Every context file pays rent on every run—even when it is nowhere near a hard limit. Measure the drag, then compact toward a target while keeping every inventoried rule recoverable. Intentional rule loss belongs to the separate guided ablation mode below.
Private by design. Your context is never sent to the leaderboard.Pasted text stays in your browser. For a public GitHub link, the server retrieves only the repository listing and file you choose. Ranked results require separate, public before/after GitHub links.
Scan a public GitHub repo
GitHub only. The scanner checks a bounded public repository tree for common agent-context files and never follows arbitrary URLs.
COMMUNITY RESULTSServer-calculated results from public GitHub before/after revisions, ranked against Lab 001's 41.6% result.
Loading…
LAB 001 BENCHMARK · CONTEXT DIET
The measured run behind the projection.
Recorded results from the original Lab 001 benchmark. The estimator above uses those measured strategies as a screening band; the installed skill audits your actual file before recommending a safe, lean, or aggressive plan.
BEFORE → AFTER · ORIGINAL LAB 001 CORPUS
The benchmark winner cut 20,647 chars (41.6%). The limit was a constraint in this run—not a prerequisite for using Context Diet.
Four historical strategies, all verified 100% faithful against the true 124-rule inventory. The current skill also treats “no change” as a valid outcome and separates faithful compaction from owner-approved retirement.
Externalize + link ★ winner41.6%
29,040 chars✓ 100% faithfulunder target
Hybrid route31.7%
33,916 chars✓ 100% faithfulunder target
Telegraphic / caveman23.6%
37,949 chars✓ 100% faithfulunder limit
Condense in place21%
39,266 chars✓ 100% faithfulunder limit
TWO MODES · DIFFERENT CLAIMS
Compaction preserves. Ablation tests an omission.
Context Diet does not treat a smaller file as proof of a better one. Choose the path that matches the change you are actually making.
Normal mode
Rule-preserving compaction
The default path inventories the original rules, generates review-only candidates, and adversarially checks the complete result—including linked files. A candidate with a weakened or missing load-bearing rule is disqualified.
Measures exact characters and estimates tokens.
Compares externalize, condense, telegraphic, and hybrid strategies.
Targets 100% recoverable rules; externalization is movement, not deletion.
Leaves application to user review through the host's normal edit flow.
Use ablation when the goal intentionally removes context. The controller checkpoints the original, derives deletion units locally, protects sensitive blocks, and stages a bounded omission without editing the live target.
Tests exact configured model routes against a sealed behavioral suite.
Blocks acceptance on regression, missing coverage, uncertain judgment, or stale evidence.
Requires an explicit trial ID and candidate SHA before the only candidate-apply path.
Journals atomic writes and supports exact, reversible rollback.
“No regression observed” applies only to that exact route, snapshot, and sealed suite. It does not prove a rule is universally safe to delete.
For ordinary compaction, Context Diet measures the file before proposing any rewrite. It then compares four strategies and qualifies only candidates that remain under the target without weakening or losing a load-bearing rule.
Measure: map exact character cost by section.
Inventory: extract atomic rules and identify protected constraints.
