pointlanguage docs0.1.1

Explanation

Agent repair walkthrough

Step-by-step: what a coding agent loads for Point check-json vs TypeScript paste — with honest CI vs estimated labels and token counts per step.

Step by step — same bug, two agent workflows

CI verified — no LLM

40/40 fixtures pass

40 fixtures · 9 feature-build · 4 repair-plan loop · check-json ~170–260 tok vs TS paste · 78–92% less context.

bun run proof:agent-repair

Full CI suite: bun test tests/agent-repair-sufficiency.test.ts tests/agent-repair-multistep.test.ts && bun run benchmark:agent-repair

View test source on GitHub →

Live model benchmark

Point 100% · TS 100%

May 22, 2026, 3:16 AM · gpt-4.1, claude-opus-4-6, claude-sonnet-4-6 on 13 fixtures (36 single-shot, 4 CI-only loop). Success = point check passes.

Latest proof run: 78/78 API runs pass point check.

  • Claude Opus 4.6Point 13/13 · TS 13/13
  • Claude Sonnet 4.6Point 13/13 · TS 13/13
  • GPT-4.1Point 13/13 · TS 13/13
Full benchmark table →

Pick a CI fixture below. Each column shows what the agent sees at every step, how many tokens that step costs, and whether we prove it in CI or illustrate it from typical chat pastes.