Fix: AttemptBuilder.buildAttempts silently skips invalid model strings (TODO flagged in source) - user typos in fallback lists produce no clear error - #5836

Fixes #5759.

Implemented the fix for #5759 on branch oss-bhai-2026-10-12 (fork SIDDARTHAREDDY8/helicone, commit cbb3eb6, base main, verified from the API). AttemptBuilder.buildAttempts now returns Promise<Result<Attempt[], string>>: on isErr(modelSpec) it keeps the console.error (renamed to 'Invalid model string') and returns err(modelSpec.error) instead of silently continuing; the maintainer's '// TODO: Return error' comment is removed; both final returns wrap in ok(...). The sole caller (verified: only SimpleAIGateway.ts:176 calls buildAttempts; code search total 3 files incl. docs) now checks isErr(attemptsResult) before the empty-array path and returns createErrorResponse with {source: 'Invalid model', message: attemptsResult.error, statusCode: 400}. ONE deliberate deviation from the issue's literal code sketch, verified by driving the real createErrorResponse: the issue sketched type 'request_failed', but the priority chain in createErrorResponse routes a 'request_failed' 400 into the firstNon429Error branch and normalizes it to HTTP 500 (confirmed by harness: request_failed+400 -> 500). The pushed error therefore uses type 'invalid_format', the only type the chain maps to HTTP 400, so the acceptance criterion 'response is a 400 naming the bad model id' actually holds at runtime. COLLISION CHECK NOTE: the step-0 check found PR #5760 ('Fixes #5759') by Harsh23Kashyap, but it is CLOSED and NOT merged (merged_at null, closed_by null = author-closed on 2026-08-18, zero maintainer reviews/comments; only a vercel bot comment). Per the check's literal stop condition it blocks only on OPEN or MERGED PRs, so the build proceeded; issue #5759 is still open and unfixed on main (TODO and silent-skip confirmed present in the fetched main source). The reviewer may still weigh #5760's history; it is flagged here rather than hidden.

Tests pass: Repo vitest suite NOT run locally: it needs the full monorepo install (node_modules + @cloudflare/vitest-pool-workers/miniflare), which the lean API workdir does not have. Instead, a focused harness (in /tmp/harness-build) transpiled the REAL edited AttemptBuilder.ts and SimpleAIGateway.ts plus the REAL parseModelString/validateProvider (packages/cost/models/provider-helpers.ts) and the REAL results modules, stubbing only heavy imports (registry data, provider keys, tracer). 21/21 checks passed: (1) valid model strings return ok wrapping the attempt array, exactly as before; (2) an invalid model string returns err naming the offending string ('Unknown model: gpt-4o-min...'); (3) mixed lists fail on the first invalid entry in both orders; (4) an invalid provider returns err naming the provider; (5) the REAL createErrorResponse priority chain maps the pushed error to HTTP 400 carrying the parser message, with details[0].source 'Invalid model'. Both edited files and the new spec transpile with zero TypeScript diagnostics. The new spec file (5 builder unit tests + 2 end-to-end gateway tests via the repo's own runGatewayTest asserting finalStatus 400 + responseContains 'Unknown model: gpt-4o-min') is written for the repo's own vitest runner; it could not be executed here.

Ticket

Fixes #5759

Component/Service

• AI Gateway

Type of Change

• Bug fix

Deployment Notes

• No special deployment steps required

Context

See description above.