headless-runner

Headless Session Runner

Overview

Launch a claude -p session in the background via the Bash tool. The wrapper filters the JSON stream to compact one-liner progress and returns a structured JSON summary with context awareness.

When to use this instead of alternatives:

Mode	Use when
Subagent-Driven	Tasks are small, need review between each, want to stay in current session
Parallel Session	Want human-in-the-loop checkpoints between batches
Headless Runner	Want to dispatch work to a background session with minimal context cost, auto context awareness, and structured results

How to Launch

Use the Bash tool with run_in_background: true:

node SKILL_DIR/claude-runner.js \
  --prompt "Your task instructions here" \
  --cwd /path/to/working/directory \
  --model sonnet

Replace SKILL_DIR with the base directory path shown at the top when this skill loads.

No CLAUDECODE= prefix needed. The script automatically unsets the CLAUDECODE environment variable before spawning claude -p, so it works correctly when called from within a Claude Code session via the Bash tool.

CLI Options

Flag	Default	Description
--prompt <text>	(required*)	Task prompt for the session
--resume <id>	(required*)	Resume an existing session
--cwd <path>	current dir	Working directory
--model <alias>	sonnet	Model: sonnet, opus, haiku
--max-turns <n>	100	Max agentic turns
--max-budget <usd>	disabled	Max dollar spend
--system-prompt <text>	none	Appended to system prompt
--allowed-tools <list>	all	Comma-separated tool list
--timeout <seconds>	disabled	Kill session after N seconds

*One of --prompt or --resume is required. Both can be used together (resume + new prompt).

Model Selection

Choose the model based on task complexity:

Model	Use when	Examples
haiku	Simple, mechanical tasks with clear instructions	Rename variables, add boilerplate, run formatters, simple tests, file moves
sonnet	Standard implementation work (default)	Implement features from a clear plan, write tests, fix known bugs, refactoring
opus	Complex tasks requiring deep reasoning	Architectural changes, cross-cutting refactors, debugging subtle issues, tasks with ambiguous requirements

Rules of thumb:

• If the plan has exact code to copy — haiku (cheapest, fastest)
• If the plan describes what to build but the engineer decides how — sonnet (best balance)
• If the task needs judgment calls, design decisions, or debugging — opus (most capable)
• When in doubt, use sonnet — it handles most implementation tasks well

Reading Results

When the background Bash task completes, read the output. The last section after ---CLAUDE-RUNNER-RESULT--- is a JSON summary:

{
  "session_id": "abc-123",
  "model": "claude-sonnet-4-5-20250929",
  "exit_code": 0,
  "duration_seconds": 120,
  "cost_usd": 0.45,
  "tokens": {
    "input": 45000,
    "output": 12000,
    "cache_read": 30000,
    "cache_creation": 15000,
    "context_window": 200000,
    "max_output": 32000,
    "context_used_pct": 38
  },
  "tool_calls": 24,
  "git": {
    "start_sha": "a1b2c3d",
    "end_sha": "f4e5d6a",
    "commits": ["f4e5d6a feat: keyword routes"],
    "changed_files": 8,
    "insertions": 245,
    "deletions": 12,
    "uncommitted_changes": 0
  },
  "errors": [],
  "context_warning": false,
  "resume_command": "node /path/to/claude-runner.js --resume abc-123 --prompt \"Continue\""
}

Context-Aware Batching

Critical: Check context_warning after each run.

• context_warning: false — safe to --resume this session for the next batch
• context_warning: true — context usage is above 60%. Start a new session with --prompt instead of --resume

How context is measured: The wrapper tracks per-turn token usage from the stream (not the aggregate which double-counts the system prompt across turns). It uses the last turn's token counts divided by the effective context window (context_window - max_output) to calculate the true point-in-time context usage. The max_output field shows how many tokens are reserved for the model's response.

Prompt Construction

Detect which mode applies and construct the prompt accordingly:

Plan-based mode (called from writing-plans workflow)

If a plan file exists (e.g., docs/plans/YYYY-MM-DD-feature.md), point the runner at it and let executing-plans handle everything:

--prompt "Use the superpowers:executing-plans skill to implement the plan at docs/plans/2026-02-15-feature.md. Work in the current directory."

The executing-plans skill handles task ordering, batching, verification, and commits internally — do NOT inline task text or micro-manage the steps.

Ad-hoc mode (called directly without a plan)

When there is no plan file, construct a self-contained prompt:

1. Describe the task clearly and completely (the session has no prior context)
2. Include specific instructions: what to build, where, how to test
3. What to report: "When done, summarize what you implemented and any issues."

--prompt "Create a spinning Hello World HTML page at spinning-hello.html. Dark background, colorful text, CSS animations. No external dependencies."

Batching Pattern

Batch 1: --prompt "Use superpowers:executing-plans to implement docs/plans/feature.md"
  Result: context_warning: false
  → Next batch can resume

Batch 2: --resume <session-id> --prompt "Continue executing the plan"
  Result: context_warning: true
  → Next batch must be a new session

Batch 3: --prompt "Use superpowers:executing-plans to implement docs/plans/feature.md. Status: Tasks 1-6 complete (auth middleware, user model, login endpoint). Continue from Task 7."
  (new session, fresh context — brief status helps avoid re-doing work)

For ad-hoc mode, follow the same context_warning logic but include full task details in each new session prompt.

Resume Prompt

• Plan-based (resume): "Continue executing the plan" — the session already has full context
• Plan-based (new session after context warning): Always instruct to use executing-plans with the plan path, plus a brief status of what was completed (task numbers and short descriptions) so the new session picks up where the last left off
• Ad-hoc: Include what was completed and what remains, since there is no plan file to reference

Error Handling

Summary field	Meaning
exit_code: 0	Success
exit_code: 1	Claude session failed — check errors array
exit_code: -1	Killed by timeout
incomplete: true	Stream ended without proper result event
killed: true	Process was killed (timeout or signal)

If exit_code != 0, report the errors to the user and ask how to proceed.

Integration

• Called by: writing-plans (option 3), or directly when user wants background execution
• Pairs with: using-git-worktrees (for isolated workspaces)
• Completes with: finishing-a-development-branch (when all batches done)