33,700 tokens on boot and nobody noticed

  BOOT SEQUENCE (read on every agent wake)

    33.7k  ████████████████████████████████████████  session-decisions.md
     6.2k  ████████                                  analyst.md
     4.6k  ██████                                    lexicon.md
     4.6k  ██████                                    quartermaster.md
     4.3k  ██████                                    AGENTS.md
     3.2k  ████                                      architect.md
     2.7k  ████                                      anotherpair.md
     2.6k  ████                                      weaver.md
     2.5k  ███                                       watchdog.md
     2.5k  ███                                       sentinel.md
     2.4k  ███                                       janitor.md
     2.3k  ███                                       dead-reckoning.md
     2.0k  ███                                       scribe.md
     1.7k  ███                                       keel.md
     1.2k  ██                                        maturin.md
  ───────
   112.9k  BOOT TOTAL

files     = git ls-files
tokenizer = cl100k_base

for file in files:
    tokens = encode(content).length + encode(path).length

sort descending
tag boot files

5 agent-minutes to build. 3 seconds to run. The fix (stop loading all 271 historical decisions on every wake) drops 30k tokens from the boot sequence.¹

Once the boot path was visible file by file, there was very little design work left. The largest source of waste was obvious, the remedy stayed small, and the same check could be rerun whenever the repository changed. Five minutes spent making the cost inspectable saved a much longer argument about where to optimise.

Provenance | Script