Summary
scan --limit 2 runs the agentic enhance phase across the entire dataset before analyzing 2 units. Intended as a cheap smoke test on a real repository, it was on track to cost roughly a tenth of a full scan and was killed partway through.
Evidence
The flag is declared as:
scan_p.add_argument("--limit", type=int, help="Max units to analyze") # openant/cli.py:1315core/scanner.py:475 documents the intent for a neighbouring phase — "--limit governs the analyze stage, not how many units the LLM reachability pass reviews — it must see the full codebase to find missed entry points" — and enhance behaves the same way. The help text says only "Max units to analyze," which reads as bounding the run.
Observed: --limit 2 produced [Enhance] Units to enhance: 1351.
Why it matters
Combined with the documented absence of a spend ceiling, the most natural way to try OpenAnt cheaply against a real codebase is also the one that incurs near-full cost. There is no warning, and the phase that runs unbounded is a high-volume per-unit LLM phase.
Suggested fix
Either have --limit bound enhance as well, or add a separate --enhance-limit, and amend the help text to state explicitly that enhance always processes the full dataset. Enhance checkpointing already works correctly, so a bounded enhance costs nothing in resumability.
Observed on one full-pipeline run at production scale: a private TypeScript monorepo (Angular front end, NestJS back end), 1,351 analysis units from 866 source files, run with --verify and no --limit, ~6.5 hours wall clock. The engine was invoked directly (python -m openant scan …, Python 3.13); the Go CLI was not built, so none of this involves the Go↔Python envelope. Stage 1 detection and Stage 2 verification ran on Claude Opus 5; application context, enhancement and reporting on Claude Sonnet 5 — all seven phases routed through OpenRouter (anthropic/claude-opus-5, anthropic/claude-sonnet-5) via a locally-added openrouter provider adapter, on a build based on upstream 2ed78f6. Cost figures are expressed as proportions of the run total; absolute amounts are omitted deliberately, as is the identity of the scanned repository.
Summary
scan --limit 2runs the agentic enhance phase across the entire dataset before analyzing 2 units. Intended as a cheap smoke test on a real repository, it was on track to cost roughly a tenth of a full scan and was killed partway through.Evidence
The flag is declared as:
core/scanner.py:475documents the intent for a neighbouring phase — "--limitgoverns the analyze stage, not how many units the LLM reachability pass reviews — it must see the full codebase to find missed entry points" — and enhance behaves the same way. The help text says only "Max units to analyze," which reads as bounding the run.Observed:
--limit 2produced[Enhance] Units to enhance: 1351.Why it matters
Combined with the documented absence of a spend ceiling, the most natural way to try OpenAnt cheaply against a real codebase is also the one that incurs near-full cost. There is no warning, and the phase that runs unbounded is a high-volume per-unit LLM phase.
Suggested fix
Either have
--limitbound enhance as well, or add a separate--enhance-limit, and amend the help text to state explicitly that enhance always processes the full dataset. Enhance checkpointing already works correctly, so a bounded enhance costs nothing in resumability.Observed on one full-pipeline run at production scale: a private TypeScript monorepo (Angular front end, NestJS back end), 1,351 analysis units from 866 source files, run with
--verifyand no--limit, ~6.5 hours wall clock. The engine was invoked directly (python -m openant scan …, Python 3.13); the Go CLI was not built, so none of this involves the Go↔Python envelope. Stage 1 detection and Stage 2 verification ran on Claude Opus 5; application context, enhancement and reporting on Claude Sonnet 5 — all seven phases routed through OpenRouter (anthropic/claude-opus-5,anthropic/claude-sonnet-5) via a locally-addedopenrouterprovider adapter, on a build based on upstream2ed78f6. Cost figures are expressed as proportions of the run total; absolute amounts are omitted deliberately, as is the identity of the scanned repository.