Uh oh!
There was an error while loading. Please reload this page.
docs(ai/models): document context_window + cost fields - #188
Merged
Conversation
Adds the two fields the open-agents frontend already reads from its
own `/api/models` endpoint, in preparation for switching its model
picker to call api's `GET /api/ai/models` directly:
- context_window (integer) — gates model selection in the picker
against per-request token budgets (lib/models.ts filters out
models without it)
- cost ({input, output}) — per-million-token pricing for the model
selector's price display (lib/model-options.ts)
Both fields are already supplied by the upstream `@ai-sdk/gateway`
metadata that api uses; the response handler currently strips them.
A follow-up api PR will plumb them through. Spec already had
`additionalProperties: true` on the entry schema, so today's clients
won't break — this just makes the contract explicit and SDK-friendly.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>Warning Rate limit exceeded
To keep reviews running without waiting, you can enable usage-based add-on for your organization. This allows additional reviews beyond the hourly cap. Account admins can enable it under billing. ⌛ How to resolve this issue?After the wait time has elapsed, a review can be triggered using the We recommend that you space out your commits to avoid hitting the rate limit. 🚦 How do rate limits work?CodeRabbit enforces hourly rate limits for each developer per organization. Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout. Please see our FAQ for further information. ✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Uh oh!
There was an error while loading. Please reload this page.
3 tasks
sweetmantech added a commit
to recoupable/api
that referenced
this pull request
May 5, 2026
* feat(ai/models): enrich response with context_window + cost from models.dev
api's GET /api/ai/models previously returned just the gateway entries.
Open-agents' frontend depends on two extra fields per model that come
from the public models.dev catalog:
- context_window (integer) — gates model selection in the picker
- cost ({input, output}) — per-million-token pricing for display
Adds three pure helpers (TDD'd individually) plus a small refactor of
the existing fetcher to merge metadata in:
- lib/ai/parseModelsDevMetadata.ts: tolerant unknown→Map parser
- lib/ai/fetchModelsDevMetadata.ts: 750ms-bounded fetch with full
error swallowing (metadata is best-effort, must never gate the
underlying gateway response)
- lib/ai/enrichGatewayModel.ts: pure, non-mutating merge
getAvailableModels now fetches gateway + metadata in parallel and
maps each non-embed model through enrichGatewayModel. If models.dev
is unreachable the response is identical to today (gateway models
unenriched).
Documented in recoupable/docs#188 (merged). Unblocks the eventual
open-agents frontend cutover for the model picker.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(ai): extract isRecord into its own lib (SRP)
Per PR feedback: each file should export one primary function.
Pulled isRecord out of parseModelsDevMetadata.ts into
lib/ai/isRecord.ts so the parser file is single-purpose.
Also includes the typecheck fix for enrichGatewayModel — the
`[key: string]: unknown` index signature on its generic constraint
was rejecting `GatewayLanguageModelEntry` and breaking the Vercel
build.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>2 tasks
sweetmantech added a commit
to recoupable/api
that referenced
this pull request
May 5, 2026
* refactor(sandbox): callers use open-agents abstraction (Phase 2.2) (#509) * refactor(sandbox): callers use open-agents abstraction (Phase 2.2) Replaces direct @vercel/sandbox SDK calls with the open-agents sandbox abstraction layer (inlined in Phase 2.1) for sandbox lifecycle (create + reconnect). HTTP response shapes preserved exactly. Per the agreed Option B (hybrid): only the lifecycle creator helpers get refactored. installClaudeCode / runClaudeCode / getSandboxStatus stay on the SDK directly because the abstraction does not cover their needs (sudo, stdout/stderr streaming, simple status reads). Those two install/run files are also dead orphans (defined but never called) and will be removed entirely after the full migration. Production refactor: createSandbox.ts Sandbox.create(...) -> VercelSandbox.create(...) Input: VercelSandboxConfig (was SDK params) Snapshot trigger: restoreSnapshotId field (was source: { type: "snapshot", ... }) Returns VercelSandbox (was SDK Sandbox) createSandboxWithFallback.ts cascade — passes restoreSnapshotId to createSandbox createSandboxFromSnapshot.ts type cascade only (Sandbox -> VercelSandbox) getActiveSandbox.ts Sandbox.get({name}) -> VercelSandbox.connect(name, {}) Status check: sandbox.status -> sandbox.sdkStatus getOrCreateSandbox.ts no code change — type cascades automatically processCreateSandbox.ts reads sandbox.sdkStatus instead of sandbox.status defensive nullish on createdAt Abstraction extension: vercel/sandbox/VercelSandbox.ts adds two readonly getters following the existing host/environmentDetails/expiresAt pattern: get sdkStatus(): string — raw SDK session status (running/pending/ stopped/failed/aborted/snapshotting), distinct from the abstraction's normalized status getter get createdAt(): Date | undefined — SDK session.createdAt These give api callers what they need to construct the existing HTTP response shape without breaking the abstraction's interface. Tests updated: createSandbox.test.ts mocks VercelSandbox.create instead of Sandbox.create; mock object uses sdkStatus instead of status createSandboxWithFallback.test.ts asserts restoreSnapshotId pass-through getActiveSandbox.test.ts mocks VercelSandbox.connect; sdkStatus on mock objects processCreateSandbox.test.ts mockSandbox uses sdkStatus Verification: - pnpm lint:check: clean - pnpm test: 2391/2391 pass - HTTP response shape unchanged: same fields, same enum values for sandboxStatus (sourced from the SDK now via sdkStatus, was directly via SDK Sandbox.status before — identical strings either way) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: address PR #509 review feedback Three real issues from CodeRabbit + cubic: 1. createdAt staleness (CodeRabbit minor) The new `createdAt` getter on VercelSandbox skipped the `refreshStateFromCurrentSession()` step that `sdkStatus` uses, so readers right after a reconnect could see stale session metadata. Add the refresh. 2. Fabricated createdAt (cubic P2) Both createSandbox.ts and processCreateSandbox.ts had a `?? new Date().toISOString()` fallback that fabricated creation metadata when sandbox.createdAt was missing. The SDK guarantees createdAt is populated for any reachable instance, so the fallback was both wrong (fabricates data) and unnecessary. Tighten the getter to return `Date` (not `Date | undefined`) and throw with an explicit "SDK contract violation" message if the field is missing — fail-fast surfaces a real contract bug instead of silently lying. Drop the `?? new Date()` fallbacks at both call sites. 3. Misleading snapshot-restore branching (CodeRabbit major) createSandbox.ts had two paths — a "snapshot" branch that omitted DEFAULT_VCPUS/DEFAULT_RUNTIME (intent: let snapshot dictate), and a "fresh" branch that applied defaults. But VercelSandbox.create internally defaults vcpus=4 and runtime="node22" regardless, so the omission was a no-op — the abstraction always forwarded those to the SDK. Drop the misleading branching. Document the actual behavior at the top of createSandbox: "VercelSandbox.create applies its own defaults regardless of source — those apply to the runtime resources of the new sandbox even when restoring from a snapshot." Updated the snapshot-restore test to assert the actual call shape (vcpus + runtime + timeout + restoreSnapshotId) instead of just the original SDK-style truncated args. Verification: - pnpm lint:check: clean - pnpm test: 2391/2391 pass Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore(sandbox): delete dead Claude Code helpers (Phase 2.3) (#512) * chore(sandbox): delete dead Claude Code helpers (Phase 2.3) installClaudeCode and runClaudeCode were defined but never imported anywhere in api production code — confirmed by grep on main: $ grep -rn "installClaudeCode\b\|runClaudeCode\b" lib/ app/ lib/sandbox/installClaudeCode.ts:9: export async function installClaudeCode(...) lib/sandbox/runClaudeCode.ts:10: export async function runClaudeCode(...) Both files were skipped during the Phase 2.2 abstraction refactor (per the agreed Option B — they used SDK features the abstraction doesn't expose: sudo, stdout/stderr streaming, batched writes). With