at agents/expert-consultant.md with these permissions:

tools: Bash(opencode *), Bash(codex *)