cafleet-model-list-refresh
Skill56 starsChanged 4 days agoMITView on GitHub
Refresh the CAFleet model list at assets/skills/cafleet/reference/coding-agents.md from the approved official pricing and capability sources. Use when a maintainer asks to refresh, update, or re-verify the model list, its token prices, its context windows, or its freshness, or when a Director reports the list was last refreshed more than 30 days ago. Maintainer-invoked only — never run automatically during a member spawn.
In this repository4 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- clean-docsSkill
- skill-authorSkill
- update-readmeSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
14 rules. About 0 are common, 7 are distinctive, and 7 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
7
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 42
Fetch every approved source: If any source cannot be fetched or parsed, stop: leave the model list unchanged, report the error, and let it …
- Very specific: line 53
Follow each dedicated procedure for its catalog.
- Distinctive: line 64
Ask which model each role should use and, for Claude and Codex, which supported effort level should accompany it.
- Very specific: line 118
Copy the Console model ID exactly: Use the *Endpoints* table under ## Endpoints, whose header is Model | Model ID | Endpoint | AI SDK …
- Distinctive: line 125
Mirror the Console catalog: Add every missing listed model ID, remove every delisted model ID or model past its published deprecation date, …
- Distinctive: line 133
Apply the row rules and maintain their notes: Keep one row per model ID, using the base (shortest-prompt) tier for tiered prices and the …
- Very specific: line 141
Order and classify every row: Keep a single table ordered most → least capable across the whole catalog, with a short Class on every row …
- Distinctive: line 144
Place unverified stealth/preview models at the bottom and state that their capability is unverified in their Class.
- Very specific: line 151
Run this procedure on every refresh.
- Distinctive: line 156
Price from the Go page only: Use the Go page's per-token amounts from its price table.
- Distinctive: line 163
Mirror the Go catalog: Add every missing listed ID, remove every delisted ID, and re-verify every remaining row's price.
- Very specific: line 167
Apply the row rules and maintain the notes: Use the Console table's columns: Model | Class | Input $/MTok | Output $/MTok, with one row per …
- Very specific: line 174
Use the currently billed base (shortest-prompt) rate for each row.
- Very specific: line 188
Order and classify every row: Keep one table ordered most → least capable, with a Class on every row.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CAFleet Model List Refresh
Refresh the repository's model catalogs and canonical Role defaults in
assets/skills/cafleet/reference/coding-agents.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.
This skill maintains each backend's Model catalog, its provenance/context notes, its canonical Role defaults table, and the shared freshness metadata. A backend whose Model catalog states "No maintained catalog" is outside the refresh scope: leave its Model catalog and its — Role defaults unchanged. Preserve Runtime bindings, bound runtime notes and Role eligibility as runtime documentation, and preserve the Director's worked model selections under its model-selection policy. The Coding-agent overlay rule owns maintenance responsibilities and backend scaffolding in its Template. Keep the catalog preamble's last-refreshed date for model data, standard provider USD per MTok as planning estimates, and reviewed capability ordering.
Approved sources (exhaustive allowlist)
Fetch only the official pages listed in this table. Every page except
Claude Code model configuration is pricing and availability authority.
Claude Code model configuration is capability authority for the claude
backend's context windows and [1m] applicability; use the pricing sources
to price models, never the capability page. Search results, third-party price
sites, and social posts are never model-list authority.
| Source | What it feeds | URL |
| Anthropic pricing | claude prices | https://platform.claude.com/docs/en/about-claude/pricing.md |
| OpenAI pricing | codex prices | https://developers.openai.com/api/docs/pricing.md |
| Codex model availability | codex availability | https://learn.chatgpt.com/docs/models.md |
| OpenCode Console models and pricing | opencode prices and availability | https://opencode.ai/v2/docs/console/models/ |
| OpenCode Go models and usage-limit pricing | opencode-go prices and availability | https://opencode.ai/v2/docs/console/go |
| Claude Code model configuration | claude context windows and [1m] applicability | https://code.claude.com/docs/en/model-config.md |
Refresh procedure
- Fetch every approved source. 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.
- 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.
- Refresh the
claudeandcodextables from their provider pricing pages, refresh theopencodetable with the OpenCode Console procedure, and refresh theopencode-gotable with the OpenCode Go procedure below. Follow each dedicated procedure for its catalog. Forcodex, 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 viacodex -m/config.toml) is removed from the table even when the pricing page still prices it; legacy models remain manual pass-through--modelvalues. - Ask the maintainer for role defaults. Show the current and available catalog tokens for reviewer and other members on each backend. Ask which model each role should use and, for Claude and Codex, which supported effort level should accompany it. Record the maintainer's choices in the proposed diff. OpenCode omits effort because its backend does not support it. A refresh keeps the existing role defaults only when the maintainer explicitly chooses to keep them.
