feat: add deployment hours availability for freebuff GLM 5.1 model

[jahooma]

Summary

Add deployment hours availability policy for freebuff models. GLM 5.1 is now only available during 9am ET - 5pm PT on weekdays, while MiniMax M2.7 remains always available.

Availability system:

• New availability field on FreebuffModelOption: 'always' or 'deployment_hours'
• New isFreebuffDeploymentHours() helper: checks Mon-Fri, 9am ET to 5pm PT
• New isFreebuffModelAvailable() and resolveAvailableFreebuffModel() utilities
• FREEBUFF_DEPLOYMENT_HOURS_LABEL constant: '9am ET-5pm PT'

CLI model selector:

• GLM 5.1 shown first with "Smartest" tagline, MiniMax as "Fast"
• Real-time availability checks via useNow hook (updates every minute)
• Unavailable models shown as "Closed", non-interactable
• Auto-resets to MiniMax if selected model becomes unavailable

Server-side:

• GLM 5.1 uses dedicated Fireworks deployment (mjb4i7ea) during deployment hours
• Session queue gates on deployment hours availability
• Outside deployment hours: GLM requests return 503 DEPLOYMENT_OUTSIDE_HOURS

Fireworks deployment:

• GLM 5.1 deployment map entry added
• Pricing: .40/M input, bash.26/M cached input, .40/M output

Test plan

• All affected test suites pass (completions, fireworks-deployment, freebuff-session, public-api)
• Deployment hours logic tested with mocked times
• No stale references to removed model IDs

[greptile-apps]

Greptile Summary

This PR replaces Kimi K2.6 with GLM 5.1 as the freebuff deployment-hours model and introduces a time-bounded availability system: GLM 5.1 is only selectable 9am ET–5pm PT on weekdays (backed by a dedicated Fireworks deployment), while MiniMax M2.7 remains always-on via serverless. The key architectural change is the removal of the serverless fallback for deployment-model failures — errors now surface explicitly as 503s so the client can redirect users to MiniMax instead of silently degrading quality.

The implementation is thorough: isFreebuffDeploymentHours correctly computes the dual-timezone window using Intl.DateTimeFormat.formatToParts, test clocks are mocked to deterministic times (8am ET for off-hours tests, explicit _tick for in-hours tests), and availability is enforced at every boundary — session admission, the Fireworks routing layer, the CLI selector, and the model store initialization. A pre-existing bug where getFireworksPricing fell back to FIREWORKS_MODEL_MAP (the model-name map) instead of FIREWORKS_PRICING_MAP is also fixed.

Confidence Score: 5/5

Safe to merge — all changes are well-guarded, test clocks are deterministic, and no P0/P1 issues found.

The PR implements a clean model swap with correct dual-timezone window logic, consistent availability enforcement at every layer (session admission, Fireworks routing, CLI store, and selector), and a bonus fix to a pre-existing getFireworksPricing fallback bug. All remaining observations are P2 (UX/design tradeoffs, not defects). Tests are deterministic thanks to injected clocks.

No files require special attention.

Important Files Changed

Filename	Overview
common/src/constants/freebuff-models.ts	Core of the change: adds availability field, isFreebuffDeploymentHours (dual-timezone window check), isFreebuffModelAvailable, and resolveAvailableFreebuffModel. Logic is correct; h23 hourCycle and formatToParts avoid AM/PM ambiguity.
web/src/llm-api/fireworks.ts	Removes serverless fallback; deployment-mapped models now return explicit 503s with structured error codes (DEPLOYMENT_OUTSIDE_HOURS, DEPLOYMENT_COOLDOWN) instead of silently falling back. Also fixes getFireworksPricing fallback to reference FIREWORKS_PRICING_MAP (was FIREWORKS_MODEL_MAP).
web/src/server/free-session/public-api.ts	Adds model_unavailable early-return before queue insertion, reuses now across the function, and returns availableHours label in the response.
web/src/server/free-session/admission.ts	Marks registered freebuff models as unhealthy during off-hours (blocking admission) without affecting unknown/unregistered models. Logic is sound.
cli/src/components/freebuff-model-selector.tsx	Adds GLM-first display ordering, isAvailable gating on interactions, useEffect to fall back to MiniMax when GLM becomes unavailable mid-session, and deployment hours label in the UI.
cli/src/hooks/use-freebuff-session.ts	Handles new model_unavailable 409 response: switches the model store to the default (MiniMax) and retries with GET, closing the loop on server-side availability enforcement.
web/src/llm-api/tests/fireworks-deployment.test.ts	Updated to use explicit now injection instead of Date.prototype.toLocaleString mocking. Tests now verify the no-fallback behavior: deployment errors return 503 directly.
web/src/app/api/v1/freebuff/session/tests/session.test.ts	Clock fixed to 2026-04-17T12:00:00Z (8am ET, outside deployment hours), making model_unavailable assertions deterministic regardless of when CI runs.
web/src/app/api/v1/chat/completions/tests/completions.test.ts	Correctly uses conditional assertions (isFreebuffDeploymentHours()) for the GLM test that runs against real clock; other freebuff tests switched to MiniMax (always-available).
web/src/server/free-session/tests/public-api.test.ts	Default test clock starts outside deployment hours; tests that need GLM available explicitly call deps._tick(new Date('2026-04-17T16:00:00Z')). New model_unavailable test is deterministic.
web/src/llm-api/fireworks-config.ts	Removes Kimi K2.6 deployment entry; adds GLM 5.1 deployment mjb4i7ea. Kimi entries moved to comments.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[User selects model in CLI] --> B{isFreebuffModelAvailable?}
    B -- "GLM 5.1 + inside 9am ET-5pm PT weekday" --> C[POST /freebuff/session with z-ai/glm-5.1]
    B -- "GLM 5.1 + outside hours" --> D[Show Closed · Switch to MiniMax]
    B -- "MiniMax M2.7 always" --> E[POST /freebuff/session with minimax/minimax-m2.7]
    C --> F{Server: isFreebuffModelAvailable?}
    E --> G[Queue / admit immediately]
    F -- "available" --> G
    F -- "unavailable 409 model_unavailable" --> H[Client: setSelectedModel DEFAULT → retry GET]
    G --> I{Admission tick}
    I -- "MiniMax: health check only" --> J[Admit from MiniMax queue]
    I -- "GLM: outside hours → unhealthy" --> K[GLM queue paused]
    I -- "GLM: inside hours + healthy" --> L[Admit from GLM queue]
    L --> M[Chat completions: createFireworksRequestWithFallback]
    J --> N[Chat completions: serverless MiniMax]
    M --> O{isDeploymentHours?}
    O -- "yes + not cooling down" --> P[Try deployment mjb4i7ea]
    O -- "no → 503 DEPLOYMENT_OUTSIDE_HOURS" --> Q[Surface error to client]
    P -- "success" --> R[Return response]
    P -- "5xx" --> S[Return 503 — no serverless fallback]

Loading

_{Reviews (1): Last reviewed commit: "feat: replace Kimi K2.6 with GLM 5.1 as ..." | Re-trigger Greptile}