tasks-toolkit 0.1.3__py3-none-any.whl

scripts/agents_sync.py

@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+"""Generate the single-file guidelines for a tool that loads exactly one instruction file (Junie).
+
+Concatenates `AGENTS.md`, the implementer subset of `agents/common/*.md` and
+`agents/junie/INSTRUCTIONS.md` into `agents/junie/GUIDELINES.generated.md`. `--check` regenerates in
+memory and fails when the file on disk differs, so `scripts/test.sh` catches drift. Standard library only.
+
+Usage: scripts/agents_sync.py [--check]
+"""
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent.parent
+TARGET = ROOT / "agents" / "junie" / "GUIDELINES.generated.md"
+SOURCES = [
+    "AGENTS.md",
+    "agents/common/layout.md",
+    "agents/common/git-and-hooks.md",
+    "agents/junie/INSTRUCTIONS.md",
+]
+HEADER = (
+    "<!-- GENERATED by scripts/agents_sync.py — do not edit by hand. "
+    "Run: python3 scripts/agents_sync.py -->\n"
+    "# VibeEmulator — Junie guidelines (generated)\n\n"
+    "The shared rules and the files a Junie implementer needs, concatenated for a tool that loads\n"
+    "exactly one guidelines file. Sources: " + ", ".join(SOURCES) + ".\n"
+)
+
+
+def render() -> str:
+    """Return the full generated guidelines text (deterministic, newline-terminated)."""
+    parts = [HEADER]
+    for rel in SOURCES:
+        text = (ROOT / rel).read_text(encoding="utf-8").rstrip("\n")
+        parts.append(f"\n---\n\n## From {rel}\n\n{text}\n")
+    return "".join(parts)
+
+
+def main(argv: list[str]) -> int:
+    """Write the guidelines file, or with --check compare it and fail on drift."""
+    expected = render()
+    rel = TARGET.relative_to(ROOT)
+    if "--check" in argv:
+        if not TARGET.exists():
+            print(f"agents_sync: {rel} is missing; run: python3 scripts/agents_sync.py", file=sys.stderr)
+            return 1
+        if TARGET.read_text(encoding="utf-8") != expected:
+            print(f"agents_sync: {rel} is stale; run: python3 scripts/agents_sync.py", file=sys.stderr)
+            return 1
+        print("agents_sync: guidelines in sync")
+        return 0
+    TARGET.parent.mkdir(parents=True, exist_ok=True)
+    TARGET.write_text(expected, encoding="utf-8")
+    print(f"wrote {rel}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv[1:]))

scripts/apply-plan.sh

@@ -0,0 +1,51 @@
+#!/usr/bin/env bash
+# Imports a validated sprint plan (tmp/plans/<name>.json) into the tracker through tasks.py.
+# Previews the plan and needs a human "y" on a TTY (otherwise --yes). A plan is applied once.
+#
+#   scripts/apply-plan.sh <name> [--yes] [--start]
+#
+# --start also creates the sprint branch (scripts/worktree.sh sprint start <n> <slug>).
+# Exit: 0 applied | 1 usage/missing plan | 2 not confirmed | 4 plan invalid | 5 already applied
+set -euo pipefail
+
+root="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$root"
+
+name="${1:-}"
+[[ "$name" =~ ^[a-z0-9][a-z0-9-]*$ ]] || { echo "usage: $0 <name> [--yes] [--start]" >&2; exit 1; }
+shift
+yes=0 start=0
+while [ $# -gt 0 ]; do
+  case "$1" in
+    --yes)   yes=1; shift ;;
+    --start) start=1; shift ;;
+    *) echo "unknown option '$1'" >&2; exit 1 ;;
+  esac
+done
+
+plan="tmp/plans/$name.json"
+marker="tmp/plans/$name.applied"
+[ -f "$plan" ] || { echo "error: no plan '$plan' (run scripts/plan-sprint.sh $name first)" >&2; exit 1; }
+[ ! -f "$marker" ] || { echo "error: '$name' was already applied ($(cat "$marker"))" >&2; exit 5; }
+
+python3 scripts/sprint_plan.py validate "$plan" --specs-written || exit 4
+python3 scripts/sprint_plan.py preview "$plan"
+echo ""
+if [ "$yes" -ne 1 ]; then
+  if [ ! -t 0 ]; then
+    echo "error: not a TTY; re-run with --yes to apply without asking" >&2
+    exit 2
+  fi
+  read -r -p "Apply this plan to the tracker? [y/N] " answer
+  [[ "$answer" =~ ^[yY]([eE][sS])?$ ]] || { echo "not applied" >&2; exit 2; }
+fi
+
+mapping="$(python3 scripts/sprint_plan.py apply "$plan")"
+printf '%s\n' "$mapping"
+sprint_id="$(printf '%s\n' "$mapping" | head -1 | sed 's/.*-> \([0-9]*\)\..*/\1/')"
+slug="$(python3 -c 'import json,sys; print(json.load(open(sys.argv[1]))["sprint"]["slug"])' "$plan")"
+printf 'sprint %s on %s\n' "$sprint_id" "$(date -u +%Y-%m-%dT%H:%M:%SZ)" >"$marker"
+echo "applied: sprint $sprint_id ($slug)"
+if [ "$start" -eq 1 ]; then
+  bash scripts/worktree.sh sprint start "$sprint_id" "$slug"
+fi