- Reapply the capability classes explicitly. The
Classdescriptions 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. Set the six Role defaults from the maintainer's choices, using local catalog entries or aliases and effort levels accepted by the selected backend. Re-verify every
clauderow'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 theclaudetable'sContextcolumn. That page is also the authority for whether a row's--modelvalue carries the[1m]suffix, which is a Claude Code model-string suffix stripped before the request reaches the provider — it is meaningless for thecodexandopencodetables, 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.- Propose, then apply atomically. Generate a concise proposed diff and
require explicit maintainer approval before applying the model-data changes in
assets/skills/cafleet/reference/coding-agents.md. Preserve the common catalog preamble contract described above. Apply catalogs, provenance/context notes, Role defaults, and the last refreshed date together after approval and successful validation. The Director's staleness check reads that date, so every successful refresh keeps it present and current. A failed refresh makes no edit.
OpenCode Console procedure (the opencode table)
The opencode catalog mirrors every model listed by the OpenCode Console,
excluding models past their published deprecation date, and uses the Console's
prices. Model IDs differ from display names, and free/preview models appear
and disappear. On every refresh, even when the claude/codex prices are
unchanged:
- Price from the Console page only. Console's own USD-per-MTok rates are
the billing rates for the
opencodebackend. Never price a Console model from the upstream vendor's page (Anthropic, OpenAI, DeepSeek, …) — the Console rate is the one cafleet members are billed at. - Copy the Console model ID exactly. Use the Endpoints table under
## Endpoints, whose header isModel | Model ID | Endpoint | AI SDK Package, to map each display name ("GLM 5.2", "Kimi K2.7 Code") to itsModel ID(glm-5.2,kimi-k2.7-code). The Pricing table keys prices by the same display name. Set the row's--modelvalue toopencode/<console-model-id>: the literalopencode/prefix followed by the ID copied verbatim from the Endpoints table. - Mirror the Console catalog. Add every missing listed model ID, remove every delisted model ID or model past its published deprecation date, and re-verify every remaining row's price. Bind each Deprecated models entry to its model ID through the Endpoints display-name mapping; a passed date excludes the row even while it appears in active pricing. A deprecated display name with no listed ID is already delisted and needs no action. A deprecated display name mapping to more than one ID is a parse failure: stop the refresh under Refresh procedure step 1.
- Apply the row rules and maintain their notes. Keep one row per model
ID, using the base (shortest-prompt) tier for tiered prices and the currently
billed Console rate for promotional discounts. Record each published free
rate as
0.00in its own column, including models with only one free rate. A model free for a limited time has0.00in both columns; update its rates when the offer ends or remove it if delisted. Keep a model with a future deprecation date and record that date beneath the table. - Order and classify every row. Keep a single table ordered most → least
capable across the whole catalog, with a short
Classon every row and the same relative order for models shared with theclaude/codexcatalogs. Place unverified stealth/preview models at the bottom and state that their capability is unverified in their Class. A new row's Class and position, and changes to existing classes or positions, are reviewed policy changes approved in the same pull request.
OpenCode Go procedure (the opencode-go table)
Run this procedure on every refresh. Maintain a separate #### OpenCode Go
section after the Console table and its notes, within the opencode backend's
### Model catalog. Point to it from the Console preamble; the shared
last-refreshed date covers both tables.
- Price from the Go page only. Use the Go page's per-token amounts from its price table. These amounts count against subscription usage limits; never price a Go row from the Console page or a vendor page.
- Copy the Go model ID exactly. Map each display name to its
Model IDthrough the Go page's Endpoints table, whose header isModel | Model ID | Endpoint | AI SDK Package. Writeopencode-go/<model-id>, with the ID copied verbatim. - Mirror the Go catalog. Add every missing listed ID, remove every delisted ID, and re-verify every remaining row's price. Apply published deprecation dates using the Console procedure's display-name binding rule: keep a future-dated model with a note and remove one whose date has passed.
Apply the row rules and maintain the notes. Use the Console table's columns:
Model | Class | Input $/MTok | Output $/MTok, with one row per Go Endpoints ID. A price-table display name with no Endpoints ID or more than one ID, or an Endpoints ID with no price-table row or without both Input and Output prices, is a parse failure: stop under Refresh procedure step 1.Use the currently billed base (shortest-prompt) rate for each row. Record a free-for-a-limited-time model with
0.00in both columns.The preamble states that this table lists every model on the Go page and that
opencode-go/is part of the--modelvalue. Explain that these tokens require configured providers that include an OpenCode Go subscription and that the Director checks availability per spawn under Director model selection.Explain beneath the table that prices are USD per MTok counted against subscription usage limits, and that Extra Usage bills usage beyond the limits to the pay-as-you-go balance. End with
Source: [OpenCode Go](https://opencode.ai/v2/docs/console/go).- Order and classify every row. Keep one table ordered most → least
capable, with a Class on every row. Reuse the Console Class for each shared
ID, and preserve the relative order of IDs shared with the Console,
claudeorcodextables. Place unverified preview models at the bottom and state that their capability is unverified in their Class. A new row's Class and position, and changes to existing ones, are reviewed policy changes approved in the same pull request.
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:
- Bumps the CAFleet release version.
- Builds the release binary with the repository
assets/skills/tree embedded, includingassets/skills/cafleet/reference/coding-agents.md, and publishes the per-target release archives. - Upgrades the installed CLI to that release and runs
cafleet setup, which installs its embedded skill payload for the selected backends, including the unified reference. The binary upgrade and setup complete the deployment.
Installing skills directly from a checkout is a separate contributor operation; it does not replace the embedded-binary build/release → upgrade → setup path.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

