content-model

Request: $ARGUMENTS ## The rule Anthropic models engineer here. Text that reaches a reader — and every verdict about such text — comes from GPT (`codex`) or Gemini (`agy`) through `.claude/scripts/text_model.py`. No engine → exit 2 → stop and report; never fill the gap yourself. This does not cover code, comments, commit messages or this harness. ## How Delegate the run to `text-broker` with: the prompt, the input file(s), the schema when the answer must be checkable, the engine and effort. One call per batch of 10–30 items, never per item. Before more than 5 calls, state the count and wait. | You want | Engine | Cost | |---|---|---| | a draft or a rewrite of a few sentences | `auto` | 1 call | | a verdict that keeps / rewrites / drops items | `both`, with a schema | 2 calls; a removal counts only where both agree | | the strongest case against a plan or a note | `both`, effort `high` | 2 calls | | structured extraction (ids in, ids out) | `gpt` or `gemini`, with a schema | 1 call | The isolation is the door's job: `codex exec --sandbox read-only --ephemeral` in an empty working directory, `agy --sandbox`; only the files you pass as `--input` are seen. Models: GPT from the user's codex config (`TEXT_MODEL_GPT` overrides), Gemini `gemini-3.1-pro-<effort>` (`TEXT_MODEL_GEMINI` overrides, `gemini-*` only). ## What comes back Mechanically validated (parses, every id returned, placeholders and backticks intact, lengths hold) — then it enters the product through the normal ...