scripts/build.sh

@@ -0,0 +1,7 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+root="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$root"
+
+exec dotnet build VibeEmulator.slnx "$@"

scripts/codex-task.sh

@@ -0,0 +1,198 @@
+#!/usr/bin/env bash
+# Delegates one task to the Codex CLI agent inside its own worktree, then verifies it.
+# Same contract as junie-task.sh / kilo-task.sh (worktree, heartbeat, verification, patch).
+#
+#   scripts/codex-task.sh <name> "<task text>" [--task-id <id>]
+#   scripts/codex-task.sh <name> --file tmp/tasks/<name>.md [--task-id <id>]
+#
+# --task-id reports progress to the task tracker (scripts/task-report.sh): the task is
+# marked in progress at the start, and after the commit and verification it is closed
+# automatically (done on success, blocked on failure); failures only warn.
+#
+# Env: CODEX_MODEL (default: Codex's own default), CODEX_SANDBOX (default workspace-write),
+#      CODEX_TIMEOUT (default 1200 s), CODEX_PING (default 30 s),
+#      CODEX_VERIFY_CMD (tests only: replaces the dotnet verification).
+# Output: tmp/runs/<name>.{status,events.jsonl,result.txt,verify.log,summary.json},
+#         tmp/patches/<name>.patch. Every status line ends with "model=<model> runner=codex".
+# Exit code: 0 verified, 1 verification failed, 3 Codex itself failed.
+#
+# Notes:
+# - Codex runs in a workspace-write sandbox with the network off by default; `dotnet` writes to
+#   ~/.nuget and ~/.dotnet, so those are added to the sandbox (--add-dir) or restore/build fails.
+# - The worktree is disposable; review the patch. Never bypass the sandbox or approval flags.
+set -euo pipefail
+
+root="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$root"
+
+runner="codex"
+name="${1:?usage: codex-task.sh <name> \"<task>\" | --file <path> [--task-id <id>]}"
+shift
+if [ "${1:-}" = "--file" ]; then
+  task="$(cat "${2:?missing task file}")"
+  shift 2
+else
+  task="${1:?missing task text}"
+  shift
+fi
+task_id=""
+if [ "${1:-}" = "--task-id" ]; then
+  task_id="${2:?--task-id needs a value}"
+fi
+
+model="${CODEX_MODEL:-codex}"
+sandbox="${CODEX_SANDBOX:-workspace-write}"
+limit="${CODEX_TIMEOUT:-1200}"
+ping_every="${CODEX_PING:-30}"
+
+if [ -d "tmp/worktrees/$name" ]; then
+  wt="$root/tmp/worktrees/$name"
+else
+  wt="$(bash scripts/worktree.sh new "$name")"
+fi
+mkdir -p tmp/runs
+
+prompt="$task
+
+Rules for this run:
+- Work only inside this project directory (a disposable git worktree). Read AGENTS.md first and follow it.
+- Do not commit, push or touch git. Do not add NuGet packages.
+- You are done only when 'bash scripts/test.sh' is green AND
+  'dotnet build VibeEmulator.slnx --no-incremental -warnaserror' reports 0 warnings and 0 errors
+  (StyleCop and mandatory XML documentation are enforced on production code).
+- Respect the sandbox; do not bypass approval or network flags.
+- Finish with a short summary of the files you created or changed."
+
+status_file="$root/tmp/runs/$name.status"
+events="$root/tmp/runs/$name.events.jsonl"
+base="$(cat "tmp/meta/$name.base")"
+started=$SECONDS
+started_epoch="$(date +%s)"
+
+report() {  # report <state> [detail]: one line to stdout and the status file
+  local line
+  line="$(printf '[%02d:%02d] %s' $(((SECONDS - started) / 60)) $(((SECONDS - started) % 60)) "$1${2:+ - $2}")"
+  echo "$line model=$model runner=$runner"
+  echo "$line model=$model runner=$runner" >"$status_file"
+}
+
+# last_activity: the last non-empty output line of the agent.
+last_activity() {
+  grep -v '^[[:space:]]*$' "$events" 2>/dev/null | tail -n 1 | cut -c1-110 || true
+}
+
+bash scripts/task-report.sh start "$task_id"
+report "starting" "worktree=$wt${task_id:+ task=$task_id}"
+
+codex_args=(exec -C "$wt" --skip-git-repo-check --sandbox "$sandbox" --json)
+[ -n "${CODEX_MODEL:-}" ] && codex_args+=(-m "$model")
+[ -d "$HOME/.nuget" ] && codex_args+=(--add-dir "$HOME/.nuget")
+[ -d "$HOME/.dotnet" ] && codex_args+=(--add-dir "$HOME/.dotnet")
+
+timeout "$limit" codex "${codex_args[@]}" "$prompt" \
+  >"$events" 2>"$root/tmp/runs/$name.stderr.log" &
+codex_pid=$!
+
+stalled=0
+while kill -0 "$codex_pid" 2>/dev/null; do
+  sleep "$ping_every" & wait $! || true
+  kill -0 "$codex_pid" 2>/dev/null || break
+  changed="$(git -C "$wt" status --porcelain | wc -l)"
+  if [ "$changed" -eq 0 ] && [ $((SECONDS - started)) -ge "${RUNNER_STALL:-420}" ]; then
+    report "stalled" "no file changed after ${RUNNER_STALL:-420}s; killing the agent"
+    kill "$codex_pid" 2>/dev/null || true
+    stalled=1
+    break
+  fi
+  last="$(last_activity)"
+  report "codex running" "$changed file(s) changed; last: ${last:-<no output yet>}"
+done
+set +e
+wait "$codex_pid"
+codex_rc=$?
+set -e
+report "codex finished" "exit $codex_rc, verifying"
+
+# With --json the last assistant text is the final message; fall back to the last output line.
+python3 - "$events" >"$root/tmp/runs/$name.result.txt" <<'PY'
+import json, sys
+out = []
+for line in open(sys.argv[1], errors="replace"):
+    line = line.strip()
+    if not line:
+        continue
+    try:
+        event = json.loads(line)
+    except ValueError:
+        out.append(line)
+        continue
+    text = event.get("text") or (event.get("message") or {}).get("content")
+    if isinstance(text, str):
+        out.append(text)
+print(out[-1] if out else "")
+PY
+
+# Commit the agent's work on task/<name> (path guard first); agents never commit themselves.
+commit_rc=0
+if [ "$codex_rc" -eq 0 ]; then
+  title="$(printf '%s\n' "$task" | grep -m1 -v '^[[:space:]]*$' | cut -c1-72)"
+  bash scripts/worktree.sh commit "$name" --runner "$runner" --model "$model" \
+    ${task_id:+--task-id "$task_id"} --title "${title:-$name}" >"tmp/runs/$name.commit.log" 2>&1 || commit_rc=$?
