config.codex.example.yml

# Perspicacité — OpenAI Codex CLI subscription preset
#
# Routes every internal LLM call through OpenAI's `codex` CLI as a
# subprocess. Your ChatGPT Plus/Pro/Team subscription (the one bundled
# with Codex CLI access) pays for inference; no OPENAI_API_KEY needed
# in this config.
#
# **Verified against codex-cli 0.130.0** (run `codex --version` to
# check yours). If your version differs, run `codex exec --help` and
# adjust the flags below.
#
# **Prerequisites:**
# 1. Install Codex CLI:  `npm install -g @openai/codex`
# 2. Sign in:  `codex login`  (browser-based ChatGPT auth)
# 3. Verify:   `echo "say hi" | codex exec --skip-git-repo-check`
# 4. Use this file as your config:
#       perspicacite -c config.codex.example.yml serve
#
# **Caveat — Codex is an agent, not a pure completion endpoint.**
# Each call spins up Codex's full session machinery (sandbox, tool
# loop, memory). Per-call latency is ~5–15 s even for tiny prompts
# (we measured ~9 s for a trivial JSON response). For heavy
# multi-paper ingests, expect minutes-to-hours and shared rate limits
# with your interactive Codex session.
#
# Same general caveats as the other agent-CLI presets: no prompt
# caching, no per-call temperature/max_tokens, no streaming.

llm:
  default_provider: "agent_cli"
  default_model:    "gpt-5.5"           # whatever your Codex subscription supports

  # Use the same provider for every stage; Codex doesn't expose
  # discrete "cheap" model tiers via the CLI today.
  providers_per_stage:
    routing:          "agent_cli"
    screening:        "agent_cli"
    rephrase:         "agent_cli"
    contextual:       "agent_cli"
    synthesis_basic:  "agent_cli"
    synthesis_heavy:  "agent_cli"

  providers:
    agent_cli:
      base_url: ""
      executable: "codex"
      timeout: 300                      # Codex is slow per call — give it room
      max_retries: 1

      # Codex's `exec` subcommand is the non-interactive entry point.
      # `--skip-git-repo-check` lets it run outside a git repo.
      # `--sandbox read-only` prevents any model-generated shell
      # commands from touching disk (we just want a completion).
      # `--ephemeral` skips persisting session files for each call.
      extra_args:
        - "exec"
        - "--skip-git-repo-check"
        - "--sandbox"
        - "read-only"
        - "--ephemeral"

      # Codex reads the prompt from stdin when the positional arg is
      # omitted. Stdin is cleaner than passing the prompt as a CLI
      # arg (avoids shell quoting hazards on long prompts).
      prompt_via: "stdin"

      # No `--system` flag in Codex 0.130. The system prompt gets
      # prepended to the body inside _flatten_messages.
      system_flag: null

      # Model selection: `codex exec --model <NAME>`. Set
      # `default_model` above (or in per-stage `models`) to whatever
      # your Codex account supports — `gpt-5.5`, `gpt-5`, `o4-mini`,
      # etc. Run `codex` interactively once to see your options.
      model_flag: "--model"

      # Cleanest extraction: `--output-last-message <FILE>` writes
      # ONLY the final assistant text to a tempfile. Avoids parsing
      # Codex's banner / token-counter / session-id stdout.
      output_file_flag: "--output-last-message"
      output_format: "text"
      result_json_path: null

      model_aliases: {}
      env_extra: {}