feat(models): integrate Kimi K3 under opencode-go

[samueltuyizere]

Summary

Integrates the Kimi K3 model as a config-driven alias served through the opencode-go provider — the provider that carries every Kimi model today (k2.5/k2.6/k2.7-code all route through Moonshot upstream via opencode-go).

Specs verified against https://opencode.ai/data/moonshot/kimi-k3:

• 1M context window
• 131K max output tokens
• Modalities: text, image, video
• Pricing: $3 / 1M input · $15 / 1M output
• Released July 2026
• OpenAI Chat Completions endpoint (endpoint classifier confirmed)
• kimi- prefix handling in the transformer (placeholder reasoning, cache_control) covers K3 automatically — no code-path changes needed
• Fallback chain: kimi-k3 → kimi-k2.7-code → kimi-k2.6

Changes

• internal/config/model_registry.go — register kimi-k3 metadata (1M / 131K / vision / tools) (+ test)
• internal/client/opencode_test.go — endpoint-classification test case
• configs/config.example.json, cmd/routatic-proxy/main.go, cmd/routatic-proxy/init_provider.go — model entry + fallback chain across all three config sources (consistent vision: true)
• MODELS.md, docs/models.md, docs/zh/MODELS.md, CLAUDE.md — docs with real specs + pricing
• Synced catalog.json intentionally left untouched (fetched from remote source; will pick up K3's full catalog entry on next sync)

Testing

• go build ./... ✅
• go vet ./... ✅
• go test ./... ✅ (full suite)
• Reviewed by phase-reviewer agent; flagged config-source vision inconsistency, fixed.

Notes

• max_tokens in the config aliases stays at a conservative per-request default (32K); the registry carries the true 131K cap for capacity filtering.
• K3's 1M context makes it a candidate for the long_context scenario, but routing logic was left unchanged (out of scope).

🤖 Generated with Claude Code

[kilo-code-bot]

Code Review Roast 🔥

Verdict: No Issues Found | Recommendation: Merge

Oh wait, this incremental commit is actually clean. I need to sit down. I had my flamethrower warmed up and everything.

This round properly resolves the previous review's kimi-k3 triple-copy finding: both getDefaultConfig() (main.go) and getOpenCodeGoConfig() (init_provider.go) now read a single //go:embed templates/default_config.json, and the two 280-line hand-synced JSON blobs are gone. The new TestDefaultAndOpenCodeGoConfigsIdentical + TestExampleConfigSharedModelsMatchDefault turn "these are the same" into a CI-enforced invariant — drift now fails the build instead of shipping silently. That's the kind of fix that respects the original roast rather than hand-waving it away.

The config.example.json superset isn't fully covered by the in-sync test (deliberate — it adds cost_routing/aws_bedrock), but the shared-model assertion guards exactly the vision-flag drift the earlier reviewer caught, so I'm calling the prior finding resolved.

📊 Overall: Like a student who actually studied for the makeup exam — the previous mistake was acknowledged and this follow-up commit fixed it properly.

Files Reviewed (4 files, incremental)

• cmd/routatic-proxy/init_provider.go - removed hand-synced JSON, now defers to getDefaultConfig() (resolved prior finding)
• cmd/routatic-proxy/main.go - extracted default config to embedded template
• cmd/routatic-proxy/init_provider_test.go - new in-sync + shared-model tests (good)
• cmd/routatic-proxy/templates/default_config.json - new single source of truth

Previous Review Summaries (3 snapshots, latest commit 031ccf7)

Current summary above is authoritative. Previous snapshots are kept for context only.

Previous review (commit 031ccf7)

Verdict: No Issues Found | Recommendation: Merge

Oh wait, this incremental diff is actually clean. I need to sit down. I had my flamethrower warmed up and everything.

