Skip to content

skill(rosetta-language-sweep): efficiency lessons from the cobol sweep (#2567) - #11

Merged
squid-protocol merged 1 commit into
mainfrom
docs/sweep-skill-efficiency-lessons
Sep 1, 2026
Merged

skill(rosetta-language-sweep): efficiency lessons from the cobol sweep (#2567)#11
squid-protocol merged 1 commit into
mainfrom
docs/sweep-skill-efficiency-lessons

Conversation

@squid-protocol

Copy link
Copy Markdown
Owner

Summary

Folds the cobol sweep's (gitgalaxy#2567) five efficiency lessons into the skill, inlined at the
point of use rather than as an appendix:

  1. Bucket 2 — the overlap-cost check is one grep against the language's other rules
    (CALL/ENTRY/INVOKE all sitting in cobol's api rule killed the args authoring idea
    in minutes; same decision shape as jcl's DISP=).
  2. Bucket 3 — run the authoring test empirically: scratch dir, git init + commit, one
    --db-only scan, read the column. Documents the engine fact that file-level signal totals
    count rule matches, not per-function derived counts (a 13-name PROCEDURE DIVISION USING still records struct_args 1).
  3. Bucket 3comment_lines is Σ(total_loc − coding_loc), so blank lines count; the
    bucket-1-vs-3 test is one blank+comment grep against morphology siblings.
  4. Phase 4 — the "anything else moving is a red flag" rule gains its known benign cause: a
    stale bias_data.json cache vs engine main (the #2618 case). Attribute movers by which DB
    columns the change feeds instead of re-scanning.
  5. Phase 5workflow_dispatch of rosetta-audit.yml on gitgalaxy main after the
    KEYWORD_ROSETTA_REF bump as the end-to-end closing proof.

Skill/docs-only; no data/ diff, so CI runs all 46 gates — expected green (engine main has
#2622, manifests re-baselined by #10).

Cross-repo

🤖 Generated with Claude Code

https://claude.ai/code/session_01BsVATdyMPhUoAUNBKbMUVi

…p (#2567)

Five instruments the second sweep proved out, inlined where they're used:
- bucket 2: the overlap-cost check is one grep against the language's other
  rules (CALL/ENTRY/INVOKE all in api killed cobol's args authoring idea)
- bucket 3: run the authoring test EMPIRICALLY via a scratch git-init +
  --db-only scan; file-level totals count rule MATCHES, not per-function
  derived counts (13-name USING still records struct_args 1)
- bucket 3: comment_lines = sum(total_loc - coding_loc), blanks count; the
  classification test is one blank+comment grep vs morphology siblings
- Phase 4: the "anything else moving is a red flag" rule has a known benign
  cause (stale bias_data cache vs engine main) -- attribute movers by which
  DB columns the change can feed, don't re-scan in circles
- Phase 5: workflow_dispatch of rosetta-audit on main after the pin bump =
  cheap end-to-end proof the loop closed

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BsVATdyMPhUoAUNBKbMUVi
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant