Skip to content

Worktrunk is a CLI for Git worktree management, designed for parallel AI agent workflows

docs: add Codex config and optimize Claude Code for LLM commits (#837)

Maximilian Rooscbe65a65ea51
diff --git a/.claude-plugin/skills/worktrunk/reference/llm-commits.md b/.claude-plugin/skills/worktrunk/reference/llm-commits.md
index ed93845f0..027aabd12 100644
--- a/.claude-plugin/skills/worktrunk/reference/llm-commits.md
+++ b/.claude-plugin/skills/worktrunk/reference/llm-commits.md
@@ -10,10 +10,10 @@ Any command that reads a prompt from stdin and outputs a commit message works. A
```toml
[commit.generation]
-command = "claude -p --model haiku"
+command = "MAX_THINKING_TOKENS=0 claude -p --model haiku --tools '' --disable-slash-commands --setting-sources '' --system-prompt ''"
```
-See [Claude Code docs](https://docs.anthropic.com/en/docs/build-with-claude/claude-code) for installation.
+The flags disable tools, skills, settings, and system prompt for fast text-only output. See [Claude Code docs](https://docs.anthropic.com/en/docs/build-with-claude/claude-code) for installation.
### llm
@@ -33,6 +33,15 @@ command = "aichat -m claude:claude-haiku-4.5"
See [aichat docs](https://github.com/sigoden/aichat).
+### Codex
+
+```toml
+[commit.generation]
+command = "codex exec -m gpt-5.1-codex-mini -c model_reasoning_effort='low' --sandbox read-only --json - | jq -sr '[.[] | select(.item.type? == \"agent_message\")] | last.item.text'"
+```
+
+Uses the fast mini model with low reasoning effort. Requires `jq` for JSON parsing. See [Codex CLI docs](https://developers.openai.com/codex/cli/).
+
## How it works
When worktrunk needs a commit message, it builds a prompt from a template and pipes it to the configured command via shell (`sh -c`). Environment variables can be set inline in the command string.
diff --git a/Taskfile.yaml b/Taskfile.yaml
index 8376de3c5..60e150489 100644
--- a/Taskfile.yaml
+++ b/Taskfile.yaml
@@ -299,6 +299,165 @@ tasks:
echo ""
echo "Done! Fixture at: tests/fixtures/standard/"
+ bench-llm-commits:
+ desc: Benchmark LLM commit message tools (claude, llm, aichat, codex)
+ silent: true
+ cmds:
+ - |
+ set -euo pipefail
+
+ # Test prompts
+ SMALL_PROMPT=$(cat <<'EOF'
+ Write a commit message for the staged changes below.
+
+ <format>
+ - Subject under 50 chars, blank line, then optional body
+ - Output only the commit message, no quotes or code blocks
+ </format>
+
+ <diffstat>
+ src/config.rs | 2 ++
+ 1 file changed, 2 insertions(+)
+ </diffstat>
+
+ <diff>
+ diff --git a/src/config.rs b/src/config.rs
+ @@ -10,6 +10,8 @@ pub struct Config {
+ pub debug: bool,
+ + /// Enable verbose logging
+ + pub verbose: bool,
+ }
+ </diff>
+
+ <context>
+ Branch: feature/logging
+ </context>
+ EOF
+ )
+
+ MEDIUM_PROMPT=$(cat <<'EOF'
+ Write a commit message for the staged changes below.
+
+ <format>
+ - Subject under 50 chars, blank line, then optional body
+ - Output only the commit message, no quotes or code blocks
+ </format>
+
+ <style>
+ - Imperative mood: "Add feature" not "Added feature"
+ - Match recent commit style (conventional commits if used)
+ </style>
+
+ <diffstat>
+ src/auth/jwt.rs | 45 +++++++++++++++++++++++++++++++++
+ src/auth/mod.rs | 2 ++
+ src/middleware.rs | 12 +++++++++
+ tests/auth_test.rs | 28 +++++++++++++++++++++
+ 4 files changed, 87 insertions(+)
+ </diffstat>
+
+ <diff>
+ diff --git a/src/auth/jwt.rs b/src/auth/jwt.rs
+ new file mode 100644
+ +use jsonwebtoken::{encode, decode, Header, Validation};
+ +
+ +pub struct Claims { pub sub: String, pub exp: usize }
+ +
+ +pub fn create_token(user_id: &str, secret: &[u8]) -> Result<String, Error> {
+ + let claims = Claims { sub: user_id.to_owned(), exp: ... };
+ + encode(&Header::default(), &claims, &EncodingKey::from_secret(secret))
+ +}
+ +
+ +pub fn validate_token(token: &str, secret: &[u8]) -> Result<Claims, Error> {
+ + decode::<Claims>(token, &DecodingKey::from_secret(secret), &Validation::default())
+ +}
+ </diff>
+
+ <context>
+ Branch: feature/auth
