Plan Sketch

Plan Sketch

Preconditions

• Confirm you are in the intended repo: git rev-parse --show-toplevel
• Confirm whether web browsing is allowed.
  • If browsing is blocked by sandbox/approvals, ask the user to enable/approve it.
  • If the user says "no web browsing", proceed but clearly label assumptions where "latest" matters.

Output contract (required)

plan-sketch must produce:

1. A sketch markdown file (usually under sketch/, or an explicit --output path if provided).
2. A small, parseable Questions for user block in the subagent's stdout output so wrapper skills can ask the user without reading the sketch content into context.
3. A machine-readable JSON Questions block in stdout so wrapper skills can parse choices programmatically.

The Questions block must cover:

• Any open questions / unknowns that change the plan materially.
• Any assumptions that should be confirmed.
• Any high-level approach forks (architecture, APIs, rollout/testing strategy, migrations).
• If there are multiple viable implementation approaches, include an explicit choose an approach question so the user picks one before planning.

Format requirements (strict):

• Do not print the sketch content in the response.
• Print the sketch path (single line) and then print a Questions block bracketed by markers:
  • BEGIN_USER_QUESTIONS
  • END_USER_QUESTIONS
• In the Questions block:
  • Start with a short instruction telling the user to answer in one batch by label/number.
  • Present all decision-steering questions at once.
  • Number questions as 1), 2), ... (not bullets).
  • For each question, include 2–4 mutually exclusive options.
  • Label options as 1a), 1b), 1c) ... (letters per question), and make the labels easy to copy/paste.
  • Put the recommended option first and include a 1-sentence rationale on the same line.
  • End the block with a single-line summary of recommended picks in the format:
    • Recommended picks (copy/paste): 1a, 2b, 3a
  • If there are no meaningful open questions, include at least 1 explicit confirmation question with options.
• After the text Questions block, print a JSON Questions block bracketed by markers:
  • BEGIN_USER_QUESTIONS_JSON
  • END_USER_QUESTIONS_JSON
• The JSON block must be valid JSON and include:
  • questions: array of objects with id (number or string) and options.
  • Each options entry includes label (e.g., 1a) and text.
  • Optional: recommended boolean on options and recommended_picks array at the top level.

Expected runtime

Typically runs 15–45 minutes. Callers should allow the full --timeout (default 3600s) before interrupting.

If you use --progress-log, do not tail -f it into the main context unless you must debug; prefer checking progress via log line counts (for example: wc -l <progress-log>) and/or the heartbeat counters.

Workflow — Option A (external script, preferred)

Run the sketch as a standalone CLI invocation:

resolve_skill_dir() {
  local name="$1"
  local repo_root=""
  repo_root="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"

  local candidates=(
    "$repo_root/.agents/skills/$name"
    "$repo_root/.claude/skills/$name"
    "$HOME/.agents/skills/$name"
    "$HOME/.codex/skills/$name"
    "$HOME/.claude/skills/$name"
  )

  for d in "${candidates[@]}"; do
    if [[ -d "$d" ]]; then
      echo "$d"
      return 0
    fi
  done

  echo "Error: skill '$name' not found in repo-scoped or user-scoped skill dirs." >&2
  return 1
}

PLAN_SKETCH_SKILL_DIR="$(resolve_skill_dir plan-sketch)"
python3 "$PLAN_SKETCH_SKILL_DIR/scripts/run_plan_sketch.py" "user's rough idea"

Flags:

• --idea-file PATH — use a file as the idea input (useful for large/multiline inputs)
• --output PATH — optional explicit sketch output path (default: model generates sketch/<meaningful-name>.md)
• --cli codex|claude — override auto-detected CLI
• --model MODEL — pass a specific model to the CLI
• --reasoning-effort LEVEL — pass explicit reasoning effort (vhigh aliases to xhigh)
• --timeout N — CLI invocation timeout in seconds (default: 3600)
• --progress-log PATH — optional path to append streamed subagent output
• --heartbeat-seconds N — heartbeat cadence in seconds (0 disables)

Default behavior:

• When using Codex and no overrides are provided, the script uses --model gpt-5.2 with xhigh reasoning effort.
• When using Codex, the script always runs with --sandbox danger-full-access and -a never.
• When using Codex, the script enables live web search (--search) so it can do bounded research for unstable facts.
• When using Claude, the script runs non-interactive with --dangerously-skip-permissions.

The script prints the output file path to stdout on success.

Workflow — Option B (inline, legacy)

1. Open references/prompts/plan-sketch.md.
2. Execute it using the user's rough idea as $ARGUMENTS.
3. Do bounded web research (official docs first) for unstable facts:
   • Current CLI/tooling behavior, deprecations, APIs, pricing, supported flags, version constraints.
   • Avoid searching any secrets/PII/proprietary code. Use generic queries (vendor + product + version + error message class).
4. Write a new sketch file under sketch/ with a meaningful kebab-case filename and report the final path.