Skill: interrupt-budget

CLI Bootstrap

在执行任何 harnessctl 命令前，先解析本地 CLI 路径：

if [ -z "${HARNESSCTL:-}" ]; then
  candidates=(
    "./stage-harness/scripts/harnessctl"
    "../stage-harness/scripts/harnessctl"
    "$(git rev-parse --show-toplevel 2>/dev/null)/stage-harness/scripts/harnessctl"
  )

  for candidate in "${candidates[@]}"; do
    if [ -n "$candidate" ] && [ -x "$candidate" ]; then
      HARNESSCTL="$candidate"
      break
    fi
  done
fi

test -n "${HARNESSCTL:-}" && test -x "$HARNESSCTL" || {
  echo "harnessctl not found. Set HARNESSCTL=/abs/path/to/stage-harness/scripts/harnessctl" >&2
  exit 1
}

Track and enforce the hard limit on user interrupts per epic.

Purpose

Autonomy is valuable only if it doesn't create more work. The Interrupt Budget is a hard constraint that forces the AI pipeline to batch questions, make reasonable assumptions, and avoid pestering the user — while preserving a guaranteed channel for genuinely critical decisions.

Budget by Risk Level

Risk Level	Total Budget	CLARIFY	SPEC	PLAN	EXECUTE
low	1	0-1	0	0	0
medium	2	1	0-1	0	0
high	3	1-2	1	0-1	0

EXECUTE never interrupts — if blocked during execution, it logs to unknowns-ledger and completes what it can.

Budget Storage

Budget state lives in state.json under the interrupt_budget field (NOT a separate file):

// .harness/features/<epic-id>/state.json (relevant excerpt)
{
  "interrupt_budget": {
    "total": 2,
    "consumed": 0,
    "remaining": 2
  }
}

Access via CLI: $HARNESSCTL budget check --epic-id <epic-id> Consume via CLI: $HARNESSCTL budget consume --epic-id <epic-id>

Budget Consumption Rules

When to consume a budget unit

• User is actually interrupted (shown a decision-packet.json)
• User responds with explicit answers or approvals

When NOT to consume

• Auto-proceeding with assumable decisions (no user shown)
• Deferring decisions to unknowns-ledger (no user shown)
• Stage gate confirmations (these are progress notifications, not decisions)

Budget Exhaustion Protocol

If remaining == 0 and a new must_confirm decision arises:

1. Attempt reclassification as assumable with safe default
2. If reclassification not safe: add to unknowns-ledger as CRITICAL-BLOCKED
3. Proceed with safe default, mark task with ⚠️ ASSUMPTION flag
4. Surface all CRITICAL-BLOCKED items in harness-status output
5. Never exceed budget — the pipeline makes a best-effort decision

Stage Gate vs Interrupt

Stage gates (end of CLARIFY, SPEC, PLAN) show summaries and ask "proceed?". These are:

• NOT counted as budget interrupts — they are progress checkpoints
• Required even at low risk level
• Can be skipped only with /harness:auto (medium risk or lower)

Usage

The Lead Orchestrator calls this skill before each potential interrupt:

Check interrupt budget before presenting Decision Packet:
- Run: $HARNESSCTL budget check --epic-id <epic-id>
- If remaining > 0: present packet, then run: $HARNESSCTL budget consume --epic-id <epic-id>
- If remaining == 0: apply Budget Exhaustion Protocol