Marcos Paulo
73ceae2aa8
Claude Code against an Ollama-backed model via the headroom hub. Model defaults to glm-5.3:cloud, overridable with OC_MODEL. Unproven here, so route it at tasks whose failure is cheap to detect.
139 lines
5.2 KiB
Bash
Executable File
139 lines
5.2 KiB
Bash
Executable File
#!/usr/bin/env bash
|
|
# Launch one refactor task in its own worktree, on the CLI its brief routes it to.
|
|
#
|
|
# .agents/scripts/launch.sh 01 scaffold
|
|
# .agents/scripts/launch.sh 02 tokens --base refactor/task-01-scaffold
|
|
# .agents/scripts/launch.sh 07 primitives --cli mm --fg
|
|
#
|
|
# Routing comes from plans/astro-refactor/MODEL-ROUTING.md. Override with --cli.
|
|
# All four CLIs are launched with their permission prompts disabled: these run
|
|
# unattended inside a worktree, and a blocked edit or bash call just hangs.
|
|
set -euo pipefail
|
|
|
|
cd "$(git rev-parse --show-toplevel)"
|
|
|
|
number=${1:?usage: launch.sh <task-number> <slug> [--base ref] [--cli codex|agy|mm|oc] [--fg]}
|
|
slug=${2:?slug, e.g. scaffold}
|
|
shift 2
|
|
|
|
base="main"
|
|
cli=""
|
|
foreground=0
|
|
while [ $# -gt 0 ]; do
|
|
case "$1" in
|
|
--base) base=$2; shift 2 ;;
|
|
--cli) cli=$2; shift 2 ;;
|
|
--fg) foreground=1; shift ;;
|
|
*) echo "unknown flag: $1" >&2; exit 2 ;;
|
|
esac
|
|
done
|
|
|
|
# Model routing. Codex takes the long iterate-until-green loops, Gemini the two
|
|
# tasks that need whole-corpus context plus visual judgement, MiniMax the rest.
|
|
if [ -z "$cli" ]; then
|
|
case "$number" in
|
|
01|03|15|16|19) cli=codex ;;
|
|
02|18) cli=agy ;;
|
|
*) cli=mm ;;
|
|
esac
|
|
fi
|
|
|
|
dir="../af-task-${number}"
|
|
branch="refactor/task-${number}-${slug}"
|
|
plan="plans/astro-refactor/task-${number}-${slug}.md"
|
|
log="$(pwd)/.agents/logs/task-${number}-${slug}.log"
|
|
|
|
[ -f "$plan" ] || { echo "no plan at $plan — check the number and slug" >&2; exit 1; }
|
|
mkdir -p .agents/logs
|
|
|
|
# The brief names its own agent; pull it out so the prompt can point at the file.
|
|
# [a-z0-9-] not [a-z-]: content-i18n-migrator has digits in it.
|
|
agent=$(sed -n 's/.*\*\*Agent\*\*: `\([a-z0-9-]*\)`.*/\1/p' "$plan" | head -1)
|
|
[ -n "$agent" ] || { echo "could not read the agent name out of $plan" >&2; exit 1; }
|
|
[ -f ".agents/agents/${agent}.md" ] || { echo "no such agent: .agents/agents/${agent}.md" >&2; exit 1; }
|
|
|
|
if [ -d "$dir" ]; then
|
|
echo "worktree $dir already exists — reusing it"
|
|
else
|
|
git worktree add "$dir" -b "$branch" "$base"
|
|
fi
|
|
|
|
# pnpm install works only once task 01 has produced a lockfile. The pre-existing root
|
|
# package.json carries two scripts and no dependencies, so before task 01 there
|
|
# is no toolchain to install and no hooks to verify.
|
|
if [ -f "$dir/pnpm-lock.yaml" ]; then
|
|
( cd "$dir" && pnpm install --frozen-lockfile && .agents/scripts/verify-hooks.sh )
|
|
fi
|
|
|
|
prompt=$(cat <<PROMPT
|
|
You are the \`${agent}\` specialist on the ai-for-dummies Astro refactor.
|
|
You are working alone in the git worktree at $(cd "$dir" && pwd), on branch ${branch}.
|
|
|
|
Read these, in this order, before you write anything:
|
|
|
|
1. AGENTS.md — project entry point and the "never touch" list
|
|
2. .agents/agents/${agent}.md — your role, what you own, what you must not do
|
|
3. ${plan} — your task brief: scope, steps, done-when, do-not
|
|
4. every file the brief and your agent definition tell you to read
|
|
(.agents/context/*, .agents/rules/*, .agents/skills/*, .agents/checklists/*)
|
|
|
|
Then do the task. Rules that override your own judgement:
|
|
|
|
- Stay inside the scope in the brief. Do not do work belonging to another task.
|
|
- Honour every "Do not" line in the brief and in your agent definition.
|
|
- The site must look and read exactly as it does today. This is a refactor,
|
|
not a redesign. If you find something that looks like a bug in the current
|
|
design, write it down in the final report — do not fix it.
|
|
- Never delete or weaken an assertion in scripts/verify.mjs to make a suite
|
|
green. Only the verification-engineer may reduce assertion count, with a
|
|
written reason per removal.
|
|
- Commit as you go with conventional-commit messages. The commit-msg and
|
|
pre-commit hooks are live; if a hook rejects you, fix the cause, never
|
|
bypass with --no-verify.
|
|
- Work through the done-when checklist at the end and actually run each check.
|
|
|
|
Finish by printing: what you changed, which done-when boxes are genuinely
|
|
ticked, which are not and why, and anything the next task needs to know.
|
|
PROMPT
|
|
)
|
|
|
|
echo "task : ${number} ${slug}"
|
|
echo "agent : ${agent}"
|
|
echo "cli : ${cli}"
|
|
echo "worktree: ${dir} (branch ${branch}, from ${base})"
|
|
echo "log : ${log}"
|
|
echo
|
|
|
|
run() {
|
|
case "$cli" in
|
|
codex)
|
|
codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check \
|
|
-C "$dir" "$prompt"
|
|
;;
|
|
agy)
|
|
( cd "$dir" && agy --dangerously-skip-permissions \
|
|
--model gemini-3.1-pro-high --print-timeout 4h -p "$prompt" )
|
|
;;
|
|
mm)
|
|
( cd "$dir" && mm --dangerously-skip-permissions \
|
|
--model opus -p "$prompt" )
|
|
;;
|
|
oc)
|
|
# Claude Code against a local-Ollama-backed model, through the headroom
|
|
# hub. Unproven on this repo — give it the task whose failure is cheapest.
|
|
( cd "$dir" && OLLAMA_CLAUDE_MODEL="${OC_MODEL:-glm-5.3:cloud}" \
|
|
ollama-claude --dangerously-skip-permissions -p "$prompt" )
|
|
;;
|
|
*) echo "unknown cli: $cli" >&2; exit 2 ;;
|
|
esac
|
|
}
|
|
|
|
if [ "$foreground" = 1 ]; then
|
|
run 2>&1 | tee "$log"
|
|
else
|
|
export cli dir prompt
|
|
nohup bash -c "$(declare -f run); run" >"$log" 2>&1 &
|
|
echo "launched in background, pid $!"
|
|
echo "follow: tail -f $log"
|
|
fi
|