the broader migration moving to Vercel Workflow + open-agents' agent package for sandbox bootstrap, these orphans have no path to being called again. Removed: lib/sandbox/installClaudeCode.ts (32 lines) lib/sandbox/runClaudeCode.ts (29 lines) lib/sandbox/__tests__/installClaudeCode.test.ts (4 tests) lib/sandbox/__tests__/runClaudeCode.test.ts (6 tests) Verification: - pnpm lint:check: clean - pnpm test: 2381/2381 pass (was 2391 — net -10 tests from the two deleted test files) Note: getOrCreateSandbox.ts also has zero importers per the audit and is similarly dead, but is intentionally NOT deleted in this PR since it was not explicitly flagged as orphan in the migration plan. Worth a separate follow-up decision. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore(sandbox): also delete getOrCreateSandbox + getActiveSandbox (YAGNI) Cascade audit found two more truly-dead helpers per YAGNI: getOrCreateSandbox.ts 0 importers (self-only references) getActiveSandbox.ts only called by getOrCreateSandbox — orphan once that goes Removed: lib/sandbox/getOrCreateSandbox.ts (39 lines) lib/sandbox/getActiveSandbox.ts (33 lines) lib/sandbox/__tests__/getOrCreateSandbox.test.ts (3 tests) lib/sandbox/__tests__/getActiveSandbox.test.ts (4 tests) Live consumers of related helpers preserved: - createSandboxFromSnapshot still used by processCreateSandbox - selectAccountSandboxes still used by aggregateAccountSandboxStats, buildGetSandboxesParams, getSandboxesHandler, validateGetSandboxesRequest Verification: - pnpm lint:check: clean - pnpm test: 2374/2374 pass (was 2381 — net -7 from the two deleted test files; -3 from getOrCreateSandbox.test.ts + -4 from getActiveSandbox.test.ts) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(sessions): port POST /api/sessions from open-agents (#515) * feat(sessions): port GET /api/sessions/[sessionId] from open-agents (Phase 2.4 — first route) First route in the route-by-route cutover plan. Strategy: open-agents frontend stays unchanged in shape; api ports each route it calls in priority order (simplest first), and the open-agents frontend gets cut over to api one route at a time. Why this route first: - Pure DB read (single-row select by id) — no agent runner, no Vercel Workflow, no sandbox runtime - Hits sessions table already migrated in database PR #20 - Frontend usage: agents-frontend hits /api/sessions/{id} on session detail page navigation - Smallest possible blast radius for proving the cutover pattern Files added: lib/supabase/sessions/selectSession.ts Single-row helper + SessionRow type (hand-typed; database.types.ts regen pending — flagged in code comment) app/api/sessions/[sessionId]/route.ts GET handler matching open-agents response shape exactly (camelCase fields, "userId" preserved on the wire even though stored as account_id internally) app/api/sessions/[sessionId]/__tests__/route.test.ts (5 tests) Auth: validateAuthContext (Privy Bearer or x-api-key). Response codes match open-agents: 200 happy path, 401 no auth, 403 not owner, 404 not found. Wire-format translation: snake_case Supabase row -> camelCase response, with account_id surfaced as userId so the existing open-agents frontend fetches with zero code changes. Translation lives at the route boundary (toSessionResponse) where it is easy to remove once chat absorbs this UI and we can switch to schema-natural naming. Verification: - pnpm lint:check: clean - pnpm test: 2379/2379 pass (5 new for this route) Up next: - Cutover step (separate PR in open-agents): point the frontend at api's URL for this single route. Validate end-to-end before porting the next route. - Next routes in priority order (still pure DB, no agent/workflow): GET /api/sessions (list with unread — needs Postgres RPC for the multi-table aggregation), GET /api/sessions/[id]/chats, GET /api/sessions/[id]/chats/[chatId]. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: address PR review — SRP splits + use Tables<\"sessions\"> from regen'd types Three review comments on PR #514: 1. SRP: extract toSessionResponse to its own file was: defined inline in app/api/sessions/[sessionId]/route.ts now: lib/sessions/toSessionResponse.ts (one exported fn per file) 2. SRP: add a handler function (mirroring api convention) was: GET handler logic inline in route.ts now: lib/sessions/getSessionByIdHandler.ts contains all the auth + ownership + DB lookup + response logic; route.ts is a thin shell that awaits options.params and delegates. Matches the pattern used by every other api route (e.g. socials/[id]/scrape, artists/[id]/...). 