Agent Council

Run the local Agent Council CLI for questions that benefit from distinct perspectives, peer critique, and synthesis.

When to Use

Use this skill when the user explicitly requests an Agent Council or broader perspectives, or when the user agrees that a paid multi-agent run is appropriate.

Do not use it by default for simple factual questions, routine edits, or high-stakes decisions where qualified human review is required.

Prerequisites

• Python 3.11 or newer.
• Completed python scripts/install.py setup.
• An OpenAI API project with access to the selected GPT-5.6 preview model.
• Sufficient limits in config/limits.yaml.

Workflow

1. Choose the Model

• gpt-5.6-sol: flagship model and default.
• gpt-5.6-terra: lower-cost balanced option.
• gpt-5.6-luna: most economical supported option.

If the user did not request a model, use gpt-5.6-sol. Tell the user that GPT-5.6 currently requires preview access before starting a run.

2. Run the Session

The session runner performs the required buffered spend check itself:

python scripts/run_session.py \
  --question "the user's question" \
  --model gpt-5.6-sol

For question text that should not appear in shell history, write it to a temporary UTF-8 text file and use:

python scripts/run_session.py \
  --question-file /path/to/question.txt \
  --model gpt-5.6-sol

With context:

python scripts/run_session.py \
  --question-file /path/to/question.txt \
  --files /path/to/requirements.md /path/to/constraints.yaml \
  --model gpt-5.6-sol

Supported context formats are .txt, .md, .json, .yaml, .yml, and .csv. Do not pass PDFs, presentations, images, audio, or other binary files.

If the spend check blocks the run, stop and report the configured limit. Do not bypass it or edit the user's budget without approval.

3. Read the Results

The runner prints the exact result path. Read results.md from that session directory. Do not search for or read the credential file.

The result contains:

• the selected council and strategy;
• each specialist's response;
• peer reviews of the other proposals;
• the final synthesis;
• the locally estimated standard token cost.

4. Report to the User

Lead with the synthesis, then call out:

1. areas of agreement;
2. material disagreements;
3. the strongest evidence or reasoning;
4. important uncertainties and missing evidence;
5. recommended next actions.

Do not present council agreement as proof. Keep factual caveats and domain-review requirements visible.

Credential and Data Rules

1. Never read, print, copy, or summarize ~/.config/agent-council/credentials.yaml.
2. Never ask the user to paste an API key into chat or a command argument.
3. Use only scripts/run_session.py and scripts/check_spend.py for sessions and estimates.
4. Treat the question, context, and session output as potentially sensitive.
5. Remember that questions and context are sent to the OpenAI API; specialists may also use hosted web search.
6. Do not commit sessions/, config/limits.yaml, credentials, or generated transcripts.

The credential file's Unix permissions reduce accidental exposure but do not isolate it from processes with the same user privileges.

Cost Boundaries

Pricing is dated 2026-07-10 and can change. The local ledger estimates standard input and output token charges. It does not include tool fees, priority processing, cached-input adjustments, or other billing modifiers. Use the API dashboard for authoritative usage and cost.

The spend check is a buffered preflight guard. It cannot stop a request already in progress and is not a billing guarantee.

Failure Handling

• Model access error: Explain that GPT-5.6 is limited preview and the API project may not be entitled.
• Missing configuration: Run or recommend python scripts/install.py; do not inspect the key.
• Unsupported context: Convert the source to a supported UTF-8 text format or omit it.
• Spend limit exceeded: Stop and show the limit message.
• API or tool failure: Report the failed stage and preserve any local output that was successfully written.

See references/TROUBLESHOOTING.md and references/EXAMPLES.md for concise operator examples.