Codex + LSP

Use this setup when you want:

• OAuth with preset: codex
• Go and Python LSP support
• an extra read-only reviewer agent

The same configuration is available as a ready-to-copy file: codex-oauth-with-lsp.yaml.

The YAML fills in a current flagship so every field is present; pick the Codex model you will actually run in Choosing models. Complete Codex sign-in and confirm the model works first, then add LSP and the reviewer as needed. Remove unused providers, pools, or roles together with their references. Install the corresponding language servers before enabling LSP.

~/.config/chord/config.yaml

providers:
  codex:
    preset: codex
    type: responses
    models:
      gpt-5.5:
        limit:
          context: 400000
          input: 272000
          output: 128000
        reasoning:
          # Recommended example when you want readable reasoning summaries from
          # OpenAI Responses reasoning models. This is not Chord's implicit default:
          # leave `summary` unset to omit the field and use provider/model behavior.
          summary: auto
        # Optional for OpenAI GPT-5 / Responses API models. Leave unset to use
        # the provider/model default; set low for shorter visible text output or high
        # when you explicitly want more detailed visible output.
        # text:
        #   verbosity: low
        variants:
          high:
            reasoning:
              effort: high
          xhigh:
            reasoning:
              effort: xhigh
        modalities:
          input: [text, image]

  fast:
    type: chat-completions
    api_url: https://api.openai.com/v1/chat/completions
    models:
      gpt-5.4-mini:
        limit:
          context: 200000
          input: 183616
          output: 128000

model_pools:
  deep:
    - codex/gpt-5.5@high
    - codex/gpt-5.5
  fast:
    - fast/gpt-5.4-mini

lsp:
  gopls:
    command: gopls
    file_types: [".go"]
    root_markers: ["go.work", "go.mod", ".git"]
  pyright:
    command: pyright-langserver
    args: ["--stdio"]
    file_types: [".py", ".pyi"]
    options:
      python.analysis:
        typeCheckingMode: standard

context:
  compaction:
    threshold: 0.8
    model_pool: fast

desktop_notification: true
prevent_sleep: true
ime_switch_target: com.apple.keylayout.ABC
log_level: info

~/.config/chord/auth.yaml

Codex OAuth credentials are usually written by chord auth codex, so you do not normally hand-edit them. They are stored under the active provider name, so with the config above auth.yaml typically contains entries like:

codex:
  - refresh: "..."
    access: "..."
    expires: 1774009702606
    account_id: acc-1

Chord can also start from a refresh-only entry and populate access, expires, and identity metadata after the first refresh:

codex:
  - refresh: "..."

~/.config/chord/agents/reviewer.md

---
name: "reviewer"
description: "Read-only code reviewer"
mode: "subagent"
model_pools:
  - deep
permission:
  "*": deny
  read: allow
  view_image: allow
  grep: allow
  glob: allow
  shell:
    "*": allow
    "rm *": deny
    "mv *": deny
    "git add *": deny
    "git commit *": deny
    "git push *": deny
    "sudo *": deny
---
## Role

- Review recent code changes for correctness, risk, and missing verification.
- Stay read-only; do not modify project files.

A few practical points matter here:

• Most models only need limit.context: keep total input + output within that window.
• Some GPT models also have a separate input cap. Set limit.input for those models so Chord knows when to compact before the prompt is too large; otherwise it derives the input budget from limit.context minus the model’s own limit.output (falling back to the global output-cap default only when the model declares none).
• limit.output is the model’s own output capacity. Chord defaults max_output_tokens to 64000, so its output budget is the smaller of that cap and the model limit; changing the global cap does not increase the provider’s 272k input cap.
• Same-named models on different providers are still tried independently in the fallback chain; Chord does not skip them just because the model name matches.

Credentials to prepare

Run chord auth codex for the codex provider. If you keep the fast OpenAI-compatible provider, also provide an OpenAI API key under fast in auth.yaml or through the environment reference you choose.

Install optional LSP servers before relying on diagnostics:

go install golang.org/x/tools/gopls@latest
npm install -g pyright

Verify

chord doctor models --pool deep
chord doctor models --pool fast

Common failures

• OAuth opens but never completes: rerun chord auth codex in a local browser-capable environment, or use the device-code login path.
• fast pool fails while deep works: the OpenAI API key for the fast provider is missing or the model name differs on your account.
• LSP indicators stay unavailable: gopls or pyright-langserver is not installed on PATH.