Re-ran the README benchmark on the current build (7 repos reindexed, median of 4): avg 35% cost / 57% tokens / 46% time / 71% tool calls saved — reproduces the published README (35/59/49/70), no regression. Adds bench-readme.sh + parse-bench-readme.mjs harness.
Fixes a token-measurement bug: result.usage is last-turn-only in current Claude Code; must sum per-turn assistant usage for cumulative tokens. Corrects the earlier excalidraw note (its '-34% tokens' was off this bug; real ~90%) and the cost MECHANISM (volume/fewer-turns, not cache-ability — the without-arm's huge token volume is mostly cheap cache-reads, so token savings 57% > cost savings 35%). Cost/time were always correct.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Fresh with-vs-without A/B on excalidraw (current build, n=3): 3x faster (49s vs 145s), 15x fewer tool calls, ~0 vs 23 reads, and -40% cost ($0.41 vs $0.68). Cost is neutral-to-lower, not flat — compact codegraph answers cache across turns while the without-arm's read/grep thrash is fresh, poorly-cacheable input. Recorded in call-sequence-analysis.md; corrected the CLAUDE.md optimization-target note (still: don't optimize for cost; target wall-clock + tool-call count).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Records why codegraph read savings (-75%) under-convert to wall-clock (-16%): the bottleneck is round-trips + the synthesis turn, not reads. Ablation (arms A-I) shows explore is 68% of payload but load-bearing, trace is path-scoped but under-adopted, instruction/description steering cannot match an append-prompt's salience (and regresses), and the shippable win is making the trace output sufficient (arm I). Adds harness: seq-matrix, run-arms/arms-*, parse-arms.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
37-cell matrix (every flow-relevant language × small/medium/large indexed
repos): a headless agent answers one canonical flow question per repo, with the
codegraph MCP vs without any MCP. Fresh re-index per cell so the with-arm
reflects current resolvers.
Result: 75% fewer file reads with codegraph (40 vs 158 across cells), ~70%
fewer greps, never more reads in any cell. Biggest wins on medium/large
backends (excalidraw 0R vs 9R, spring-halo 0R vs 9R+8 Bash, jellyfin 4R vs 13R+
21 Bash + a spawned sub-agent); tie zone on tiny repos where the flow fits in
1-2 files.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Colby McHenry