cli-model-chat

name	cli-model-chat
description	Use when Codex needs to ask DeepSeek v4 Pro a focused one-to-one question through a local Python wrapper instead of MCP. Default to DeepSeek for second-model consultation unless the user explicitly names Codex, codex exec, or the Codex provider; only then use the isolated Codex CLI backend.

CLI Model Chat

Overview

Use scripts/cli_model_chat.py to send a prompt to a single external model and print only the assistant response. The default provider is DeepSeek's OpenAI-compatible API with model deepseek-v4-pro, thinking mode enabled, reasoning_effort=high, max_tokens=384000, non-thinking temperature=1.3, and local env loading from skills/cli-model-chat/.env.

Required Rules

• Never hardcode API keys in repo files, skill files, shell history, or documentation.
• Pass DeepSeek credentials through DEEPSEEK_API_KEY, preferably in the local ignored env file skills/cli-model-chat/.env.
• Keep this as a one-to-one chat bridge. Do not add MCP orchestration, multi-agent workflows, consensus, or background subagents unless the user asks.
• If the user does not explicitly mention Codex, codex exec, or the Codex provider, use DeepSeek. Omit --provider or pass --provider deepseek.
• Use --provider codex only when the user explicitly asks for Codex as the second model/backend.
• Treat --provider codex as a single-turn isolated conversation, not as a repo agent. It must run in a temporary empty directory with --ephemeral, --skip-git-repo-check, --ignore-rules, --ignore-user-config, and --sandbox read-only.
• The Codex provider must also inject a safety instruction telling the child Codex not to inspect the filesystem, run commands, edit files, open network resources, or use tools.
• Do not pass the current repo path, project files, secrets, or broad filesystem context to the Codex provider. If file context is needed, paste the exact excerpt into the prompt.
• Do not use --codex-allow-user-config unless the user explicitly accepts loading their Codex config/profiles for this one run.
• Prefer --list-models before changing the default model.
• Use exact V4 model ids: deepseek-v4-pro for hard reasoning/coding review, deepseek-v4-flash only when speed/cost matter more.
• Do not use legacy aliases deepseek-chat or deepseek-reasoner; DeepSeek says they currently route to V4-Flash modes and are scheduled for retirement.
• Do not expose or tune a --max-tokens CLI option. Always pass DeepSeek V4's maximum available output budget, currently max_tokens=384000, so reasoning and final output are not clipped by this wrapper.
• Use --json only when the caller needs structured metadata; otherwise stdout should be the model's text answer.

DeepSeek V4 Usage Guidance

Use deepseek-v4-pro when the second model is expected to find mistakes, reason through code, review architecture, compare alternatives, or answer from large pasted context. Official DeepSeek docs describe V4-Pro as a 1.6T total / 49B active MoE model with stronger agentic coding, world knowledge, and Math/STEM/Coding reasoning than current open models; both V4-Pro and V4-Flash support 1M context and the OpenAI ChatCompletions API.

Use deepseek-v4-flash for quick sanity checks, simple summarization, and low-cost drafts. DeepSeek describes Flash as faster and more economical, with reasoning close to Pro and simple agent-task performance on par with Pro, but this skill defaults to Pro because its purpose is high-signal second-model consultation.

For DeepSeek V4:

• Keep thinking mode enabled for review, debugging, technical planning, and non-trivial reasoning.
• Use --reasoning-effort max only for difficult code review, ambiguous debugging, or prompts where paying extra latency/tokens is justified.
• Use --thinking disabled for deterministic extraction, short rewrites, or when temperature should actually affect sampling.
• Do not tune --temperature while thinking is enabled; DeepSeek documents that temperature/top_p/presence/frequency penalties have no effect in thinking mode.
• In non-thinking mode, default to temperature=1.3 for ordinary one-to-one conversation because DeepSeek's temperature guide recommends 1.3 for general conversation.
• Override non-thinking temperature by task: 0.0 for coding/math, 1.0 for data cleaning/analysis, 1.3 for general conversation/translation, and 1.5 for creative writing/poetry.
• Leave output budget to the wrapper. It always sends the maximum V4 output allowance instead of a caller-controlled cap.
• Ask one focused question per call. Provide the relevant files, error, constraints, and desired output shape directly in the prompt.
• For long sessions, prefer explicit --session names and periodically restart with --new-session when old context is no longer useful.
• Check --json usage data on expensive prompts; DeepSeek bills by actual input/output tokens, not by the requested maximum output cap.

