Repository navigation
Description
Summary
#18788 added a weekly job that keeps the Claude model allowlist in step with the Claude Code version pinned in the image. Every other coding CLI still relies on hand-edited lists. This issue extends the same treatment to every backend, so all providers work the same way:
- CI keeps the upper limit: the models the CLI version pinned in
src/Dockerfilecan actually launch. - Each hive shows what its own login can use, filtered by that limit. CI never pushes a model list to running hives.
Do it in parts. Each part below is self-contained and can be its own PR. Write it for v5; it will be forward-ported to v6.
All links are pinned to v5 at c30d0a91dc2e66aef1eb5cf09b22c2124290e5c7. Re-check them against a fresh pull before editing.
Background: how the Claude version works today
discoverClaudeModels(cli_models.go:1979) calls Anthropic's/v1/modelswith the hive's own credential, then keeps only ids inclaudePinnedCLIModels(cli_models.go:461).pinnedCLIModelAllowlist(cli_models.go:1210) has allowlists for claude, copilot and gemini only..github/workflows/claude-model-freshness.ymlplus.github/scripts/claude-model-freshness.shandcheck-static-model-freshness.py --apply claudefind new candidates, install the pinned CLI, ask it offline whether it recognises each id (with control checks), and open one PR on a bot branch. That PR always carriesneeds-humanand is never auto-merged. Read ✨ Weekly PR adds new Claude models to the lists and price table once the pinned Claude Code accepts them #18788's description before starting; it explains the gate and the control checks..github/workflows/static-model-freshness.ymlonly files a drift report issue for the other lists. It does not change them.
Current state per backend
| Backend | Pin in src/Dockerfile |
What the hive asks at runtime | Allowlist | CI keeps it fresh |
|---|---|---|---|---|
| claude | CLAUDE_CODE_VERSION=2.1.327 |
Anthropic /v1/models with the hive's own login |
claudePinnedCLIModels |
yes (#18788) |
| copilot | COPILOT_VERSION=1.0.88 |
Copilot SDK catalogue with the hive's login | copilotPinnedCLIModels |
no (report only) |
| gemini | none (no Gemini CLI installed) | Google models API with GEMINI_API_KEY |
geminiPinnedCLIModels |
no (report only) |
| codex | CODEX_VERSION=0.170.0 |
codex app-server model/list from the local binary |
none | no; list comment says 0.166.0/0.167.0 |
| agy | AGY_VERSION=1.5.0 |
agy models from the local binary |
none | no; static floor from agy 1.1.18 |
| omp | OMP_VERSION=18.22.0 |
omp models --json from the local binary |
none | no |
| goose | GOOSE_VERSION=1.60.1 |
goose ACP probe, per configured provider | none | no |
Part 1: copilot (same pattern as Claude)
Where: copilotPinnedCLIModels (cli_models.go:342), copilotStaticModels (:304), the matching JS list in static-src/js/backends-tokens.js, and copilotCLIAcceptedModels in src/crux/copilot_models.go (the comment at :331 says to keep them in sync).
Do:
- Find an offline gate first. Install
@github/copilot@<pin>the same way the Dockerfile does (npm --ignore-scripts) and find a signal that separates "id the CLI knows" from "id it doesn't" without a login. Note 🐛 copilot: picking claude-sonnet-5-5 or claude-opus-5-5 runs the agent on claude-sonnet-5 (Copilot CLI 1.0.88) #9927: the CLI silently swaps an unknown dashed id for another model (Model "..." from --model flag is not available. Using "claude-sonnet-5" instead.). That message may itself be the signal. Record the exact command and output in the PR. - If there is a reliable offline signal, add
--apply copilotnext to--apply claude, using the same three control checks (version equals the pin, a listed id is recognised, a made-up id is not). Copilot ids are dotted (claude-sonnet-5.5), not dashed. Keep that spelling. - If there is no offline signal, stop after step 1 and write the finding on this issue. Don't build a gate that guesses.
Done when: a dry run against the real pin prints the edits it would make, a re-run changes nothing, and every edited list stays in sync.
Part 2: codex (refresh the static list from the pinned binary)
Codex has no per-account catalogue: the model list is built into the CLI binary. So the fix is not an allowlist. It is keeping codexStaticModels (cli_models.go:429) equal to what the pinned binary reports. That list is what hosted hives without the binary always serve.
