Interactive · paste your own numbers
One file instead of
twenty-nine thousand tokens.
Every cold agent session starts the same way: list the repo to find where work lives, search for active tasks, readcandidate files to reconstruct state — before the first real edit. barkcli's auto-maintained project_context.md is that state, compiled, verified against git, and ~515 tokens. The page below is the measurement.
What the benchmark measures
A deterministic proxy for agent cold-start — no LLM runs, no vendors involved. It simulates what an agent must do without the context file, tokenized the same way the file itself is counted:
the agent finds where work lives — every file path in the repo
top hit files opened line-by-line to reconstruct state
searches over card titles + code to guess active work
Generated, staleness-checked, and reconciled against git by barkcli context digest / verify. Hand-edited sections are preserved and flagged in a verification footer when drifted.
Trust it on your own repo
The proxy is deterministic — same repo, same numbers. Run it, then paste the output below to visualize your own ratio (everything stays in your browser):
barkcli context benchmark --jsonHonest limits
- It's a proxy: no LLM is in the loop, so it measures token payloads, not comprehension. A 57× cheaper payload is necessary — an agent still has to read the file.
- Thin repos under-report: no cards, no files to read → low cold numbers. Try it on a repo with a real board.
- Lockfiles, build output and page HTML are excluded from scanning; titles used for searching come from your actual cards.
- Claimed on the barkcli repo itself @
339f66a, Oct 2026 (29,402 → 515 = 57.1×; a conservative 3-candidate variant reads 44× less).