taskforge-cost

name	taskforge-cost
description	Displays a per-model cost summary. Use when the user says "/taskforge-cost", "cost", "how much did I spend", "token usage", "how much has this cost", or similar. v2 aggregates from telemetry.jsonl.

Cost — Cost Summary (v2)

Summarize costs by model and phase for the project execution, and project remaining costs.

Prerequisites

• A project with execution state must exist under _workspace/projects/{projectId}/
• Project selection: Same as /taskforge-execute — auto-select if only one, ask if multiple
• All file paths below are relative to _workspace/projects/{projectId}/

Data Sources (v2)

Primary: telemetry.jsonl — one JSON object per line. Relevant events:

Event	Fields used
task_end	model, costUSD, tokens, wallMin, outcome, taskId
sprint_complete	sprintId, totalCostUSD, guardrailEvents
guardrail_triggered	taskId, type, value, limit
plan_approved	milestoneId, estimatedCostUSD
quick_task	costUSD, tokens, wallMin

Fallback (if telemetry.jsonl is missing or empty): execution-state.json per-task cost fields.

PM calls (plan) and validation (validate) costs are tracked as separate telemetry events and bucketed into "Other" in the summary.

Estimating Remaining Cost

Project remaining task costs using per-model average rates from completed tasks:

Estimated remaining = Σ (average costUSD per model × remaining tasks per model)

Per-model average:
  - Average costUSD of completed haiku tasks × remaining haiku tasks
  - Average costUSD of completed sonnet tasks × remaining sonnet tasks

For models with no completed tasks yet, use default estimates:
  - haiku:  $0.002/task
  - sonnet: $0.04/task

Note: code-writing (execution) tasks only use haiku or sonnet. Opus is reserved for judgment work — planning, plan-edit, discover, retro, milestone QA, and harness orchestrators — and is tracked separately under "Other".

Output Format

Template — fill from telemetry/state. Do not copy literal labels.

비용 요약 — {projectName}

모델별 (실행):
  haiku:  {n} runs,  ${cost}  ({pct}%)   | avg ${avg}/task
  sonnet: {n} runs,  ${cost}  ({pct}%)   | avg ${avg}/task
  ─────────────────────────
  Subtotal: {n} runs, ${cost}

단계별:
  {milestoneId} ({title}):   ${cost} ({done}/{total} tasks)    {statusLabel}
  {milestoneId} ({title}):   ${cost} ({done}/{total} tasks)    {statusLabel}
  {milestoneId} ({title}):   미시작
  {milestoneId} ({title}):   미시작

기타 (PM/검증, opus):
  계획 (plan):       ${cost}
  마일스톤 QA:        ${cost}
  스프린트 검증:      ${cost}  (자동, sonnet)
  ─────────────────────────
  Subtotal:          ${cost}

총 비용:             ${cost}

가드레일 발동: {n}회 (이번 마일스톤)
  - {taskId}: {limitName} 초과 ({actualValue})

예상 남은 비용:
  haiku  × {n} = ~${cost}
  sonnet × {n} = ~${cost}
  PM/검증    = ~${cost}
  ─────────────────────────
  남은 예상:   ~${cost}
  총 예상:     ~$0.84

Aggregation Rules

1. By model (execution): group all task_end events by model, sum costUSD, count runs.
2. By milestone: group task_end events by the milestoneId derived from the task's taskId prefix (e.g. M1-S1-T3 → M1). Cross-check with project-plan.json for total task counts.
3. Other: aggregate non-execution events — plan runs, validate runs, milestone QA. These can use opus.
4. Guardrail summary: count guardrail_triggered events for the current milestone. Show the top 3 offenders (taskId + trigger type).

When Data Is Unavailable

If neither telemetry.jsonl nor execution-state.json has cost data yet:

• "아직 실행 내역이 없어서 비용 데이터가 없어요. 태스크가 실행되면 자동으로 기록됩니다."
• Show the plan's estimatedCostUSD (from plan_approved event) if available, as a rough upfront estimate.

Notes

• telemetry.jsonl is the source of truth. Only fall back to execution-state.json when telemetry is missing.
• Opus usage in "Other" is expected and normal — it's the PM model. An opus entry under "모델별 (실행)" would be a bug.
• Guardrail events do not directly add cost, but a high count usually correlates with higher-than-estimated cost for that milestone.