Cafleet Model List Refresh

CAFleet Model List Refresh

Refresh the repository's model list — the per-backend model tables of skills/cafleet/reference/model-list.md — from the approved official sources, under explicit maintainer review. The repository file is the release source: a running Director reads only its deployed skill replica, so a refresh reaches Directors exclusively through the release/deployment transaction below.

Approved sources (exhaustive allowlist)

Fetch only these five official pages. Search results, third-party price sites, and social posts are never model-list authority. The first four are pricing and availability authority; the fifth is capability authority for the claude backend's context windows and Claude Code model-string syntax, and is never used to price a model.

Refresh procedure

1. Fetch all five approved sources. If any source cannot be fetched or parsed, stop: leave the model list unchanged, report the error, and let it become stale rather than fabricating values.

2. Extract pricing facts only: the currently effective standard input and output prices (USD per MTok) and availability/deprecation qualifiers. Do not copy long provider text into the repository, and do not present the pricing pages as benchmark evidence — they are pricing/availability sources only.

3. Refresh the claude and codex tables from their provider pricing pages, and refresh the opencode table with the dedicated procedure below — the Zen catalog behaves differently from the provider pages, so follow that procedure exactly. For codex, pricing and availability come from different pages: price each model from the OpenAI pricing page, but take Codex availability from the Codex models page (https://learn.chatgpt.com/docs/models.md) — the table carries only the models Codex currently offers in its model picker. A model absent there or demoted to a legacy model (reachable only via codex -m / config.toml) is removed from the table even when the pricing page still prices it; legacy models remain manual pass-through --model values.

4. Reapply the capability classes explicitly. The Class descriptions and the most-to-least-capable row ordering are reviewed maintainer judgment, not provider benchmark claims; changing a class or the ordering requires reviewed policy approval in the same pull request. Every row carries both prices from its approved source. Re-derive the Monitor and reviewer defaults table from the refreshed backend tables, and mirror its values into each backend section's {monitor_model} and {reviewer_model} rows (skills/cafleet/reference/coding-agent-overlays.md).

5. Re-verify every claude row's context window on every refresh, even when the prices are unchanged. Take each window from the Claude Code model configuration page's Extended context section and record it in the claude table's Context column. That page is also the authority for whether a row's --model value carries the [1m] suffix, which is a Claude Code model-string suffix stripped before the request reaches the provider — it is meaningless for the codex and opencode tables, whose values never carry it. Apply the suffix only where it changes the window a member actually gets:

   • A model the page lists as always running at 1M on the Anthropic API carries no suffix — the window is already 1M and the suffix is inert.
   • A model the page gives no 1M variant carries no suffix.
   • A model without 1M support carries no suffix; appending it there is a malformed value.
   • Where the page shows the window is opt-in for the reader's plan, the suffix belongs in the page's note rather than the row, so the Director applies it per spawn against the operator's plan instead of every spawn inheriting a credit-billed window.

   A suffixed value must be single-quoted everywhere it appears in a command, because [1m] is a glob pattern that fails an unquoted zsh invocation.

6. Propose, then apply atomically. Generate a concise proposed diff and require explicit maintainer approval before rewriting the tables in skills/cafleet/reference/model-list.md. Preserve the prescribed preamble verbatim; update the preamble's last refreshed date only after approval. That date is part of the page contract — the Director's staleness check reads it, so every refresh keeps it present and current. A failed refresh makes no edit.

OpenCode Zen procedure (the opencode table)

The opencode backend is priced by the OpenCode Zen page, and its catalog behaves differently from the two provider pages: model IDs differ from display names, the page lists many models while cafleet curates a few, and free stealth/preview models appear and disappear. On every refresh, even when the claude/codex prices are unchanged:

1. Price from the Zen page only. Zen's own USD-per-MTok rates are the billing rates for the opencode backend. Never price a Zen model from the upstream vendor's page (Anthropic, OpenAI, DeepSeek, …) — the Zen rate is the one cafleet members are billed at.
2. Copy the Zen model ID exactly; never slugify the display name. The page shows a display name ("GLM 5.2", "Kimi K2.7 Code") and a distinct model ID (glm-5.2, kimi-k2.7-code). The row's --model value is opencode/<zen-model-id> — the literal opencode/ prefix followed by the ID copied verbatim from the page. A hand-derived slug that does not match the Zen ID fails at spawn time.
3. Re-verify every curated row against the page. For each existing row, confirm the model is still listed on Zen and its price is current. A delisted model's row is removed in the same refresh.
4. Handle free models by their published price. A model Zen offers free for a limited time is priced 0.00 in both columns. When the free offer ends: update the row to the newly published Zen price, or remove the row if the model disappears or has no published price.
5. Keep the table a curated subset. Do not mirror the full Zen catalog — keep a handful of reviewed models spanning the price range. Adding or removing a curated model is a reviewed policy change, approved in the same pull request like any class or ordering change.

Cadence and staleness

Refresh the model list at least every 30 days and whenever the user asks for a refresh. A stale list disables cost efficiency mode — the Director relays an operator choice for those spawns — until a maintainer refreshes the model list, commits the repository source, and completes the release/deployment transaction.

Release-coupled deployment

There is no model-list-only sync path: a committed source file alone does not refresh a running Director's asset copy. To deploy a refreshed model list, the maintainer:

1. Bumps the CAFleet release version.
2. Builds the wheel and cafleet-assets-v<version>.zip containing the repository skills/ tree — the model list rides inside skills/cafleet/reference/ like every other reference page.
3. Publishes the release. Each active backend upgrades to that CLI version and runs cafleet setup, which overwrites its installed cafleet skill replica, model list included.

Ownership boundary

This skill owns model-list maintenance; the CAFleet Director owns per-spawn selection by reading the list; cafleet member create remains the execution boundary.