Alternative LLM Providers (MiniMax, Kimi, GLM everywhere)

Status: Partially implemented — provider catalog, Claude Code routing, MiniMax Codex routing, OpenCode routing, Auth-tab key support for Z.AI/MiniMax/Kimi, N-provider launch picker, provider tab labels, per-provider role-manifest model overrides, and operator guides for Z.AI/MiniMax/Kimi have shipped. Codex GLM/Z.AI and Codex Kimi remain deferred until those providers expose a Responses-compatible endpoint. Remaining work is provider-catalog polish: remove known config-generation drift, tighten availability/update detection, and keep provider docs aligned with upstream endpoint/model changes.

Registry note. Provider routing now lives behind the workspace crate registry. The durable design record for the runtime/provider adapter work lives in Architecture.

Shipped baseline

jackin❯ now has one provider catalog (Provider plus sealed ProviderAdapter) shared by the console picker, capsule spawn protocol, Claude env overrides, Codex profile setup, and OpenCode config generation. The catalog contains native Anthropic for Claude Code/OpenCode, native OpenAI for Codex, and alternative providers Z.AI/GLM, MiniMax, and Kimi.

Alternative-provider auth is key-only. The operator configures provider keys through the jackin console Auth tab; jackin❯ stores them in the normal env map and injects them only into the selected agent process. The shipped env keys are ZAI_API_KEY, MINIMAX_API_KEY, and KIMI_CODE_API_KEY. Z.AI and MiniMax have dedicated env-only Auth rows; Kimi uses the existing Kimi Auth row because the same KIMI_CODE_API_KEY also serves the Kimi agent runtime.

Current provider matrix:

Claude Code routing emits Anthropic-compatible env overrides: ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, Claude model-tier vars, API_TIMEOUT_MS, and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC. Z.AI maps tiers to GLM models; MiniMax maps all tiers to MiniMax-M3; Kimi maps all tiers to kimi-for-coding.

Codex routing currently ships only for MiniMax. jackin-capsule runtime-setup writes [model_providers.minimax] into the container-local Codex config, writes minimax.config.toml, writes a minimax.models.json catalog with a 512k context window, and the runtime entrypoint launches Codex with --profile minimax plus -c model_catalog_json=<file> when MiniMax is selected.

OpenCode routing writes a container-local opencode.json provider block for each configured alternative-provider key. The blocks are self-contained and include SDK package, base URL, API key, and model id; provider selection supplies -m <provider>/<model> at spawn.

Role manifests can override models per provider with [claude.providers.<id>], [codex.providers.<id>], and [opencode.providers.<id>]; unsupported agents ignore provider tables.

Operator guides now live at Z.AI (GLM Coding Plan), MiniMax (Token Plan), and Kimi. Role-author provider-model override docs live in Role Manifest.

Remaining plan

Deferred Codex GLM/Z.AI and Kimi cells

Keep (Z.AI, Codex) and (Kimi, Codex) out of the provider matrix until those providers document and test cleanly against a Responses-compatible endpoint. Current code deliberately excludes them from ProviderAdapter::supports_agent("codex") because Codex requires Responses API routing and the known GLM/Kimi coding endpoints are not wired as deliverable Responses surfaces.

When an upstream Responses-compatible endpoint exists, the work should be catalog/config data rather than a new auth model: add the provider support in the adapter matrix, generate the matching Codex provider/profile/catalog artifacts, and extend tests that bind available_for, codex_profile, and launch args.

Config-generation drift cleanup

Generalize the MiniMax-specific Codex setup names (write_codex_provider_config, codex_minimax_provider_toml, codex_minimax_profile_toml, write_codex_minimax_catalog) into a provider-driven renderer before adding any second Codex alternative provider. Today the implementation is correct for the only shipped Codex alt cell, but the names and shape are intentionally MiniMax-specific.

Do the same for OpenCode generation after the catalog grows: build_opencode_config still takes explicit Z.AI/MiniMax/Kimi key arguments and hardcodes each block. Keep the self-contained block shape, but drive it from provider adapters or an adjacent provider-renderer table so adding a provider does not require another closed match in runtime setup.

Availability and update detection

Provider::available_for and its available_for_provider_matrix test already cover the operator-facing cases: native-only agents skip the picker, a single non-native provider auto-routes without an extra choice, and configured-but-unsupported cells (e.g. Z.AI/Kimi under Codex) stay hidden. No telemetry emits around picker/auto-route decisions today; add it only if operators report a provider-routing mismatch that logs can't currently explain.

Provider docs and constants should be audited when upstream changes plan names, endpoint hosts, key shapes, or model ids. The current documented subscription flows are key-copy flows: Z.AI API key, MiniMax Token Plan sk-cp-… subscription key, and Kimi Code sk-ki… key.

Source evidence

• crates/jackin-protocol/src/lib.rs — Provider variants, labels, env overrides, available_for, key env vars, Codex profile hook, provider-model map in CapsuleConfig.
• crates/jackin-protocol/src/provider_adapter.rs — sealed ProviderAdapter, native provider behavior, Z.AI/MiniMax/Kimi support matrix, Codex profile support only for MiniMax, OpenCode model strings.
• crates/jackin-capsule/src/runtime_setup.rs — container-local Codex MiniMax provider/profile/catalog generation and OpenCode provider JSON generation for Z.AI/MiniMax/Kimi.
• crates/jackin-capsule/src/session.rs — env passthrough for provider keys and selected Codex profile.
• docker/runtime/entrypoint.sh — Codex launch converts JACKIN_CODEX_PROFILE into --profile and model_catalog_json; OpenCode launch accepts the selected provider/model flag.
• crates/jackin-console/src/tui/auth.rs and crates/jackin-console/src/tui/auth_config.rs — Auth-tab rows and env persistence for provider keys.
• crates/jackin-console/src/services/launch.rs and crates/jackin-console/src/tui/components/provider_picker.rs — launch-time provider availability and picker UI.
• crates/jackin-core/src/manifest.rs — role-manifest provider model overrides for Claude Code, Codex, and OpenCode.
• Z.AI (GLM Coding Plan), MiniMax (Token Plan), Kimi, and Role Manifest — shipped operator and role-author docs.

Cross-references

• Multi-runtime support — agent runtimes this routing targets.
• AgentRuntime and Provider registry — provider adapter architecture.
• Agent launch flags API — shares the jackin-capsule runtime-setup dispatch surface.
• Reliable Claude authentication strategy — parallel Anthropic-native auth strategy.