Skip to content

[M4-3a] Record and ship an agent proof session (asciinema replay) #99

Description

@williamdemeo

Feeds #25 (M4-3) directly: ADR-008 conditions the portfolio's whole ordering on the AI page shipping a checkable worked example. This issue produces that example and its player.

Record a real session of an agent (Claude) driving Agda through agda-native-air to a finished proof; ship it as a self-hosted asciinema replay on the M4-3 page, annotated beat by beat.

Why

  • A typechecker as dense, unfakeable reward signal, and an agent loop shaped around it — this is precisely the intersection the target roles sit at. Watching it run says more than any paragraph.
  • asciinema-player is two files (JS + CSS) plus the .cast recording, all servable from this origin (https://docs.asciinema.org/manual/server/embedding/) — make offline-audit stays green by construction.
  • The same recording seeds the strongest M6-4 blog-post candidate ("watch Claude prove a theorem"), and the .cast file is itself a checkable artifact, checked into the repo.

Tasks

  • Choose the lemma: small enough to replay in ~60–90 seconds, real enough to be honest work (a kernel-is-a-congruence-shaped result reads well).
  • Record the session (asciinema rec), trim dead air, keep the recovery-from-error beat if one occurs — that beat is the most credible part.
  • Vendor asciinema-player.min.js + CSS under docs/assets/, pinned by hash like the fonts; wire the page embed.
  • Annotate alongside: what the model saw, which tool it called, what the typechecker said back.
  • Check the .cast into the repo next to the proof it produced; link both from the page.
  • Confirm player weight is paid only on the M4-3 page.

Acceptance criteria

Metadata

Metadata

Assignees

No one assigned

    Labels

    careerDirectly serves the research-role application goal.contentMigrating, triaging, or restructuring existing content.infrastructureBuild system, tooling, and repository plumbing.milestone-4-portfolioMilestone 4: Portfolio of flagship work.

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions