Permafrost 0.3.0 — optimal pass, fully live-validated

Optimal pass — everything now validated against real Claude Code + real DeepSeek in one live suite (15/15 phases pass; session total 71% hit / 68% cost saved).

Update inside Claude Code: /plugin marketplace update permafrost → /plugin install permafrost@permafrost

New since 0.2.0

Coalescing release policy — DeepSeek's cache write is async (probed: ~4s after first byte still misses, ~6s hits). Default PERMAFROST_COALESCE_SETTLE_MS is now 2500 (live-validated: cold 3-burst followers hit in full vs 0% at settle 0). New PERMAFROST_COALESCE_RELEASE=completion holds followers until the leader fully streams, for max hit odds.

Self-debugging doctor — when an anchor changes within a lineage, /permafrost:doctor now shows the exact byte where it diverged plus was/now excerpts. A future Claude Code release that introduces a new cache-buster shows up as a readable diff, not a sinking hit rate (this is the workflow that originally caught CC's cch= billing nonce).

Real-CC fixture regression — tests/fixtures/cc_request.json reproduces a real CC 2.1.x request shape; CI now asserts the anchor survives inter-turn drift (new nonce, changed gitStatus, reshuffled tools, grown messages) with no API key.

Engineering — pooled upstream connections (no per-request TLS handshake); per-session keepalive slots (parallel conversations all stay warm, not just the latest); /permafrost/* control endpoints are loopback-only even if you bind the proxy wide.

Full live e2e suite — e2e/run_full_suite.sh asserts every feature end-to-end: alignment+freeze, coalescing (real subagent fan-out and deterministic cold burst), keepalive+resume (resumed session read 96% from warm cache), warm endpoint (99%), session/lineage accounting, doctor anchor-diff.

Full details and measured DeepSeek cache behavior: docs/e2e-findings.md.

Permafrost 0.2.0 — installable Claude Code plugin

First marketplace release. Install inside Claude Code:

/plugin marketplace add jianzhichun/permafrost
/plugin install permafrost@permafrost

What it does: freezes Claude Code's prompt prefix so DeepSeek's automatic cache keeps hitting — measured 64% cost reduction on real Claude Code traffic (66% cache hit over a 4-turn agentic task), validated against the live api.deepseek.com/anthropic endpoint.

Highlights

• Alignment proxy: deterministic tool sort, cache_control strip, env freeze+delta, billing-nonce stabilization, canonical serialization
• Cold-anchor coalescing for parallel subagent fan-out (live: 0% → 73% hit)
• Opt-in idle keepalive (live: 99.9% hit on unchanged replay) + permafrost warm
• Per-session & per-lineage stats, /permafrost:status|doctor|benchmark|wrap commands, savings statusline
• Zero dependencies (stdlib Python); offline benchmark + isolated e2e harness (e2e/run_claude_code.sh)

Measured DeepSeek cache behavior (see docs/e2e-findings.md): the endpoint re-renders to canonical order before caching; cache identity includes request params and the client header fingerprint; partial-prefix (anchor-only) warms never hit.