Fresh Slack thread · fixed context budget

ForgeBot Starts 32–37K Tokens Deep

A new thread begins with roughly one-tenth of the 372K-token window already occupied. Most of that is tool descriptions and the 336-skill catalog—not loaded skill documents or conversation history.

32.1Krecent live Slack first-call measurement
37.1Kcurrent offline fresh-session estimate
8.6–10%of the 372K-token context window at startup
334–340Ktokens remain before work begins

Top 10 startup payloads

1
Skills index
7,945 tok
2
Core system guidance
6,443
3
Computer-use schema
2,676
4
BFL video schemas
2,621
5
Scheduled-job schema
2,459
6
Session-search schema
1,749
7
File-tool schemas
1,676
8
Delegation schema
1,446
9
Skill-tool schemas
1,414
10
Terminal schemas
1,319

These ten account for 29.7K tokens—80% of the fixed startup estimate.

What loads by default

TOOLSSKILLS INDEXCORE
Tool schemas: ~21.1K
Skills index: ~7.9K
Core guidance: ~6.4K
Memory, profile, rules: ~1.6K
Important distinction. Full SKILL.md files do not load at startup. Only each skill’s name and short description load. The recent 81% reduction in large skill documents helps after a skill is opened, not at thread creation.

Why two totals?

The current offline inspector estimates 37.1K tokens from 64KB of system text plus 84KB of tool definitions. Recent live Slack thread logs started at about 32.1K.

The estimator and running gateway do not expose exactly the same gated tool set. The honest startup figure is therefore a 32–37K range, not fake single-token precision.

Both measurements agree on the decision: startup occupies about one-tenth of the window. Forced compaction comes later from oversized tool results and repeated workflow pivots, not from startup alone.

Best reduction targets

1

Trim the 336-skill catalog

Its index costs 7.9K tokens every thread. Consolidating aliases and shortening descriptions cuts startup directly.

2

Stop shipping inactive video schemas

Six BFL video tools cost 2.6K tokens even though video generation is currently unavailable. This is the cleanest obvious cut.

3

Gate heavyweight specialist tools

Computer use, scheduled jobs, and session search cost 6.9K combined. Load them only where the thread needs them if platform routing can preserve reliability.

What the fixed payload contains

System identity and operating rules define ForgeBot’s role, safety, execution discipline, communication style, project rules, environment, and platform behavior. Tool schemas describe the arguments for every callable tool. The skills index lists 336 available procedures. Memory and user profile add durable ForgeFX and Adam preferences. The first user message and thread history come afterward.

Prompt caching reduces repeated billing/latency for this fixed prefix, but it does not make these tokens disappear from the model’s context window.