Github's new AI Credit Usage Report — DevLoop Session
Session: waymark-devloop-controlled
Date: 2026-06-03 (22:18–01:14 KST / 13:18–16:14 UTC)
Model: claude-sonnet-4.6
Extra cost incurred: ~$10 USD overage
Devloop: Custom multi-agent AI development pipeline for software development tasks, built on top of Copilot and Claude.
Teststop: A tool that runs the full test suite and returns results as context for the next assistant turn using Claude.
Summary: This report analyzes the root causes of the unexpectedly high AI credit usage during a single DevLoop session, which ran for ~3 hours and consumed approximately 36.4 million tokens, leading to a ~$10 overage charge on top of the monthly Copilot plan.
Total Token Consumption
| Metric | Value |
|---|---|
| Input tokens | 35,952,200 |
| Output tokens | 451,307 |
| Total tokens | 36,403,507 (~36.4M) |
| User messages sent | 7 |
| Assistant events generated | 870 |
| Session duration | ~3 hours |
Tool Call Breakdown
| Tool | Calls | Impact |
|---|---|---|
bash |
766 | Command output added to context each time |
view |
233 | Large source files read into context |
task (subagents) |
32 | Each returned 500–2,000 lines of output |
read_agent |
31 | Full reviewer/architect output injected back |
read_bash |
36 | Worker output read back into context |
grep |
21 | File search results added to context |
glob |
16 | File listings added to context |
Root Causes (Ranked by Impact)
🔴 1. Context Snowball — ~80% of tokens
Copilot re-sends the entire accumulated conversation history on every assistant turn. With 870 events over 3 hours, the context grew exponentially:
- Turn 1: ~10K tokens in context
- Turn 100: ~100K tokens in context
- Turn 870: millions of tokens re-sent every time
A single never-reset session running 10+ architect → implement → review → fix loops caused this.
🔴 2. Verbose Subagent Output Injected Back (32 task calls)
Each @devloop-architect and @devloop-reviewer subagent returned 500–2,000 lines of output, which was read back via read_agent (31 calls) and added permanently to the growing context.
Reviewer verdicts alone averaged ~200 lines each across 10+ reviews.
🟡 3. 766 Bash Calls
Every devloop work, devloop fix, npm run build, git command, and teststop run returned output that was kept in the ever-growing context window.
🟡 4. Large Source Files Read 233 Times
Files like database.ts (3,000+ lines) and server.ts (1,300+ lines) were repeatedly read via view calls and added to context.
🟡 5. Long Session Summary at Start
The session started with a ~300-line prior checkpoint summary injected into the system prompt, re-sent on every single one of the 870 turns.
Tasks Completed During This Session
| Task | Result | Approx. Loops |
|---|---|---|
| Session rollback races | ✅ APPROVED 9/10 | 3 fix loops |
| EventTopic build fix | ✅ Direct commit | — |
| Config trailing newline trim | ✅ Direct commit | — |
| Slack dedup key fix | ✅ Direct commit | — |
| Token approve-after-reject | ✅ Direct commit | — |
| PID reuse race (TASK-20260604-004255) | ✅ APPROVED 9/10 | 1 fix loop |
| Hub stop inflight guard (TASK-20260604-004538) | 🔄 Fix in progress | 1+ fix loops |
Why It Cost ~$10 Extra
Claude Sonnet 4.6 pricing is approximately $3 per million input tokens.
At 35.9M input tokens:
- Estimated gross cost: ~$108 token-equivalent
- Covered by Copilot plan: monthly included quota
- Overage charged: ~$10 USD (tokens beyond plan limit billed pay-as-you-go)
How to Prevent This in Future Sessions
| Action | Savings Estimate |
|---|---|
Reset session every 2–3 tasks (/clear or new session) |
60–70% reduction |
| Truncate reviewer output — verdict + bullet points only, not 200-line reports | 15–20% reduction |
Don't read_agent full output — check verdict line only, skip verbose body |
10% reduction |
| Run teststop earlier — catch failures before running multiple fix loops | Fewer loops = less context |
| Avoid reading large files repeatedly — cache content in session state | 5–10% reduction |
One-Line Root Cause
DevLoop ran 10+ architect → implement → review → fix loops in a single never-reset 3-hour session, causing the context to balloon to 36.4 million tokens and exceed the monthly Copilot plan quota. [planning to work for the devloop improvements]
Based on above usage, there are also place of improvements for copilot cli too.
Specially heavily token usage alert
Auto context compact for repetitive context
Top comments (0)