feat: fall back to OpenCode when Anthropic is unavailable

[vinicius91carvalho]

Closes #106

Summary

• add opt-in Anthropic-first passthrough with OAuth/capability header preservation
• fall back on 408, 429, 5xx, and transport failures before response streaming starts
• honor Retry-After with adaptive single-request recovery probes and no synthetic health calls
• skip usage-limited OpenCode Go models and continue through working Zen-free fallbacks
• refresh generated defaults and document Claude subscription configuration

Validation

• gofmt check
• go vet ./...
• go test ./... -race
• go build ./cmd/routatic-proxy
• golangci-lint (0 issues)
• live Claude Code -> Anthropic passthrough
• forced Anthropic 429 -> real OpenCode fallback
• resumed Claude Code session retained context across Anthropic -> OpenCode failover
• live one-token checks for configured Zen-free models

Notes

The free Zen catalog still lists MiniMax M3 Free and Qwen3.6 Plus Free, but both report that their promotions ended. Defaults therefore use the currently working Nemotron 3 Ultra Free, MiMo V2.5 Free, and DeepSeek V4 Flash Free endpoints.

[kilo-code-bot]

Code Review Roast 🔥

Verdict: No Issues Found | Recommendation: Merge

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

📊 Overall: Like finding a unicorn in production — I didn't think clean PRs existed anymore, but here we are.

Files Reviewed (2 files)

• internal/handlers/anthropic_first.go
• internal/handlers/anthropic_first_test.go

Previous Review Summaries (2 snapshots, latest commit ef6bfef)

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

Previous review (commit ef6bfef)

Verdict: 1 Issue Found | Recommendation: Address before merge

Overview

Severity	Count
⚠️ warning	1

Issue Details (click to expand)

File	Line	Roast
internal/handlers/anthropic_first.go	120	Retry-After: 0 gets floor'd to 1 second instead of immediate retry

---

🔥 The incremental fix added a 1-second minimum floor to parseRetryAfter, but "0" per HTTP spec means "retry immediately" — not "wait one second." Past-dated HTTP dates also incorrectly return 1 second instead of 0. This could delay recovery when Anthropic signals readiness.

🏆 Best part: The context-aware streaming with sync.Pool buffer reuse is actually clever. Memory allocation optimization while maintaining cancel propagation? I need to sit down.

💀 Worst part: Retry-After: 0 treated as 1 second. Nothing says "production ready" like reinterpreting spec-compliant retry directives.

📊 Overall: Like a second pancake — better than the first but still slightly misshapen.

Files Reviewed (2 files)

• internal/handlers/anthropic_first.go - 1 issue
• internal/handlers/anthropic_first_test.go - no issues (fix correct)

Fix these issues in Kilo Cloud

Previous review (commit e1b12f8)

Verdict: No Issues Found | Recommendation: Merge

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

📊 Overall: Like finding a unicorn in production — I didn't think clean PRs existed anymore, but here we are.

The Anthropic-first passthrough implementation is solid:

• Proper circuit breaker pattern with exponential backoff up to 15 minutes
• Retry-After header support for HTTP-date format parsing (though now is passed as parameter, not captured internally)
• Zero-byte stream handling with 32KB buffer
• Provider-wide usage limit skipping works correctly in both streaming and non-streaming paths
• Tests cover concurrent probe scenarios, transport failures, and redirect handling without credential leaks

The only thing that made me raise an eyebrow was copyStreamingResponse silently swallowing write errors — but that's expected for streaming where client disconnects are normal.

Files Reviewed (11 files)

• internal/handlers/anthropic_first.go - 287 lines (new)
• internal/handlers/anthropic_first_test.go - 228 lines (new)
• internal/handlers/messages.go - 6 changed lines
• internal/handlers/messages_test.go - 87 new lines
• internal/server/server.go - 2 changed lines
• internal/config/config.go - 5 new lines
• internal/config/loader.go - 4 new lines
• internal/config/loader_test.go - 22 new lines
• internal/router/fallback.go - 9 changed lines
• internal/router/fallback_test.go - 48 new lines
• internal/provider/opencode_zen.go - 2 new lines
• cmd/routatic-proxy/main.go - 20 changed lines

---

_{Reviewed by step-3.7-flash-20260528 · Input: 341.5K · Output: 4.3K · Cached: 5.9K}

[samueltuyizere]

LGTM!