kyleandClaude Opus 5.5
89a50bc3c6
Three changes from the M5b run (implementer lessons T47, T48): - each session runs `opencode run --pure` with OPENCODE_DISABLE_CLAUDE_CODE=1, so no global plugin or borrowed rules file reaches the implementer; PLUGINS=1 restores both; - with RESUME=1, a session stopped at SESSION_LIMIT_MIN with work left uncommitted is resumed once, as one that ended by itself already was; an idle stop is still not resumed; - the prompt names the log row's Task cell, so a task's heading cannot mislead the row the driver then looks for. Checked against a stub opencode in scratch repositories (17 checks: the flags and variable, PLUGINS=1, a resume after the limit, no resume without work or without RESUME). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
160 lines
7.3 KiB
Bash
Executable File
160 lines
7.3 KiB
Bash
Executable File
#!/bin/sh
|
|
# Runs the tasks of a plan one after another, each in its own fresh OpenCode session, and stops
|
|
# at the first task that does not end cleanly. This is the owner's tool; the implementing model
|
|
# never runs it.
|
|
#
|
|
# Usage: tools/run-plan.sh docs/plans/M2a [FIRST] FIRST is a task number such as 04
|
|
# Options (environment):
|
|
# OPENCODE_MODEL model to use (default straylight/laguna-s-2.1; OpenCode's own default differs).
|
|
# The part after the provider (`ornith-1.5-35b-a3b`) names the model in each
|
|
# commit's `Co-Authored-By: <model> <agent@wntrmute.dev>` trailer and in the log.
|
|
# OPENCODE_FLAGS extra flags for `opencode run`
|
|
# PLUGINS=1 load OpenCode's external plugins and let it read ~/.claude/CLAUDE.md. By
|
|
# default each session runs with `--pure` and OPENCODE_DISABLE_CLAUDE_CODE=1:
|
|
# a global plugin that rewrites the prompt forces a full re-read on a model
|
|
# that cannot rewind its cache (implementer lessons, T48).
|
|
# SESSION_LIMIT_MIN minutes a session may run before it is stopped (default 90)
|
|
# SLOTS_URL llama-server's slots URL, for example
|
|
# http://127.0.0.1:8081/slots?model=ornith-1.5-35b-a3b. When set, a session is
|
|
# stopped once no slot has been busy for IDLE_LIMIT_MIN minutes (default 10).
|
|
# Trial: not yet used on a run.
|
|
# RESUME=1 when a session ends, or is stopped at SESSION_LIMIT_MIN, with work left
|
|
# uncommitted, resume it once, in a new session with the same limit.
|
|
# DRY_RUN=1 print what would be run and stop
|
|
# Each session runs as the `implement` agent of ./opencode.json: its permissions deny edits to the
|
|
# plans and design documents, reads outside the repository, subagents and pushes.
|
|
set -eu
|
|
|
|
plan="${1:?usage: tools/run-plan.sh docs/plans/<milestone> [FIRST]}"
|
|
first="${2:-00}"
|
|
model="${OPENCODE_MODEL:-straylight/laguna-s-2.1}"
|
|
agent="${model##*/}"
|
|
trailer="Co-Authored-By: $agent <agent@wntrmute.dev>"
|
|
limit="${SESSION_LIMIT_MIN:-90}"
|
|
idle="${IDLE_LIMIT_MIN:-10}"
|
|
log=docs/implementer-log.md
|
|
milestone=$(basename "$plan")
|
|
runs=".state/runs/$milestone"
|
|
|
|
[ -d "$plan" ] || { echo "run-plan: $plan is not a directory" >&2; exit 2; }
|
|
[ -f "$log" ] || { echo "run-plan: run this from the repository root" >&2; exit 2; }
|
|
mkdir -p "$runs"
|
|
|
|
# One OpenCode session with prompt $1, appended to the task's run log. Returns opencode's status,
|
|
# 124 when the session outlived $limit minutes, or 125 when SLOTS_URL showed no busy slot for $idle
|
|
# minutes. A stopped session is killed by process group (it may have started cargo), never with
|
|
# `pkill`, which would match this script too.
|
|
run_session() {
|
|
set -m
|
|
if [ "${PLUGINS:-0}" = 1 ]; then
|
|
pure=""
|
|
else
|
|
pure="--pure"
|
|
export OPENCODE_DISABLE_CLAUDE_CODE=1
|
|
fi
|
|
# shellcheck disable=SC2086
|
|
opencode run $pure --agent implement -m "$model" ${OPENCODE_FLAGS:-} --title "$milestone/$name" \
|
|
"$1" >> "$runs/$name.log" 2>&1 &
|
|
pid=$!
