L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Productivité · SKILL.md

council-execution

council-execution

Executes council queries by running the query pipeline across selected AI providers (Gemini, OpenAI, Grok, Perplexity), displaying formatted responses verbatim, and generating a synthesis of consensus, divergence, and recommendations. Invoked by the ask command during standard (non-agent) council queries.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: council-execution description: Executes council queries by running the query pipeline across selected AI providers (Gemini, OpenAI, Grok, Perplexity), displaying formatted responses verbatim, and generating a synthesis of consensus, divergence, and recommendations. Invoked by the ask command during standard (non-agent) council queries.

Council Query Execution

Step 1: Run Query and Save to File

bash ${CLAUDE_PLUGIN_ROOT}/scripts/run-council.sh --providers=gemini,openai -- "Your question"

This outputs the path to the saved file (e.g., .claude/council-cache/council-1734567890.md).

Flag syntax: Use = with no spaces: --providers=gemini,openai

Pass only the flags the user asked for. Never add --no-pane, --quiet, --no-cache or --no-auto-context on your own judgment; they are the user's to choose, and the script already skips a pane that cannot open.

CRITICAL: Always place -- before the prompt to prevent prompt text containing dashes from being parsed as flags.

Run this Bash call with timeout: 600000 (ten minutes). The Bash tool's two-minute default is shorter than a slow council round, and inside tmux a provider failure makes the streaming pane offer the user a retry for COUNCIL_RETRY_WAIT seconds (default 45) and then runs another provider round if accepted. A call cut off mid-run loses the whole transcript, the answers that did arrive included.

Step 2: Read and Display the Output VERBATIM

Use the Read tool to read the output file path returned by Step 1.

If the file's first line is <!-- council: round 1 was shown in the pane -->, the user already read every round-1 answer in the streaming pane. Do not reprint them. Instead, output one line per provider in round 1, taken from its ## header (name, role, model, and any fallback note) followed by answered or the Error: line under it. When the header says the seat fell back, add the > <provider>: <reason> line printed under it. Then display VERBATIM the rest of the file from its ## Round 2 heading, or from ## Synthesis when there is no round 2 (debate rebuttals never reach the pane), and continue with Step 3 under the ## Synthesis header. Without that first line, display the whole file as below.

CRITICAL: Display the file content EXACTLY as written. Do NOT:

  • Reformat or reinterpret any text
  • Add your own headers or structure
  • Summarize or abbreviate responses
  • Skip any lines including separator lines (---)

Simply copy-paste the entire file content into your response.

Step 3: Complete the Synthesis Section

The file ends with a ## Synthesis header. Read ${CLAUDE_PLUGIN_ROOT}/prompts/synthesis.md and write your synthesis UNDER that header following its structure (Consensus / Divergence / Recommendation) and calibration rules.

Step 4: Notify User of Saved Output

After displaying the synthesis, tell the user:


(use this emoji 💾) Full output saved to .claude/council-cache/council-TIMESTAMP.md (use the actual filename)

This lets them review the complete responses later.

Output Format: Provider Names

Each provider section opens with a ## ● <Provider> (<role>) - <model> header, where the dot is the provider's colour swatch. Display the header as written.

Skills enregistrées