Every token in the first request is paid again on every call. MendCode measures its prompt per tier, lets you pick how much harness the model sees, defers secondary tools behind tool_search and shows compaction as it happens.
Generated editorial cover. Product screenshots inside the post are unedited.
Measured, not guessed
The system prompt and tool list are sent on every model call, so their size is a cost on every turn, not once. MendCode counts the input tokens of the first request of a fresh session for each tier, with gpt-6-luna on the ChatGPT backend, no MCP and no skills. Your AGENTS.md and skills add to it.
Data visual
Input tokens on the first request
gpt-6-luna on the ChatGPT backend, fresh session, no MCP, no skills.
bar view
Input tokens
Input tokens on the first request
setup
Input tokens
minimalist
1064
pi 0.85.1
1188
balanced
2421
full
3433
pi 0.85.1 with its four tools, measured the same way.
Prompt modes
/prompt-mode chooses how much of MendCode’s harness the model reads, separately from the tier:
Modes
Mode
The model reads
minimal
A small boundary, a docs pointer, your project and user instructions and the selected memories
focus
The native coding harness, your instructions, the skill catalog, four working habits and a family note (the lean default)
guide
Focus with a longer guide to how the model’s family uses MendCode’s tools; about 560 to 690 more tokens per call
full
Guide plus a short operator’s guide to MendCode and the family’s context files, such as CLAUDE.md or GEMINI.md
Instruction files may include others with @relative/path lines, confined to the project, up to depth 8, 100 files and 64,000 characters.
Tools on demand
Every tool schema in the request costs tokens too. With tool discovery on (the default in the full tier when MCP is on), core tools stay visible and the rest (other built-ins, MCP, custom and extension tools) are found with tool_search, which searches names and descriptions locally and returns 1 to 8 tools.
The context page, drawn from MendCode’s own render code.
Compaction you can see
When a session nears its window, compaction folds the history into a memory. /compaction opens its progress, the folded memory and a scratchpad, and a live panel above the prompt shows the phase and the tokens before and after. The footer says how many turns are left before it happens.
mendcode — compaction
fix/coupon-cap ~/Code/shop-apiCap coupon discounts at t+33-2› Orders with a capped coupon get the full percentage off. Find why…• The discount is now min(subtotal × rate, cap).• The new test covers the cap: 3 passed.┌─ Context packed ───────────────────────────────────────────────────┐│ ││ auto/compaction open ││ ││ ▸ Compacted memory · packed ││ ││ Follow-up ││ Orders with a capped coupon get the full percentage off. Find why, ││ fix it and add a test. ││ │└────────────────────────────────────────────────────────────────────┘────────────────────────────────────────────────────────────────────❭ ────────────────────────────────────────────────────────────────────Build · GPT-6-Luna · native↻ 1 loop