+ <recent_commits>
+ - feat(api): add user registration endpoint
+ - fix(db): handle connection pool exhaustion
+ </recent_commits>
+ </context>
+ EOF
+ )
+
+ # Documented commands from llm-commits.md
+ declare -A COMMANDS
+ COMMANDS["claude"]='MAX_THINKING_TOKENS=0 claude -p --model haiku --tools '"'"''"'"' --disable-slash-commands --setting-sources '"'"''"'"' --system-prompt '"'"''"'"''
+ COMMANDS["llm"]='llm -m claude-haiku-4.5'
+ COMMANDS["aichat"]='aichat -m claude:claude-haiku-4.5'
+ COMMANDS["codex"]='codex exec -m gpt-5.1-codex-mini -c model_reasoning_effort='"'"'low'"'"' --sandbox read-only --json - | jq -sr '"'"'[.[] | select(.item.type? == "agent_message")] | last.item.text'"'"''
+
+ # Check which tools are available
+ echo "=== LLM Commit Message Benchmark ==="
+ echo ""
+ echo "Checking available tools..."
+ AVAILABLE=()
+ for tool in claude llm aichat codex; do
+ if command -v "$tool" &>/dev/null; then
+ echo " ✓ $tool"
+ AVAILABLE+=("$tool")
+ else
+ echo " ✗ $tool (not installed)"
+ fi
+ done
+ if command -v jq &>/dev/null; then
+ echo " ✓ jq (required for codex)"
+ else
+ echo " ✗ jq (required for codex)"
+ AVAILABLE=("${AVAILABLE[@]/codex}")
+ fi
+ echo ""
+
+ if [ ${#AVAILABLE[@]} -eq 0 ]; then
+ echo "No tools available. Install at least one:"
+ echo " claude: https://docs.anthropic.com/en/docs/claude-code"
+ echo " llm: uv tool install llm llm-anthropic"
+ echo " aichat: https://github.com/sigoden/aichat"
+ echo " codex: npm install -g @openai/codex"
+ exit 1
+ fi
+
+ # Benchmark function
+ bench() {
+ local name="$1"
+ local cmd="$2"
+ local prompt="$3"
+ local start end elapsed output first_line
+
+ start=$(date +%s)
+ output=$(echo "$prompt" | eval "$cmd" 2>/dev/null) || output="[error]"
+ end=$(date +%s)
+ elapsed=$((end - start))
+ first_line=$(echo "$output" | head -1 | cut -c1-50)
+
+ printf " %-8s %3ds %s\n" "$name:" "$elapsed" "$first_line"
+ }
+
+ # Run benchmarks
+ for size in small medium; do
+ if [ "$size" = "small" ]; then
+ PROMPT="$SMALL_PROMPT"
+ echo "### Small diff (add verbose flag) ###"
+ else
+ PROMPT="$MEDIUM_PROMPT"
+ echo "### Medium diff (JWT auth module) ###"
+ fi
+ echo ""
+
+ for tool in "${AVAILABLE[@]}"; do
+ bench "$tool" "${COMMANDS[$tool]}" "$PROMPT"
+ done
+ echo ""
+ done
+
+ echo "Commands used (from docs/content/llm-commits.md):"
+ for tool in "${AVAILABLE[@]}"; do
+ echo " $tool: ${COMMANDS[$tool]}"
+ done
+
generate-logo:
desc: Generate logo using Gemini AI
dir: "{{.REPO_ROOT}}"
diff --git a/docs/content/llm-commits.md b/docs/content/llm-commits.md
index ba57e4bfb..8aa456a65 100644
--- a/docs/content/llm-commits.md
+++ b/docs/content/llm-commits.md
@@ -23,10 +23,10 @@ Any command that reads a prompt from stdin and outputs a commit message works. A
```toml
[commit.generation]
-command = "claude -p --model haiku"
+command = "MAX_THINKING_TOKENS=0 claude -p --model haiku --tools '' --disable-slash-commands --setting-sources '' --system-prompt ''"
```
-See [Claude Code docs](https://docs.anthropic.com/en/docs/build-with-claude/claude-code) for installation.
+The flags disable tools, skills, settings, and system prompt for fast text-only output. See [Claude Code docs](https://docs.anthropic.com/en/docs/build-with-claude/claude-code) for installation.
### llm
@@ -46,6 +46,15 @@ command = "aichat -m claude:claude-haiku-4.5"
See [aichat docs](https://github.com/sigoden/aichat).
+### Codex
+
+```toml
+[commit.generation]
+command = "codex exec -m gpt-5.1-codex-mini -c model_reasoning_effort='low' --sandbox read-only --json - | jq -sr '[.[] | select(.item.type? == \"agent_message\")] | last.item.text'"
+```
+
+Uses the fast mini model with low reasoning effort. Requires `jq` for JSON parsing. See [Codex CLI docs](https://developers.openai.com/codex/cli/).
+
## How it works
When worktrunk needs a commit message, it builds a prompt from a template and pipes it to the configured command via shell (`sh -c`). Environment variables can be set inline in the command string.