|
|
set +m
|
|
started=$(date +%s)
|
|
busy=$started
|
|
why=0
|
|
while kill -0 "$pid" 2> /dev/null; do
|
|
sleep 10
|
|
now=$(date +%s)
|
|
if [ $((now - started)) -ge $((limit * 60)) ]; then
|
|
why=124
|
|
break
|
|
fi
|
|
if [ -n "${SLOTS_URL:-}" ]; then
|
|
if curl -s -m 10 "$SLOTS_URL" | grep -Eq '"is_processing": *true'; then
|
|
busy=$now
|
|
fi
|
|
if [ $((now - busy)) -ge $((idle * 60)) ]; then
|
|
why=125
|
|
break
|
|
fi
|
|
fi
|
|
done
|
|
if [ "$why" != 0 ]; then
|
|
kill -TERM "-$pid" 2> /dev/null || kill -TERM "$pid" 2> /dev/null || true
|
|
sleep 30
|
|
kill -KILL "-$pid" 2> /dev/null || true
|
|
wait "$pid" 2> /dev/null || true
|
|
return "$why"
|
|
fi
|
|
status=0
|
|
wait "$pid" || status=$?
|
|
return "$status"
|
|
}
|
|
|
|
for task in "$plan"/[0-9][0-9]-*.md; do
|
|
name=$(basename "$task" .md)
|
|
number=$(echo "$name" | cut -c1-2)
|
|
[ "$number" -ge "$first" ] || continue
|
|
row="Your log row's Task cell is \`$milestone/$name\`, and you are $agent: write $agent in its Model column, and end the commit message with this line, exactly: $trailer"
|
|
prompt="Read \`$task\` and do exactly that task. $row"
|
|
|
|
if [ "${DRY_RUN:-0}" = 1 ]; then
|
|
echo "opencode run --agent implement -m $model ${OPENCODE_FLAGS:-} --title $milestone/$name \"$prompt\""
|
|
continue
|
|
fi
|
|
if [ -n "$(git status --porcelain)" ]; then
|
|
echo "run-plan: the working tree is not clean before $name; stopping" >&2
|
|
exit 1
|
|
fi
|
|
|
|
before=$(git rev-parse HEAD)
|
|
: > "$runs/$name.log"
|
|
echo "=== $milestone/$name: started $(date '+%H:%M:%S')"
|
|
# A new `opencode run` is a new session: nothing carries over from the task before.
|
|
rc=0
|
|
run_session "$prompt" || rc=$?
|
|
# A session that ended, or ran out of time, part-way through a task is resumed once. One stopped
|
|
# for idling (125) is not: the server, not the task, is the likely cause.
|
|
if { [ "$rc" = 0 ] || [ "$rc" = 124 ]; } && [ "${RESUME:-0}" = 1 ] \
|
|
&& [ "$(git rev-parse HEAD)" = "$before" ] && [ -n "$(git status --porcelain)" ]; then
|
|
if [ "$rc" = 124 ]; then
|
|
echo "=== $milestone/$name: stopped at $limit minutes with work uncommitted"
|
|
fi
|
|
echo "=== $milestone/$name: work left uncommitted; resuming once $(date '+%H:%M:%S')"
|
|
rc=0
|
|
run_session "The previous session on \`$task\` left uncommitted work (see \`git diff\`). Do not start over. Finish the task, run the gate, commit, and write the log row. $row" || rc=$?
|
|
fi
|
|
case "$rc" in
|
|
0) ;;
|
|
124) echo "run-plan: $name ran past $limit minutes and was stopped; see $runs/$name.log" >&2; exit 1 ;;
|
|
125) echo "run-plan: no llama slot was busy for $idle minutes in $name, so it was stopped; see $runs/$name.log" >&2; exit 1 ;;
|
|
*) echo "run-plan: opencode exited with status $rc in $name; see $runs/$name.log" >&2; exit 1 ;;
|
|
esac
|
|
echo "=== $milestone/$name: session ended $(date '+%H:%M:%S')"
|
|
|
|
# The task must have left exactly what AGENTS.md asks for: a commit carrying the trailer, a
|
|
# clean tree, and a `done` row in the log. Anything else needs a person, so stop.
|
|
if [ "$(git rev-parse HEAD)" = "$before" ]; then
|
|
echo "run-plan: $name made no commit; stopping" >&2
|
|
exit 1
|
|
fi
|
|
if [ -n "$(git status --porcelain)" ]; then
|
|
echo "run-plan: $name left uncommitted changes; stopping" >&2
|
|
exit 1
|
|
fi
|
|
if ! git log -1 --format=%B | grep -qxF "$trailer"; then
|
|
echo "run-plan: $name's commit lacks the line '$trailer'; stopping" >&2
|
|
exit 1
|
|
fi
|
|
if git show HEAD -- "$log" | grep -Eq "^\+\| *$milestone/$name *\|[^|]*\| *stopped *\|"; then
|
|
echo "run-plan: $name stopped and said why in its log row; stopping" >&2
|
|
git show HEAD -- "$log" | grep -E "^\+\| *$milestone/$name *\|" | cut -c2- >&2
|
|
exit 1
|
|
fi
|
|
if ! grep -Eq "^\| *$milestone/$name *\|[^|]*\| *done *\|" "$log"; then
|
|
echo "run-plan: no 'done' row for $milestone/$name in $log; stopping" >&2
|
|
grep -E "^\| *$milestone/$name *\|" "$log" >&2 || true
|
|
exit 1
|
|
fi
|
|
echo "=== $milestone/$name: done ($(git log -1 --format=%h))"
|
|
done
|
|
echo "run-plan: every task from $first on ended cleanly. Ready for review."
|