fix(ai): default Anthropic cache retention to 5m

[code-yeongyu]

Summary

• use Anthropic's provider-default 5-minute prompt-cache retention when direct Anthropic requests omit a retention setting
• preserve explicit one-hour retention through cacheRetention: "long", model configuration, or PI_CACHE_RETENTION=long
• keep the browser-safe TTL resolver, provider docs, fork change records, and AI/coding-agent changelogs aligned

Why

Direct Anthropic requests currently opt into ttl: "1h" by default even though Anthropic's own default is five minutes. One-hour cache writes cost more and should be an explicit choice rather than Senpi's implicit default.

RED -> GREEN

• RED payload proof: default canonical Anthropic request expected bare { type: "ephemeral" } but received { type: "ephemeral", ttl: "1h" }
• RED runtime proof: resolvePromptCacheTtlSeconds() expected 300 but received 3600
• GREEN focused suite: 3 files passed, 58 tests passed, 4 credential-gated tests skipped

Verification

• TypeScript LSP: 0 errors, 0 warnings on both changed source files and both changed tests
• npm run build --workspace packages/ai
• npm run test --workspace packages/ai - 196 files passed, 1,872 tests passed
• npm run check
• node scripts/check-pr-changelog.mjs --base origin/main --labels ""
• Senpi Anthropic mock-loop self-test - 22/22 passed, localhost-only, real auth unchanged
• real Senpi CLI capture - five cache-control markers, all bare ephemeral objects with no ttl
• built SDK driver - default: 300 seconds/no ttl; explicit long: 3600 seconds/ttl: "1h"

Evidence

Local QA receipts are retained under:

local-ignore/qa-evidence/20260811-anthropic-default-5m/

The real CLI request capture is under:

local-ignore/qa-evidence/20260812-anthropic-cache-default/

Cleanup

• no mock-loop/fake-model/driver process remains
• no mock-loop temporary sandbox remains
• no QA listener remains

---

Summary by cubic

Defaulted direct Anthropic prompt caching to 5 minutes to match Anthropic’s default and lower costs; 1-hour retention is now opt-in. coding-agent keepalive now reports a 300s TTL by default.

• Bug Fixes

  • Direct Anthropic requests omit ttl by default, sending { type: "ephemeral" } instead of ttl: "1h".
  • resolvePromptCacheTtlSeconds() now returns 300s for Anthropic’s default; precedence stays provider-specific.
  • coding-agent keepalive and tests now expect 300s; docs and changelogs in packages/ai and packages/coding-agent updated.

• Migration

  • To keep 1-hour retention, set PI_CACHE_RETENTION=long, cacheRetention: "long", or configure it on the model.

^{Written for commit cc10b74. Summary will update on new commits.}