Skip to content

Explicit selection on addressed comments

For the structured Web composer contract and direct-comment YAML defaults, see Structured Comment Execution.

For a daemon started with the Codex executor, place this directive on the first line of an addressed comment:

@MBPBot [dispatch model=gpt-6.1-sol effort=high]
Implement the requested change. Keep this task text exactly as written.

The entire first line is the header. Put the task on the next line. Both model and effort are required; each may appear once, in either order. The header is removed before prompt construction, while the task body and its whitespace are preserved. A plain mention has no override and continues to use the Codex CLI's configured defaults. Text such as model=... in the task, quoted examples, and directives on later lines are ordinary prompt text. This syntax selects the executor model; it does not switch an agent from Claude, Goose, or Pi to Codex.

Supported direct comment model IDs are gpt-6.1-sol, gpt-6-sol, gpt-6-astra, and gpt-6-luna. Supported efforts are low, medium, high, xhigh, and max. The exact OpenAI model page lists gpt-6.1-sol and confirms high. Codex's configuration reference documents model_reasoning_effort. The daemon rejects malformed, duplicate, unknown, or unsupported selections with a comment on the card. It never substitutes a different model.

This feature requires a Codex CLI and account that actually offer the selected model. The official OpenAI API model page establishes gpt-6.1-sol as an API model ID and lists high reasoning effort; it does not establish availability through Codex using a ChatGPT account. The repository Dockerfile currently pins Codex CLI 0.144.5. In this development environment, CLI 0.156.0 did not list gpt-6.1-sol in its model catalog, and a direct probe returned HTTP 400: “The 'gpt-6.1-sol' model is not supported when using Codex with a ChatGPT account.” The container's model support has therefore not been verified. A container rebuild alone is not a demonstrated fix for this account-level rejection. The agent's request and continuation keep the selected model and effort; a provider rejection remains a visible failure.

Operator migration checklist

Complete these checks with the operator before upgrading or starting a daemon with this feature:

  1. Identify the actual live rules file from the daemon's KARDBRD_AGENT_RULES_FILE setting, or its configured working directory's kardbrd.yml. Inspect its rules and schedules. The bundled file in this repository may differ from the live copy.
  2. For the applicable MBPBot workflow rules, stage a change from reasoning: xhigh to reasoning: high, retaining their current model IDs and all rule conditions/actions. Preserve intentional effort settings outside this MBPBot scope. Validate the staged file with kardbrd agent validate path/to/kardbrd.yml using the reviewed binary. Coordinate any live file replacement because ordinary rules hot-reload.
  3. In the actual target runtime and account, check codex --version, codex exec --help, the available model catalog (codex debug models where supported), and the exact probe below. A successful probe must finish without turn.failed or an HTTP error. Do not infer target-model access from the official API catalog or a newer CLI version alone.
  4. If the probe still reports the ChatGPT-account rejection, stop before enabling a rule or dispatching a comment that selects gpt-6.1-sol. The remaining operator decision is whether to obtain verified access to that exact model or defer its use. Do not silently substitute a model, change credentials/account, or retarget existing rules as part of this migration.
  5. After independent review, an approved release, successful compatibility verification, and the MBPBot live-config change to high, follow the normal deployment procedure. No rule activation or runtime restart is implied by this document.

Exact compatibility probe (run in the target container/account before pulling or restarting it):

codex --version
codex exec --help
codex exec --ephemeral --skip-git-repo-check --model gpt-6.1-sol --config 'model_reasoning_effort="high"' --json 'Reply OK only.'

The command selects the exact requested model and effort. The local CLI 0.156.0 probe failed with HTTP 400 as described above; the operator must resolve that compatibility decision before using this model in a live dispatch.