Phase 3: Tier 3-9 Providers - Context

Gathered: 2026-04-05 Status: Ready for planning Mode: Auto-generated (infrastructure phase — discuss skipped)

## Phase Boundary

All 108+ LLM provider definitions exist — specialized models, Chinese/regional providers, infrastructure gateways, emerging tools, code assistants, self-hosted runtimes, and enterprise platforms. Completes the provider library started in Phase 2 (26 Tier 1-2 providers → 108+ total).

## Implementation Decisions

Claude's Discretion

All implementation choices are at Claude's discretion — pure infrastructure/data phase. Use existing YAML schema and patterns from Phase 2.

Lessons from Phase 2 (CRITICAL)

Generic regex patterns cause false positives at scale. Phase 2 Tier 2 providers without distinctive prefixes produced false positives on synthetic test fixtures. With 82 more providers in Phase 3 (mostly generic Chinese/regional/emerging providers), this risk compounds.
Mitigation strategy:
1. Providers WITHOUT distinctive key prefixes MUST rely on keyword pre-filter alone, NOT on overly-broad regex matching.
2. Set confidence to "low" or omit regex entirely and use keyword-only detection where no documented key format exists.
3. Use HIGH entropy_min (≥4.0) for generic alphanumeric patterns.
4. Test each provider's regex against existing testdata/samples/ fixtures to prevent regression.
Category field: still not in schema.go — YAMLs should OMIT category field unless schema is updated this phase.

<code_context>

Existing Code Insights

Reusable Assets

Phase 2 established dual-location pattern: providers/X.yaml + pkg/providers/definitions/X.yaml
pkg/providers/tier12_test.go — guardrail test pattern to replicate for Tier 3-9
pkg/providers/schema.go — Provider struct (Name, DisplayName, Tier, FormatVersion, LastVerified, Keywords, Patterns, Verify)
27 existing provider YAMLs in providers/ as templates

Established Patterns

Go embed.FS via go:embed definitions/*.yaml
RE2-only regex (no lookahead, no backrefs)
Keywords list non-empty (required for Aho-Corasick pre-filter)
patterns with regex, entropy_min, confidence fields

Integration Points

providers/ and pkg/providers/definitions/ — add N new YAML files (likely 82-90)
pkg/providers/tier12_test.go — extend or mirror for tier3+ guardrail
keyhunter providers list --tier=enterprise — requires tier filter support in cmd/providers.go (may need to be added this phase)

</code_context>

## Specific Ideas

Provider groups to cover (per ROADMAP success criteria):

Chinese/regional: DeepSeek, Zhipu, Moonshot (Kimi), Qwen (Alibaba), Baidu (Wenxin/ERNIE), ByteDance (Doubao), 01.AI (Yi), MiniMax, StepFun, SenseTime
Self-hosted runtimes: Ollama, vLLM, LocalAI, llama.cpp, LMStudio, text-generation-webui, Jan, GPT4All
Code assistants: GitHub Copilot, Cursor, Tabnine, Codeium, Cody (Sourcegraph), Continue, Aider, Phind
Enterprise: Salesforce (Einstein GPT), ServiceNow (Now Assist), SAP (Joule), Palantir (AIP), Databricks (DBRX), Snowflake (Cortex), Oracle (Generative AI), HPE, IBM Watsonx
Infrastructure gateways: OpenRouter, LiteLLM, LangSmith, Helicone, PortKey, LangFuse, Traceloop
Specialized: Perplexity (pplx-), ElevenLabs, AssemblyAI, Deepgram, Speechmatics, Runway, Pika, Stability AI, Midjourney, Ideogram, Leonardo AI, Suno, Udio, PlayHT, Descript
Emerging: Writer, Jasper, Copy.ai, WordAI, ContentBot, Rytr
Aim for 82-90 additional providers to reach 108+ total

## Deferred Ideas

--tier=enterprise CLI filter: add to Phase 3 if schema already has Tier field, otherwise defer to later
Provider "category" enum: defer to a separate phase (schema change)
Keyword-only detection mode for providers without documented key formats: lightweight enhancement to pkg/engine/detector.go — consider including if time permits

3.9 KiB Raw Blame History