feat: 默认 Opus 模型从 4.8 升级到 Opus 5

[lishuceo]

背景

将 anycode 的默认 Opus 模型从 claude-opus-4-8 升级到 claude-opus-5（2026-05 GA，$5/$25 per MTok，1M 上下文，同价位替换）。

改动

• src/config.ts — CLAUDE_MODEL 默认值 claude-opus-4-8 → claude-opus-5（仅在 agent 未显式指定模型时作为兜底）
• src/claude/executor.ts — MODEL_PRICING 新增 claude-opus-5 定价条目（input 5 / output 25 / cacheWrite 6.25 / cacheRead 0.50，与 Opus 4.8 一致）；保留 claude-opus-4-8 条目，确保 resume 历史会话的成本计算仍准确

注：运行中 bot 的实际模型由部署侧 config/agents.json（未纳入 git、按部署隔离）控制，已在部署环境同步改为 Opus 5 并热重载生效。本 PR 只固化仓库内的默认值与定价表。

测试

• npm run typecheck ✅
• npx vitest run：1776 passed，唯一失败项 memory/quality.test.ts > 当前工作重点 为向量检索质量的既有 flaky 项（已验证在未含本改动的干净基线上同样失败），与本次改动无关
• 纯配置值与定价常量变更，无新增逻辑分支，未新增单测

🤖 Generated with Claude Code

[github-actions]

Claude finished @lishuceo's task in 2m 36s —— View job

---

Code Review — feat: 默认 Opus 模型从 4.8 升级到 Opus 5

• Check for previous Claude inline comments on this PR (none found)
• Read PR intent + diff (git diff origin/main...HEAD)
• Review src/config.ts (default model change)
• Review src/claude/executor.ts (pricing table entry + lookup path)
• Sweep repo for other hardcoded model references / model-name-based branching
• Check existing tests for assertions on the old default
• Run npm run typecheck / npx vitest run — blocked, see note below

---

✅ Approved

No blocking issues. 3-line constant change, correct in both files, and I couldn't find anything it breaks.

What I verified:

1. No test breakage from the default change. No test asserts 'claude-opus-4-8'; the test fixtures all pin claude-opus-4-6 explicitly in their own config objects (src/claude/__tests__/executor.test.ts:15, src/__tests__/config-loader.test.ts, etc.), so they're unaffected by the config.claude.model fallback.

2. No model-name-based feature gating anywhere. Grepped for model.startsWith(...) / model.includes(...) / regexes over claude-opus-* across src/ — zero hits. So claude-opus-5 won't fall on the wrong side of any version check. src/session/model-override.ts only handles the claude-fable-5 / [1m] path and is orthogonal.

3. Pricing lookup is safe even on a key miss. MODEL_PRICING[model] (src/claude/executor.ts:77) is an exact-match lookup against SDK modelUsage keys. If the SDK ever reports a dated ID (e.g. claude-opus-5-20260501) rather than the alias, the entry misses and falls through to DEFAULT_PRICING — which is claude-opus-4-6 at the same 5 / 25 / 6.25 / 0.50, so the computed cost is still correct. Keeping the claude-opus-4-8 row for resumed sessions is the right call.

Optional (non-blocking) follow-ups:

• 🟡 src/claude/executor.ts:41 — DEFAULT_PRICING = MODEL_PRICING['claude-opus-4-6']! now anchors the fallback to a legacy alias while the default model is Opus 5. Numerically identical today, so no behavior change, but repointing it to 'claude-opus-5' keeps the fallback tracking the actual default and avoids a silent drift if Opus-tier pricing ever diverges. Fix this →
• Docs/examples still advertise claude-opus-4-6: README.md:142, config/agents.example.json:22 and :41. Pre-existing staleness (not introduced here), but since deployment config/agents.json is now on Opus 5, refreshing the tracked example would keep first-time setup aligned.
• CLAUDE.md's testing policy asks each feat commit to carry a test. A one-line assertion that MODEL_PRICING['claude-opus-5'] exists with input 5 / output 25 would satisfy it cheaply — though I agree a pure constant change is a reasonable exception, and the PR body states this explicitly.

Verification I could not run: npx tsc --noEmit and npx vitest run both require approval in this sandbox, so I'm relying on the PR body's reported results (typecheck ✅, 1776 passed, one pre-existing flaky memory/quality.test.ts case verified against a clean baseline). Static risk here is nil — one added Record entry plus a changed string literal.

One thing worth confirming outside my reach: I have no network access in this run, so I could not check claude-opus-5's published rates against docs.anthropic.com. The 5 / 25 / 6.25 / 0.50 row is internally consistent (cacheWrite = 1.25× input, cacheRead = 0.1× input, matching every other Opus row), so if the $5/$25 base is right the derived values are right too.

--- • Branch: feat/upgrade-opus-5