The previous review flagged the triple-copied kimi-k3 template block across getOpenCodeGoConfig, getDefaultConfig, and config.example.json (init_provider.go:512). This incremental round answered that feedback properly: both Go template functions now carry explicit SYNC doc comments pointing at the other two sources, and a new TestDefaultAndOpenCodeGoConfigsInSync turns the "these two are the same" claim into a CI-enforced invariant — drift now fails the build instead of shipping silently. That's the kind of fix that respects the original roast rather than hand-waving it away.

📊 Overall: Like a student who actually studied for the makeup exam — the previous mistake was acknowledged and this follow-up commit fixed it properly. The remaining config.example.json isn't covered by the in-sync test, but the SYNC comment covers the paper-trail requirement, so I'm calling it resolved.

Files Reviewed (3 files, incremental)

• cmd/routatic-proxy/init_provider.go - added SYNC comment (resolved prior finding)
• cmd/routatic-proxy/main.go - added SYNC comment (resolved prior finding)
• cmd/routatic-proxy/init_provider_test.go - new in-sync + presence tests (good)

Previous review (commit d3d0d7a)

Verdict: No Issues Found | Recommendation: Merge

Oh wait, this incremental diff is actually clean. I need to sit down. I had my flamethrower warmed up and everything.

Overview

Severity	Count
🚨 critical	0
⚠️ warning	0
💡 suggestion	0
🤏 nitpick	0

Issue Details (click to expand)

No new issues found in changed code.

---

🏆 Best part: The spec fix was thorough — registry metadata, test expectations, and all four doc files (English, Chinese, MODELS.md, CLAUDE.md) are now consistent with the real 1M/131K figures and $3/$15 pricing. That's the kind of complete sync the previous reviewer was begging for.

💀 Worst part: Nothing. I'm actually disappointed. I prepared three different metaphors about triple-copied JSON and I didn't get to use any of them.

📊 Overall: Like a student who actually studied for the makeup exam — the previous mistakes were acknowledged and this follow-up commit fixed them properly.

Files Reviewed (6 files)

• internal/config/model_registry.go - updated kimi-k3 metadata (1M/131K)
• internal/config/model_registry_test.go - updated test expectations
• CLAUDE.md - updated model table
• MODELS.md - updated pricing/details
• docs/models.md - updated quick reference table
• docs/zh/MODELS.md - updated Chinese docs

Previous review (commit ef4ea83)

Verdict: 1 Issue Found | Recommendation: Address before merge

Overview

Severity	Count
🚨 critical	0
⚠️ warning	0
💡 suggestion	1
🤏 nitpick	0

Issue Details (click to expand)

File	Line	Roast
cmd/routatic-proxy/init_provider.go	508	kimi-k3 model + fallback chain duplicated across 3 hardcoded JSON templates (main.go, init_provider.go, config.example.json) — drift hazard the PR's own history already demonstrated

---

🏆 Best part: Genuinely clean execution — the registry entry, test coverage (endpoint classification and metadata resolution), and the docs/markdown/CLAUDE.md updates are all consistent and complete across every source. The author even caught and fixed their own vision inconsistency mid-PR. That's discipline.

💀 Worst part: The one finding is the triple-copied kimi-k3 block. Three hand-synced JSON strings is a maintainability landmine waiting for the next Kimi release — the exact defect class the PR already admitted to once.

📊 Overall: Like a tidy desk with three identical sticky notes on it — everything's in the right place, but you only needed one.

Files Reviewed (10 files)

• internal/config/model_registry.go - registered kimi-k3 metadata
• internal/config/model_registry_test.go - metadata resolution test
• internal/client/opencode_test.go - endpoint classification test
• cmd/routatic-proxy/main.go - default config template
• cmd/routatic-proxy/init_provider.go - opencode-go config template
• configs/config.example.json - example config
• CLAUDE.md - model table
• MODELS.md - model reference + endpoint tables
• docs/models.md - model reference
• docs/zh/MODELS.md - Chinese model reference

Fix these issues in Kilo Cloud

---

_{Reviewed by hy3:free · Input: 81.9K · Output: 2.6K · Cached: 275.3K}

_{Review guidance: REVIEW.md from base branch main (truncated)}