Repository navigation
Description
Target: v5.
Target branch:
v5. Open the PR with--base v5. The code paths are the same onv6(line numbers can differ), and the regular v5 → v6 top-up carries it there, so do not open a second PR forv6. Citations arepath:lineonv5@7ca94dd, plus the function name, so the pointers still work after the lines drift.In the PR body write
Refs #<this issue>. Do not use a closing keyword: a PR based onv5cannot close an issue that way, so a maintainer closes this one after the merge.
Executive summary
Every time Anthropic releases a model, someone has to hand-edit five places before the hive's dashboard offers it and prices it correctly. Nobody is told to, so the lists fall behind. Fable 5.2, Fable 5.1 and Haiku 5.5 have all been missed.
The hive already has weekly jobs that notice part of this:
- one reports newer models as an issue;
- one updates prices, but only for models already in the table.
This issue turns that into a weekly pull request that adds new Claude models to the lists, their prices and their Claude Code version requirement. A reviewer then only has to check and merge it.
The closed allowlist stays exactly as it is: a model is added only once the pinned Claude Code version accepts it.
Problem
Adding a Claude model takes edits in five places, all by hand:
claudePinnedCLIModels, the closed allowlist (🐛 bug: agent model dropdown offers models the installed CLI can't run (all backends, not just claude) #8418);claudeStaticModels;CLAUDE_CLI_MODELS, plus a rebuild ofstatic/index.html;src/scripts/cli-model-requirements.txt(✨ [v5] cli-pin-bump: check Hive's model lists and prices against the new CLI pin (and bump patch-versioned CLIs like Claude Code) #9805);- a
modelPricesrow.
What exists today covers only parts of it:
static-model-freshness.yml(🌱 Weekly static model list freshness check and list-shape test #18474, Mondays 07:17 UTC) runs.github/scripts/check-static-model-freshness.py. It reports "newer models of the same family on OpenRouter that the list lacks" in one tracking issue, and changes nothing. It hasissues: writeonly.price-table-refresh.yml(Mondays 06:41 UTC) runssrc/internal/tools/pricetable, which reprices existing rows from OpenRouter and opens a PR. It never adds a row for a model the table lacks.cli-pin-bump.ymlbumps the Claude Code pin and runscheck-cli-model-lists.sh. That check only compares againstcli-model-requirements.txt, which has one line (Sonnet 5.5), and it runs withHIVE_MODEL_CHECK_STRICT: '0', so drift is a warning.
So a new model is detected at best as an issue line, and nothing adds it.
Evidence
- Fable 5.2, Fable 5.1 and Haiku 5.5 are all served by Anthropic (https://platform.claude.com/docs/en/about-claude/models/overview) and missing from every list. 🐛 [v5] Claude model dropdown and price table: add Fable 5.2, Fable 5.1 and Haiku 5.5 #18753 adds them by hand.
- The lists' comments still say "Refreshed 2026-08-03 from a live /v1/models response". The only addition since is Sonnet 5.5 (✨ [v5] model lists are out of date across CLI backends (Sonnet 5.5 unselectable, stale fallbacks, missing prices) #9804).
cli-model-requirements.txthas one line in total.
Root cause
.github/scripts/check-static-model-freshness.py@7ca94dd:main(:112) buildsnewer(:147-167) and only writes report text..github/workflows/static-model-freshness.yml:34-36@7ca94dd:contents: read,issues: write.src/internal/tools/pricetable/reconcile.go@7ca94dd: reconciles rows that exist, recordingchangeandunverified. There is no path that adds a row..github/workflows/cli-pin-bump.yml:105@7ca94dd:HIVE_MODEL_CHECK_STRICT: '0'.- The lists:
src/pkg/dashboard/cli_models.go:450(claudePinnedCLIModels),:272(claudeStaticModels),:392(claudeAlwaysIncludeModels), andsrc/pkg/dashboard/static-src/js/backends-tokens.js:53(CLAUDE_CLI_MODELS), all@7ca94dd.
Wanted behavior
A weekly job that opens, or updates, one PR adding the Claude models the lists lack:
- Find candidates. Reuse the freshness script's comparison. Candidates are Claude ids newer than the lists' entries in the same line (Fable, Opus, Sonnet, Haiku), from OpenRouter's public catalog. Use Anthropic's
/v1/modelsonly if a key is configured for CI, and never require one. Map each to the Claude CLI's dashed spelling (anthropic/claude-fable-5.2→claude-fable-5-2). - Gate on the pinned Claude Code version. Add a candidate only when the Claude Code version in
src/Dockerfileaccepts it, and record that version incli-model-requirements.txt. Work out how to decide "accepts" without credentials, and document it in the PR:- First choice: in the pin-smoke image
cli-pin-bump.ymlalready builds, check whether an unauthenticatedclaude --model <id> -p xanswers with the CLI's "version X or newer is required" refusal, as distinct from an authentication error. If it reliably does, that is the gate. - Fallback: if the CLI does not tell these apart offline, open the PR as a draft, listing each candidate under "needs a check against the Claude Code changelog". Never add an id as accepted on a guess.
