fix(providers): MiniMax system prompt fix + live integration tests for 11 providers

[penso]

Summary

Bug fix: MiniMax silently ignores the top-level "system" JSON field — system prompts were completely lost. Now prepends system content to the first user message wrapped in [System Instructions] markers. Fixes #578, #592.

Bug fix: GPT-5/reasoning model probe failed with max_completion_tokens: 1 rejected as 400. Bumped to 16 tokens.

Live API integration tests for 11 providers (all #[ignore]d, run with -- --ignored):

Provider	Tests	Env var	Status
MiniMax	12	MINIMAX_API_KEY	12/12
Moonshot	10	MOONSHOT_API_KEY	10/10
Anthropic	9	ANTHROPIC_API_KEY	9/9
OpenAI	9	OPENAI_API_KEY	9/9
DeepSeek	9	DEEPSEEK_API_KEY	9/9
Z.AI	9	Z_API_KEY	9/9
Gemini	9	GEMINI_API_KEY	9/9
OpenRouter	8	OPENROUTER_API_KEY	8/8
Fireworks	9	FIREWORKS_API_KEY	8/9 (rate limit)
Mistral	9	MISTRAL_API_KEY	transient 503s
Kimi Code	8	KIMI_API_KEY	needs membership

Each suite covers: system prompt delivery, tool calling (streaming + non-streaming), multi-turn tool use, probe, streaming, model catalog liveness, and model discovery.

Daily CI workflow (provider-integration.yml): runs all 11 providers sequentially at 06:00 UTC with max-parallel: 1 to avoid rate limits. Each provider is continue-on-error so one outage doesn't fail the whole run. Missing secrets emit a warning and skip.

Findings from integration tests

• MiniMax: top-level system field silently ignored (fixed), all 7 catalog models alive
• Moonshot: 13 new models not in catalog, reasoning_content required for multi-turn
• Fireworks: 4 of 7 catalog models dead/undeployed, 11 new models
• Anthropic: claude-3-7-sonnet-20250219 returns 404
• OpenAI: 102 new chat-capable models, GPT-5 probe bug (fixed)
• Z.AI: glm-4.7-flash and glm-4.7-flashx return 500, 2 new models (glm-5-turbo, glm-5.1)
• Gemini: 4 older preview models dead, 30 new models, thought_signature required for multi-turn

Validation

Completed

• cargo +nightly-2025-11-30 fmt --all -- --check
• cargo +nightly-2025-11-30 clippy -p moltis-providers --all-targets --features local-llm-metal -- -D warnings
• cargo test --workspace (all 270 provider unit tests + full workspace pass)
• cargo test --test minimax_integration -- --ignored (12/12)
• cargo test --test moonshot_integration -- --ignored (10/10)
• cargo test --test anthropic_integration -- --ignored (9/9)
• cargo test --test openai_integration -- --ignored (9/9)
• cargo test --test deepseek_integration -- --ignored (9/9)
• cargo test --test zai_integration -- --ignored (9/9)
• cargo test --test gemini_integration -- --ignored (9/9)
• cargo test --test openrouter_integration -- --ignored (8/8)

Remaining

• ./scripts/local-validate.sh
• Add API keys to CI secrets and trigger gh workflow run provider-integration.yml

Manual QA

1. Deploy with MiniMax provider — verify system prompt delivered via Telegram/Discord
2. Verify no invalid message role: system (2013) errors
3. Verify tool calls work (not returned as text)

Fixes #578, #592

🤖 Generated with Claude Code

[codspeed-hq]

Merging this PR will degrade performance by 46.32%

❌ 1 regressed benchmark
✅ 38 untouched benchmarks
⏩ 5 skipped benchmarks¹

⚠️ Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	env_substitution	10 µs	18.7 µs	-46.32%

---

_{Comparing nervous-ironclad (4dcaaf5) with main (50f621b)}

Footnotes

1. 5 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[greptile-apps]

Greptile Summary

This PR delivers two targeted bug fixes — MiniMax system prompt delivery (now correctly inlined into the first user message, with multimodal content handled) and the GPT-5/reasoning-model probe rejecting max_completion_tokens: 1 (bumped to 16) — plus live #[ignore]d integration tests for 11 providers and a daily CI workflow.

