Uh oh!
There was an error while loading. Please reload this page.
docs(m163): --provider and --model are checked against the model catalogue - #176
Merged
Merged
Conversation
A new how-to, 'Use your own model provider', covers the whole flow: store a credential with the typed secret form, select it as the tenant provider, point at an openai-compatible endpoint, and return to the platform default. The CLI reference gains the typed-form row, and the changelog records the closure: 116 accepted ids, exit 2 with the full list on an unknown name, the CLI-engine refusals, the required --provider pairing, case folding, and the credential store reading the caller's environment. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…n the client The previous entry described a design that did not ship. Closing --provider to a 116-id copy of the runtime's dial tables made the terminal disagree with the dashboard, which builds its provider dropdown from GET /v1/models. The client now reads that same endpoint, so a provider becomes available by being added to the catalogue rather than by upgrading the client. - changelog.mdx: the Aug 15 entry rewritten — catalogue-checked rather than 116 ids, the new `agentsfleet models` command, and the empty/unreachable degradations. No exit-code-2 claim: the check is no longer at parse time. - cli/agentsfleet.mdx: a Model catalogue section for `models`, and the secret create row points at it. - fleets/model-providers.mdx: the how-to opens with `agentsfleet models` so both identifiers are looked up rather than guessed; the troubleshooting entries name the custom-endpoint route, except for the CLI engines, which have no endpoint to point at. - changelog.mdx: entries before Jul 1, 2026 removed (8 of 59). They described pre-release behaviour later entries supersede. An Archive entry points at the git history, which keeps the full record. make lint passes; no broken links. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
--model was validated nowhere, so a typo stored a credential that reported success and failed at the first event. It is now checked against the same catalogue read that resolves --provider, scoped to that provider's models. - changelog.mdx: the check, and why model ids are matched exactly where provider ids fold case — model ids belong to the provider and several are case-sensitive. - cli/agentsfleet.mdx: the secret create row covers both flags, and notes that openai-compatible accepts any model. - fleets/model-providers.mdx: the matching step and a troubleshooting entry. - fleets/model-providers.mdx: the worked example used `--model gpt-5`, which is not a catalogue id — the new check would refuse the page's own example. Now gpt-5.6-sol, in both the command and the `tenant provider show` output. make lint passes; no broken links. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Tightened the eight longest lead paragraphs and the eight heaviest bullets: 982 -> 944 lines, 13,514 -> 13,295 words (-1.6%). Every <Update> label and tag set is byte-identical, all 52 entries remain, and no Breaking-changes bullet moved sections. The pass was scoped against a ~35-40% reduction target. That target was wrong, and the measurement is why: bullets 77% of the file 320 bullets, median 28 words lead paragraphs 14% 3+ sentence bullets 11% of bullets The eight heaviest bullets shed 72 words between them — about nine each. What remains in them is not padding: the 146-word lease-pruning bullet carries seven distinct operator facts (the window measured from settlement not start, the two activity records, running leases untouched, the Activity feed kept at any age, expired-lease handling, lifetime counts unaffected, and the export warning with its endpoint). Reaching 35% would mean deleting the "what was broken before" clause from most bullets. That clause is what tells a reader whether they were affected — for example, a metered stage now charged the catalogue rate reads very differently without "instead of resolving to zero". Cutting it is a decision about what the changelog is for, not a copy-edit, so it is left for Indy. make lint passes; no broken links. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
One conflict, in changelog.mdx: both sides added an entry labelled Aug 15, 2026. They are different releases, not two versions of one — ours is M163's `--provider` catalogue check, main's is M164's "Sign-in events record the caller's address, not the proxy's". Resolved by keeping both, ours first. No entry was dropped, no label or tag set was altered, and the pre-Jul-1 cut plus its Archive entry are unchanged: 53 entries, oldest Jul 02, 2026. M165's changelog work did not touch this file, so it merged clean. make lint passes; no broken links. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Resolving the Aug 15 conflict sliced the "ours" side at the ======= marker, which sat immediately after the last bullet — so the entry's closing tag went with the marker. Mintlify caught it: "Expected a closing tag for <Update> (10:1)". 53 open tags, 53 closing tags now. I pushed the broken state first: the verify chain read `echo`'s exit code rather than `make lint`'s, so a red lint printed exit=2 and the push ran anyway. The lint was never green on 173e50b. make lint exit 0; no broken links. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Uh oh!
There was an error while loading. Please reload this page.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Documents the behaviour that shipped in agentsfleet#604 (merged), and brevity-passes the changelog.
What changed on the product
--providerused to take any string; a credential naming a provider the runner could not dial was stored, reported stored, and surfaced later as a fleet that could not reach any model. Both--providerand--modelare now checked againstGET /v1/models— the same catalogue the dashboard's model picker reads — so the terminal and the dashboard offer the same set, and a provider becomes available by being added to the catalogue rather than by upgrading the client.Pages
changelog.mdx2at parse time). Now: catalogue-checked, the newagentsfleet modelscommand, and the empty/unreachable degradations.cli/agentsfleet.mdxmodels/models --provider; thesecret createrow covers both flags and notes thatopenai-compatibleaccepts any model.fleets/model-providers.mdxagentsfleet modelsso both identifiers are looked up rather than guessed; troubleshooting entries for an unpriced provider and an unserved model.One correction worth calling out: the how-to's worked example used
--model gpt-5, which is not a catalogue id. The new check would have refused the page's own example. Nowgpt-5.6-sol, in both the command and thetenant provider showoutput.Changelog brevity pass
Two parts, in separate commits.
Cut to Jul 1, 2026 — 8 of 59 entries removed. They covered the pre-release period and described behaviour later entries supersede. An
Archiveentry points at the git history, which keeps the full record.Brevity pass on what remains — the eight longest lead paragraphs and eight heaviest bullets. 982 → 944 lines, 13,514 → 13,295 words (-1.6%). All 52 entries remain, every label and tag set is byte-identical, and no Breaking-changes bullet moved sections.
That is well short of the ~35-40% this was scoped against, and the measurement is why:
The eight heaviest bullets shed 72 words between them — about nine each. What remains is not padding. The 146-word lease-pruning bullet carries seven distinct operator facts: the window measured from settlement rather than start, the two activity records, running leases untouched, the Activity feed kept at any age, expired-lease handling, lifetime counts unaffected, and the export warning with its endpoint.
Reaching 35% would mean deleting the "what was broken before" clause from most bullets — the clause that tells a reader whether they were affected. "A metered platform stage is charged the catalogue rate" reads very differently without "instead of resolving to zero". That is a decision about what the changelog is for, not a copy-edit, so it is left open rather than taken unilaterally.
Verification
make lintpasses on every commit; no broken links; 52<Update>entries before and after.Greptile Summary
The PR documents catalogue-backed provider and model validation, adds a model-provider setup guide to navigation, and condenses older changelog content.
agentsfleet modelsand provider-credential behavior to the CLI reference.Confidence Score: 5/5
The PR appears safe to merge because no eligible or outstanding blocking failure remains.
No blocking failure remains.
Important Files Changed
Flowchart
%%{init: {'theme': 'neutral'}}%% flowchart LR A[Run agentsfleet models] --> B[Choose provider and model] B --> C[Create workspace provider secret] C --> D[Select tenant provider] D --> E[Run fleets with self-managed provider] E --> F[Delete tenant provider to restore platform default]Reviews (3): Last reviewed commit: "fix(docs): restore the </Update> the mer..." | Re-trigger Greptile