- First choice: in the pin-smoke image
- Make all five edits for each accepted candidate:
- the three lists, with a display label ("Fable 5.2");
- a rebuilt
static/index.html(make dashboard-gen); - the requirements line;
- a
modelPricesrow from the OpenRouter quote, through thepricetabletool's existing plausibility checks. A quote that fails them is left out and flagged, not guessed.
- One PR, kept current. Reuse the price-table PR pattern (
price-table-refresh.ymlalready holdscontents: writeandpull-requests: write). Use a fixed branch and update the open PR in place on later runs. With no candidates, do nothing. - Make the requirements check strict once the list is current:
HIVE_MODEL_CHECK_STRICT: '1'incli-pin-bump.yml. A pin bump then fails while a model it newly accepts is still missing from the allowlist.
Acceptance criteria
- With a fixture catalog containing a newer Fable and Haiku than the lists hold, a dry run prints the five edits per model and opens no PR.
- A candidate the pinned CLI is not shown to accept is never added as accepted; it appears under "needs a check".
- A real run opens one PR on a fixed branch. A second run with the same candidates updates that PR and does not open another.
- The generated PR passes
make dashboard-gen-check,TestStaticModelListsShapeandcheck-cli-model-lists.shwithHIVE_MODEL_CHECK_STRICT=1. - Price rows come only through the
pricetableplausibility checks. - Mythos (limited-access) ids are never proposed. Retired ids are never added.
- With no candidates, no PR and no branch.
Test plan (CI decides)
- Python unit tests for the freshness script's new candidate and edit logic, against fixture catalogs: same-line newer detection, the dotted→dashed mapping, Mythos excluded, idempotent edits (running twice changes nothing more). Put them next to the script, in the style of
.github/scripts/test-ci-infra.py. src/internal/tools/pricetable/main_test.go: adding a missing row from a quote; refusing an implausible one.src/scripts/test-check-cli-model-lists.sh: strict mode fails on a requirements line whose model is missing from the allowlist, and passes when present.- A
workflow_dispatchdry run on a fork, linked from the PR.
Suggested implementation
- Split the freshness script's comparison into a function that returns structured candidates. The report and the new
--apply claudemode share it. --applyedits the Go and JS lists as text, keeping order and comments, and runsmake dashboard-gen.- Extend
pricetablewith an "add missing ids" mode that takes the accepted ids. - Add the PR step, either to
static-model-freshness.ymlwith the price workflow's permissions, or as a job inprice-table-refresh.ymlso lists and prices land in one PR. Say which in the PR and why. - Flip
HIVE_MODEL_CHECK_STRICTonly after the sibling issue's refresh has merged, so strict mode starts green.
Rejected options
- Drop the closed allowlist and trust the live
/v1/modelslist. The list is closed because the installed CLI refuses models newer than its version (🐛 bug: pinned agent CLIs are stale — Claude Code 2.1.226 rejects claude-opus-5-5, Codex 0.153.4 lacks GPT-6 Sol/Luna #8417, 🐛 bug: agent model dropdown offers models the installed CLI can't run (all backends, not just claude) #8418). This issue keeps that guarantee and automates the upkeep instead. - Auto-merge the PR. Model and price changes reach every hive's dashboard and cost figures. A reviewer checks them.
Out of scope / do not change
- Copilot, Codex, Gemini and other backends' lists. Same pattern, but do Claude first. The freshness report keeps covering the rest.
- The Claude Code pin bump itself (
cli-pin-bump.yml), other than strict mode.
Note for the implementer
This touches .github/workflows/, so the PR needs a lane that may push workflow files.
Dependencies
- Builds on 🐛 [v5] Claude model dropdown and price table: add Fable 5.2, Fable 5.1 and Haiku 5.5 #18753, which adds Fable 5.2, Fable 5.1 and Haiku 5.5 by hand.
- Turn strict mode on only after 🐛 [v5] Claude model dropdown and price table: add Fable 5.2, Fable 5.1 and Haiku 5.5 #18753's PR has merged.
Activity
🌱 Re-pin TRUSTED_SCRIPTS_TREE for the src/scripts test change
hivecommons-hive commentedon Nov 17, 2026
Hive status — updated in place; edits do not notify. Full history is in the Activity Monitor.
PR #18788 covers this issue: #18788
The PR body says this is partial/non-closing work (
Refs/Part ofor leave-open wording), so Hive will not add a closing keyword. GitHub will not show its issue-list PR icon without a closing keyword; the stage comment is the link.✨ Weekly PR adds new Claude models to the lists and price table once …
clubanderson commentedon Nov 18, 2026
Closing by hand: delivered by #18788 (merged to v5). v5 merges with Refs do not auto-close, so a maintainer closes it after the merge.
github-actions commentedon Nov 18, 2026
@Danathar this issue was closed by @clubanderson 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.