Quick Start

Ask DeepSeek, the default provider. Prefer omitting --provider when DeepSeek is intended:

python3 skills/cli-model-chat/scripts/cli_model_chat.py "Reply with OK only."

Ask an isolated single-turn Codex CLI process:

python3 skills/cli-model-chat/scripts/cli_model_chat.py --provider codex "Reply with OK only."

Use a Codex model without loading user config:

python3 skills/cli-model-chat/scripts/cli_model_chat.py \
  --provider codex \
  --model gpt-5 \
  "Give a second opinion on this design decision: ..."

List available DeepSeek models:

python3 skills/cli-model-chat/scripts/cli_model_chat.py --list-models

Ask the default DeepSeek deepseek-v4-pro model:

python3 skills/cli-model-chat/scripts/cli_model_chat.py "Reply with OK only."

Ask V4-Pro with maximum thinking effort:

python3 skills/cli-model-chat/scripts/cli_model_chat.py \
  --reasoning-effort max \
  "Review this patch for correctness risks: ..."

Use non-thinking mode for simple short-form output where sampling controls matter:

python3 skills/cli-model-chat/scripts/cli_model_chat.py \
  --thinking disabled \
  --temperature 0 \
  "Rewrite this sentence in plain English: ..."

Use a system prompt:

python3 skills/cli-model-chat/scripts/cli_model_chat.py \
  --system "You are a concise reviewer." \
  "Review this decision in three bullets: ..."

Keep a lightweight local conversation:

python3 skills/cli-model-chat/scripts/cli_model_chat.py \
  --session deepseek-review \
  "Remember: we are evaluating a local CLI wrapper."

Session files are stored under ~/.local/state/cli-model-chat/deepseek/ and should not be committed.

Provider Selection

Default:

deepseek

Use DeepSeek when the task needs a strong non-Codex second opinion with long context, thinking mode, and explicit model/token accounting.

Use Codex only when the user explicitly asks for Codex, codex exec, or the Codex provider. The wrapper invokes codex exec, writes the final answer through --output-last-message, and runs from a temporary empty directory. By default it ignores Codex user config so MCP servers, repo rules, and profiles are not loaded.

Codex CLI does not currently expose a wrapper-level switch that proves "no filesystem tools exist" inside the child agent. This skill therefore uses layered containment: empty temporary cwd, read-only sandbox, ignored rules, ignored user config, ephemeral session, and an explicit prompt instruction forbidding filesystem, shell, network, edit, or tool use. Treat this as a plain-chat wrapper, not an execution agent.

Codex profile usage is intentionally gated:

python3 skills/cli-model-chat/scripts/cli_model_chat.py \
  --provider codex \
  --codex-allow-user-config \
  --codex-profile my-profile \
  "..."

Only use this if the user explicitly wants profile behavior and accepts the weaker isolation boundary.

DeepSeek Model Selection

Default:

deepseek-v4-pro

Override per run:

python3 skills/cli-model-chat/scripts/cli_model_chat.py --model deepseek-v4-flash "..."

Override by environment:

DEEPSEEK_MODEL=deepseek-v4-pro python3 skills/cli-model-chat/scripts/cli_model_chat.py "..."

Failure Handling

• If DEEPSEEK_API_KEY is missing, stop and ask the user to provide it through the environment or a local secret mechanism.
• If model selection fails, run --list-models and use an exact returned model id.
• If the API returns a non-2xx response, report the status and response body without printing credentials.

Source Notes

• DeepSeek V4 API release: https://api-docs.deepseek.com/news/news260424
• DeepSeek model ids, context, features, and pricing: https://api-docs.deepseek.com/quick_start/pricing
• DeepSeek thinking mode semantics: https://api-docs.deepseek.com/guides/thinking_mode
• DeepSeek chat completion parameters: https://api-docs.deepseek.com/api/create-chat-completion