Repository navigation
docs(llm-commits): bump Codex commit-generation model to gpt-5.6-luna - #3430
Conversation
worktrunk-bot
left a comment
There was a problem hiding this comment.
Verified the model swap against OpenAI's Codex models page: gpt-5.6-luna exists and is described as "Fast and affordable GPT-5.6 model that delivers strong capability at the lowest cost in the family" — the fastest/lowest-cost of the current 5.6 family (Luna/Terra/Sol), and gpt-5.4-mini now sits under the previous-generation "Other models" tier. So the pick and the drift claim both check out, and the 12→12-char literal swap preserves snapshot alignment. All 11 active references updated; the two remaining gpt-5.4-mini hits are the new changelog entry (describing the bump) and the historical #2949 entry (correctly left alone).
One nit in scope: the prose right after the code block still calls it "the fast mini model", but gpt-5.6-luna is the Luna variant, not a mini model — "mini" was tied to the old gpt-5.4-mini name. Suggestion below on the primary source (docs/content/llm-commits.md); the same line in skills/worktrunk/reference/llm-commits.md is an auto-synced mirror, so it'll pick this up when test_docs_are_in_sync regenerates.
| command = "codex exec -m gpt-5.6-luna -c model_reasoning_effort='low' -c system_prompt='' --sandbox=read-only --json - | jq -sr '[.[] | select(.item.type? == \"agent_message\")] | last.item.text'" | ||
| ``` | ||
|
|
||
| Uses the fast mini model with low reasoning effort and an empty system prompt for faster output. Requires `jq` for JSON parsing. See [Codex CLI docs](https://developers.openai.com/codex/cli/). |
There was a problem hiding this comment.
| Uses the fast mini model with low reasoning effort and an empty system prompt for faster output. Requires `jq` for JSON parsing. See [Codex CLI docs](https://developers.openai.com/codex/cli/). | |
| Uses the fast, low-cost model with low reasoning effort and an empty system prompt for faster output. Requires `jq` for JSON parsing. See [Codex CLI docs](https://developers.openai.com/codex/cli/). |
gpt-5.6-luna is the Luna variant, not a mini model, so "fast mini model" is now stale — this keeps it accurate without re-pinning to a name that drifts again next bump.
…#3867) The Codex setup section in the LLM-commits docs describes the pinned model as "the fast mini model", but the command it sits under pins `gpt-5.6-luna` — which is not a `-mini` model. The prose was written in #837 when the pin was a mini variant and was left behind when #3430 bumped `gpt-5.4-mini` → `gpt-5.6-luna`. This rewords the sentence to describe what the command actually pins, following #3430's own framing of `gpt-5.6-luna` as "the fast/low-cost variant of the current recommended (5.6) family". `docs/src/content/docs/llm-commits.md` is the primary (non-command doc, per the sync taxonomy in `docs/CLAUDE.md`); the two mirrors under `skills/` and `plugins/` were regenerated by `cargo test --test integration test_docs_are_in_sync`, which now passes. The branch was rebased onto `main` after #3866 relocated the docs tree, so the edit lands at the post-#3866 path rather than the old `docs/content/` one. No regression test: the change is a single prose sentence with no behavior behind it. The sync test already pins the mirrors to the primary, which is the only mechanical invariant here. <details><summary>Nightly sweep context</summary> Found during the rolling survey (bucket 23/28), reviewing `skills/worktrunk/reference/llm-commits.md`. Provenance: ``` $ git log --oneline -1 -S 'Uses the fast mini model' -- docs/content/llm-commits.md cbe65a6 docs: add Codex config and optimize Claude Code for LLM commits (#837) $ git log --oneline -1 -S 'gpt-5.6-luna' -- docs/content/llm-commits.md c532d2d docs(llm-commits): bump Codex commit-generation model to gpt-5.6-luna (#3430) ``` A repo-wide `grep -i mini` turned up no other stale reference to the old pin — the model string itself is consistent across `src/cli/mod.rs`, `dev/config.example.toml`, `Taskfile.yaml`, and `src/output/commit_generation.rs`; only this one sentence describes it. </details> Co-authored-by: worktrunk-bot <254187624+worktrunk-bot@users.noreply.github.com>
Summary
Weekly LLM-model-name check found drift in the recommended Codex commit-generation command. OpenAI's models page has moved
gpt-5.4-miniinto the previous-generation "Other models" tier; the current recommended (5.6) family's fastest/lowest-cost variant isgpt-5.6-luna("strong capability at the lowest cost in the family"). This bumps the recommendationgpt-5.4-mini→gpt-5.6-luna.Anthropic's Haiku pin (
--model=haiku/claude-haiku-4.5) is not drifted — Haiku 4.5 is still the current fastest Claude model — so it's left unchanged.Changes
Source of truth (
after_long_helpinsrc/cli/mod.rs) plus the runtime source (dev/config.example.toml) andTaskfile.yaml, with the generated mirrors and snapshots updated to match:src/cli/mod.rs,dev/config.example.toml,Taskfile.yamldocs/content/{config,llm-commits}.md,skills/worktrunk/reference/{config,llm-commits}.mdsrc/output/commit_generation.rssnapshot,tests/snapshots/…help_config_{create,long}.snapgpt-5.4-miniandgpt-5.6-lunaare both exactly 12 characters, so the snapshot table alignment is preserved by the literal swap. The historical CHANGELOG entry for the earliergpt-5.4-minibump (#2949) is intentionally left as-is.The "fastest/smallest current variant" pick is a judgment, and I could not verify
gpt-5.6-lunaagainst an installedcodexCLI — this CI sandbox has nocodexand no Rust toolchain, so neithercodex -m gpt-5.6-lunanor the doc-sync/insta regeneration (test_docs_are_in_sync,cargo insta) was run locally. The mirrors and snapshots were hand-edited with an identical, same-length literal replacement so they stay consistent (CI'stest_docs_are_in_syncand insta tests are the backstop). Please confirmgpt-5.6-lunais the intended model name before merge.🤖 Weekly LLM-model-name maintenance