Do:
- Install
@openai/codex@<pin>in CI and run the sameapp-servermodel/listexchange thatdiscoverCodexModels(cli_models_codex.go:100) speaks. Check first whethermodel/listanswers without a ChatGPT login. If it doesn't, stop and write the finding here. - Rewrite
codexStaticModelsandCODEX_CLI_MODELS(JS) in the orderorderCodexServedModelsuses: visible ids first, then hidden ones. Update the "matches codex X" comment to the pin. - Open or update one PR on a bot branch with
needs-human, same as ✨ Weekly PR adds new Claude models to the lists and price table once the pinned Claude Code accepts them #18788.
Done when: HIVE_MODEL_CHECK_STRICT=1 src/scripts/check-cli-model-lists.sh no longer reports "codex static list comment says … but src/Dockerfile pins …". That finding is live today (0.166.0 vs 0.170.0).
Part 3: agy and omp (refresh the static floor)
Both are probed from the local binary at runtime (agy models; omp models --json), so the running CLI already sets the limit and no allowlist is needed. Their static floors are stale, though: agyStaticModels (cli_models_agy.go:28) is a snapshot from agy 1.1.18, and the pin is 1.5.0.
Do: for each one, check whether the subcommand runs without a login on the pinned binary. If it does, refresh the static list (Go and JS) from it in the same weekly PR. If it doesn't, list it as "needs a manual capture" in the PR body. Don't guess.
Part 4: gemini and goose (report only)
- gemini: no Gemini CLI is installed in
src/Dockerfileorsrc/Dockerfile.contributor, so there is no pinned binary to check against. LeavegeminiPinnedCLIModelshand-kept, and say so insrc/docs/cli-pins.md. - goose: its models depend on whichever provider an operator configures. There is no single list to keep fresh. Document that it is out of scope.
How to ship
- One weekly workflow, or extend
claude-model-freshness.yml. Either is fine, but keep one bot PR per run, always withneeds-humanand never auto-merged. Give it write permissions only on the job that opens the PR, as ✨ Weekly PR adds new Claude models to the lists and price table once the pinned Claude Code accepts them #18788 does. - Keep
static-model-freshness.yml's report working. It can stop reporting a list once that list is maintained automatically. - Update
src/docs/cli-pins.mdwith one short section per backend: how it is kept fresh, or why it isn't.
Tests
- Hermetic Python tests with a fake CLI per backend, like
test-claude-model-freshness.py: candidate detection, gate pass/fail, control-check failures add nothing, dry run writes nothing, and a re-apply is a no-op. - The Go tests that check list shape and Go/JS sync (
TestStaticModelListsShape) still pass. - The PR's self-test job dry-runs the whole pipeline against the real pins.
- Locally, only
gofmt -landnode --check(see AGENTS.md). CI decides the rest.
Out of scope
- The Claude backend. It already works this way (✨ Weekly PR adds new Claude models to the lists and price table once the pinned Claude Code accepts them #18788); use it only as the reference pattern.
- Changing runtime discovery or the per-hive credential logic. Hives already ask with their own login, and that stays.
- Bumping any CLI pin (that is
cli-pin-bump.yml). - Price-table changes for non-Claude models.
- Turning on
HIVE_MODEL_CHECK_STRICTincli-pin-bump.yml. Do that as a follow-up once Part 2 removes the codex finding.
Activity
hivecommons-hive commentedon Nov 18, 2026
🔓 Claim by agent
scannerreleased (covering PR merged or closed).Claim history
scanner(agent on hive hosted-available-oke-11-placeholder-r05x)scanner(agent; covering PR merged or closed)docs: say how gemini and goose model lists are kept (part of #19028) (#…
github-actions commentedon Nov 18, 2026
@Danathar this issue was closed by @hivecommons-hive[bot] as completed.
If this was closed by a merged PR, the fix only takes effect once it ships in a tagged release and your deployment updates to it — check your running version against the fix before concluding the fix failed.
If you are still seeing the problem, comment
/reopenand it will be reopened for you — you do not need write access. Please add what you are still observing (running version, logs, steps) so it is not closed the same way again.🌱 chore: top up v6 from v5 (2026-11-18) (#19100)