Extraction paused: main's extractor undercounts from v2.1.218 on #18

deterministic-extract is disabled until the extractor can count correctly again. It is currently wrong in both directions.

Main's method undercounts (v2.1.218 onward)

Main's extractor greps process.env.<NAME> literals. Starting with v2.1.218, Claude Code reads many variables through a declared env registry instead, and those reads are not literals. Evidence:

• On main: v2.1.197 = 655 vars, v2.1.218 = 538. 122 of the 126 names that v2.1.218 records as "removed" still appear verbatim in the v2.1.296 binary. So v2.1.218 (538) and v2.1.222 (539) are lower bounds, not counts.
• Bot PRs Document v2.1.263 #12..Document v2.1.296 #17 (v2.1.263..v2.1.296) each report 149..166 removals vs v2.1.222. For Document v2.1.296 #17, 153 of its 166 "removed" names are still in the v2.1.296 binary. ANTHROPIC_DEFAULT_FABLE_MODEL is one of them.

#7's registry-aware method overcounts on new releases

#7 reads the registry. It gave a plausible 994 on v2.1.234, but on v2.1.302 it reports 2,917 names, 1,309 of which are not env-var shaped (return, takeDown, EmptyResultSchema), and every validation gate still passed. Details: #7 (comment)

To resume

1. Fix Document v2.1.234; re-measure v2.1.170+ against the declared env registry #7's extractor and add the two plausibility gates its validation lacks:
   • registry names must not exceed typed getters by an implausible margin
   • the total must not move more than a set fraction vs the previous documented release
2. Re-extract v2.1.218 and v2.1.222 with the fixed method, then the newest release.
3. Re-enable deterministic-extract. The one-PR-at-a-time guard from claude-code-http-spec#21 should come with it, because this workflow also opens a new PR every week whether or not the last one merged.

Until then, the README's last trustworthy count is v2.1.197 (655).

🤖 Generated with Claude Code