• The apply_system_prompt_rewrite fix in openai.rs correctly handles string, array (multimodal), and absent content; unit tests and live tests confirm the regression is closed.
• The CI workflow scopes each job to a single API key via GITHUB_ENV, addressing the previous all-secrets-exposed concern.
• lib.rs was updated to add new provider catalogs but still contains claude-3-7-sonnet-20250219 (confirmed 404 by the new tests) and reportedly dead Fireworks entries; consider removing them here rather than deferring to a follow-up.

Confidence Score: 5/5

Safe to merge — both bug fixes are correct and well-tested; the sole remaining finding is a P2 cleanup of dead catalog entries that don't block the fix.

All P0/P1 concerns from prior review rounds (Secret for API keys, multimodal content drop, all-secrets-in-env) are resolved. The only open item is a P2 suggestion to remove dead Anthropic/Fireworks catalog entries that the tests discovered; this doesn't affect correctness of the fixes or the test infrastructure.

crates/providers/src/lib.rs — ANTHROPIC_MODELS and FIREWORKS_MODELS contain confirmed-dead model IDs that should be pruned.

Important Files Changed

Filename	Overview
crates/providers/src/openai.rs	MiniMax system prompt rewrite fix handles both string and array (multimodal) content correctly; GPT-5 probe token cap bumped to 16 — both fixes are well-tested with unit and integration coverage.
.github/workflows/provider-integration.yml	Adds daily provider integration CI; correctly exports only the relevant API key per matrix job using the GITHUB_ENV pattern, addressing the previous all-secrets-exposed concern.
crates/providers/tests/minimax_integration.rs	12 integration tests covering system prompt delivery, tool calling, multi-turn, streaming, and model discovery; api_key() returns Secret throughout including the raw-reqwest model-endpoint test.
crates/providers/tests/anthropic_integration.rs	9 tests pass; KNOWN_MODELS still includes claude-3-7-sonnet-20250219 which the tests confirmed returns 404 — catalog_models_are_live won't fail on it but the production catalog in lib.rs still serves this dead entry to users.
crates/providers/tests/fireworks_integration.rs	KNOWN_MODELS mirrors the lib.rs catalog; PR description notes 4 of 7 models are dead/undeployed but the catalog wasn't pruned.
crates/providers/tests/openai_integration.rs	9 tests; correctly uses streaming probe for OpenAI; Secret handled consistently.
crates/providers/tests/moonshot_integration.rs	10 tests covering reasoning_content requirement for multi-turn tool use; all patterns correct.
crates/providers/tests/gemini_integration.rs	9 tests for Gemini; handles thought_signature requirement for multi-turn; Secret correct.
crates/providers/tests/deepseek_integration.rs	9 tests; consistent structure and Secret usage.
crates/providers/tests/zai_integration.rs	9 tests for Z.AI; notes glm-4.7-flash / glm-4.7-flashx return 500 but continues-on-error.
crates/providers/tests/mistral_integration.rs	9 tests; transient 503s noted in PR description but test logic is correct.
crates/providers/tests/openrouter_integration.rs	8 tests; no static catalog (BYOM provider via discovery); patterns consistent.
crates/providers/tests/kimi_code_integration.rs	8 tests; provider uses OAuth or API key; Secret correct, kimi-code-specific base URL hardcoded appropriately.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[OpenAiProvider::serialize_messages_for_request] --> B[build request body]
    B --> C{apply_system_prompt_rewrite}
    C -->|rejects_system_role = false| D[No-op: pass body through]
    C -->|rejects_system_role = true| E[Extract & remove system messages]
    E --> F{First user message content type?}
    F -->|String| G[Prepend system block to string]
    F -->|Array multimodal| H[Insert text element at index 0]
    F -->|Missing / other| I[Replace with synthetic system block]
    E --> J{No user message at all?}
    J -->|true| K[Insert synthetic user message with system block]
    G & H & I & K --> L[Send to MiniMax API]

Loading

_{Reviews (3): Last reviewed commit: "fix(providers): handle multimodal conten..." | Re-trigger Greptile}

[codecov]

Codecov Report

❌ Patch coverage is 97.34513% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/providers/src/openai.rs	97.34%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!