[ne0x86]

What and why

A usability round from the owner's test with dozens of projects and containers.

Models and effort

• Model suggestions. Each Model field has a list to pick from, and any other name can still be typed, because catalogs change every month.

  • Claude Code and Codex. The list comes from data/tiers.json: the aliases haiku, sonnet and opus for Claude Code, and the models of the Codex CLI that the runner pins. fable is left out: claude -p can bill it to usage credits without asking, and lastcall never spends paid usage.
  • API keys. The list is the provider's models from models.dev's public catalog (lastcall/catalog.py): those that call tools, have a price and are not deprecated.
  • The catalog request. The Accounts page asks for it once a day, waiting 5 s at most. It is bounded at 32 MB packed or unpacked, and at 1,000 models per provider. When the request fails, the page shows no suggestions and tries again after 10 minutes. The README says that the page asks models.dev.

• Default efforts that fit the task, for both engines:

  • fast: medium. Only a repository's own reviewers ask for it.
  • balanced: high. The reviewers.
  • deep: xhigh. The judge, the implementers and the plans, where quality matters most and the quota would expire anyway.

  Codex used to say "Default (high)" for every tier.

Clearer texts on a project's settings

• "Writes in" / "The account's" becomes "Language of its pull requests and issues" / "Same as the account", with a hint.
• The four options say plainly what they add: "Tests too", "Dependencies too", "Merge after my approval" and "Watch after the merge".
• The watch's text says what it does not do. It looks only at the checks of the merged commit and at the health address. Container logs, Loki and the error tracker feed the audits, not the watch.

Lists that scale

• Settings › Container logs and Logs in Loki come in three parts:

  • Read now: a compact list, where each chosen row can stop being read.
  • Suggested: checkboxes, with "Read the N suggested" and "Skip the ticked" above the list and "Read the ticked" at the bottom of the screen.
  • Everything else: a fold with a filter by name, where a project is typed into one <input list> with type-ahead instead of a select of every project in every row.

  lastcall's own containers and sources are hidden. The server checks every typed project against the projects, in any case, and a form is applied whole or not at all. A choice for a removed project still survives every save.

• Projects.

  • Search and filters. The page has a search by name, and filters with counts: All, Active, Inactive, and Waiting for you (pull requests, decisions or ideas). They show from the seventh project on, or while one is set.
  • Shorter cards. With more than 6 projects, each card's "Run now" starts folded.

Docs

• README.md: Models and effort, the privacy note about models.dev, and the Container logs and Loki paragraphs.
• DESIGN.md: section 6 (tiers, suggestions, why the efforts differ) and section 13 (the lists and the Projects page).
• CONTRIBUTING.md: models and Provider.catalog.

How it was verified

• The checks. uv run ruff check && uv run ruff format --check && uv run mypy && uv run pytest passes: 832 tests pass and 33 are skipped. 114 tests are new. They cover:

  • the datalists, and models.dev with its cache, timeout, failure, bounds and gzip bomb;
  • the new defaults, and the texts in English and Spanish;
  • the three parts for containers and sources: ticked, all suggested, skip, stop, the typed project checked, and forged forms refused;
  • the filters, and the search and filters of the Projects page with their counts;
  • Run now folded.

  Seven existing assertions changed on purpose: they pinned the old efforts, or the old rows of the logs lists.

• Screenshots at 390 and 1280 px, in English and Spanish, light and dark, of:

  • the Accounts page for Claude Code, Codex and OpenRouter accounts;
  • a project's settings;
  • the logs and Loki lists, with 40 containers and 30 sources;
  • the Projects page, with 30 projects.

  Every page answers 200, and nothing is wider than the screen. The real lastcall web was checked with 30 projects too.

🤖 Generated with Claude Code