+fi
+patch="$(bash scripts/worktree.sh patch "$name")"
+
+verify_rc=0
+{
+  if [ -n "${CODEX_VERIFY_CMD:-}" ]; then
+    echo "== verify (CODEX_VERIFY_CMD override) =="
+    (cd "$wt" && bash -c "$CODEX_VERIFY_CMD") || verify_rc=1
+  elif ! bash "$root/scripts/needs-dotnet-checks.sh" "$wt" "$base" HEAD; then
+    echo "== conditional verification (no .NET files changed): script tests =="
+    (cd "$wt" && bash scripts/test_worktree.sh && bash scripts/test_runners.sh \
+       && python3 -m unittest discover -s scripts -p "test_*.py") || verify_rc=1
+  else
+    echo "== format =="; (cd "$wt" && dotnet format VibeEmulator.slnx --verify-no-changes) || verify_rc=1
+    echo "== build ==";  (cd "$wt" && dotnet build VibeEmulator.slnx --no-incremental -warnaserror) || verify_rc=1
+    echo "== test ==";   (cd "$wt" && dotnet test VibeEmulator.slnx --no-build) || verify_rc=1
+  fi
+} >"tmp/runs/$name.verify.log" 2>&1
+
+if [ "$commit_rc" -ne 0 ]; then
+  verify_rc=1
+  echo "== commit failed (exit $commit_rc): see tmp/runs/$name.commit.log ==" >>"tmp/runs/$name.verify.log"
+fi
+finished="$(date +%s)"
+files_changed="$(git -C "$wt" diff --cached --name-only "$base" | wc -l)"
+if [ "$codex_rc" -ne 0 ]; then
+  verdict="agent_failed"
+elif [ "$verify_rc" -ne 0 ]; then
+  verdict="verify_failed"
+else
+  verdict="verified"
+fi
+if ! python3 scripts/run_summary.py "$root/tmp/runs/$name.summary.json" name="$name" runner="$runner" model="$model" \
+    task_id="$task_id" started="$started_epoch" finished="$finished" \
+    agent_exit="$codex_rc" verify_exit="$verify_rc" verdict="$verdict" \
+    patch="$patch" files_changed="$files_changed"; then
+  echo "warning: could not write run summary (tmp/runs/$name.summary.json)" >&2
+fi
+
+report "done" "codex exit $codex_rc, verification $([ "$verify_rc" -eq 0 ] && echo PASSED || echo FAILED)"
+echo "runner:       $runner"
+echo "model:        $model"
+echo "codex exit:   $codex_rc"
+echo "verification: $([ "$verify_rc" -eq 0 ] && echo PASSED || echo FAILED) (tmp/runs/$name.verify.log)"
+echo "patch:        $patch"
+git -C "$wt" diff --cached --stat "$base" | tail -n 15
+if [ "$codex_rc" -ne 0 ]; then
+  bash scripts/task-report.sh note "$task_id" "codex '$name' failed (exit 3) model=$model; see tmp/runs/$name.verify.log"
+  bash scripts/task-report.sh close "$task_id" red "agent exit $codex_rc; see tmp/runs/$name.verify.log"
+  exit 3
+elif [ "$verify_rc" -ne 0 ]; then
+  bash scripts/task-report.sh note "$task_id" "codex '$name' failed (exit 1) model=$model; see tmp/runs/$name.verify.log"
+  bash scripts/task-report.sh close "$task_id" red "verification failed; see tmp/runs/$name.verify.log"
+  exit 1
+else
+  bash scripts/task-report.sh note "$task_id" "codex '$name' verified model=$model; patch tmp/patches/$name.patch"
+  bash scripts/task-report.sh close "$task_id" green
+fi

scripts/install-hooks.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+root="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$root"
+
+chmod +x .githooks/*
+git config core.hooksPath .githooks
+echo "Git hooks installed (core.hooksPath=.githooks)."

scripts/junie-status.sh

@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+# Shows the progress of a delegated Junie or Kilo run: scripts/junie-status.sh <name>
+set -euo pipefail
+
+root="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$root"
+name="${1:?usage: junie-status.sh <name>}"
+
+[ -f "tmp/runs/$name.stdout.log" ] || [ -f "tmp/runs/$name.events.jsonl" ] || { echo "no run named '$name'" >&2; exit 1; }
+echo "status:  $(cat "tmp/runs/$name.status" 2>/dev/null || echo "(no heartbeat recorded)")"
+# runner/model come from the run summary once the run has finished, and from the
+# last status line while it is still running; runs that predate model recording
+# simply have neither and print nothing here.
+runner=""
+model=""
+if [ -f "tmp/runs/$name.summary.json" ]; then
+  runner="$(python3 -c 'import json,sys; print(json.load(open(sys.argv[1])).get("runner", ""))' "tmp/runs/$name.summary.json" 2>/dev/null || true)"
+  model="$(python3 -c 'import json,sys; print(json.load(open(sys.argv[1])).get("model", ""))' "tmp/runs/$name.summary.json" 2>/dev/null || true)"
+else
+  last="$(cat "tmp/runs/$name.status" 2>/dev/null || true)"
+  runner="$(printf '%s\n' "$last" | sed -n 's/.*runner=\([^ ]*\).*/\1/p')"
+  model="$(printf '%s\n' "$last" | sed -n 's/.*model=\([^ ]*\).*/\1/p')"
+fi
+[ -z "$runner" ] || echo "runner:   $runner"
+[ -z "$model" ] || echo "model:    $model"
+if pgrep -f "(junie|kilo) .*$root/tmp/worktrees/$name" >/dev/null; then
+  echo "process: running"
+else
+  echo "process: not running"
+fi
+if [ -d "tmp/worktrees/$name" ]; then
+  echo "changed files:"
+  git -C "tmp/worktrees/$name" status --porcelain | sed 's/^/  /'
+fi
+if [ -f "tmp/runs/$name.events.jsonl" ]; then
+  echo "recent activity (kilo):"
+  python3 - "tmp/runs/$name.events.jsonl" <<'PY' | tail -n 8
+import json, sys
+for line in open(sys.argv[1], errors="replace"):
+    try:
+        part = json.loads(line).get("part", {})
+    except ValueError:
+        continue
+    if part.get("type") == "tool":
+        print("  tool " + str(part.get("tool")))
+    elif part.get("type") == "text" and part.get("text"):
+        print("  " + " ".join(part["text"].split())[:130])
+PY
+fi
+echo "recent output:"
+tail -n 8 "tmp/runs/$name.stdout.log" 2>/dev/null | sed -E 's/\x1b\[[0-9;]*m//g' | cut -c1-140 | sed 's/^/  /'
+[ ! -f "tmp/runs/$name.verify.log" ] || { echo "verify log tail:"; tail -n 5 "tmp/runs/$name.verify.log" | sed 's/^/  /'; }

scripts/junie-task.sh

@@ -0,0 +1,183 @@
+#!/usr/bin/env bash
+# Delegates one task to the cheap Junie CLI agent inside its own worktree, then verifies it.
+#
+#   scripts/junie-task.sh <name> "<task text>" [--task-id <id>]
+#   scripts/junie-task.sh <name> --file tmp/task.md [--task-id <id>]
+#
+# --task-id reports progress to the task tracker (scripts/task-report.sh): the task is
+# marked in progress at the start, and after the commit and verification it is closed
+# automatically (done on success, blocked on failure); failures only warn.
+#
+# Env: RUNNER_STALL (default 420 s: kill an agent that changed no file by then), JUNIE_MODEL (default junie-lite), JUNIE_TIMEOUT (default 1200 s), JUNIE_PING
+#      (default 30 s), JUNIE_VERIFY_CMD (tests only: replaces the dotnet verification).
+#
+# Output (stdout): a short report. Artifacts: tmp/runs/<name>.json (Junie result),
+# tmp/runs/<name>.verify.log (build + tests), tmp/runs/<name>.summary.json (machine
+# readable run summary) and tmp/patches/<name>.patch.
+# Prints a heartbeat line every JUNIE_PING seconds (default 30) and mirrors it to
+# tmp/runs/<name>.status; every line ends with "model=<model> runner=junie" so it can
+# be parsed. Check it any time with scripts/junie-status.sh <name>.
+# Exit code: 0 verified, 1 verification failed, 3 Junie itself failed.
+set -euo pipefail
+
+root="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+cd "$root"
+
+runner="junie"
+name="${1:?usage: junie-task.sh <name> \"<task>\" | --file <path> [--task-id <id>]}"
+shift
+if [ "${1:-}" = "--file" ]; then
+  task="$(cat "${2:?missing task file}")"
+  shift 2
+else
+  task="${1:?missing task text}"
+  shift
+fi
+task_id=""
+if [ "${1:-}" = "--task-id" ]; then
+  task_id="${2:?--task-id needs a value}"
+fi
+
+model="${JUNIE_MODEL:-junie-lite}"
+limit="${JUNIE_TIMEOUT:-1200}"
+
+if [ -d "tmp/worktrees/$name" ]; then
+  wt="$root/tmp/worktrees/$name"
+else
+  wt="$(bash scripts/worktree.sh new "$name" ${task_id:+--task-id "$task_id"})"
+fi
+mkdir -p tmp/runs
+
+run_file="$root/agents/$runner/RUN.md"
+[ -f "$run_file" ] || { echo "error: missing $run_file" >&2; exit 1; }
+prompt="$task
+
+$(cat "$run_file")"
+
+status_file="$root/tmp/runs/$name.status"
+log="$root/tmp/runs/$name.stdout.log"
+ping_every="${JUNIE_PING:-30}"
+base="$(cat "tmp/meta/$name.base")"
+started=$SECONDS
+started_epoch="$(date +%s)"
+
+report() {  # report <state> [detail]: one line to stdout and the status file
+  local line
+  line="$(printf '[%02d:%02d] %s' $(((SECONDS - started) / 60)) $(((SECONDS - started) % 60)) "$1${2:+ - $2}")"
+  echo "$line model=$model runner=$runner"
+  echo "$line model=$model runner=$runner" >"$status_file"
+}
+
+bash scripts/task-report.sh start "$task_id"
+report "starting" "worktree=$wt${task_id:+ task=$task_id}"
+GIT_CONFIG_COUNT=1 \
+GIT_CONFIG_KEY_0=core.hooksPath \
+GIT_CONFIG_VALUE_0="$root/scripts/coder-git-hooks" \
+  timeout "$limit" junie --skip-update-check --model="$model" --output-format=json \
+  --json-output-file="$root/tmp/runs/$name.json" --project="$wt" "$prompt" >"$log" 2>&1 &
+junie_pid=$!
+
+# Heartbeat: while Junie works, report elapsed time, changed files and its latest output line.
+stalled=0
+while kill -0 "$junie_pid" 2>/dev/null; do
+  sleep "$ping_every" & wait $! || true
+  kill -0 "$junie_pid" 2>/dev/null || break
+  changed="$(git -C "$wt" status --porcelain | wc -l)"
+  # Stall detector: an agent that has changed nothing after RUNNER_STALL seconds is stuck
+  # (seen with Kilo: 8+ minutes, 0 files). Kill it so the planner can switch runner early.
+  if [ "$changed" -eq 0 ] && [ $((SECONDS - started)) -ge "${RUNNER_STALL:-420}" ]; then
+    report "stalled" "no file changed after ${RUNNER_STALL:-420}s; killing the agent"
+    kill "$junie_pid" 2>/dev/null || true
+    stalled=1
+    break
+  fi
+  last="$(tail -n 5 "$log" | sed -E 's/\x1b\[[0-9;]*m//g' | grep -v '^\s*$' | tail -n 1 | cut -c1-110 || true)"
+  report "junie running" "$changed file(s) changed; last: ${last:-<no output yet>}"
+done
+set +e
+wait "$junie_pid"
+junie_rc=$?
+set -e
+report "junie finished" "exit $junie_rc, verifying"
+
+# Commit the agent's work on task/<name> (path guard first); agents never commit themselves.
+commit_rc=0
+if [ "$junie_rc" -eq 0 ]; then
+  task_head="$(git -C "$wt" rev-parse HEAD)"
+  if [ "$task_head" != "$base" ]; then
+    commit_rc=1
+    printf 'Junie changed task HEAD itself; only the parent runner may commit.\n' >"tmp/runs/$name.commit.log"
+  elif ! git -C "$wt" add -A; then
+    commit_rc=1
+    printf 'Could not stage the Junie task diff for guarded commit.\n' >"tmp/runs/$name.commit.log"
+  else
+    changed_lines="$(git -C "$wt" diff --cached --numstat "$base" \
+      | awk '$1 ~ /^[0-9]+$/ && $2 ~ /^[0-9]+$/ { total += $1 + $2 } END { print total + 0 }')"
+    if [ "$changed_lines" -gt 300 ]; then
+      commit_rc=1
+      printf 'Junie task diff has %s added/deleted lines (limit 300); split the task before committing.\n' \
+        "$changed_lines" >"tmp/runs/$name.commit.log"
+    else
+      title="$(printf '%s\n' "$task" | grep -m1 -v '^[[:space:]]*$' | cut -c1-72)"
+      bash scripts/worktree.sh commit "$name" --runner "$runner" --model "$model" \
+        ${task_id:+--task-id "$task_id"} --title "${title:-$name}" >"tmp/runs/$name.commit.log" 2>&1 || commit_rc=$?
+    fi
+  fi
+fi
+patch="$(bash scripts/worktree.sh patch "$name")"
+
+verify_rc=0
+{
+  if [ -n "${JUNIE_VERIFY_CMD:-}" ]; then
+    echo "== verify (JUNIE_VERIFY_CMD override) =="
+    (cd "$wt" && bash -c "$JUNIE_VERIFY_CMD") || verify_rc=1
+  elif ! bash "$root/scripts/needs-dotnet-checks.sh" "$wt" "$base" HEAD; then
+    echo "== conditional verification (no .NET files changed): script tests =="
+    (cd "$wt" && bash scripts/test_worktree.sh && bash scripts/test_runners.sh \
+       && python3 -m unittest discover -s scripts -p "test_*.py") || verify_rc=1
+  else
+    echo "== format =="; (cd "$wt" && dotnet format VibeEmulator.slnx --verify-no-changes) || verify_rc=1
+    echo "== build ==";  (cd "$wt" && dotnet build VibeEmulator.slnx --no-incremental -warnaserror) || verify_rc=1
+    echo "== test ==";   (cd "$wt" && dotnet test VibeEmulator.slnx --no-build) || verify_rc=1
+  fi
+} >"tmp/runs/$name.verify.log" 2>&1
+
+if [ "$commit_rc" -ne 0 ]; then
+  verify_rc=1
+  echo "== commit failed (exit $commit_rc): see tmp/runs/$name.commit.log ==" >>"tmp/runs/$name.verify.log"
+fi
+finished="$(date +%s)"
+files_changed="$(git -C "$wt" diff --cached --name-only "$base" | wc -l)"
+if [ "$junie_rc" -ne 0 ]; then
+  verdict="agent_failed"
+elif [ "$verify_rc" -ne 0 ]; then
+  verdict="verify_failed"
+else
+  verdict="verified"
+fi
+if ! python3 scripts/run_summary.py "$root/tmp/runs/$name.summary.json" name="$name" runner="$runner" model="$model" \
+    task_id="$task_id" started="$started_epoch" finished="$finished" \
+    agent_exit="$junie_rc" verify_exit="$verify_rc" verdict="$verdict" \
+    patch="$patch" files_changed="$files_changed"; then
+  echo "warning: could not write run summary (tmp/runs/$name.summary.json)" >&2
+fi
+
+report "done" "junie exit $junie_rc, verification $([ "$verify_rc" -eq 0 ] && echo PASSED || echo FAILED)"
+echo "runner:       $runner"
+echo "model:        $model"
+echo "junie exit:   $junie_rc"
+echo "verification: $([ "$verify_rc" -eq 0 ] && echo PASSED || echo FAILED) (tmp/runs/$name.verify.log)"
+echo "patch:        $patch"
+git -C "$wt" diff --cached --stat "$base" | tail -n 15
+if [ "$junie_rc" -ne 0 ]; then
+  bash scripts/task-report.sh note "$task_id" "junie '$name' failed (exit 3) model=$model; see tmp/runs/$name.verify.log"
+  bash scripts/task-report.sh close "$task_id" red "agent exit $junie_rc; see tmp/runs/$name.verify.log"
+  exit 3
+elif [ "$verify_rc" -ne 0 ]; then
+  bash scripts/task-report.sh note "$task_id" "junie '$name' failed (exit 1) model=$model; see tmp/runs/$name.verify.log"
+  bash scripts/task-report.sh close "$task_id" red "verification failed; see tmp/runs/$name.verify.log"
+  exit 1
+else
+  bash scripts/task-report.sh note "$task_id" "junie '$name' verified model=$model; patch tmp/patches/$name.patch"
+  bash scripts/task-report.sh close "$task_id" green
+fi