Skip to content

EXAMPLE.md

Latest commit

 

History

History
82 lines (60 loc) · 3.99 KB

EXAMPLE.md

File metadata and controls

82 lines (60 loc) · 3.99 KB
Raw

bean — a worked example

A short walkthrough of what bean changes. The point isn't the task; it's that the answer converges — bean re-decides as it learns, and it doesn't stop at "looks done with a few caveats."

The task

Replace the deprecated parseDate() helper with parseDateUtc() everywhere it's used in this service, and make sure nothing breaks.

Without bean (one-shot)

The model greps parseDate(, find-and-replaces the hits, runs the test suite (green), and reports success. It missed two call sites — one imported under an alias, one invoked through a thin wrapper — that the literal grep never matched. Tests stayed green because those paths weren't covered. It breaks in production. The wrong belief — "the helper only appears as a literal parseDate( call" — was never challenged.

With bean (convergence loop)

Frame. Goal: every call resolves to parseDateUtc, nothing breaks. Seed the ledger:

c1 constraint  no behavior change for callers     (stated)
c2 factual     helper appears as literal calls     (stated)

Round 1 — survey, investigate the most decisive front

c2 is stated and load-bearing, so prove the set of call forms before changing anything. Fan out: one search agent per package, briefed to find the symbol in any form — literal call, aliased import, re-export, dynamic/wrapped invocation. Grounding beats assumption: they read the actual files, not a single grep.

  • Record: c3 factual — usages found per package, at documented (read from source).
  • Compile: conflict — c3 shows an aliased import and a wrapper call, contradicting c2's "literal only"; one package's count came from a single agent (single-source).
  • Revise belief: supersede c2 → c3 ("4 call forms, incl. alias + wrapper"). This is the belief the one-shot run never killed.

Round 2 — drive to the evidence bar

Investigate the compiler's next front: corroborate the single-source package, apply the change to every form, draft per-call edits.

  • Record: c4 recommendation — exact change per call form (documented).
  • Compile: no conflicts, but one dynamic call is still stated — static search can't prove it's covered. Not converged.

Round 3 — don't stop at the caveat (verify by running)

The one-shot run would have shipped here with a footnote. bean doesn't: it drives the open front to closure. Add a test that actually exercises the dynamic call, then run the suite — execution is the check, not a re-read. The wrapper path fails → fix it → re-run green. The stated caveat becomes a tested fact instead of a flagged risk.

  • Record: c5 factual — all call forms resolve to parseDateUtc, suite green (tested).
  • Cross-model blindspot: for a behavior-changing sweep, have Codex review the raw diff independently (it judges the diff, not your summary). It flags a comment still referencing the old name — cosmetic, fixed.
  • Compile: no unresolved conflicts; every load-bearing claim tested; a full round added only the cosmetic note. Converged.

Deliver (the honest residual)

Converged in 3 rounds. Revised 1 belief (call forms: literal-only → 4 forms incl. alias + wrapper — the ones a naive grep misses). Drove every gap to a tested fix. The only thing I could not drive to closure is genuinely out of reach: one downstream service consumes this helper's output format and I can't run it from here — flagged for its owner to confirm.

That last line is the distinction bean enforces: a residual is something genuinely unreachable, not something you stopped short of. "Here are some caveats" is not done; "here is the one thing I cannot verify, and why" is.

The difference

Same model, same capability. bean didn't make it smarter — the compiler surfaced the "literal-only" belief as a conflict in round 1 and forced its revision, and the never-stop-at-caveats rule turned the round-2 caveat into a round-3 verified fix instead of a production break.