3. DRY: use existing db schema type was: hand-typed SessionRow interface in selectSession.ts now: Tables<\"sessions\"> from types/database.types.ts (regenerated via npx supabase gen types typescript --project-id ... --schema public) The types regen also resolved the preview-build failure (\"Type instantiation is excessively deep and possibly infinite\") on the .from(\"sessions\") call — Supabase's type inference was choking because the table was unknown to the generic. Files added: lib/sessions/toSessionResponse.ts lib/sessions/getSessionByIdHandler.ts Files modified: app/api/sessions/[sessionId]/route.ts thin shell now app/api/sessions/[sessionId]/__tests__/ route.test.ts type alias updated lib/supabase/sessions/selectSession.ts Tables<\"sessions\"> types/database.types.ts Supabase regen Verification: - pnpm lint:check: clean - pnpm test: 2379/2379 pass (no test changes; same 5 route tests) - tsc compile clean (the local pnpm build progresses past compile into page-data collection where it fails on missing local env vars — Vercel preview will have those set, so the preview rebuild should now succeed) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(sessions): make 404/403 errors emit status:"error" for shape consistency The 401 returned by validateAuthContext shaped like {status:"error", error:"..."} but 404/403 from this handler returned {error:"..."} only. Same endpoint, two error shapes — inconsistent for clients. Align all error responses on the validateAuthContext shape. Tests now assert the full error body, not just the status code. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(sessions): port POST /api/sessions from open-agents Implements the POST /api/sessions contract documented in recoupable/docs PR #186 + #187. Creates a session row and an initial chat row; rolls back the session if chat insert fails so callers never observe an orphaned session. Auth: validateAuthContext (Privy Bearer or x-api-key). Validation: Zod schema + GitHub repo segment regex. Body is optional — empty body creates a session with sensible defaults (status=running, lifecycle_state=provisioning, sandbox_state.type= vercel, title="New session"). Out of scope (will follow once database catches up): auto_commit_push_override, auto_create_pr_override, pr_number, pr_status — these columns don't yet exist on api's sessions table, so the docs spec was trimmed accordingly in docs PR #187. TDD: 9 handler tests cover 401, 400 (sandboxType / repoOwner / repoName), 200 happy path, branch generation, title pass-through, 500 (insertSession failure), and 500-with-rollback (insertChat failure). Plus 1 thin test on the route shell. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(sessions): add OPTIONS handler + cache directives to POST route Match the convention from app/api/sessions/[sessionId]/route.ts: - OPTIONS handler returning 200 + CORS headers (preflight) - dynamic="force-dynamic", fetchCache="force-no-store", revalidate=0 POST routes that mutate DB shouldn't be cached, and browsers issuing preflight checks (POST with JSON body + custom auth headers) need OPTIONS to respond. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(sessions): address PR review feedback - SRP: extract insert-row construction to lib/sessions/buildSessionInsertRow.ts - YAGNI: drop generateSessionBranchName + isNewBranch handling (sessions commit to whatever branch the client provides; auto-generation was speculative) - Tighten isValidGitHubRepoOwner: GitHub's actual rules are alphanumeric + hyphen only (no `_` or `.`), 1-39 chars, no leading/trailing or consecutive hyphens - Tighten isValidGitHubRepoName: reject reserved `.` and `..`, reject `.git` suffix, cap at 100 chars - Add unit tests for both validators (15 cases) and for the new buildSessionInsertRow (4 cases) - Split createSessionHandler tests into auth/validation + persistence files; share fixtures via createSessionHandlerFixtures.ts. All test files now under 100 lines. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(sessions): address second round of PR review - 500 message: "Failed to create session" → "Internal server error" (per cubic.dev standardized 500 envelope feedback) - SRP: extract failedToCreateSession to lib/sessions/failedToCreateSession.ts - YAGNI: drop repoOwner from request body and remove isValidGitHubRepoOwner helper entirely (recoupable is the only owner; no need to validate) - YAGNI: drop repoName from request body and remove isValidGitHubRepoName helper (repo identity is derived server-side from the authenticated account, not accepted from user input) - Single-export per file: split createSessionHandlerFixtures.ts into makeCreateSessionReq.ts, baseSessionRow.ts, baseChatRow.ts. okAuth constant inlined where used. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(sessions): port random-city title fallback from open-agents Generated session titles now match the open-agents UX — names like "Anchorage", "Vienna", "Philadelphia" — instead of every untitled session being called "New session". Closes a wire-shape gap with open-agents production identified by the head-to-head test on PR. Pieces: - lib/sessions/cityNames.ts: ~200-city curated list (verbatim port) - lib/sessions/getRandomCityName.ts: pick a city not in `usedNames`, numeric-suffix fallback when the curated list is exhausted - lib/supabase/sessions/selectSessionTitlesByAccountId.ts: Supabase helper for collision avoidance - lib/sessions/resolveSessionTitle.ts: orchestrates provided title (trimmed) > random city fallback. Async. Kept separate from the insert-row builder so that stays synchronous + pure. - buildSessionInsertRow now takes `title` as a parameter - createSessionHandler awaits resolveSessionTitle before building the row TDD: 4 tests for getRandomCityName, 4 for resolveSessionTitle. Handler tests updated to mock resolveSessionTitle. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: remove GET-only files (scope this PR to POST) The GET endpoint + handler + tests live in PR #514 and were inadvertently brought in when this branch was rebased after #514's work. This PR is scoped to POST only; GET ships in #514. Shared infrastructure stays (types/database.types.ts regen + lib/sessions/toSessionResponse.ts) — both are required by the POST handler too. When either #514 or this PR merges to test first, the other will see those files already present and resolve cleanly. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(sessions): consolidate request validation + DRY supabase select Two reviewer asks rolled into one commit: SRP — validateCreateSessionBody now owns the full validation flow. The handler used to call safeParseJson, validateAuthContext, and the Zod body schema separately; that was three places to short-circuit and three places to duplicate the error envelope. Folded them into validateCreateSessionBody so the handler does one call → success or NextResponse error. Returns { body, auth } on success. DRY — replaced lib/supabase/sessions/selectSession.ts and selectSessionTitlesByAccountId.ts with a single selectSessions({ id?, accountId? }) that supports both call sites. resolveSessionTitle now derives titles from the general fetch. Tests: - New validateCreateSessionBody.test.ts covers auth-failure / 400 / success / malformed-JSON tolerance (4 cases) - Handler tests now mock validateCreateSessionBody (single mock surface instead of three) - resolveSessionTitle tests mock selectSessions Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(sessions): address automated review feedback Four small fixes from the latest round: 1. Zod v4 migration: { message } → { error } on the sandboxType literal. v4 unified the error customization API; { message } is deprecated. 2. Orphan rollback observability: when insertChat fails AND the session-rollback delete also fails, log the session id so ops can detect orphaned rows. New persistence test asserts the log. 3. Defensive try/catch in selectSessions so a thrown exception (network-level rejection, not a Supabase {error} return) doesn't bubble up and 500 the entire session-creation flow. 4. Deterministic test for getRandomCityName suffix-increment: pin Math.random instead of looping until the random pick lands on baseCity. Previous test could pass without ever asserting if the loop cap was hit. Skipped: cubic-dev-ai's note about logging raw sessionId in selectSession.ts — that file was deleted earlier in this PR. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: prettier format fix on persistence test The new orphan-session test had a line that exceeded prettier's wrap width. Auto-format fixed it; format-check now clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(sessions): port GET /api/sessions/[sessionId] from open-agents (#514) Rebased onto current main (which now has the POST endpoint + shared infra from PR #515). Three pieces of GET-specific work: - app/api/sessions/[sessionId]/route.ts: thin shell delegating to the handler, plus OPTIONS for CORS preflight + cache directives - lib/sessions/getSessionByIdHandler.ts: validates auth via validateAuthContext, reads via selectSessions({id}), enforces ownership (403 if account_id mismatch), 404 if missing - app/api/sessions/[sessionId]/__tests__/route.test.ts: 5 cases — 401 / 404 / 403 / 200 happy path / OPTIONS smoke Uses the new general selectSessions({id}) reader rather than the deleted single-purpose selectSession helper. All other shared infra (types, toSessionResponse) is already on main from #515. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(ai/models): enrich response with context_window + cost (#518) * feat(ai/models): enrich response with context_window + cost from models.dev api's GET /api/ai/models previously returned just the gateway entries. Open-agents' frontend depends on two extra fields per model that come from the public models.dev catalog: - context_window (integer) — gates model selection in the picker - cost ({input, output}) — per-million-token pricing for display Adds three pure helpers (TDD'd individually) plus a small refactor of the existing fetcher to merge metadata in: - lib/ai/parseModelsDevMetadata.ts: tolerant unknown→Map parser - lib/ai/fetchModelsDevMetadata.ts: 750ms-bounded fetch with full error swallowing (metadata is best-effort, must never gate the underlying gateway response) - lib/ai/enrichGatewayModel.ts: pure, non-mutating merge getAvailableModels now fetches gateway + metadata in parallel and maps each non-embed model through enrichGatewayModel. If models.dev is unreachable the response is identical to today (gateway models unenriched). Documented in recoupable/docs#188 (merged). Unblocks the eventual open-agents frontend cutover for the model picker. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(ai): extract isRecord into its own lib (SRP) Per PR feedback: each file should export one primary function. Pulled isRecord out of parseModelsDevMetadata.ts into lib/ai/isRecord.ts so the parser file is single-purpose. Also includes the typecheck fix for enrichGatewayModel — the `[key: string]: unknown` index signature on its generic constraint was rejecting `GatewayLanguageModelEntry` and breaking the Vercel build. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
4 tasks
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
1 participant
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.
Summary
Adds two response fields to
GatewayLanguageModelEntryin the AI/models OpenAPI spec:context_window(integer) — max tokens per request. Frontend clients gate model selection on this.cost({input, output}) — per-million-token pricing. Used for the model picker's price display.Why
Pre-step for cutting open-agents' model picker over to api's
GET /api/ai/models. The frontend already reads both fields from its own/api/modelsendpoint:apps/web/lib/models.ts:38-39— filters out models wherecontext_windowis missing/non-numericapps/web/lib/model-options.ts:36-37— surfacescostto the model selector UIThe api handler upstreams these from
@ai-sdk/gatewayalready; it just doesn't currently forward them. A follow-up api PR will plumb them through.Compatibility
Spec already has
additionalProperties: trueonGatewayLanguageModelEntry, so existing clients aren't affected. This change just makes the contract explicit and friendly to SDK generators.Test plan
npx mintlify@latest devrenders the AI Models reference with the two new properties🤖 Generated with Claude Code
Summary by cubic
Add
context_window(max tokens) andcost({input, output} per-million pricing) toGatewayLanguageModelEntryin the AI/models OpenAPI spec. This lets clients gate model selection and show pricing, matches data from@ai-sdk/gateway, and prepares the model picker to callGET /api/ai/models.Written for commit de301f0. Summary will update on new commits.