| name | harness-work |
| description | Executes plans.json tasks — solo, parallel, or breezing team mode. Use when implementing tasks or running the work loop. |
| when_to_use | implement task, run task, execute plans, work on task, run all tasks, parallel workers, breezing |
| allowed-tools | ["Read","Write","Edit","Grep","Glob","Bash","Task"] |
| argument-hint | [all|task-number|range|--codex|--opencode|--parallel N|--commit|--resume id|--breezing|--auto-mode] |
| effort | high |
| model | sonnet |
Harness Work
Unified execution skill for Harness.
Consolidates the following legacy skills:
work — plans.json task implementation (auto scope detection)impl — Feature implementation (task-based)breezing — Full team auto-executionparallel-workflows — Parallel workflow optimizationci — CI failure recovery
Quick Reference
| User Input | Subcommand | Behavior |
|---|
harness-work | auto | Auto-selects based on task count (see below) |
harness-work all | auto | Executes all incomplete tasks in auto mode |
harness-work 3 | solo | Immediately executes task 3 only |
harness-work --parallel 5 | parallel | Forces parallel execution with 5 workers |
harness-work --codex | codex | Delegates to Codex CLI (explicit only) |
harness-work --opencode | opencode | Delegates to OpenCode CLI (explicit only) |
harness-work --breezing | breezing | Forces team execution |
Execution Mode Auto Selection (auto-selection when no flag is specified)
When no explicit mode flag (--parallel, --breezing, --codex, --opencode) is provided,
the optimal mode is automatically selected based on the number of target tasks:
| Target Task Count | Auto-Selected Mode | Reason |
|---|
| 1 task | Solo | Minimal overhead. Direct implementation is fastest |
| 2-3 tasks | Parallel (Task tool) | Threshold where Worker isolation benefits emerge |
| 4+ tasks | Breezing | Lead coordination + Worker parallelism + Reviewer independence is effective |
Rules
- Explicit flags always override auto mode
--parallel N → Parallel mode (regardless of task count)--breezing → Breezing mode (regardless of task count)--codex → Codex mode (regardless of task count)--opencode → OpenCode mode (regardless of task count)
--codex activates only when explicitly specified. Not auto-selected because Codex CLI may not be installed in some environments--codex can be combined with other modes: --codex --breezing → Codex + Breezing--opencode activates only when explicitly specified. Not auto-selected because opencode-plugin-cc may not be installed in some environments
Options
| Option | Description | Default |
|---|
all | Target all incomplete tasks | - |
N or N-M | Task number/range specification | - |
--parallel N | Number of parallel workers | auto |
--sequential | Force sequential execution | - |
--codex | Delegate implementation to Codex CLI (explicit only, not auto-selected) | false |
--opencode | Delegate implementation to OpenCode via opencode-plugin-cc (explicit only, not auto-selected) | false |
--commit | Commit after each task is done (default: off) | false |
--resume <id|latest> | Resume previous session | - |
--breezing | Lead/Worker/Reviewer team execution | false |
--no-tdd | Skip TDD phase | false |
--no-simplify | Skip Auto-Refinement | false |
--auto-mode | Explicitly enable Auto Mode rollout. Only considered when the parent session's permission mode is compatible | false |
Token Optimization (v2.1.69+): For lightweight tasks that don't involve git operations,
enable includeGitInstructions: false in plugin settings to
reduce prompt token usage.
1-Hour Prompt Cache (breezing sessions > 30 min): source "${CLAUDE_PLUGIN_ROOT}/scripts/enable-1h-cache.sh" before starting. Exports ANTHROPIC_CACHE_CONTROL=max-age=3600 so all spawned Workers, Reviewers, and hooks share a cached system prompt — reduces token cost and latency across long sessions.
Scope Dialog (when no arguments provided)
Note: A lightweight drift check (bash "${CLAUDE_PLUGIN_ROOT}/scripts/plans-drift-check.sh") runs before the scope dialog regardless of which option the user chooses. If stale markers are detected, the dialog is preceded by a drift summary and a confirmation prompt.
harness-work
How far do you want to go?
1) Next task: The next incomplete task from plans.json → Execute in Solo mode
2) All (recommended): Complete all remaining tasks → Auto-select mode based on task count
3) Specify numbers: Enter task numbers (e.g., 3, 5-7) → Auto-select mode based on count
If arguments are provided, execute immediately (skip dialog):
harness-work all → All tasks, auto mode selectionharness-work 3-6 → 4 tasks, so Breezing is auto-selected
Effort Level Control (v2.1.68+, simplified in v2.1.72)
Claude Code v2.1.68 sets medium effort (◐) as default for Opus 4.6.
v2.1.72 removed the max level, simplifying to 3 levels: low(○)/medium(◐)/high(●).
/effort auto resets to default.
For complex tasks, use the ultrathink keyword to enable high effort (●).
Multi-Factor Scoring
At task start, the following scores are summed, and ultrathink is injected when the threshold reaches 3 or above:
| Factor | Condition | Score |
|---|
| File count | 4+ files to be changed | +1 |
| Directory | Includes core/, guardrails/, security/ | +1 |
| Keywords | Contains architecture, security, design, migration | +1 |
| Failure history | Agent memory contains failure records for the same task | +2 |
| Explicit specification | PM template includes ultrathink notation | +3 (auto-adopted) |
Injection Method
When score >= 3, prepend ultrathink to the Worker spawn prompt.
The same logic applies in breezing mode (managed centrally by harness-work).
Execution Modes
[ralph] marker pre-dispatch rule: Before any standard worker dispatch (solo or breezing),
harness-work checks the task description for the [ralph] marker. If present, it delegates to
harness-ralph-loop instead of the standard worker flow. Ralph tasks serialize within a session
(only one Ralph loop runs at a time). See references/solo-mode.md
and references/breezing-mode.md for dispatch details.
Worker Contracts
Failure & Review
Templates
Task Status SSOT
.claude/harness/plans.json is the sole source of truth for task status. All reads via harness plan-cli list/get, all writes via harness plan-cli update. See harness/references/cli-reference.md. The native Claude Code task list is not mirrored.
Related Skills
harness-plan — Plan the tasks to executeharness-sync — Sync implementation with plans.jsonharness-review — Review implementationsharness-release — Version bump and release
